diff --git a/.gitignore b/.gitignore
index ae935a783..eccb91e00 100644
--- a/.gitignore
+++ b/.gitignore
@@ -93,3 +93,7 @@ AgroBase/AgroBase/bin/x64/Debug/Python/venv/
 /Exemplos
 /Python/raspi/cam_2/imx296_pi/
 /Python/raspi/cam_3/imx296_pi/
+Python/OAK/datasets/oak-fcc-3/audit/
+/Python/OAK/datasets/oak-fcc-3/benchmarks_async
+/Python/OAK/datasets/oak-d/benchmark_corridor
+/AgroBase/livox_visual_debugger/x64/Debug
diff --git a/AgroBase/AgroBase/Forms/IHM/Operacao/Parametros/ucParametrosMapa.cs b/AgroBase/AgroBase/Forms/IHM/Operacao/Parametros/ucParametrosMapa.cs
index de1e42264..1012f467e 100644
--- a/AgroBase/AgroBase/Forms/IHM/Operacao/Parametros/ucParametrosMapa.cs
+++ b/AgroBase/AgroBase/Forms/IHM/Operacao/Parametros/ucParametrosMapa.cs
@@ -105,7 +105,7 @@ namespace AgroBase.Forms.IHM.Operacao.Parametros
 
                 _mapa.TipoMapa = op.Parametros.TipoMapa;
 
-                await _mapa.DefinirRuasSelecionadasAsync(ruasSelecionadas, false);
+                await _mapa.DefinirRuasSelecionadasAsync(op, ruasSelecionadas, false);
 
                 lblMapaPlaceholder.Visible = false;
 
diff --git a/AgroBase/AgroBase/Forms/IHM/Operacao/ucOperacaoNavegacao.cs b/AgroBase/AgroBase/Forms/IHM/Operacao/ucOperacaoNavegacao.cs
index 3e9188490..4c36e710d 100644
--- a/AgroBase/AgroBase/Forms/IHM/Operacao/ucOperacaoNavegacao.cs
+++ b/AgroBase/AgroBase/Forms/IHM/Operacao/ucOperacaoNavegacao.cs
@@ -169,7 +169,7 @@ namespace AgroBase.Forms.IHM.Operacao
 
                 mapa.TipoMapa = op.Parametros?.TipoMapa ?? Enums.TipoMapaOperacao.Indefinido;
 
-                await mapa.DefinirRuasSelecionadasAsync(ruasSelecionadas, false);
+                await mapa.DefinirRuasSelecionadasAsync(op, ruasSelecionadas, false);
             }
             else
             {
diff --git a/AgroBase/AgroBase/Forms/IHM/Operacao/ucOperacaoParametros.cs b/AgroBase/AgroBase/Forms/IHM/Operacao/ucOperacaoParametros.cs
index 8b8186250..7bbaf041a 100644
--- a/AgroBase/AgroBase/Forms/IHM/Operacao/ucOperacaoParametros.cs
+++ b/AgroBase/AgroBase/Forms/IHM/Operacao/ucOperacaoParametros.cs
@@ -197,7 +197,7 @@ namespace AgroBase.Forms.IHM.Operacao
             if (mapaAtual != null)
             {
                 novaOperacao.Mapa = mapaAtual;
-                await mapaAtual.LimparParametrizacaoMapaAsync();
+                await mapaAtual.LimparParametrizacaoMapaAsync(novaOperacao);
             }
 
             /*
diff --git a/AgroBase/AgroBase/Forms/Operacoes/frmResultadosOperacao.Designer.cs b/AgroBase/AgroBase/Forms/Operacoes/frmResultadosOperacao.Designer.cs
index f36dac0c9..602aa38c2 100644
--- a/AgroBase/AgroBase/Forms/Operacoes/frmResultadosOperacao.Designer.cs
+++ b/AgroBase/AgroBase/Forms/Operacoes/frmResultadosOperacao.Designer.cs
@@ -367,6 +367,11 @@
             this.tabOrientacao = new System.Windows.Forms.TabPage();
             this.pnlInclinacao = new System.Windows.Forms.Panel();
             this.pnl3D = new System.Windows.Forms.Panel();
+            this.pnlOrientacaoDados = new System.Windows.Forms.Panel();
+            this.txtImuLateral = new System.Windows.Forms.TextBox();
+            this.txtImuFrontal = new System.Windows.Forms.TextBox();
+            this.label147 = new System.Windows.Forms.Label();
+            this.label146 = new System.Windows.Forms.Label();
             this.tabSonar = new System.Windows.Forms.TabPage();
             this.chbSonarDeteccoes = new System.Windows.Forms.CheckBox();
             this.tabControl3 = new System.Windows.Forms.TabControl();
@@ -531,11 +536,7 @@
             this.label36 = new System.Windows.Forms.Label();
             this.txtMomento = new System.Windows.Forms.TextBox();
             this.btnPlay = new System.Windows.Forms.Button();
-            this.pnlOrientacaoDados = new System.Windows.Forms.Panel();
-            this.label146 = new System.Windows.Forms.Label();
-            this.label147 = new System.Windows.Forms.Label();
-            this.txtImuFrontal = new System.Windows.Forms.TextBox();
-            this.txtImuLateral = new System.Windows.Forms.TextBox();
+            this.btnCarregarMapa = new System.Windows.Forms.Button();
             ((System.ComponentModel.ISupportInitialize)(this.trbMomento)).BeginInit();
             this.panel1.SuspendLayout();
             this.tabControl4.SuspendLayout();
@@ -590,6 +591,7 @@
             ((System.ComponentModel.ISupportInitialize)(this.picRobo)).BeginInit();
             this.tabOrientacao.SuspendLayout();
             this.pnl3D.SuspendLayout();
+            this.pnlOrientacaoDados.SuspendLayout();
             this.tabSonar.SuspendLayout();
             this.tabControl3.SuspendLayout();
             this.tabSonarRGB.SuspendLayout();
@@ -616,7 +618,6 @@
             this.panel6.SuspendLayout();
             this.pnlInformacoesOperacao.SuspendLayout();
             ((System.ComponentModel.ISupportInitialize)(this.picCameraCaminho)).BeginInit();
-            this.pnlOrientacaoDados.SuspendLayout();
             this.SuspendLayout();
             // 
             // trbMomento
@@ -4131,6 +4132,51 @@
             this.pnl3D.Size = new System.Drawing.Size(486, 512);
             this.pnl3D.TabIndex = 1;
             // 
+            // pnlOrientacaoDados
+            // 
+            this.pnlOrientacaoDados.Controls.Add(this.txtImuLateral);
+            this.pnlOrientacaoDados.Controls.Add(this.txtImuFrontal);
+            this.pnlOrientacaoDados.Controls.Add(this.label147);
+            this.pnlOrientacaoDados.Controls.Add(this.label146);
+            this.pnlOrientacaoDados.Location = new System.Drawing.Point(1, 362);
+            this.pnlOrientacaoDados.Name = "pnlOrientacaoDados";
+            this.pnlOrientacaoDados.Size = new System.Drawing.Size(150, 150);
+            this.pnlOrientacaoDados.TabIndex = 0;
+            // 
+            // txtImuLateral
+            // 
+            this.txtImuLateral.Location = new System.Drawing.Point(6, 87);
+            this.txtImuLateral.Name = "txtImuLateral";
+            this.txtImuLateral.Size = new System.Drawing.Size(100, 20);
+            this.txtImuLateral.TabIndex = 3;
+            this.txtImuLateral.TextAlign = System.Windows.Forms.HorizontalAlignment.Right;
+            // 
+            // txtImuFrontal
+            // 
+            this.txtImuFrontal.Location = new System.Drawing.Point(6, 32);
+            this.txtImuFrontal.Name = "txtImuFrontal";
+            this.txtImuFrontal.Size = new System.Drawing.Size(100, 20);
+            this.txtImuFrontal.TabIndex = 1;
+            this.txtImuFrontal.TextAlign = System.Windows.Forms.HorizontalAlignment.Right;
+            // 
+            // label147
+            // 
+            this.label147.AutoSize = true;
+            this.label147.Location = new System.Drawing.Point(3, 71);
+            this.label147.Name = "label147";
+            this.label147.Size = new System.Drawing.Size(39, 13);
+            this.label147.TabIndex = 2;
+            this.label147.Text = "Lateral";
+            // 
+            // label146
+            // 
+            this.label146.AutoSize = true;
+            this.label146.Location = new System.Drawing.Point(3, 16);
+            this.label146.Name = "label146";
+            this.label146.Size = new System.Drawing.Size(39, 13);
+            this.label146.TabIndex = 1;
+            this.label146.Text = "Frontal";
+            // 
             // tabSonar
             // 
             this.tabSonar.Controls.Add(this.chbSonarDeteccoes);
@@ -5875,10 +5921,10 @@
             // 
             // btnCarregarOperacao
             // 
-            this.btnCarregarOperacao.Location = new System.Drawing.Point(1180, 587);
+            this.btnCarregarOperacao.Location = new System.Drawing.Point(1138, 587);
             this.btnCarregarOperacao.Margin = new System.Windows.Forms.Padding(2);
             this.btnCarregarOperacao.Name = "btnCarregarOperacao";
-            this.btnCarregarOperacao.Size = new System.Drawing.Size(128, 28);
+            this.btnCarregarOperacao.Size = new System.Drawing.Size(115, 28);
             this.btnCarregarOperacao.TabIndex = 4;
             this.btnCarregarOperacao.Text = "Carregar Opereação";
             this.btnCarregarOperacao.UseVisualStyleBackColor = true;
@@ -5907,7 +5953,7 @@
             // 
             // btnPlay
             // 
-            this.btnPlay.Location = new System.Drawing.Point(1120, 587);
+            this.btnPlay.Location = new System.Drawing.Point(1078, 587);
             this.btnPlay.Margin = new System.Windows.Forms.Padding(2);
             this.btnPlay.Name = "btnPlay";
             this.btnPlay.Size = new System.Drawing.Size(56, 28);
@@ -5916,56 +5962,23 @@
             this.btnPlay.UseVisualStyleBackColor = true;
             this.btnPlay.Click += new System.EventHandler(this.btnPlay_Click);
             // 
-            // pnlOrientacaoDados
+            // btnCarregarMapa
             // 
-            this.pnlOrientacaoDados.Controls.Add(this.txtImuLateral);
-            this.pnlOrientacaoDados.Controls.Add(this.txtImuFrontal);
-            this.pnlOrientacaoDados.Controls.Add(this.label147);
-            this.pnlOrientacaoDados.Controls.Add(this.label146);
-            this.pnlOrientacaoDados.Location = new System.Drawing.Point(1, 362);
-            this.pnlOrientacaoDados.Name = "pnlOrientacaoDados";
-            this.pnlOrientacaoDados.Size = new System.Drawing.Size(150, 150);
-            this.pnlOrientacaoDados.TabIndex = 0;
-            // 
-            // label146
-            // 
-            this.label146.AutoSize = true;
-            this.label146.Location = new System.Drawing.Point(3, 16);
-            this.label146.Name = "label146";
-            this.label146.Size = new System.Drawing.Size(39, 13);
-            this.label146.TabIndex = 1;
-            this.label146.Text = "Frontal";
-            // 
-            // label147
-            // 
-            this.label147.AutoSize = true;
-            this.label147.Location = new System.Drawing.Point(3, 71);
-            this.label147.Name = "label147";
-            this.label147.Size = new System.Drawing.Size(39, 13);
-            this.label147.TabIndex = 2;
-            this.label147.Text = "Lateral";
-            // 
-            // txtImuFrontal
-            // 
-            this.txtImuFrontal.Location = new System.Drawing.Point(6, 32);
-            this.txtImuFrontal.Name = "txtImuFrontal";
-            this.txtImuFrontal.Size = new System.Drawing.Size(100, 20);
-            this.txtImuFrontal.TabIndex = 1;
-            this.txtImuFrontal.TextAlign = System.Windows.Forms.HorizontalAlignment.Right;
-            // 
-            // txtImuLateral
-            // 
-            this.txtImuLateral.Location = new System.Drawing.Point(6, 87);
-            this.txtImuLateral.Name = "txtImuLateral";
-            this.txtImuLateral.Size = new System.Drawing.Size(100, 20);
-            this.txtImuLateral.TabIndex = 3;
-            this.txtImuLateral.TextAlign = System.Windows.Forms.HorizontalAlignment.Right;
+            this.btnCarregarMapa.Location = new System.Drawing.Point(1257, 587);
+            this.btnCarregarMapa.Margin = new System.Windows.Forms.Padding(2);
+            this.btnCarregarMapa.Name = "btnCarregarMapa";
+            this.btnCarregarMapa.Size = new System.Drawing.Size(77, 28);
+            this.btnCarregarMapa.TabIndex = 96;
+            this.btnCarregarMapa.Text = "Mapa";
+            this.btnCarregarMapa.UseVisualStyleBackColor = true;
+            this.btnCarregarMapa.Click += new System.EventHandler(this.btnCarregarMapa_Click);
             // 
             // frmResultadosOperacao
             // 
             this.AutoScaleDimensions = new System.Drawing.SizeF(6F, 13F);
             this.AutoScaleMode = System.Windows.Forms.AutoScaleMode.Font;
             this.ClientSize = new System.Drawing.Size(1345, 632);
+            this.Controls.Add(this.btnCarregarMapa);
             this.Controls.Add(this.btnPlay);
             this.Controls.Add(this.txtMomento);
             this.Controls.Add(this.label36);
@@ -6061,6 +6074,8 @@
             ((System.ComponentModel.ISupportInitialize)(this.picRobo)).EndInit();
             this.tabOrientacao.ResumeLayout(false);
             this.pnl3D.ResumeLayout(false);
+            this.pnlOrientacaoDados.ResumeLayout(false);
+            this.pnlOrientacaoDados.PerformLayout();
             this.tabSonar.ResumeLayout(false);
             this.tabSonar.PerformLayout();
             this.tabControl3.ResumeLayout(false);
@@ -6096,8 +6111,6 @@
             this.pnlInformacoesOperacao.ResumeLayout(false);
             this.pnlInformacoesOperacao.PerformLayout();
             ((System.ComponentModel.ISupportInitialize)(this.picCameraCaminho)).EndInit();
-            this.pnlOrientacaoDados.ResumeLayout(false);
-            this.pnlOrientacaoDados.PerformLayout();
             this.ResumeLayout(false);
             this.PerformLayout();
 
@@ -6588,5 +6601,6 @@
         private System.Windows.Forms.TextBox txtImuFrontal;
         private System.Windows.Forms.Label label147;
         private System.Windows.Forms.Label label146;
+        private System.Windows.Forms.Button btnCarregarMapa;
     }
 }
\ No newline at end of file
diff --git a/AgroBase/AgroBase/Forms/Operacoes/frmResultadosOperacao.cs b/AgroBase/AgroBase/Forms/Operacoes/frmResultadosOperacao.cs
index 9be4a4842..bd4445d75 100644
--- a/AgroBase/AgroBase/Forms/Operacoes/frmResultadosOperacao.cs
+++ b/AgroBase/AgroBase/Forms/Operacoes/frmResultadosOperacao.cs
@@ -1,5 +1,6 @@
 using AgroBase.Models;
 using AgroBase.Models.Modules;
+using AgroBase.Models.Operacoes;
 using AgroBase.Models.Operadores;
 using AgroBase.Properties;
 using AgroBase.Services;
@@ -12,6 +13,7 @@ using System.Drawing.Drawing2D;
 using System.Globalization;
 using System.IO;
 using System.Linq;
+using System.Text;
 using System.Threading.Tasks;
 using System.Windows.Forms;
 
@@ -53,9 +55,14 @@ namespace AgroBase.Forms.Operacoes
         
         DateTime MomentoAtual = DateTime.Now;
         AsyncTaskTimerModel tmrPlay;
+
+        private List<List<GPSModel>> _ruasMapaReplay = new List<List<GPSModel>>();
+        private List<PontoTrajetoriaModel> _trajetoriaFixaReplay = new List<PontoTrajetoriaModel>();
+
         MapasModel Mapa = new MapasModel();
-        private Visualizador3DService visualizador3D;
+        //TrajetoriaMapaOperacaoModel Trajetoria = new TrajetoriaMapaOperacaoModel(new List<List<GPSModel>>());
         private MapaDinamicoModel MapaDinamico;
+        private Visualizador3DService visualizador3D;
         private int idxMomentoAtual
         {
             get
@@ -165,12 +172,46 @@ namespace AgroBase.Forms.Operacoes
                 LogsOperacao = JsonConvert.DeserializeObject<OperacaoSensoriamentoLogModel[]>(OperacaoModel.DeserializarDadosOperacao(ofd.FileName, data.operacao)).ToList();
                 LogsGPS = JsonConvert.DeserializeObject<GPSModel[]>(OperacaoModel.DeserializarDadosOperacao(ofd.FileName, Enums.T_Code.Gps.ToString())).ToList();
 
-                GPSModel pos = GPSService.UltimaLeitura.Clone();
-                GPSService.UltimaLeitura = LogsGPS.FirstOrDefault();
-                bool sucesso = op.CarregarParametrizacaoOperacao(null, null, ofd.FileName.Replace("data.lgop", "operacao.opr"));
-                bool condicaoOperacaoCarregada() => op.Trajetoria?.CorredorAtual != null;
-                bool resultado = await FuncoesGlobais.AguardarCondicaoAsync(condicaoOperacaoCarregada, 15000, 200);
-                GPSService.UltimaLeitura = pos;
+
+                string caminhoOpr = Path.Combine(
+                    Path.GetDirectoryName(ofd.FileName),
+                    "operacao.opr"
+                );
+
+                OperacaoParametrosModel parametrosOperacao = null;
+
+                if (File.Exists(caminhoOpr))
+                {
+                    parametrosOperacao =
+                        JsonConvert.DeserializeObject<OperacaoParametrosModel>(
+                            File.ReadAllText(caminhoOpr, Encoding.UTF8)
+                        );
+
+                    if (parametrosOperacao != null)
+                    {
+                        GPSModel posAnterior =
+                            GPSService.UltimaLeitura?.Clone();
+
+                        try
+                        {
+                            // Importantíssimo:
+                            // a trajetória será projetada tomando como referência
+                            // a posição inicial real daquela operação.
+                            GPSService.UltimaLeitura =
+                                LogsGPS.FirstOrDefault()?.Clone();
+
+                            Variaveis.OperacaoEmAndamento = OperacaoModel.CarregarParamerosOperacaoBase(
+                                parametrosOperacao
+                            );
+                        }
+                        finally
+                        {
+                            if (posAnterior != null)
+                                GPSService.UltimaLeitura = posAnterior;
+                        }
+                    }
+                }
+
 
                 LogsControle = JsonConvert.DeserializeObject<OperacaoControleModel[]>(OperacaoModel.DeserializarDadosOperacao(ofd.FileName, "controle")).ToList();
                 LogsCooler = JsonConvert.DeserializeObject<CoolerControlDataModel[]>(OperacaoModel.DeserializarDadosOperacao(ofd.FileName, "cooler")).ToList();
@@ -239,7 +280,7 @@ namespace AgroBase.Forms.Operacoes
                 };
                 MapasService.CriarArquivoMapa(new List<List<GPSModel>>() { LogsGPS.ToList() }, 1517, data.id);
                 Mapa.btnCarregar_Click(sender, e, Path.Combine(Variaveis.CaminhoSistema, Variaveis.CaminhoMapasConvertidos + data.id + ".json"));
-
+                
                 trbMomento.Maximum = LogsOperacao.Count - 1;
                 trbMomento.Minimum = 0;
                 trbMomento.Value = 0;
@@ -1309,28 +1350,60 @@ namespace AgroBase.Forms.Operacoes
 
             pnlOrientacaoGPS.Invalidate();
 
-            var _trajaetoria = LogsGPS.Select(x => new PontoTrajetoriaModel(Enums.TipoPontoRua.Indefinido)
-            {
-                Posicao = x
-            })
-            .ToList();
-
-
             var LogOpe = LogsOperacao[idxMomentoAtual];
             var LogSnr = LogsVisualWorker[idxMomentoAtual];
             var LogTrj = LogsTrajetoria[idxMomentoAtual];
             var LogCtrl = LogsControle[idxMomentoAtual];
 
+
+            var trajetoriaFixa =
+                (_trajetoriaFixaReplay?.Count ?? 0) > 0
+                    ? _trajetoriaFixaReplay
+                    : Variaveis.OperacaoEmAndamento
+                        ?.Trajetoria
+                        ?._TrajetoriaFixa
+                        ?? new List<PontoTrajetoriaModel>();
+
+            var trajetoriaDinamica =
+                MontarTrajetoriaDinamicaReplay(
+                    Log,
+                    LogTrj
+                );
+
+            var ruasMapa =
+                (_ruasMapaReplay?.Count ?? 0) > 0
+                    ? _ruasMapaReplay
+                    : Variaveis.OperacaoEmAndamento
+                        ?.Mapa
+                        ?.TrajetoriaMapa
+                        ?? new List<List<GPSModel>>();
+
+            var rastroRover =
+                LogsGPS
+                    .Take(idxMomentoAtual + 1)
+                    .ToList();
+
+
+            Variaveis.OperacaoEmAndamento.Trajetoria.LoopAtualizaDados();
             MapaDinamico.AtualizarDados(
-                //(LogSnr?.Iniciado ?? false) && LogSnr.Resumo.DirecaoDesvio != Enums.Direcao.Frente ? LogSnr.Resumo.ObstaculoCritico : null,
-                null, 
+                null,
                 (float)LogTrj.AnguloCaminho,
                 (float)Log.AnguloCarroDefinido,
                 Log,
-                _trajaetoria,
-                Variaveis.OperacaoEmAndamento.Trajetoria?._TrajetoriaFixa,
-                new List<GPSModel>(LogsGPS.Take(LogsGPS.IndexOf(Log) + 1)),
-                Variaveis.OperacaoEmAndamento.Mapa.TrajetoriaMapa,
+
+                // Roxo: trajetória ainda restante naquele instante
+                trajetoriaDinamica,
+
+                // Verde/azul: trajetória fixa completa projetada
+                trajetoriaFixa,
+
+                // Laranja: caminho realmente percorrido até aquele instante
+                rastroRover,
+
+                // Azul/preto: linhas originais do mapa
+                ruasMapa,
+
+                // Verde tracejado: simulação MPC daquele frame
                 LogCtrl.SimulacaoMPC
             );
         }
@@ -2334,6 +2407,48 @@ namespace AgroBase.Forms.Operacoes
                 }
             }
         }
+
+        private async void btnCarregarMapa_Click(object sender, EventArgs e)
+        {
+            
+        }
+
+        private List<PontoTrajetoriaModel> MontarTrajetoriaDinamicaReplay(GPSModel posicaoAtual, OperacaoSensoriamentoLogTrajetoriaModel logTrajetoria)
+        {
+            var resultado = new List<PontoTrajetoriaModel>();
+
+            if (posicaoAtual != null)
+            {
+                resultado.Add(
+                    new PontoTrajetoriaModel(Enums.TipoPontoRua.PosicaoRobo)
+                    {
+                        Posicao = posicaoAtual,
+                        LarguraCorredor = 1.0,
+                        idxCorredor = logTrajetoria?.CorredorAtual?.Idx ?? 0,
+                        Visitado = true
+                    }
+                );
+            }
+
+            if (_trajetoriaFixaReplay == null ||
+                _trajetoriaFixaReplay.Count == 0)
+            {
+                return resultado;
+            }
+
+            int idxProximo =
+                logTrajetoria?.ProximoPonto?.idxPonto ?? 0;
+
+            resultado.AddRange(
+                _trajetoriaFixaReplay
+                    .Where(x =>
+                        x != null &&
+                        x.idxPonto >= idxProximo)
+            );
+
+            return resultado;
+        }
+
     }
 
 }
diff --git a/AgroBase/AgroBase/Forms/Simuladores/frmTreinamentoIA.cs b/AgroBase/AgroBase/Forms/Simuladores/frmTreinamentoIA.cs
index d34104928..9d220bddc 100644
--- a/AgroBase/AgroBase/Forms/Simuladores/frmTreinamentoIA.cs
+++ b/AgroBase/AgroBase/Forms/Simuladores/frmTreinamentoIA.cs
@@ -88,7 +88,7 @@ namespace AgroBase.Forms
             txtAnguloGPS.Text = anguloInicial.ToString("0.00");
             txtDistanciaGPS.Text = distanciaInicial.ToString("0.00");
             Variaveis.OperacaoEmAndamento.GPSTrajetoria = new System.Collections.Generic.List<GPSModel>();
-            Variaveis.OperacaoEmAndamento.Trajetoria = new TrajetoriaMapaOperacaoModel(Variaveis.OperacaoEmAndamento.Mapa.TrajetoriaMapa);
+            Variaveis.OperacaoEmAndamento.Trajetoria = new TrajetoriaMapaOperacaoModel(Variaveis.OperacaoEmAndamento, Variaveis.OperacaoEmAndamento.Mapa.TrajetoriaMapa);
             Variaveis.OperacaoEmAndamento.Trajetoria.ProjetarTrajetoriaFixa();
             _robotState = new TreinamentoIAModel();
         }
diff --git a/AgroBase/AgroBase/Models/MapasModel.cs b/AgroBase/AgroBase/Models/MapasModel.cs
index ae9cb907a..8b83102db 100644
--- a/AgroBase/AgroBase/Models/MapasModel.cs
+++ b/AgroBase/AgroBase/Models/MapasModel.cs
@@ -344,7 +344,7 @@ namespace AgroBase.Models
             _mapaEnviado = true;
         }
 
-        public void DefinirRuasSelecionadas(IEnumerable<string> ruas, bool popularTrajetoria = false)
+        public void DefinirRuasSelecionadas(OperacaoModel op, IEnumerable<string> ruas, bool popularTrajetoria = false)
         {
             RuasPercorrer =
                 ruas == null
@@ -361,12 +361,12 @@ namespace AgroBase.Models
             }
 
             if (popularTrajetoria)
-                PopularTrajetoriaMapa();
+                PopularTrajetoriaMapa(op);
         }
 
-        public async Task DefinirRuasSelecionadasAsync(IEnumerable<string> ruas, bool popularTrajetoria = false)
+        public async Task DefinirRuasSelecionadasAsync(OperacaoModel op, IEnumerable<string> ruas, bool popularTrajetoria = false)
         {
-            DefinirRuasSelecionadas(ruas, popularTrajetoria);
+            DefinirRuasSelecionadas(op, ruas, popularTrajetoria);
 
             if (!_paginaCarregada || browser?.CoreWebView2 == null)
             {
@@ -378,11 +378,11 @@ namespace AgroBase.Models
             await browser.CoreWebView2.ExecuteScriptAsync("definirSelecaoRuas(" + ruasJson + ");");
         }
 
-        public Task DefinirRuasSelecionadasAsync(IEnumerable<int> ruas)
+        public Task DefinirRuasSelecionadasAsync(OperacaoModel op, IEnumerable<int> ruas)
         {
             IEnumerable<string> convertidas = ruas == null ? null : ruas.Select(x => x.ToString());
 
-            return DefinirRuasSelecionadasAsync(convertidas);
+            return DefinirRuasSelecionadasAsync(op, convertidas);
         }
 
         private bool PodeSelecionarRuas()
@@ -424,7 +424,7 @@ namespace AgroBase.Models
                             .Distinct()
                             .ToList();
 
-                PopularTrajetoriaMapa();
+                PopularTrajetoriaMapa(Variaveis.OperacaoEmAndamento);
                 RuasSelecionadasAlteradas?.Invoke(this, EventArgs.Empty);
             }
             catch (Exception ex)
@@ -433,7 +433,7 @@ namespace AgroBase.Models
             }
         }
 
-        public void PopularTrajetoriaMapa(MapaFeatureCollectionModel dados = null)
+        public void PopularTrajetoriaMapa(OperacaoModel op, MapaFeatureCollectionModel dados = null)
         {
             if (dados != null)
                 mapaService.DefinirMapa(dados, mapaService.NomeArquivos ?? "Mapa");
@@ -512,14 +512,12 @@ namespace AgroBase.Models
                 ruasMontadas.Add(rua);
             }
 
-            OperacaoModel op = Variaveis.OperacaoEmAndamento;
-
             if (op == null)
                 throw new InvalidOperationException("Operação indisponível para instalar a trajetória.");
 
             var trajetoriaAnterior = op.Trajetoria;
             var mapaAnterior = TrajetoriaMapa;
-            var novaTrajetoria = new TrajetoriaMapaOperacaoModel(ruasMontadas);
+            var novaTrajetoria = new TrajetoriaMapaOperacaoModel(op, ruasMontadas);
 
             try
             {
@@ -647,9 +645,9 @@ namespace AgroBase.Models
             pnlMapa = null;
         }
 
-        public async Task LimparParametrizacaoMapaAsync()
+        public async Task LimparParametrizacaoMapaAsync(OperacaoModel op)
         {
-            DefinirRuasSelecionadas(new List<string>(), false);
+            DefinirRuasSelecionadas(op, new List<string>(), false);
             if (_paginaCarregada && browser?.CoreWebView2 != null)
             {
                 await browser.CoreWebView2.ExecuteScriptAsync("definirSelecaoRuas([]);");
diff --git a/AgroBase/AgroBase/Models/Operacoes/OperacaoModel.cs b/AgroBase/AgroBase/Models/Operacoes/OperacaoModel.cs
index 8686b9426..9e6c6ebee 100644
--- a/AgroBase/AgroBase/Models/Operacoes/OperacaoModel.cs
+++ b/AgroBase/AgroBase/Models/Operacoes/OperacaoModel.cs
@@ -476,7 +476,6 @@ namespace AgroBase.Models
                 Simulando = false,
                 GPSTrajetoria = new List<GPSModel>(),
                 Mapa = new MapasModel(),
-                Trajetoria = new TrajetoriaMapaOperacaoModel(new List<List<GPSModel>>()),
                 ControleAnterior = new OperacaoControleModel(),
                 Parametros = new OperacaoParametrosModel()
                 {
@@ -499,10 +498,12 @@ namespace AgroBase.Models
                     }
                 }
             };
+            op.Trajetoria = new TrajetoriaMapaOperacaoModel(op, new List<List<GPSModel>>());
             op.Sensoriamento = new OperacaoSensoriamentoConjuntoModel(op)
             {
                 Operacao = new OperacaoSensoriamentoLogModel()
                 {
+                    Modo = Modo,
                     OperacaoIniciada = false,
                     DataInicio = DateTime.MinValue,
                     DataFim = DateTime.MinValue
@@ -559,7 +560,7 @@ namespace AgroBase.Models
                         AtuPercentualErvasBicoOff = 2,
                         AtuPercentualErvasBicoOn = 1,
                         AtuPercentualInicioPulverizacao = 85,
-                        AtuPressaoLinha = 18,
+                        AtuPressaoLinha = 22,
                         AtuAgitadorModo = ModoAgitadorCalda.SemAgitacao,
                         AtuModoControle = ModoControleBomba.PID,
                         AtuModeloCabeca = CabecasModelo.Target,
@@ -666,14 +667,14 @@ namespace AgroBase.Models
                         DirVelocidadeMovimento = 80,
                         DirTipoMovimento = TiposControladorDirecional.MPC,
 
-                        DistanciaManobra = 3.0,
+                        DistanciaManobra = 2.2,
 
                         AtuAlturaAreaPulverizacao = 10,
                         AtuDuracaoAtuacao = 0,
                         AtuPercentualErvasBicoOff = 2,
                         AtuPercentualErvasBicoOn = 1,
                         AtuPercentualInicioPulverizacao = 85,
-                        AtuPressaoLinha = 18,
+                        AtuPressaoLinha = 22,
                         AtuAgitadorModo = ModoAgitadorCalda.Continuo,
                         AtuModoControle = ModoControleBomba.PID,
                         AtuModeloCabeca = CabecasModelo.Target,
@@ -1312,9 +1313,9 @@ namespace AgroBase.Models
             if (parametros.Mapa != null)
             {
                 novaOp.Mapa = new MapasModel();
-                novaOp.Mapa.DefinirRuasSelecionadas(ruasSelecionadas);
+                novaOp.Mapa.DefinirRuasSelecionadas(novaOp, ruasSelecionadas);
                 novaOp.Mapa.TipoMapa = novaOp.Parametros.TipoMapa;
-                novaOp.Mapa.PopularTrajetoriaMapa(parametros.Mapa);
+                novaOp.Mapa.PopularTrajetoriaMapa(novaOp, parametros.Mapa);
             }
 
             novaOp?.DispAtu?.Dados?.PrepararAtuadorParaOperacao();
diff --git a/AgroBase/AgroBase/Models/TrajetoriaMapaOperacaoModel.cs b/AgroBase/AgroBase/Models/TrajetoriaMapaOperacaoModel.cs
index 1b9d1c39e..a539806c2 100644
--- a/AgroBase/AgroBase/Models/TrajetoriaMapaOperacaoModel.cs
+++ b/AgroBase/AgroBase/Models/TrajetoriaMapaOperacaoModel.cs
@@ -16,8 +16,9 @@ namespace AgroBase.Models
 {
     public class TrajetoriaMapaOperacaoModel
     {
-        public TrajetoriaMapaOperacaoModel(List<List<GPSModel>> RuasMapa)
+        public TrajetoriaMapaOperacaoModel(OperacaoModel _op, List<List<GPSModel>> RuasMapa)
         {
+            op = _op;
             RuasPlantacao = ClonarRuasMapa(RuasMapa);
             AutonomiaCorredor = new AutonomiaCorredorModel();
         }
@@ -47,12 +48,15 @@ namespace AgroBase.Models
         private const double ToleranciaCoordenada = 1e-12;
         private const double DistanciaMaximaSaltoMapaM = 25.0;
 
+        [JsonIgnore]
+        private readonly OperacaoModel op;
+
         #region PARAMETROS
 
         public double AnguloAberturaCurva { get; set; } = 25; // Angulo usado para deslocar o ponto de curva
         public double DistanciaProjecaoRua =>
             (VariaveisEquipamento.DistanciaEntreEixosCm / 100.0 / 2.0) +
-            (Variaveis.OperacaoEmAndamento?.Parametros?.Controle?.DistanciaManobra ?? 3.0); // Distancia para projetar o primeiro ponto para fora do corredor
+            (op?.Parametros?.Controle?.DistanciaManobra ?? 3.0); // Distancia para projetar o primeiro ponto para fora do corredor
         public static double DistanciaEntrePontos { get; set; } = 0.8; // Distancia entre os pontos dentro do corredor
         public static double DistanciaEntrePontosCurva { get; set; } = 0.25; // Distancia entre os pontos durante a curva entre corredores
         private double DistanciaManobraEntreRuas { get; set; } = 3.0; // Distancia máxima para gerar a curva de conexão entre os corredores
@@ -318,7 +322,7 @@ namespace AgroBase.Models
         public static double DistanciaMaximaEntreLeituras { get; private set; }
         public void AtualizarDistanciaMaximaEntreLeituras()
         {
-            double percentualVelocidade = Variaveis.OperacaoEmAndamento?.Controle?.PercentualVelocidadeSP ?? 0;
+            double percentualVelocidade = op?.Controle?.PercentualVelocidadeSP ?? 0;
             double velocidadeCarroMs = FuncoesMatematicas.CalculaVelocidadeMsPercentual(percentualVelocidade);
             double distanciaMaxima = velocidadeCarroMs / Math.Max(1, GPSService.TaxaAmostragemHz);
 
@@ -673,7 +677,7 @@ namespace AgroBase.Models
                 if (CorredorAtual != null)
                 {
                     bool operacaoEmAndamento =
-                        Variaveis.OperacaoEmAndamento?.Sensoriamento?.Operacao?.StatusOperacaoAtual ==
+                        op?.Sensoriamento?.Operacao?.StatusOperacaoAtual ==
                         StatusOperacao.EmAndamento;
 
                     if ((RetornandoBase && CorredorAtual.Dentro) || (!RetornandoBase && operacaoEmAndamento))
@@ -784,8 +788,6 @@ namespace AgroBase.Models
         public string TempoEstimadoRestante { get; private set; }
         public void AtualizarTempoEstimado(double? velocidadeSemErvasMs = null, double? velocidadeComErvasMs = null)
         {
-            var op = Variaveis.OperacaoEmAndamento;
-
             if (op?.Sensoriamento == null || op?.Parametros?.Controle == null)
             {
                 TempoEstimadoOperacao = "00:00:00";
@@ -871,8 +873,6 @@ namespace AgroBase.Models
             if (CorredorAtual.Idx > 0) return;
             if (!CorredorAtual.Dentro) return;
 
-            var op = Variaveis.OperacaoEmAndamento;
-
             if (op?.Sensoriamento == null)
                 return;
 
@@ -952,7 +952,7 @@ namespace AgroBase.Models
             {
                 MarcarVisitadoAte(_TrajetoriaFixa.Count - 1);
 
-                Variaveis.OperacaoEmAndamento.Sensoriamento?.InserirLog(
+                op.Sensoriamento?.InserirLog(
                     T_Code.Trj,
                     StatusModulo.Operante,
                     100,
@@ -976,7 +976,6 @@ namespace AgroBase.Models
         }
         private void AtualizarErroCombinado()
         {
-            var op = Variaveis.OperacaoEmAndamento;
             var gps = GPSPosicaoAtual;
 
             if (op?.Parametros?.Controle == null || gps == null || PontoAtual?.Posicao == null || ProximoPonto?.Posicao == null)
@@ -1139,6 +1138,10 @@ namespace AgroBase.Models
 
             for (int i = idxInicial; i <= idxFinal; i++)
             {
+                if (PontoAtual.idxPonto == _TrajetoriaFixa[i].idxPonto)
+                {
+
+                }
                 _TrajetoriaFixa[i].AtualizarPropriedades(
                     _gpsAtualCiclo,
                     _gpsAnteriorCiclo,
@@ -1261,7 +1264,7 @@ namespace AgroBase.Models
             if (!confirmou)
                 return;
 
-            Variaveis.OperacaoEmAndamento.Sensoriamento?.InserirLog(
+            op.Sensoriamento?.InserirLog(
                 T_Code.Trj,
                 StatusModulo.Operante,
                 100,
@@ -1468,7 +1471,7 @@ namespace AgroBase.Models
 
         private bool VerificaInicioOperacaoMeioRua(bool considerarStatus)
         {
-            if (_TrajetoriaFixaDefinida && !VerificacaoInicialMeioRuaConcluida && ((considerarStatus && new List<StatusOperacao>() { StatusOperacao.Aguardando, StatusOperacao.EmAndamento }.Contains(Variaveis.OperacaoEmAndamento.Sensoriamento.Operacao.StatusOperacaoAtual)) || !considerarStatus))
+            if (_TrajetoriaFixaDefinida && !VerificacaoInicialMeioRuaConcluida && ((considerarStatus && new List<StatusOperacao>() { StatusOperacao.Aguardando, StatusOperacao.EmAndamento }.Contains(op.Sensoriamento.Operacao.StatusOperacaoAtual)) || !considerarStatus))
             {
                 var _posicaoAtual = GPSPosicaoAtual;
                 if (_posicaoAtual == null)
@@ -1487,7 +1490,7 @@ namespace AgroBase.Models
                 if (!double.IsNaN(distanciaUltimoPonto) && !double.IsInfinity(distanciaUltimoPonto) && distanciaUltimoPonto < 5.0)
                 {
                     msg = $"Início no meio da rua ignorado: robô está a {distanciaUltimoPonto:0.00}m do fim. Nova operação começará do zero.";
-                    Variaveis.OperacaoEmAndamento?.Sensoriamento?.InserirLog(T_Code.Trj, StatusModulo.Operante, 100, msg);
+                    op?.Sensoriamento?.InserirLog(T_Code.Trj, StatusModulo.Operante, 100, msg);
                     Variaveis.MostrarLog($"[TRJ] {msg}");
 
                     return false;
@@ -1504,7 +1507,7 @@ namespace AgroBase.Models
                     )
                     {
                         msg = "Inicio no meio da rua rejeitado: geometria e trajetoria discordam sobre o corredor atual.";
-                        Variaveis.OperacaoEmAndamento?.Sensoriamento?.InserirLog(T_Code.Trj, StatusModulo.Falha, 0, msg);
+                        op?.Sensoriamento?.InserirLog(T_Code.Trj, StatusModulo.Falha, 0, msg);
                         Variaveis.MostrarLog("[TRJ] " + msg);
                         throw new InvalidOperationException(msg);
                     }
@@ -1517,14 +1520,14 @@ namespace AgroBase.Models
                     CorredorAtual?.AtualizarDados();
 
                     msg = $"Início no meio da rua confirmado: idx={idxPontoMaisProximo}/{_TrajetoriaFixa.Count - 1}, marcados={idxPontoMaisProximo + 1}/{_TrajetoriaFixa.Count}.";
-                    Variaveis.OperacaoEmAndamento?.Sensoriamento?.InserirLog(T_Code.Trj, StatusModulo.Alerta, 100, msg);
+                    op?.Sensoriamento?.InserirLog(T_Code.Trj, StatusModulo.Alerta, 100, msg);
                     Variaveis.MostrarLog($"[TRJ] {msg}");
 
                     return true;
                 }
 
                 msg = $"Início no meio da rua ignorado: robô está fora do corredor. Nova operação começará do zero.";
-                Variaveis.OperacaoEmAndamento?.Sensoriamento?.InserirLog(T_Code.Trj, StatusModulo.Operante, 100, msg);
+                op?.Sensoriamento?.InserirLog(T_Code.Trj, StatusModulo.Operante, 100, msg);
                 Variaveis.MostrarLog($"[TRJ] {msg}");
             }
 
@@ -1534,7 +1537,6 @@ namespace AgroBase.Models
 
         private void VerificaFimCorredorSegmentacao()
         {
-            var op = Variaveis.OperacaoEmAndamento;
             double limiteFim = op?.Parametros?.Controle?.AnteciparManobraCorredorM ?? 0;
 
             if (
@@ -1972,7 +1974,7 @@ namespace AgroBase.Models
                 catch (Exception ex)
                 {
                     Variaveis.MostrarLog("[TRJ] Falha no ciclo da trajetoria: " + ex.Message);
-                    Variaveis.OperacaoEmAndamento?.Sensoriamento?.InserirLog(
+                    op?.Sensoriamento?.InserirLog(
                         T_Code.Trj,
                         StatusModulo.Falha,
                         0,
@@ -2005,8 +2007,6 @@ namespace AgroBase.Models
 
         private void LoopAtualizaDadosCore()
         {
-            var op = Variaveis.OperacaoEmAndamento;
-
             if (!(op?.Parametros?.ControleAutomatico ?? false))
                 return;
 
@@ -2023,8 +2023,6 @@ namespace AgroBase.Models
 
         private void AtualizarDadosControleRedis()
         {
-            var op = Variaveis.OperacaoEmAndamento;
-
             if (op == null || !op.Sensoriamento.Operacao.OperacaoIniciada)
             {
                 return;
@@ -2270,7 +2268,7 @@ namespace AgroBase.Models
                 DefinirPontoAtual();
                 DefinirProximoPonto();
 
-                Variaveis.OperacaoEmAndamento.Sensoriamento?.InserirLog(
+                op.Sensoriamento?.InserirLog(
                     T_Code.Trj,
                     StatusModulo.Operante,
                     100,
@@ -3203,7 +3201,7 @@ namespace AgroBase.Models
 
             List<List<GPSModel>> ruasParaProjetar;
 
-            if (Variaveis.OperacaoEmAndamento?.Mapa?.TipoMapa == TipoMapaOperacao.Corredores)
+            if (op?.Mapa?.TipoMapa == TipoMapaOperacao.Corredores)
             {
                 throw new NotSupportedException(
                     "O tipo de mapa Corredores ainda nao possui geracao validada para campo. " +
@@ -3679,7 +3677,7 @@ namespace AgroBase.Models
 
         private void ProjetarTrajetoriaFixaCore(GPSModel posicaoRobo)
         {
-            if (Variaveis.OperacaoEmAndamento?.Sensoriamento?.Operacao?.Modo != ModoOperacao.MapaGPS)
+            if (op?.Parametros?.Modo != ModoOperacao.MapaGPS)
                 return;
 
             List<PontoTrajetoriaModel> _trajetoriaFixa = new List<PontoTrajetoriaModel>();
@@ -3693,7 +3691,7 @@ namespace AgroBase.Models
             catch (Exception ex)
             {
                 Variaveis.MostrarLog($"[TRJ] Trajetória rejeitada: {ex.Message}");
-                Variaveis.OperacaoEmAndamento?.Sensoriamento?.InserirLog(T_Code.Trj, StatusModulo.Falha, 0, $"Trajetória rejeitada: {ex.Message}");
+                op?.Sensoriamento?.InserirLog(T_Code.Trj, StatusModulo.Falha, 0, $"Trajetória rejeitada: {ex.Message}");
 
                 throw;
             }
@@ -3810,7 +3808,8 @@ namespace AgroBase.Models
                 Corredores[idx] = CorredorAtual;
 
                 double anguloProjetar1 = GPSUtils.CalcularOrientacao(CorredorAtual.Skip(1).FirstOrDefault(), CorredorAtual.FirstOrDefault());
-                GPSModel PrimeiroPonto = GPSUtils.ProjetarPontoDeslocado(CorredorAtual.FirstOrDefault(), DistanciaProjecaoRua * 0.8, anguloProjetar1);
+                double percentualAcrescimoPontoAberturaCurvaPrimeiroPonto = 0.7;
+                GPSModel PrimeiroPonto = GPSUtils.ProjetarPontoDeslocado(CorredorAtual.FirstOrDefault(), DistanciaProjecaoRua * percentualAcrescimoPontoAberturaCurvaPrimeiroPonto, anguloProjetar1);
 
                 var ultimoPontoTrajetoria = _trajetoriaFixa.LastOrDefault();
 
@@ -3823,7 +3822,7 @@ namespace AgroBase.Models
                     Direcao = direcaoAtual,
                     Orientacao = GPSUtils.CalcularOrientacao(ultimoPontoTrajetoria.Posicao, PrimeiroPonto),
                     Visitado = false,
-                    LarguraCorredor = larguraCorredorMenor
+                    LarguraCorredor = LarguraCorredorPadrao // larguraCorredorMenor
                 };
 
                 _trajetoriaFixa.Add(PontoInicial);
@@ -3849,8 +3848,8 @@ namespace AgroBase.Models
                     _trajetoriaFixa.Add(PontoTrajetoria);
                 }
 
-                double percentualAcrescimoPontoAberturaCurva = 0.8;
-                double distancia_projetar = !ultimoCorredor ? (DistanciaProjecaoRua * percentualAcrescimoPontoAberturaCurva) : DistanciaProjecaoRua;
+                double percentualAcrescimoPontoAberturaCurvaUltimoPonto = 0.8;
+                double distancia_projetar = !ultimoCorredor ? (DistanciaProjecaoRua * percentualAcrescimoPontoAberturaCurvaUltimoPonto) : DistanciaProjecaoRua;
 
                 GPSModel ultimoPontoCorredorAtual = CorredorAtual.Last();
                 double anguloProjetar2 = GPSUtils.CalcularOrientacao(CorredorAtual.Skip(CorredorAtual.Count() - 2).FirstOrDefault(), ultimoPontoCorredorAtual);
@@ -3882,7 +3881,7 @@ namespace AgroBase.Models
                     Direcao = direcaoAtual,
                     Orientacao = GPSUtils.CalcularOrientacao(_trajetoriaFixa.Last().Posicao, UltimoPonto),
                     Visitado = false,
-                    LarguraCorredor = larguraCorredorMenor // ultimoCorredor ? larguraCorredorMenor : (larguraCorredorMenor * 0.8)
+                    LarguraCorredor = LarguraCorredorPadrao * 1.1 // larguraCorredorMenor // ultimoCorredor ? larguraCorredorMenor : (larguraCorredorMenor * 0.8)
                 };
                 _trajetoriaFixa.Add(PontoFinal);
 
@@ -4376,7 +4375,6 @@ namespace AgroBase.Models
                 double distanciaTotalAnterior = DistanciaTotal;
                 double distanciaPercorridaAnterior = DistanciaPercorrida;
                 var autonomiaAnterior = AutonomiaCorredor?.Clone();
-                var op = Variaveis.OperacaoEmAndamento;
                 var parametrosAnteriores = op?.Parametros;
                 bool simulandoAnterior = op?.Simulando ?? false;
                 bool? operacaoIniciadaAnterior = op?.Sensoriamento?.Operacao?.OperacaoIniciada;
@@ -4407,7 +4405,7 @@ namespace AgroBase.Models
                     }
 
                     Variaveis.MostrarLog("[RETORNO] Trajetoria rejeitada: " + ex.Message);
-                    Variaveis.OperacaoEmAndamento?.Sensoriamento?.InserirLog(
+                    op?.Sensoriamento?.InserirLog(
                         T_Code.Trj,
                         StatusModulo.Falha,
                         0,
@@ -4649,7 +4647,7 @@ namespace AgroBase.Models
             AtualizarTrajetoriaDinamicaCore();
             AtualizarDistanciaRestante();
 
-            Variaveis.OperacaoEmAndamento.Sensoriamento?.InserirLog(T_Code.Trj, StatusModulo.Operante, 100, $"[RETORNO] Trajetória aplicada ({motivo}) - Pontos: {traj.Count}");
+            op.Sensoriamento?.InserirLog(T_Code.Trj, StatusModulo.Operante, 100, $"[RETORNO] Trajetória aplicada ({motivo}) - Pontos: {traj.Count}");
         }
 
         private List<GPSModel> CortarFinalDaRota(List<GPSModel> pontos, double distanciaAntesDoFimM)
@@ -4689,8 +4687,6 @@ namespace AgroBase.Models
 
         private void AtualizarDadosOperacaoParaRetorno(List<GPSModel> waypointsManual)
         {
-            var op = Variaveis.OperacaoEmAndamento;
-
             if (op == null)
                 throw new InvalidOperationException("Operacao indisponivel ao iniciar retorno.");
 
@@ -4784,7 +4780,7 @@ namespace AgroBase.Models
                  */
                 if (!ReferenceEquals(
                     op,
-                    Variaveis.OperacaoEmAndamento))
+                    op))
                 {
                     Variaveis.MostrarLog(
                         "[RETORNO/MAPSYNC] Operação mudou durante sincronização."
@@ -5971,7 +5967,7 @@ namespace AgroBase.Models
                  */
                 limiteDistancia = Math.Min(limiteDistancia, 0.80);
 
-                if (NaMargem && DistanciaAtual <= limiteDistancia)
+                if (NaMargem) // && DistanciaAtual <= limiteDistancia
                 {
                     Visitado = true;
                 }
diff --git a/AgroBase/AgroBase/Services/Operadores/HealthWorkerService.cs b/AgroBase/AgroBase/Services/Operadores/HealthWorkerService.cs
index 14b8bac52..b5da57be5 100644
--- a/AgroBase/AgroBase/Services/Operadores/HealthWorkerService.cs
+++ b/AgroBase/AgroBase/Services/Operadores/HealthWorkerService.cs
@@ -324,7 +324,7 @@ namespace AgroBase.Services.Operadores
                         {
                             percent_vel_max = pControle.MovVelocidadeSErvasPercent,
                             percent_vel_min = pControle.MovVelocidadeCErvasPercent,
-                            percent_vel_curva = 20.0,
+                            percent_vel_curva = 15.0,
                             rampa_partida_pct = VariaveisEquipamento.PercentualVelMin,
                             rampa_aceleracao_pct_s = 12.0,
                             rampa_desaceleracao_pct_s = 18.0,
@@ -548,7 +548,7 @@ namespace AgroBase.Services.Operadores
                         $"[MAPSYNC] Tentativa={tentativa} | " +
                         $"cmd={commandId} | " +
                         $"rev={mapRevision} | " +
-                        $"pontos={op.Trajetoria._TrajetoriaFixa.Count}"
+                        $"pontos={op.Trajetoria._TrajetoriaFixa?.Count}"
                     );
 
                     bool atualizado =
diff --git a/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/camera_worker/camera_oak.py b/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/camera_worker/camera_oak.py
index 4cdb02b73..8ad2dfdbb 100644
--- a/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/camera_worker/camera_oak.py
+++ b/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/camera_worker/camera_oak.py
@@ -419,6 +419,7 @@ class CameraOak:
                             t0 = time.time()
                             frame = pkt.getCvFrame()
                             dur = time.time() - t0
+                            frame_ts_host = time.time()
 
                             resultado = {
                                 "erro": None,
@@ -450,7 +451,7 @@ class CameraOak:
                                 self.perf.tick(
                                     "camera_rgb",
                                     latencia_ms=dur * 1000.0,
-                                    frame_ts_host=self._rgb_cache_ts,
+                                    frame_ts_host=frame_ts_host,
                                     frame_ts_device=pkt_ts_device,
                                     seq=pkt_seq,
                                     seq_delta=seq_delta,
@@ -461,7 +462,7 @@ class CameraOak:
 
                             with self._cache_lock:
                                 self._rgb_cache = frame
-                                self._rgb_cache_ts = time.time()
+                                self._rgb_cache_ts = frame_ts_host
                                 self._rgb_cache_resultado = resultado
                                 self.timestamp_ultimo_frame_rgb = self._rgb_cache_ts
                                 self._rgb_seq = pkt_seq
@@ -478,6 +479,7 @@ class CameraOak:
                             t0 = time.time()
                             frame = pkt.getFrame()
                             dur = time.time() - t0
+                            frame_ts_host = time.time()
 
                             resultado = {
                                 "erro": None,
@@ -509,7 +511,7 @@ class CameraOak:
                                 self.perf.tick(
                                     "camera_depth",
                                     latencia_ms=dur * 1000.0,
-                                    frame_ts_host=self._depth_cache_ts,
+                                    frame_ts_host=frame_ts_host,
                                     frame_ts_device=pkt_ts_device,
                                     seq=pkt_seq,
                                     seq_delta=seq_delta,
@@ -520,7 +522,7 @@ class CameraOak:
 
                             with self._cache_lock:
                                 self._depth_cache = frame
-                                self._depth_cache_ts = time.time()
+                                self._depth_cache_ts = frame_ts_host
                                 self._depth_cache_resultado = resultado
                                 self.timestamp_ultimo_frame_depth = self._depth_cache_ts
                                 self.ultimo_resultado_depth = resultado
@@ -1062,6 +1064,50 @@ class CameraOak:
         self._definir_heartbeat()
         return frame, resultado
 
+    def requisitar_frame_rgb_ref(self):
+        """
+        Retorna referência READ-ONLY ao latest frame RGB cacheado.
+
+        Diferente de requisitar_frame_rgb(), não copia ~6 MB por frame 1080p.
+        O cache da câmera apenas substitui a referência por um ndarray novo; não
+        modifica o ndarray anterior. Portanto o consumidor pode manter a referência
+        durante resize/inferência sem segurar o lock.
+
+        REGRA: o consumidor não deve escrever no ndarray retornado.
+        """
+        with self._cache_lock:
+            resultado = dict(self._rgb_cache_resultado)
+            erro = resultado.get("erro") if self._is_erro_fatal_depthai(resultado.get("erro")) else None
+            frame = self._rgb_cache
+
+        if erro:
+            self._falha_fatal_depthai(erro)
+            return None, resultado
+
+        self._definir_heartbeat()
+        return frame, resultado
+
+    def requisitar_frame_depth_ref(self):
+        """Mesmo contrato read-only do RGB, aplicado ao depth latest-frame."""
+        if not self.tem_depth:
+            return None, {
+                "erro": "Camera nao possui sensor de profundidade",
+                "duracao": 0,
+                "frame_valido": False,
+            }
+
+        with self._cache_lock:
+            resultado = dict(self._depth_cache_resultado)
+            erro = resultado.get("erro") if self._is_erro_fatal_depthai(resultado.get("erro")) else None
+            frame = self._depth_cache
+
+        if erro:
+            self._falha_fatal_depthai(erro)
+            return None, resultado
+
+        self._definir_heartbeat()
+        return frame, resultado
+
     def requisitar_frame_depth(self):
         if not self.tem_depth:
             return None, {
diff --git a/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/manager_worker/modulos/movimentacao.py b/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/manager_worker/modulos/movimentacao.py
index f5719cf0e..9486e6031 100644
--- a/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/manager_worker/modulos/movimentacao.py
+++ b/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/manager_worker/modulos/movimentacao.py
@@ -1,6 +1,9 @@
 import time
 
-from shared.enums import ModoOperacao, StatusModulo, StatusCarroMapa, T_Code
+from shared.enums import (
+    ModoOperacao, StatusModulo, StatusCarroMapa, T_Code,
+    TipoMovimentoDirecional,
+)
 from shared.contexto_global_redis import ContextoGlobalRedis, CtxKey
 from manager_worker.config import mostrar_log
 from manager_worker.filtros import FiltroVelocidade
@@ -67,6 +70,44 @@ def definir_comando(filtro_vel: FiltroVelocidade, erro_orientacao: float = None)
             abs(_float(dir_cfg.get("angulo_max", 30.0), 30.0))
         )
 
+        # V5 - velocidade coerente com a política geométrica do MPC.
+        # Até ~5° o carro pode manter velocidade de cruzeiro. Entre 5° e 15°
+        # reduz progressivamente; a partir de 15° (faixa de MovimentoArco) usa
+        # velocidade de curva. Os parâmetros podem ser publicados em Mov sem
+        # exigir mudança imediata do contrato C#.
+        erro_heading_curva_forte_graus = _clamp(
+            _float(
+                mov_cfg.get("erro_heading_curva_forte_graus", 15.0),
+                15.0,
+            ),
+            3.0,
+            ang_max,
+        )
+        erro_heading_curva_deadband_graus = _clamp(
+            _float(
+                mov_cfg.get("erro_heading_curva_deadband_graus", 5.0),
+                5.0,
+            ),
+            0.0,
+            max(0.0, erro_heading_curva_forte_graus - 0.5),
+        )
+
+        # Movimento/ângulo atualmente aplicados. Como o Processador calcula MOV
+        # antes de DIR, isto funciona como memória segura do ciclo anterior. O
+        # erro geométrico abaixo continua sendo o gatilho primário do ciclo atual.
+        tipo_direcional_atual = _enum_or(
+            TipoMovimentoDirecional,
+            controle.get(
+                "tipo_movimento_direcional",
+                TipoMovimentoDirecional.RodasDianteiras.value,
+            ),
+            TipoMovimentoDirecional.RodasDianteiras,
+        )
+        angulo_direcional_atual = abs(_float(
+            controle.get("angulo_sp", 0.0),
+            0.0,
+        ))
+
         pontos_min_fim_corredor = max(
             0,
             _int(mov_cfg.get("pontos_fim_corredor_reduzir", 3), 3)
@@ -105,12 +146,43 @@ def definir_comando(filtro_vel: FiltroVelocidade, erro_orientacao: float = None)
 
         status_carro = estado["status_carro"]
         erro_angular = estado["erro_orientacao"]
+        erro_angular_caminho = float(
+            estado.get("erro_angular_caminho", erro_angular)
+        )
+        erro_angular_combinado = float(
+            estado.get("erro_angular_combinado", erro_angular)
+        )
         pontos_fim_corredor = estado["pontos_fim_corredor"]
         estado_valido = estado["valido"]
 
         debug["status_carro"] = status_carro.name
         debug["erro_orientacao"] = erro_angular
+        debug["erro_orientacao_fonte"] = estado.get(
+            "erro_orientacao_fonte",
+            "nao_informada"
+        )
+        debug["erros_orientacao"] = {
+            "selecionado": round(float(erro_angular), 4),
+            "combinado": round(
+                float(estado.get("erro_angular_combinado", erro_angular)),
+                4
+            ),
+            "caminho": round(
+                float(estado.get("erro_angular_caminho", erro_angular)),
+                4
+            ),
+            "proximo_ponto": round(
+                float(estado.get("erro_angular_proximo_ponto", erro_angular)),
+                4
+            ),
+        }
         debug["pontos_fim_corredor"] = pontos_fim_corredor
+        debug["geometria_curva"] = {
+            "erro_heading_curva_forte_graus": round(erro_heading_curva_forte_graus, 3),
+            "erro_heading_curva_deadband_graus": round(erro_heading_curva_deadband_graus, 3),
+            "tipo_direcional_atual": tipo_direcional_atual.name,
+            "angulo_direcional_atual": round(angulo_direcional_atual, 3),
+        }
 
         if not estado_valido:
             motivo = estado.get("motivo", "Estado operacional inválido para movimento")
@@ -133,10 +205,45 @@ def definir_comando(filtro_vel: FiltroVelocidade, erro_orientacao: float = None)
         resetar_filtro = False
 
         reduzir_para_pulverizar = False
+        reduzir_por_weed_indisponivel = False
+        limitar_velocidade_por_weed = False
         reduzir_para_fim_corredor = False
         reduzir_por_curva = False
+        teto_rigido_curva = False
+        teto_curva_pct = None
         reduzir_por_ipb = False
 
+        # Erro geométrico que conversa com o seletor do MPC. Em tracking de
+        # caminho e no retorno usamos o heading da centerline, evitando que um
+        # simples offset lateral pareça uma curva. Em Direcionando preservamos
+        # o erro combinado porque ali a aquisição do alvo ainda é relevante.
+        if status_carro in [
+            StatusCarroMapa.CaminhandoRua,
+            StatusCarroMapa.Direcionando,
+            StatusCarroMapa.RetornandoBase,
+        ]:
+            # Mesma referência geométrica do MPC V5: heading da centerline.
+            # Assim, rover paralelo porém deslocado não é confundido com curva.
+            erro_curva_graus = abs(erro_angular_caminho)
+            fonte_curva = "erro_angular_caminho"
+        else:
+            erro_curva_graus = abs(erro_angular_combinado)
+            fonte_curva = "erro_angular_combinado"
+
+        arco_ja_aplicado = (
+            tipo_direcional_atual == TipoMovimentoDirecional.MovimentoArco
+        )
+        curva_forte_geometrica = (
+            erro_curva_graus >= erro_heading_curva_forte_graus
+        )
+
+        debug["geometria_curva"].update({
+            "erro_usado_graus": round(float(erro_curva_graus), 3),
+            "fonte_erro": fonte_curva,
+            "arco_ja_aplicado": bool(arco_ja_aplicado),
+            "curva_forte_geometrica": bool(curva_forte_geometrica),
+        })
+
         if status_carro == StatusCarroMapa.Parado:
             velocidade_sp = 0.0
             hard_stop = True
@@ -151,6 +258,8 @@ def definir_comando(filtro_vel: FiltroVelocidade, erro_orientacao: float = None)
         ]:
             velocidade_sp = vel_curva
             reduzir_por_curva = True
+            teto_rigido_curva = True
+            teto_curva_pct = vel_curva
             debug["motivos"].append("movimento em curva/manobra")
 
         elif status_carro in [
@@ -158,26 +267,64 @@ def definir_comando(filtro_vel: FiltroVelocidade, erro_orientacao: float = None)
             StatusCarroMapa.RetornandoBase
         ]:
             velocidade_sp = calcular_velocidade_relativa(
-                vel_min=vel_com_ervas,
+                vel_min=vel_curva,
                 vel_max=vel_sem_ervas,
-                ang_max=ang_max,
-                erro_orientacao=erro_angular,
-                k=0.85,
+                ang_max=erro_heading_curva_forte_graus,
+                erro_orientacao=erro_curva_graus,
+                k=1.0,
                 curva=0.80,
-                dead=10.0
+                dead=erro_heading_curva_deadband_graus
+            )
+            reduzir_por_curva = bool(
+                erro_curva_graus > erro_heading_curva_deadband_graus
+                or arco_ja_aplicado
+            )
+            teto_rigido_curva = bool(reduzir_por_curva)
+            if teto_rigido_curva:
+                teto_curva_pct = (
+                    vel_curva
+                    if (curva_forte_geometrica or arco_ja_aplicado)
+                    else velocidade_sp
+                )
+            debug["motivos"].append(
+                f"{status_carro.name}: velocidade pela geometria do caminho/alvo"
             )
-            debug["motivos"].append("direcionando/retornando com redução por erro angular")
 
         elif status_carro == StatusCarroMapa.CaminhandoRua:
             info_ervas = _ler_weed_worker()
 
-            ervas_no_radar = (
-                pulverizador_automatico and
-                info_ervas["atualizado"] and
-                info_ervas["ervas_no_radar"]
-            )
+            # O WeedWorker é a autoridade para dizer se há ervas no radar.
+            #
+            # Contrato atual:
+            # DadosWeedWorker.analise = envelope
+            # envelope.analise = dados_visuais reais
+            #
+            # _ler_weed_worker() normaliza esse contrato e também aceita o
+            # formato legado (campos diretamente em DadosWeedWorker.analise).
+            #
+            # Se o campo booleano não existir em uma versão legada, usamos
+            # ema_global apenas como fallback.
+            if info_ervas["campo_ervas_no_radar_presente"]:
+                ervas_detectadas = bool(info_ervas["ervas_no_radar"])
+            else:
+                ervas_detectadas = (
+                    info_ervas["percentual_ervas"] >= ervas_min_percent
+                )
 
-            reduzir_para_pulverizar = bool(ervas_no_radar)
+            if pulverizador_automatico:
+                if info_ervas["atualizado"]:
+                    reduzir_para_pulverizar = bool(ervas_detectadas)
+                else:
+                    # Falha conservadora de qualidade de aplicação:
+                    # percepção desconhecida NÃO significa "sem ervas".
+                    # Mantém a velocidade de pulverização até a análise voltar,
+                    # sem parar a operação por si só.
+                    reduzir_por_weed_indisponivel = True
+
+            limitar_velocidade_por_weed = bool(
+                reduzir_para_pulverizar
+                or reduzir_por_weed_indisponivel
+            )
 
             reduzir_para_fim_corredor = (
                 pontos_fim_corredor >= 0 and
@@ -185,20 +332,57 @@ def definir_comando(filtro_vel: FiltroVelocidade, erro_orientacao: float = None)
             )
 
             debug["weed"] = info_ervas
+            debug["weed"]["pulverizador_automatico"] = bool(
+                pulverizador_automatico
+            )
+            debug["weed"]["ervas_detectadas_efetivo"] = bool(
+                ervas_detectadas
+            )
+            debug["weed"]["limitar_velocidade"] = bool(
+                limitar_velocidade_por_weed
+            )
 
             velocidade_livre = calcular_velocidade_relativa(
-                vel_min=vel_com_ervas,
+                vel_min=vel_curva,
                 vel_max=vel_sem_ervas,
-                ang_max=ang_max,
-                erro_orientacao=erro_angular,
+                ang_max=erro_heading_curva_forte_graus,
+                erro_orientacao=erro_curva_graus,
                 k=1.0,
-                curva=0.70,
-                dead=5.0
+                curva=0.80,
+                dead=erro_heading_curva_deadband_graus
+            )
+
+            # Mesmo que o C# publique CaminhandoRua durante RetornoBase, uma
+            # curva forte da centerline não pode ser tratada como reta rápida.
+            # O heading do caminho é a mesma referência usada pelo MPC V5.
+            reduzir_por_curva = bool(
+                erro_curva_graus > erro_heading_curva_deadband_graus
+                or arco_ja_aplicado
+            )
+            teto_rigido_curva = bool(reduzir_por_curva)
+            if teto_rigido_curva:
+                teto_curva_pct = (
+                    vel_curva
+                    if (curva_forte_geometrica or arco_ja_aplicado)
+                    else velocidade_livre
+                )
+
+            debug["motivos"].append(
+                "caminhando rua: velocidade por erro angular do caminho"
             )
 
             if reduzir_para_pulverizar:
                 velocidade_sp = min(velocidade_livre, vel_com_ervas)
-                debug["motivos"].append("ervas no radar: reduzindo para pulverizar")
+                debug["motivos"].append(
+                    "ervas no radar: reduzindo para pulverizar"
+                )
+
+            elif reduzir_por_weed_indisponivel:
+                velocidade_sp = min(velocidade_livre, vel_com_ervas)
+                debug["motivos"].append(
+                    "WeedWorker ausente/desatualizado: "
+                    "mantendo velocidade conservadora de pulverização"
+                )
 
             elif reduzir_para_fim_corredor:
                 velocidade_sp = min(velocidade_livre, vel_com_ervas)
@@ -217,6 +401,16 @@ def definir_comando(filtro_vel: FiltroVelocidade, erro_orientacao: float = None)
                 f"status_carro não tratado com segurança: {status_carro.name}"
             )
 
+        debug["geometria_curva"].update({
+            "reduzir_por_curva": bool(reduzir_por_curva),
+            "teto_rigido_curva": bool(teto_rigido_curva),
+            "teto_curva_pct": (
+                None if teto_curva_pct is None
+                else round(float(teto_curva_pct), 3)
+            ),
+            "vel_curva_pct": round(float(vel_curva), 3),
+        })
+
         # Se já decidiu parar, não faz limitador tentar reviver velocidade.
         if hard_stop:
             _resetar_rampa_velocidade(filtro_vel)
@@ -284,7 +478,8 @@ def definir_comando(filtro_vel: FiltroVelocidade, erro_orientacao: float = None)
 
         if usar_imu_movimento:
             vel_imu, imu_resetar, imu_stop, imu_debug = _aplicar_limitador_imu(
-                velocidade_atual=velocidade_sp
+                velocidade_atual=velocidade_sp,
+                vel_min_equip=vel_min_equip
             )
 
             debug["imu"] = imu_debug
@@ -329,7 +524,7 @@ def definir_comando(filtro_vel: FiltroVelocidade, erro_orientacao: float = None)
         # Aceleração normal é filtrada para não dar tranco.
 
         reducao_seguranca = bool(
-            reduzir_para_pulverizar
+            limitar_velocidade_por_weed
             or reduzir_para_fim_corredor
             or reduzir_por_curva
             or reduzir_por_ipb
@@ -351,23 +546,59 @@ def definir_comando(filtro_vel: FiltroVelocidade, erro_orientacao: float = None)
                 debug=debug,
             )
 
-        # Correção crítica:
-        # havendo erva no radar, a saída nunca pode superar a mínima definida.
-        if reduzir_para_pulverizar:
-            velocidade_filtrada = min(
-                velocidade_filtrada,
-                vel_com_ervas
-            )
+        # Teto rígido de curva/Arco. A rampa de segurança continua suavizando
+        # reduções moderadas, mas uma geometria já classificada como curva forte
+        # não pode passar um ciclo em velocidade de cruzeiro enquanto o MPC
+        # coloca o rover em MovimentoArco.
+        if teto_rigido_curva and teto_curva_pct is not None:
+            teto_curva_pct = _clamp(teto_curva_pct, vel_min_equip, vel_sem_ervas)
+            velocidade_filtrada = min(velocidade_filtrada, teto_curva_pct)
 
-            # Sincroniza o estado da rampa com a saída realmente aplicada.
             try:
                 filtro_vel._ultimo_sp_campo = velocidade_filtrada
                 filtro_vel._ultimo_ts_campo = time.monotonic()
             except Exception:
                 pass
 
+            motivo_curva = (
+                "curva forte/Arco"
+                if (curva_forte_geometrica or arco_ja_aplicado)
+                else "curva moderada"
+            )
             debug["limitadores"].append(
-                f"ervas no radar: teto rígido em {vel_com_ervas:.1f}%"
+                f"{motivo_curva}: teto geométrico em {teto_curva_pct:.1f}%"
+            )
+
+        # Teto rígido do WeedWorker:
+        #
+        # - ervas detectadas: nunca ultrapassa vel_com_ervas;
+        # - WeedWorker desatualizado com pulverizador automático ligado:
+        #   também não assume que a rua está livre.
+        #
+        # Esse clamp acontece DEPOIS da rampa para não existir overshoot
+        # transitório acima da velocidade de aplicação.
+        if limitar_velocidade_por_weed:
+            velocidade_filtrada = min(
+                velocidade_filtrada,
+                vel_com_ervas
+            )
+
+            # Sincroniza o estado da rampa com a saída realmente aplicada.
+            # Assim, quando a condição desaparecer, a aceleração volta a
+            # acontecer pela rampa normal em vez de saltar para o SP antigo.
+            try:
+                filtro_vel._ultimo_sp_campo = velocidade_filtrada
+                filtro_vel._ultimo_ts_campo = time.monotonic()
+            except Exception:
+                pass
+
+            if reduzir_para_pulverizar:
+                motivo_teto = "ervas no radar"
+            else:
+                motivo_teto = "WeedWorker indisponível/desatualizado"
+
+            debug["limitadores"].append(
+                f"{motivo_teto}: teto rígido em {vel_com_ervas:.1f}%"
             )
 
         velocidade_filtrada = round(
@@ -415,6 +646,33 @@ def _resolver_estado_operacional(
     contexto,
     erro_orientacao
 ):
+    """
+    Resolve o estado operacional e escolhe a referência angular usada
+    EXCLUSIVAMENTE para a política de velocidade.
+
+    Política alinhada ao MPC Path Tracking V2:
+
+    - CaminhandoRua:
+        usa erro_angular_caminho.
+        O que importa para liberar velocidade é o corpo do rover estar
+        alinhado com a orientação local da passada. Deslocamento lateral
+        pode ser corrigido pelo MovimentoDiagonal sem penalizar a velocidade
+        como se o rover estivesse "torto".
+
+    - Direcionando / RetornandoBase:
+        usa erro_angular combinado.
+        Fora da passada, a aquisição geométrica do alvo continua relevante.
+
+    - EntrandoRua / SaindoRua / Manobrando:
+        a velocidade base é vel_curva, portanto o erro selecionado não
+        altera a velocidade nesses estados. Mantemos o combinado para
+        diagnóstico/coerência.
+
+    - erro_orientacao explícito:
+        continua tendo prioridade para preservar compatibilidade com
+        chamadas externas/legadas.
+    """
+
     if op_modo in [ModoOperacao.MapaGPS, ModoOperacao.RetornoBase]:
         trajetoria = _dict(contexto.get("Trajetoria", {}))
 
@@ -424,6 +682,10 @@ def _resolver_estado_operacional(
                 "motivo": "Trajetoria ausente no contexto",
                 "status_carro": StatusCarroMapa.Parado,
                 "erro_orientacao": 0.0,
+                "erro_orientacao_fonte": "trajetoria_ausente",
+                "erro_angular_combinado": 0.0,
+                "erro_angular_caminho": 0.0,
+                "erro_angular_proximo_ponto": 0.0,
                 "pontos_fim_corredor": -1,
             }
 
@@ -433,33 +695,86 @@ def _resolver_estado_operacional(
             StatusCarroMapa.Parado
         )
 
-        if erro_orientacao is None:
-            erro_orientacao = _float(trajetoria.get("erro_angular", 0.0), 0.0)
+        erro_combinado = _float(
+            trajetoria.get("erro_angular", 0.0),
+            0.0
+        )
+        erro_caminho = _float(
+            trajetoria.get(
+                "erro_angular_caminho",
+                erro_combinado
+            ),
+            erro_combinado
+        )
+        erro_proximo_ponto = _float(
+            trajetoria.get(
+                "erro_angular_proximo_ponto",
+                erro_combinado
+            ),
+            erro_combinado
+        )
+
+        if erro_orientacao is not None:
+            erro_selecionado = _float(
+                erro_orientacao,
+                erro_combinado
+            )
+            fonte_erro = "override_explicito"
+
+        elif status_carro == StatusCarroMapa.CaminhandoRua:
+            # Na passada, o heading local da rua é a referência correta.
+            # O bearing para waypoint não deve reduzir velocidade só porque
+            # o MPC está corrigindo offset lateral em MovimentoDiagonal.
+            erro_selecionado = erro_caminho
+            fonte_erro = "erro_angular_caminho"
+
+        else:
+            # Aquisição, retorno e demais estados mantêm a referência
+            # combinada que incorpora a geometria do alvo.
+            erro_selecionado = erro_combinado
+            fonte_erro = "erro_angular_combinado"
 
         corredor = _dict(trajetoria.get("CorredorAtual", {}))
-        pontos_fim_corredor = _int(corredor.get("pontos_restantes", -1), -1)
+        pontos_fim_corredor = _int(
+            corredor.get("pontos_restantes", -1),
+            -1
+        )
 
         return {
             "valido": True,
             "motivo": "",
             "status_carro": status_carro,
-            "erro_orientacao": _float(erro_orientacao, 0.0),
+            "erro_orientacao": float(erro_selecionado),
+            "erro_orientacao_fonte": fonte_erro,
+            "erro_angular_combinado": float(erro_combinado),
+            "erro_angular_caminho": float(erro_caminho),
+            "erro_angular_proximo_ponto": float(erro_proximo_ponto),
             "pontos_fim_corredor": pontos_fim_corredor,
         }
 
     if op_modo == ModoOperacao.MapeamentoVisual:
-        dados_vw = _dict(ContextoGlobalRedis.get(CtxKey.DadosVisualWorker, {}))
-        segmentacao = _dict(dados_vw.get("segmentacao", {}))
+        dados_vw = _dict(
+            ContextoGlobalRedis.get(CtxKey.DadosVisualWorker, {})
+        )
+        segmentacao = _dict(
+            dados_vw.get("segmentacao", {})
+        )
 
         seg_ts = _float(
             segmentacao.get(
                 "ts",
-                dados_vw.get("ts_segmentacao", dados_vw.get("momento", 0.0))
+                dados_vw.get(
+                    "ts_segmentacao",
+                    dados_vw.get("momento", 0.0)
+                )
             ),
             0.0
         )
 
-        seg_atualizada, idade_ms = _timestamp_atualizado(seg_ts, max_idade_s=1.0)
+        seg_atualizada, idade_ms = _timestamp_atualizado(
+            seg_ts,
+            max_idade_s=1.0
+        )
 
         if not segmentacao or not seg_atualizada:
             motivo = "segmentacao visual ausente/desatualizada"
@@ -472,31 +787,61 @@ def _resolver_estado_operacional(
                 "motivo": motivo,
                 "status_carro": StatusCarroMapa.Parado,
                 "erro_orientacao": 0.0,
+                "erro_orientacao_fonte": "segmentacao_invalida",
+                "erro_angular_combinado": 0.0,
+                "erro_angular_caminho": 0.0,
+                "erro_angular_proximo_ponto": 0.0,
                 "pontos_fim_corredor": -1,
             }
 
         status_carro = _enum_or(
             StatusCarroMapa,
-            segmentacao.get("status_corredor", StatusCarroMapa.Parado.value),
+            segmentacao.get(
+                "status_corredor",
+                StatusCarroMapa.Parado.value
+            ),
             StatusCarroMapa.Parado
         )
 
-        if erro_orientacao is None:
-            erro_orientacao = _float(segmentacao.get("erro_angular", 0.0), 0.0)
+        erro_visual = _float(
+            segmentacao.get("erro_angular", 0.0),
+            0.0
+        )
+
+        if erro_orientacao is not None:
+            erro_selecionado = _float(
+                erro_orientacao,
+                erro_visual
+            )
+            fonte_erro = "override_explicito"
+        else:
+            erro_selecionado = erro_visual
+            fonte_erro = "segmentacao_visual"
 
         return {
             "valido": True,
             "motivo": "",
             "status_carro": status_carro,
-            "erro_orientacao": _float(erro_orientacao, 0.0),
+            "erro_orientacao": float(erro_selecionado),
+            "erro_orientacao_fonte": fonte_erro,
+            "erro_angular_combinado": float(erro_visual),
+            "erro_angular_caminho": float(erro_visual),
+            "erro_angular_proximo_ponto": float(erro_visual),
             "pontos_fim_corredor": -1,
         }
 
     return {
         "valido": False,
-        "motivo": f"Modo de operação sem estratégia de movimento: {op_modo.name}",
+        "motivo": (
+            "Modo de operação sem estratégia de movimento: "
+            f"{op_modo.name}"
+        ),
         "status_carro": StatusCarroMapa.Parado,
         "erro_orientacao": 0.0,
+        "erro_orientacao_fonte": "modo_sem_estrategia",
+        "erro_angular_combinado": 0.0,
+        "erro_angular_caminho": 0.0,
+        "erro_angular_proximo_ponto": 0.0,
         "pontos_fim_corredor": -1,
     }
 
@@ -506,34 +851,119 @@ def _resolver_estado_operacional(
 # ============================================================
 
 def _ler_weed_worker():
-    try:
-        dados = _dict(ContextoGlobalRedis.get(CtxKey.DadosWeedWorker, {}))
-        analise = _dict(dados.get("analise", {}))
+    """
+    Lê e normaliza o contrato do WeedWorker.
 
+    Contrato atual publicado pelo CameraManager:
+
+        DadosWeedWorker = {
+            "ts_publicacao": ...,
+            "ts_analise": ...,
+            "analise": {
+                "ts_analise": ...,
+                "fps_model": ...,
+                "fps_inferencia": ...,
+                "infer_ms": ...,
+                "infer_gpu_ms": ...,
+                "analise": {
+                    "ervas_no_radar": ...,
+                    "estatisticas": {
+                        "erva": {
+                            "ema_global": ...
+                        }
+                    },
+                    ...
+                }
+            }
+        }
+
+    Também aceita o formato legado, no qual ervas_no_radar e estatisticas
+    ficam diretamente em DadosWeedWorker["analise"].
+    """
+    try:
+        dados = _dict(
+            ContextoGlobalRedis.get(CtxKey.DadosWeedWorker, {})
+        )
+
+        envelope = _dict(dados.get("analise", {}))
+
+        # Contrato atual possui envelope["analise"].
+        # No contrato legado, o próprio envelope já é a análise.
+        analise_interna = _dict(envelope.get("analise", {}))
+
+        if analise_interna:
+            analise = analise_interna
+            esquema = "nested_v2"
+        else:
+            analise = envelope
+            esquema = "flat_legacy"
+
+        # Preferência de timestamp:
+        # 1) timestamp do envelope da análise atual;
+        # 2) timestamp dentro da análise legada;
+        # 3) timestamp raiz do DadosWeedWorker.
         ts = _float(
-            analise.get(
-                "ts",
-                analise.get(
-                    "timestamp",
-                    dados.get("ts_analise", dados.get("momento", 0.0))
+            envelope.get(
+                "ts_analise",
+                envelope.get(
+                    "ts",
+                    envelope.get(
+                        "timestamp",
+                        analise.get(
+                            "ts",
+                            analise.get(
+                                "timestamp",
+                                dados.get(
+                                    "ts_analise",
+                                    dados.get(
+                                        "ts_publicacao",
+                                        dados.get("momento", 0.0)
+                                    )
+                                )
+                            )
+                        )
+                    )
                 )
             ),
             0.0
         )
 
-        atualizado, idade_ms = _timestamp_atualizado(ts, max_idade_s=1.5)
+        atualizado, idade_ms = _timestamp_atualizado(
+            ts,
+            max_idade_s=1.5
+        )
 
-        estatisticas = _dict(analise.get("estatisticas", {}))
-        erva = _dict(estatisticas.get("erva", {}))
+        estatisticas = _dict(
+            analise.get("estatisticas", {})
+        )
+        erva = _dict(
+            estatisticas.get("erva", {})
+        )
 
-        ervas_no_radar = _bool(analise.get("ervas_no_radar", False))
-        percentual_ervas = _float(erva.get("ema_global", 0.0), 0.0)
+        campo_ervas_presente = "ervas_no_radar" in analise
+
+        ervas_no_radar = _bool(
+            analise.get("ervas_no_radar", False)
+        )
+
+        percentual_ervas = _float(
+            erva.get(
+                "ema_global",
+                analise.get("percentual_ervas_no_radar", 0.0)
+            ),
+            0.0
+        )
 
         return {
-            "atualizado": atualizado,
+            "atualizado": bool(atualizado),
             "idade_ms": idade_ms,
-            "ervas_no_radar": ervas_no_radar,
+            "ervas_no_radar": bool(ervas_no_radar),
+            "campo_ervas_no_radar_presente": bool(
+                campo_ervas_presente
+            ),
             "percentual_ervas": percentual_ervas,
+            "timestamp_analise": ts,
+            "esquema": esquema,
         }
 
     except Exception as e:
@@ -541,7 +971,10 @@ def _ler_weed_worker():
             "atualizado": False,
             "idade_ms": None,
             "ervas_no_radar": False,
+            "campo_ervas_no_radar_presente": False,
             "percentual_ervas": 0.0,
+            "timestamp_analise": 0.0,
+            "esquema": "erro",
             "erro": str(e),
         }
 
@@ -643,7 +1076,7 @@ def _aplicar_limitador_visual(
 # Limitador IMU
 # ============================================================
 
-def _aplicar_limitador_imu(*, velocidade_atual):
+def _aplicar_limitador_imu(*, velocidade_atual, vel_min_equip):
     debug = {
         "habilitado": True,
         "aplicado": False,
@@ -693,7 +1126,41 @@ def _aplicar_limitador_imu(*, velocidade_atual):
         if risk_level_num <= 0:
             return velocidade_atual, False, False, debug
 
-        vel_nova = float(velocidade_atual) * velocidade_factor
+        velocidade_atual = max(0.0, float(velocidade_atual))
+        vel_min_equip = max(0.0, float(vel_min_equip))
+
+        # Attention/risk são REDUÇÕES, não ordens de parada.
+        #
+        # Se a IMU, sozinha, levar o SP abaixo da velocidade mínima útil do
+        # equipamento, limita no piso operacional em vez de transformar um
+        # nível attention/risk em parada.
+        #
+        # O min(..., velocidade_atual) é proposital: a IMU nunca pode elevar
+        # uma velocidade que já tenha sido reduzida por outro limitador
+        # (Visual Worker, IPB, fim de corredor, pulverização etc.).
+        vel_calculada = velocidade_atual * velocidade_factor
+
+        if velocidade_atual >= vel_min_equip and vel_calculada > 0.0:
+            vel_nova = min(
+                velocidade_atual,
+                max(vel_min_equip, vel_calculada)
+            )
+        else:
+            # Se outro limitador já entregou algo abaixo do piso, a IMU não
+            # "ressuscita" a velocidade. A regra geral posterior decide se
+            # esse valor deve virar parada.
+            vel_nova = min(velocidade_atual, vel_calculada)
+
+        debug["velocidade_entrada"] = round(velocidade_atual, 3)
+        debug["velocidade_calculada"] = round(vel_calculada, 3)
+        debug["velocidade_saida"] = round(vel_nova, 3)
+        debug["vel_min_equip"] = round(vel_min_equip, 3)
+        debug["piso_operacional_aplicado"] = bool(
+            velocidade_atual >= vel_min_equip
+            and 0.0 < vel_calculada < vel_min_equip
+            and abs(vel_nova - vel_min_equip) < 1e-9
+        )
+
         resetar_filtro = False
 
         return vel_nova, resetar_filtro, False, debug
diff --git a/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/manager_worker/modulos/mpc.py b/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/manager_worker/modulos/mpc.py
index 8ccf37476..ee460033f 100644
--- a/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/manager_worker/modulos/mpc.py
+++ b/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/manager_worker/modulos/mpc.py
@@ -341,7 +341,156 @@ class ControladorMPC:
             max_value=20.0,
         )
 
-        # Amortecimento aplicado somente durante CaminhandoRua. Nas curvas o
+        # ------------------------------------------------------------------
+        # PATH TRACKING V2 - referência geométrica da passada
+        # ------------------------------------------------------------------
+        # Na passada o rover NÃO mira um ponto. O ponto/progresso continua
+        # sendo autoridade do C#, mas a direção segue: (1) heading médio do
+        # caminho nos próximos metros + (2) cross-track lateral. O alvo
+        # pure-pursuit fica reservado para entrada/saída/manobra.
+        self._heading_lookahead_base_m = _safe_float(
+            parametros_mpc.get("heading_lookahead_base_m", 1.60),
+            1.60, min_value=0.40, max_value=8.0,
+        )
+        self._heading_lookahead_velocidade_s = _safe_float(
+            parametros_mpc.get("heading_lookahead_velocidade_s", 0.55),
+            0.55, min_value=0.0, max_value=4.0,
+        )
+        self._heading_lookahead_min_m = _safe_float(
+            parametros_mpc.get("heading_lookahead_min_m", 1.20),
+            1.20, min_value=0.25, max_value=8.0,
+        )
+        self._heading_lookahead_max_m = _safe_float(
+            parametros_mpc.get("heading_lookahead_max_m", 3.00),
+            3.00, min_value=self._heading_lookahead_min_m, max_value=12.0,
+        )
+
+        # Heading primeiro, lateral depois. A diagonal só pode existir quando
+        # o corpo já está praticamente paralelo ao caminho; assim ela
+        # recentraliza sem congelar um erro de heading de 5-10 graus.
+        self._heading_deadband_graus = _safe_float(
+            parametros_mpc.get("heading_deadband_graus", 0.45),
+            0.45, min_value=0.0, max_value=3.0,
+        )
+        self._diagonal_entrada_heading_graus = _safe_float(
+            parametros_mpc.get("diagonal_entrada_heading_graus", 1.25),
+            1.25, min_value=0.2, max_value=8.0,
+        )
+        self._diagonal_saida_heading_graus = _safe_float(
+            parametros_mpc.get("diagonal_saida_heading_graus", 2.50),
+            2.50, min_value=self._diagonal_entrada_heading_graus, max_value=12.0,
+        )
+        self._lateral_deadband_m = _safe_float(
+            parametros_mpc.get("lateral_deadband_m", 0.05),
+            0.05, min_value=0.0, max_value=0.30,
+        )
+        self._ganho_heading_reta = _safe_float(
+            parametros_mpc.get("ganho_heading_reta", 1.00),
+            1.00, min_value=0.1, max_value=3.0,
+        )
+        self._ganho_lateral_dianteira = _safe_float(
+            parametros_mpc.get("ganho_lateral_dianteira", 0.32),
+            0.32, min_value=0.0, max_value=3.0,
+        )
+        self._ganho_lateral_diagonal = _safe_float(
+            parametros_mpc.get("ganho_lateral_diagonal", 0.85),
+            0.85, min_value=0.0, max_value=4.0,
+        )
+        self._velocidade_offset_lateral = _safe_float(
+            parametros_mpc.get("velocidade_offset_lateral", 0.35),
+            0.35, min_value=0.05, max_value=2.0,
+        )
+        self._angulo_diagonal_max_graus = _safe_float(
+            parametros_mpc.get("angulo_diagonal_max_graus", 16.0),
+            16.0, min_value=1.0, max_value=self.angulo_max_graus,
+        )
+        self._heading_reta_aquisicao_max_graus = _safe_float(
+            parametros_mpc.get("heading_reta_aquisicao_max_graus", 15.0),
+            15.0, min_value=3.0, max_value=35.0,
+        )
+        self._lateral_reta_aquisicao_max_m = _safe_float(
+            parametros_mpc.get("lateral_reta_aquisicao_max_m", 0.80),
+            0.80, min_value=0.15, max_value=2.0,
+        )
+        self._penalidade_troca_tipo = _safe_float(
+            parametros_mpc.get("penalidade_troca_tipo", 0.18),
+            0.18, min_value=0.0, max_value=3.0,
+        )
+
+        # Direcionando fora da rua: usa Arco para a correção grosseira de
+        # heading e volta para RodasDianteiras somente quando já estiver bem
+        # alinhado. Os dois limiares criam histerese e evitam troca de modo
+        # perto da fronteira.
+        self._direcionando_arco_entrada_graus = _safe_float(
+            parametros_mpc.get("direcionando_arco_entrada_graus", 15.0),
+            15.0, min_value=3.0, max_value=60.0,
+        )
+        self._direcionando_arco_saida_graus = _safe_float(
+            parametros_mpc.get("direcionando_arco_saida_graus", 8.0),
+            8.0, min_value=1.0, max_value=self._direcionando_arco_entrada_graus,
+        )
+
+        # --------------------------------------------------------------
+        # V5 - Política geométrica universal do path tracking.
+        #
+        # A lógica vale tanto para a passada agrícola quanto para a rota geral
+        # de RetornoBase:
+        #   - heading muito errado -> MovimentoArco;
+        #   - heading moderadamente errado -> RodasDianteiras;
+        #   - heading praticamente alinhado + cross-track -> MovimentoDiagonal.
+        #
+        # A histerese 15°/8° evita ficar trocando Arco <-> Dianteira na
+        # fronteira. Os defaults herdam a política já validada em Direcionando.
+        # Manobras estruturais reais continuam com Arco obrigatório.
+        # --------------------------------------------------------------
+        self._tracking_arco_entrada_graus = _safe_float(
+            parametros_mpc.get(
+                "tracking_arco_entrada_graus",
+                self._direcionando_arco_entrada_graus,
+            ),
+            self._direcionando_arco_entrada_graus,
+            min_value=3.0,
+            max_value=60.0,
+        )
+        self._tracking_arco_saida_graus = _safe_float(
+            parametros_mpc.get(
+                "tracking_arco_saida_graus",
+                self._direcionando_arco_saida_graus,
+            ),
+            self._direcionando_arco_saida_graus,
+            min_value=1.0,
+            max_value=self._tracking_arco_entrada_graus,
+        )
+
+        # --------------------------------------------------------------
+        # V4 - Entrada/saída de rua: Arco + RodasDianteiras competem.
+        #
+        # A preferência usa o HEADING DO CAMINHO, e não o bearing até o
+        # alvo. Assim, quando o corpo já está alinhado com a próxima rua,
+        # RodasDianteiras ganham vantagem e terminam a aquisição de forma
+        # suave. Arco continua favorecido quando o heading ainda está muito
+        # errado.
+        #
+        # Entre os dois limiares a preferência varia continuamente; a
+        # penalidade normal de troca de tipo fornece histerese adicional.
+        # --------------------------------------------------------------
+        self._transicao_rua_dianteira_heading_graus = _safe_float(
+            parametros_mpc.get("transicao_rua_dianteira_heading_graus", 7.0),
+            7.0, min_value=0.5, max_value=30.0,
+        )
+        self._transicao_rua_arco_heading_graus = _safe_float(
+            parametros_mpc.get("transicao_rua_arco_heading_graus", 20.0),
+            20.0,
+            min_value=self._transicao_rua_dianteira_heading_graus + 0.5,
+            max_value=60.0,
+        )
+        self._transicao_rua_penalidade_modo = _safe_float(
+            parametros_mpc.get("transicao_rua_penalidade_modo", 1.0),
+            1.0, min_value=0.0, max_value=5.0,
+        )
+
+        # Amortecimento aplicado ao tracking de passada. MovimentoArco em
+        # manobra preserva autoridade; RodasDianteiras/Diagonal são limitados.
         # MPC preserva toda a autoridade angular necessaria para a manobra.
         self._filtro_angulo_reta_tau_s = _safe_float(
             parametros_mpc.get("filtro_angulo_reta_tau_s", 0.25),
@@ -562,6 +711,11 @@ class ControladorMPC:
         return (float(xy[0]), float(xy[1])), i
 
     def _distancia_lookahead(self, status_carro, velocidade):
+        """Look-ahead do alvo de manobra/pure-pursuit.
+
+        Em CaminhandoRua esse alvo continua existindo para debug/progresso,
+        mas NÃO define o heading desejado do rover.
+        """
         v = max(0.0, _safe_float(velocidade, 0.0))
 
         if _status_in(status_carro, [StatusCarroMapa.CaminhandoRua]):
@@ -580,7 +734,7 @@ class ControladorMPC:
         ))
 
     def _construir_alvo_direcional(self, x, y, idx_base, status_carro, velocidade):
-        """Cria alvo pure-pursuit continuo sem alterar o progresso visitado."""
+        """Cria alvo contínuo para manobra/debug sem alterar progresso visitado."""
         n = len(self.pontos_info)
         if n <= 0:
             return {"xy": (float(x), float(y)), "_idx_base": 0}
@@ -606,21 +760,285 @@ class ControladorMPC:
         alvo["_lookahead_m"] = float(max(0.0, s_alvo - s_proj))
         return alvo
 
+    def _ponto_estrutural(self, idx):
+        if not self.pontos_info:
+            return True
+        i = max(0, min(_safe_int(idx, 0), len(self.pontos_info) - 1))
+        p = self.pontos_info[i]
+        return bool(p.get("estrutural", p.get("PontoEstrutural", False)))
+
+    def _idx_corredor(self, idx):
+        if not self.pontos_info:
+            return -1
+        i = max(0, min(_safe_int(idx, 0), len(self.pontos_info) - 1))
+        p = self.pontos_info[i]
+        return _safe_int(p.get("idxCorredor", p.get("IdxCorredor", -1)), -1)
+
+    def _segmento_operacional_mesmo_corredor(self, idx_base):
+        """True quando há um segmento de passada válido junto ao índice base.
+
+        Aceita a borda de entrada como início da aquisição se o segmento à
+        frente já pertence ao mesmo corredor e o próximo ponto é operacional.
+        Isso evita obrigar Arco só porque o estado C# ainda diz Manobrando.
+        """
+        n = len(self.pontos_info)
+        if n < 2:
+            return False
+        idx = max(0, min(_safe_int(idx_base, 0), n - 1))
+
+        pares = []
+        if idx + 1 < n:
+            pares.append((idx, idx + 1))
+        if idx - 1 >= 0:
+            pares.append((idx - 1, idx))
+
+        for a, b in pares:
+            ca, cb = self._idx_corredor(a), self._idx_corredor(b)
+            if ca < 0 or cb < 0 or ca != cb:
+                continue
+            # Um ponto estrutural isolado na entrada pode iniciar a reta, mas
+            # dois pontos estruturais seguidos caracterizam ligação/manobra.
+            if self._ponto_estrutural(a) and self._ponto_estrutural(b):
+                continue
+            return True
+        return False
+
+    def _segmento_rua_puro_mesmo_corredor(self, idx_base):
+        """
+        True somente quando a referência local está realmente em um trecho
+        operacional Rua -> Rua do MESMO corredor.
+
+        Diferente de _segmento_operacional_mesmo_corredor(), aqui não aceitamos
+        ponto estrutural isolado. Isso é importante para distinguir:
+
+        - recuperação dentro/ao lado de uma passada reta;
+        - entrada, saída, ligação ou cabeceira real.
+
+        Na recuperação de uma passada, MovimentoArco é indesejado: ele dobra a
+        autoridade de yaw e pode transformar um desvio lateral em uma inversão
+        brusca de heading.
+        """
+        n = len(self.pontos_info)
+        if n < 2:
+            return False
+
+        idx = max(0, min(_safe_int(idx_base, 0), n - 1))
+
+        pares = []
+        if idx + 1 < n:
+            pares.append((idx, idx + 1))
+        if idx - 1 >= 0:
+            pares.append((idx - 1, idx))
+
+        for a, b in pares:
+            ca, cb = self._idx_corredor(a), self._idx_corredor(b)
+
+            if ca < 0 or cb < 0 or ca != cb:
+                continue
+
+            if self._ponto_estrutural(a) or self._ponto_estrutural(b):
+                continue
+
+            return True
+
+        return False
+
+    def _limitar_s_ao_corredor(self, s_proj, s_desejado, idx_base):
+        """Não deixa o heading look-ahead enxergar a rua seguinte pela cabeceira."""
+        if self._s_nodes is None or len(self._s_nodes) == 0:
+            return float(s_desejado)
+
+        n = len(self.pontos_info)
+        idx = max(0, min(_safe_int(idx_base, 0), n - 1))
+        corredor = self._idx_corredor(idx)
+        if corredor < 0:
+            return float(s_desejado)
+
+        s_lim = float(self._s_nodes[-1])
+        for j in range(idx + 1, n):
+            if self._idx_corredor(j) != corredor or self._ponto_estrutural(j):
+                s_lim = float(self._s_nodes[j])
+                break
+
+        # precisa sobrar um pequeno vetor para formar direção; se já estamos
+        # no fim, a tangente local será usada pelo chamador.
+        return float(min(float(s_desejado), s_lim))
+
+    def _referencia_caminho_lookahead(self, x, y, idx_referencia, velocidade):
+        """Retorna heading médio dos próximos metros + cross-track.
+
+        A orientação é a corda da centerline entre a projeção atual e um ponto
+        adiante. Em reta ela é constante; em curva antecipa suavemente o arco.
+        Nunca aponta do ROBÔ para o alvo, portanto deslocamento lateral não
+        contamina o erro de heading.
+        """
+        e_lat, s_proj, i_seg, proj, theta_local, margem = self._projetar_na_trajetoria_local(
+            x, y, idx_referencia=idx_referencia,
+        )
+
+        v = max(0.0, _safe_float(velocidade, 0.0))
+        look = self._heading_lookahead_base_m + self._heading_lookahead_velocidade_s * v
+        look = float(np.clip(look, self._heading_lookahead_min_m, self._heading_lookahead_max_m))
+
+        s_fim = min(float(self._s_nodes[-1]), float(s_proj) + look)
+        s_fim = self._limitar_s_ao_corredor(s_proj, s_fim, idx_referencia)
+
+        p0, _ = self._xy_na_abscissa(s_proj)
+        p1, _ = self._xy_na_abscissa(s_fim)
+        dx = float(p1[0] - p0[0])
+        dy = float(p1[1] - p0[1])
+        dist = math.hypot(dx, dy)
+
+        if dist >= 0.20:
+            theta_ref = float(math.atan2(dx, dy))
+            look_real = float(s_fim - s_proj)
+        else:
+            theta_ref = float(theta_local)
+            look_real = 0.0
+
+        return {
+            "e_lat": float(e_lat),
+            "s_proj": float(s_proj),
+            "i_seg": int(i_seg),
+            "projecao": proj,
+            "theta_local": float(theta_local),
+            "theta_ref": float(theta_ref),
+            "margem": float(margem),
+            "lookahead_heading_m": float(max(0.0, look_real)),
+        }
+
+    def _erro_heading_caminho(self, theta_robo, theta_ref):
+        return float(self._wrap_pi(float(theta_ref) - float(theta_robo)))
+
+    def _tracking_reta_permitido(self, contexto, idx_base, e_lat, erro_heading_rad):
+        """
+        Decide se o path tracking geométrico deve assumir o controle.
+
+        Regras V5:
+
+        1) CaminhandoRua / EntrandoRua / SaindoRua / Direcionando /
+           RetornandoBase:
+           usam a mesma política geométrica universal. O estado operacional
+           continua útil para velocidade e semântica, mas não deve aprisionar
+           a direção em uma geometria específica.
+
+        2) Manobrando em trecho Rua -> Rua do MESMO corredor:
+           é recuperação de trajetória e também pode usar a política universal.
+
+        3) Manobra estrutural real:
+           fica fora deste ramo e preserva MovimentoArco obrigatório, adequado
+           à curva de cabeceira/180 graus.
+        """
+        carro = _as_dict(_as_dict(contexto).get("Carro", {}))
+        status = carro.get("Status", StatusCarroMapa.Parado.value)
+
+        if _status_in(
+            status,
+            [
+                StatusCarroMapa.CaminhandoRua,
+                StatusCarroMapa.EntrandoRua,
+                StatusCarroMapa.SaindoRua,
+                StatusCarroMapa.Direcionando,
+                StatusCarroMapa.RetornandoBase,
+            ],
+        ):
+            return True
+
+        if _status_in(status, [StatusCarroMapa.Manobrando]):
+            return self._segmento_rua_puro_mesmo_corredor(idx_base)
+
+        return False
+
+    def _referencia_direcional_reta(self, e_lat, erro_heading_rad, velocidade, tipo_preferido=None):
+        """Escolhe a geometria pelo estado geométrico rover x caminho.
+
+        Política universal V5:
+        - heading muito errado -> MovimentoArco para recuperar orientação com
+          menor raio;
+        - heading moderadamente errado -> RodasDianteiras;
+        - heading praticamente alinhado + cross-track -> MovimentoDiagonal;
+        - heading alinhado e centralizado -> RodasDianteiras praticamente reta.
+
+        A decisão usa heading do CAMINHO, nunca bearing rover->waypoint.
+        """
+        e_lat = float(e_lat)
+        v = max(0.0, float(velocidade))
+        erro_heading_rad = float(erro_heading_rad)
+        e_head_deg = abs(math.degrees(erro_heading_rad))
+        tipo_prev = _movimento_from_value(tipo_preferido)
+
+        # 1) Erro grande de heading tem prioridade absoluta sobre cross-track.
+        # A histerese mantém Arco até o erro cair bem abaixo do limiar de
+        # entrada, evitando caça de geometria em torno de 15 graus.
+        manter_arco = (
+            tipo_prev == TipoMovimentoDirecional.MovimentoArco
+            and e_head_deg > self._tracking_arco_saida_graus
+        )
+        entrar_arco = e_head_deg >= self._tracking_arco_entrada_graus
+
+        if entrar_arco or manter_arco:
+            delta = self._ganho_heading_reta * erro_heading_rad
+            lim = math.radians(self.angulo_max_graus)
+            return (
+                TipoMovimentoDirecional.MovimentoArco,
+                float(np.clip(delta, -lim, lim)),
+            )
+
+        # 2) Só usa diagonal quando o corpo já está praticamente paralelo ao
+        # caminho. Ela corrige deslocamento lateral sem criar yaw desnecessário.
+        precisa_recentralizar = abs(e_lat) > self._lateral_deadband_m
+        manter_diag = (
+            tipo_prev == TipoMovimentoDirecional.MovimentoDiagonal
+            and e_head_deg <= self._diagonal_saida_heading_graus
+        )
+        entrar_diag = (
+            precisa_recentralizar
+            and e_head_deg <= self._diagonal_entrada_heading_graus
+        )
+        usar_diag = bool(manter_diag or entrar_diag)
+
+        if usar_diag:
+            if abs(e_lat) <= self._lateral_deadband_m:
+                delta = 0.0
+            else:
+                delta = math.atan2(
+                    self._ganho_lateral_diagonal * e_lat,
+                    v + self._velocidade_offset_lateral,
+                )
+            lim = math.radians(min(self._angulo_diagonal_max_graus, self.angulo_max_graus))
+            return TipoMovimentoDirecional.MovimentoDiagonal, float(np.clip(delta, -lim, lim))
+
+        # 3) Faixa intermediária: RodasDianteiras têm UMA responsabilidade,
+        # corrigir heading. O erro lateral fica reservado para a diagonal quando
+        # o corpo voltar à janela estreita de alinhamento.
+        if e_head_deg <= self._heading_deadband_graus:
+            e_head_eff = 0.0
+        else:
+            e_head_eff = erro_heading_rad
+
+        delta = self._ganho_heading_reta * e_head_eff
+        lim = math.radians(self.angulo_max_graus)
+        return TipoMovimentoDirecional.RodasDianteiras, float(np.clip(delta, -lim, lim))
+
     def _estabilizar_angulo_saida(
         self,
         angulo_desejado_rad,
         status_carro,
         comando_anterior,
         dt_controle,
+        tipo_movimento=None,
     ):
-        """Filtra/rate-limita esterco somente na passada reta."""
+        """Rate-limit de saída para modos de tracking; Arco mantém autoridade."""
         desejado = float(np.clip(
             angulo_desejado_rad,
             -np.radians(self.angulo_max_graus),
             np.radians(self.angulo_max_graus),
         ))
 
-        if not _status_in(status_carro, [StatusCarroMapa.CaminhandoRua]):
+        tipo = _movimento_from_value(tipo_movimento) if tipo_movimento is not None else None
+        if tipo == TipoMovimentoDirecional.MovimentoArco:
+            return desejado
+        if tipo is None and not _status_in(status_carro, [StatusCarroMapa.CaminhandoRua]):
             return desejado
 
         anterior = np.radians(_safe_float(
@@ -631,21 +1049,51 @@ class ControladorMPC:
         ))
         dt = _safe_float(dt_controle, 0.5, min_value=0.05, max_value=1.5)
 
-        tau = float(self._filtro_angulo_reta_tau_s)
-        alpha = 1.0 if tau <= 1e-6 else dt / (tau + dt)
-        filtrado = anterior + alpha * (desejado - anterior)
-
-        max_delta = np.radians(self._taxa_angulo_reta_graus_s) * dt
-        filtrado = anterior + float(np.clip(
-            filtrado - anterior,
-            -max_delta,
-            max_delta,
-        ))
-
         dead = np.radians(self._deadband_angulo_reta_graus)
+
+        # Guarda de fase:
+        #
+        # Um filtro/rate-limit não pode continuar esterçando para a direita
+        # quando o MPC já pediu correção significativa para a esquerda, ou
+        # vice-versa. Nos logs reais isso chegou a persistir por vários ciclos
+        # e o rover atravessou a centerline antes de o comando cruzar zero.
+        #
+        # Na inversão significativa fazemos uma passagem deliberada por zero.
+        # É mais suave mecanicamente do que saltar direto de +A para -B e,
+        # principalmente, jamais aplica correção no sentido sabidamente errado.
+        inversao_significativa = bool(
+            abs(desejado) > dead
+            and abs(anterior) > dead
+            and (desejado * anterior) < 0.0
+        )
+
+        if inversao_significativa:
+            filtrado = 0.0
+        else:
+            tau = float(self._filtro_angulo_reta_tau_s)
+            alpha = 1.0 if tau <= 1e-6 else dt / (tau + dt)
+            filtrado = anterior + alpha * (desejado - anterior)
+
+            max_delta = np.radians(self._taxa_angulo_reta_graus_s) * dt
+            filtrado = anterior + float(np.clip(
+                filtrado - anterior,
+                -max_delta,
+                max_delta,
+            ))
+
         if abs(desejado) <= dead and abs(anterior) <= 2.0 * dead:
             filtrado = 0.0
 
+        # Última barreira: mesmo diante de qualquer alteração futura no filtro,
+        # uma saída não nula nunca pode permanecer no hemisfério oposto ao
+        # comando desejado.
+        if (
+            abs(desejado) > dead
+            and abs(filtrado) > dead
+            and (desejado * filtrado) < 0.0
+        ):
+            filtrado = 0.0
+
         return float(np.clip(
             filtrado,
             -np.radians(self.angulo_max_graus),
@@ -980,23 +1428,37 @@ class ControladorMPC:
 
     def _kappa_from_angle(self, tipo, a_rad, L, k_r=0.5, crab_in_phase=False, beta_crab=0.0):
         """
-        Mapeia ângulo de direção 'a_rad' -> curvatura κ (1/m) de acordo com o tipo.
-        - crab_in_phase=True: movimento 'crab' (quase κ=0), usa desloc. lateral beta_crab (rad).
+        Curvatura GEOMÉTRICA usada pela LUT do costmap.
+
+        Importante: não usa calcular_omega(v=1). O GPSHandler aplica um fator
+        dinâmico dependente da velocidade (Ku), então omega(1)/1 não representa
+        uma curvatura geométrica independente da velocidade. A LUT é deliberadamente
+        geométrica/conservadora; a simulação pesada do MPC usa calcular_omega com
+        a velocidade real do ciclo.
         """
         if crab_in_phase:
-            return 0.0  # tratamos crab fora (x ≈ y * tan(beta_crab))
-        if tipo == TipoMovimentoDirecional.RodasDianteiras:
-            return math.tan(a_rad) / L
-        if tipo == TipoMovimentoDirecional.RodasTraseiras:
-            return math.tan(a_rad) / L
-        elif tipo == TipoMovimentoDirecional.MovimentoArco:
-            return math.tan((1.0 - k_r) * a_rad) / L
-        elif tipo == TipoMovimentoDirecional.MovimentoDiagonal:
             return 0.0
-        elif tipo == TipoMovimentoDirecional.MovimentoLateral:
+
+        tipo_enum = _movimento_from_value(tipo)
+        L_eff = max(float(L), 1e-6)
+        t = math.tan(float(a_rad))
+
+        if abs(t) < 1e-9:
             return 0.0
-        else:
-            return math.tan(a_rad) / L  # default
+        if tipo_enum == TipoMovimentoDirecional.RodasDianteiras:
+            return t / L_eff
+        if tipo_enum == TipoMovimentoDirecional.RodasTraseiras:
+            # Mantém a convenção atualmente usada pelo GPSHandler/C#.
+            return t / L_eff
+        if tipo_enum == TipoMovimentoDirecional.MovimentoArco:
+            # df=+delta, dr=-delta => (tan(df)-tan(dr))/L = 2*tan(delta)/L
+            return (2.0 * t) / L_eff
+        if tipo_enum in [
+            TipoMovimentoDirecional.MovimentoDiagonal,
+            TipoMovimentoDirecional.MovimentoLateral,
+        ]:
+            return 0.0
+        return t / L_eff
 
     def _calcula_erro_posicao(self, x, y, ponto_alvo, d_sat=3.0):
         try:
@@ -1007,16 +1469,20 @@ class ControladorMPC:
             _log(f"Erro ao calcular erro de posicao: {e}")
             return 1.0, float("inf")
 
-    def _calcula_erro_orientacao(self, x, y, theta, ponto_alvo, angulo_caminho, peso_proximo_ponto=0.5):
+    def _calcula_erro_orientacao(self, x, y, theta, ponto_alvo, angulo_caminho, peso_proximo_ponto=0.5, usar_apenas_caminho=False):
         try:
-            orient_sim = self.gps_handler.calcular_orientacao((x, y), ponto_alvo)
-            orient_sim = (orient_sim + np.pi) % (2 * np.pi)
-            erro_ori_sim = self.gps_handler.erro_angular(orient_sim, theta, angulo_caminho, peso_proximo_ponto)
-            #erro_ori_sim = abs(self._wrap_pi(orient_sim - theta_sim))
-            #erro_ori_sim = abs(self._wrap_pi(theta_path - theta_sim))
-            o_norm = (erro_ori_sim / np.pi)
-            erro_ori_sim = np.degrees(erro_ori_sim)
-            return (o_norm, erro_ori_sim)
+            if usar_apenas_caminho:
+                erro = abs(self._wrap_pi(float(angulo_caminho) - float(theta)))
+            else:
+                # Compatibilidade para manobras: mantém o blend histórico
+                # entre bearing do alvo e direção local do caminho.
+                orient_sim = self.gps_handler.calcular_orientacao((x, y), ponto_alvo)
+                orient_sim = (orient_sim + np.pi) % (2 * np.pi)
+                erro = self.gps_handler.erro_angular(
+                    orient_sim, theta, angulo_caminho, peso_proximo_ponto
+                )
+            o_norm = abs(float(erro)) / np.pi
+            return (o_norm, abs(float(np.degrees(erro))))
         except Exception as e:
             _log(f"Erro ao calcular erro de orientacao: {e}")
             return 1.0, 180.0
@@ -1172,7 +1638,7 @@ class ControladorMPC:
                 (theta_trecho - theta_robo + math.pi) % (2.0 * math.pi)
                 - math.pi
             )
-            erro_maximo = math.radians(35.0 if estrutural else 45.0)
+            erro_maximo = math.radians(40.0 if estrutural else 50.0) # 35.0 e 45.0
 
             ponto_ficou_para_tras = (
                 math.isfinite(avanco)
@@ -1191,7 +1657,14 @@ class ControladorMPC:
 
         return self._proximo_nao_visitado(pontos_visitados)
 
-    def _calcular_pesos_movimento(self, contexto, erro_ori_graus, erro_lat_m):
+    def _calcular_pesos_movimento(
+        self,
+        contexto,
+        erro_ori_graus,
+        erro_lat_m,
+        tracking_reta=False,
+        erro_heading_caminho_graus=None,
+    ):
         # helpers simples
         def _clamp(x, lo, hi): 
             return lo if x < lo else hi if x > hi else x
@@ -1241,7 +1714,11 @@ class ControladorMPC:
             peso_suavidade  = _w_from_err_inv(erro_lat_m, w_min=pesos.get("suavidade_min", 0.5), w_max=pesos.get("suavidade_max", 3.6), deadband=0.05, tol=0.25, power=1.3, smooth=True)
             peso_fator_re   = pesos.get("fator_re", 0.05)
             peso_ideal      = pesos.get("ideal", 1.4)
-            peso_lateral    = _w_from_err(erro_lat_m, w_min=pesos.get("lateral_min", 1.8), w_max=pesos.get("lateral_max", 8.5), deadband=0.05, tol=0.7, power=1.3, smooth=True) if dentro else 0.0
+            peso_lateral    = _w_from_err(erro_lat_m, w_min=pesos.get("lateral_min", 1.8), w_max=pesos.get("lateral_max", 8.5), deadband=self._lateral_deadband_m, tol=0.7, power=1.3, smooth=True)
+            if not dentro:
+                # Se saiu da faixa, recuperar a centerline fica MAIS importante,
+                # nunca menos. A versão anterior zerava este custo.
+                peso_lateral *= 1.20
 
             # Penalidade proporcional à velocidade
             penalidade_por_velocidade = (velocidade / self.velocidade_max) / 10.0
@@ -1255,14 +1732,67 @@ class ControladorMPC:
             erro_ori_abs = abs(float(erro_ori_graus))
             erro_lat_abs = abs(float(erro_lat_m))
 
-            if status in [ StatusCarroMapa.EntrandoRua, StatusCarroMapa.SaindoRua, StatusCarroMapa.Manobrando ]:
+            if tracking_reta:
+                # V5: os pesos acompanham a mesma política geométrica usada na
+                # geração de candidatos. Embora normalmente apenas um tipo seja
+                # gerado por ciclo, manter o custo coerente evita preferências
+                # contraditórias nas simulações do horizonte.
+                if erro_heading_caminho_graus is None:
+                    erro_heading_modo = erro_ori_abs
+                else:
+                    erro_heading_modo = abs(float(erro_heading_caminho_graus))
+
+                if erro_heading_modo >= self._tracking_arco_entrada_graus:
+                    custo_movimento[TipoMovimentoDirecional.MovimentoArco] += 0.0
+                    custo_movimento[TipoMovimentoDirecional.RodasDianteiras] += 1.5
+                    custo_movimento[TipoMovimentoDirecional.MovimentoDiagonal] += 3.0
+                elif erro_heading_modo <= self._diagonal_saida_heading_graus:
+                    custo_movimento[TipoMovimentoDirecional.MovimentoArco] += 3.0
+                    custo_movimento[TipoMovimentoDirecional.MovimentoDiagonal] += 0.0
+                    custo_movimento[TipoMovimentoDirecional.RodasDianteiras] += 0.25
+                else:
+                    custo_movimento[TipoMovimentoDirecional.MovimentoArco] += 1.5
+                    custo_movimento[TipoMovimentoDirecional.RodasDianteiras] += 0.0
+                    custo_movimento[TipoMovimentoDirecional.MovimentoDiagonal] += 3.0
+
+            elif status == StatusCarroMapa.Manobrando:
+                # Cabeceira/ligação estrutural real: preserva a autoridade do
+                # Arco. Na geração de candidatos V4 ele continua sendo o único
+                # tipo permitido neste estado quando não estamos em recuperação
+                # Rua -> Rua da V3.
                 custo_movimento[TipoMovimentoDirecional.MovimentoArco] += 0.0
-                custo_movimento[TipoMovimentoDirecional.RodasDianteiras] += 1.0
-                custo_movimento[TipoMovimentoDirecional.MovimentoDiagonal] += 2.0
+                custo_movimento[TipoMovimentoDirecional.RodasDianteiras] += 1.5
+                custo_movimento[TipoMovimentoDirecional.MovimentoDiagonal] += 3.0
+
+            elif status in [StatusCarroMapa.EntrandoRua, StatusCarroMapa.SaindoRua]:
+                # V4: usa o heading do CAMINHO para decidir qual geometria
+                # merece preferência. O bearing/erro combinado continua nos
+                # demais custos da manobra, mas não decide sozinho o modo.
+                if erro_heading_caminho_graus is None:
+                    erro_heading_modo = erro_ori_abs
+                else:
+                    erro_heading_modo = abs(float(erro_heading_caminho_graus))
+
+                e_lo = float(self._transicao_rua_dianteira_heading_graus)
+                e_hi = float(self._transicao_rua_arco_heading_graus)
+
+                t = (erro_heading_modo - e_lo) / max(e_hi - e_lo, 1e-6)
+                t = _smoothstep01(_clamp(t, 0.0, 1.0))
+
+                penal = float(self._transicao_rua_penalidade_modo)
+
+                # t=0: alinhado    -> dianteira preferida
+                # t=1: desalinhado -> arco preferido
+                custo_movimento[TipoMovimentoDirecional.RodasDianteiras] += penal * t
+                custo_movimento[TipoMovimentoDirecional.MovimentoArco] += penal * (1.0 - t)
+
+                # Diagonal não pertence a esta política de transição.
+                custo_movimento[TipoMovimentoDirecional.MovimentoDiagonal] += 3.0
+
             else:
-                custo_movimento[TipoMovimentoDirecional.MovimentoArco] += 0.0 if erro_ori_abs >= 10.0 else 1.0
-                custo_movimento[TipoMovimentoDirecional.RodasDianteiras] += 1.2 if erro_ori_abs >= 10.0 else 0.0
-                custo_movimento[TipoMovimentoDirecional.MovimentoDiagonal] += 0.0 if erro_ori_abs <= 2.0 and erro_lat_abs <= 0.4 else 2.0
+                custo_movimento[TipoMovimentoDirecional.MovimentoArco] += 1.5
+                custo_movimento[TipoMovimentoDirecional.RodasDianteiras] += 0.0
+                custo_movimento[TipoMovimentoDirecional.MovimentoDiagonal] += 0.5
 
             return (peso_erro_pos, peso_erro_ori, peso_suavidade, peso_fator_re, peso_ideal, peso_lateral, custo_movimento)
         except Exception as e:
@@ -1769,7 +2299,11 @@ class ControladorMPC:
                     #candidatos_ativos = self._selecionar_melhores(candidatos_ativos, N=N_top)
                     beam_topN_max = int(getattr(self, "_beam_topN_max", 5))
                     beam_topN_min = int(getattr(self, "_beam_topN_min", 2))
-                    candidatos_ativos = self._filtrar_por_margem_angular(novos_candidatos, margem_graus=2.0)
+                    candidatos_ativos = self._filtrar_por_margem_angular(
+                        novos_candidatos,
+                        margem_graus=2.0,
+                        por_tipo=True,
+                    )
                     # número alvo de trajetórias ativas
                     #N_alvo = min(beam_topN_max, max(beam_topN_min, beam_traj_min))
                     beam_traj_min = max(1, min(beam_traj_min, beam_topN_max))  # sanity
@@ -1814,15 +2348,24 @@ class ControladorMPC:
                     status_carro,
                     comando_anterior,
                     self.tempo_execucao_local,
+                    tipo_movimento=tipo_final,
                 )
 
             ponto_alvo_primario = ponto_alvo_real["xy"]
-            erro_lateral, _, _, _, theta_path, _ = self.cross_track_error_point(
-                x,
-                y,
-                idx_referencia=idx_alvo_correcao,
+            ref_saida = self._referencia_caminho_lookahead(
+                x, y, idx_alvo_correcao, velocidade
             )
-            _, erro_orientacao = self._calcula_erro_orientacao(x, y, theta, ponto_alvo_primario, theta_path, 0.5)
+            erro_lateral = float(ref_saida["e_lat"])
+            erro_heading_saida = self._erro_heading_caminho(theta, ref_saida["theta_ref"])
+            tracking_reta_saida = self._tracking_reta_permitido(
+                contexto, idx_alvo_correcao, erro_lateral, erro_heading_saida
+            )
+            if tracking_reta_saida:
+                erro_orientacao = abs(float(np.degrees(erro_heading_saida)))
+            else:
+                _, erro_orientacao = self._calcula_erro_orientacao(
+                    x, y, theta, ponto_alvo_primario, ref_saida["theta_local"], 0.5
+                )
 
             _cmd = {
                 "enviar_comando": True,
@@ -1837,6 +2380,11 @@ class ControladorMPC:
                 "idx_alvo_autoritativo": int(idx_alvo_correcao),
                 "alvo_direcional_xy": ponto_alvo_primario,
                 "lookahead_direcional_m": float(ponto_alvo_real.get("_lookahead_m", 0.0)),
+                "tracking_reta": bool(tracking_reta_saida),
+                "heading_caminho_lookahead_graus": float(np.degrees(ref_saida["theta_ref"])),
+                "heading_caminho_local_graus": float(np.degrees(ref_saida["theta_local"])),
+                "heading_lookahead_m": float(ref_saida["lookahead_heading_m"]),
+                "erro_heading_caminho_graus": abs(float(np.degrees(erro_heading_saida))),
                 "debug_custo": debug_custo,
                 "candidatos_testados": K,
                 "motivos": []
@@ -1946,61 +2494,129 @@ class ControladorMPC:
         """
         try:
             now = time.perf_counter
+            CARRO = _as_dict(contexto.get("Carro", {}))
+            status = CARRO.get("Status", StatusCarroMapa.Parado.value)
+            velocidade = max(0.0, _safe_float(CARRO.get("Velocidade", 0.0), 0.0))
+            idx_base = _safe_int(ponto_alvo.get("_idx_base", 0), 0)
 
-            # 1) Ângulo ideal (wrap [-pi,pi] e clamp ao máx)
-            orient = self.gps_handler.calcular_orientacao(ponto_alvo["xy"], ponto_atual)
-            delta_theta = (orient - ponto_atual[2] + np.pi) % (2*np.pi) - np.pi
-
-            #delta_theta, theta_ref, dbg = self.delta_por_dist_esq_dir(
-            #    estado_atual=ponto_atual,
-            #    ponto_alvo_xy=ponto_alvo["xy"],
-            #    orient_corredor=np.radians(contexto["Carro"]["AnguloCaminho"]),  # em rad
-            #    dist_esq=contexto["Carro"]["DistanciaEsquerda"],
-            #    dist_dir=contexto["Carro"]["DistanciaDireita"],
-            #    angulo_max_graus=self.angulo_max_graus,
-            #    sigma_factor=0.6,
-            #    w_min=0.15,
-            #    usar_nudge_centro=True,
-            #    k_e=1.0, v=contexto["Carro"].get("Velocidade", 0.0), v0=0.3
-            #)
-
-            ang_max_rad = np.radians(self.angulo_max_graus)
-            angulo_ideal_rad = float(np.clip(delta_theta, -ang_max_rad, ang_max_rad))
-            angulo_ideal_deg = float(np.degrees(angulo_ideal_rad))
-            #print(f"Orient: {np.degrees(orient)}, delta theta: {np.degrees(delta_theta)}, angulo ideal: {angulo_ideal_deg}")
-
-            e_lat, _, _, _, theta_path, _ = self.cross_track_error_point(
-                ponto_atual[0],
-                ponto_atual[1],
-                idx_referencia=ponto_alvo.get("_idx_base"),
+            # --------------------------------------------------------------
+            # PATH TRACKING: heading do CAMINHO + cross-track, sem mirar ponto.
+            # A geometria é escolhida por erro de heading/cross-track (V5).
+            # MANOBRA estrutural: mantém pure-pursuit e Arco obrigatório.
+            # --------------------------------------------------------------
+            ref_path = self._referencia_caminho_lookahead(
+                ponto_atual[0], ponto_atual[1], idx_base, velocidade
             )
-            (_, e_ori) = self._calcula_erro_orientacao(ponto_atual[0], ponto_atual[1], ponto_atual[2], ponto_alvo["xy"], theta_path, 0.5)
+            e_lat = float(ref_path["e_lat"])
+            theta_ref = float(ref_path["theta_ref"])
+            erro_heading = self._erro_heading_caminho(ponto_atual[2], theta_ref)
+            e_ori = abs(math.degrees(erro_heading))
 
-            # 2) Tipos válidos conforme contexto
-            CARRO = contexto.get("Carro", {})
-            status = CARRO.get("Status", 0)
-            if _status_in(status, [StatusCarroMapa.EntrandoRua, StatusCarroMapa.SaindoRua, StatusCarroMapa.Manobrando]):
-                tipos_validos = [TipoMovimentoDirecional.MovimentoArco]
-            elif abs(e_ori) > 15.0:
-                tipos_validos = [TipoMovimentoDirecional.RodasDianteiras, TipoMovimentoDirecional.MovimentoArco]
-            else:
-                # Histerese de modo na passada. Antes, 0,20 m/3 graus era
-                # uma fronteira seca: qualquer ruido alternava Diagonal e
-                # RodasDianteiras. Entrar exige uma faixa estreita; depois de
-                # entrar, a faixa de permanencia e um pouco maior.
-                tipo_preferido_enum = _movimento_from_value(tipo_preferido)
-                manter_diagonal = (
-                    tipo_preferido_enum == TipoMovimentoDirecional.MovimentoDiagonal
-                    and abs(e_ori) <= 4.0
-                    and abs(e_lat) <= 0.30
+            tracking_reta = self._tracking_reta_permitido(
+                contexto, idx_base, e_lat, erro_heading
+            )
+
+            if tracking_reta:
+                tipo_ref, angulo_ideal_rad = self._referencia_direcional_reta(
+                    e_lat,
+                    erro_heading,
+                    velocidade,
+                    tipo_preferido=tipo_preferido,
                 )
-                entrar_diagonal = abs(e_ori) <= 1.5 and abs(e_lat) <= 0.12
+                tipos_validos = [tipo_ref]
+            else:
+                orient = self.gps_handler.calcular_orientacao(ponto_alvo["xy"], ponto_atual)
+                delta_theta = self._wrap_pi(orient - ponto_atual[2])
+                angulo_ideal_rad = float(np.clip(
+                    delta_theta,
+                    -np.radians(self.angulo_max_graus),
+                    np.radians(self.angulo_max_graus),
+                ))
+
+                tipo_preferido_enum = _movimento_from_value(tipo_preferido)
+                erro_heading_alvo_deg = abs(float(math.degrees(delta_theta)))
+
+                if _status_in(status, [StatusCarroMapa.Manobrando]):
+                    # Manobra estrutural real continua com Arco obrigatório.
+                    # A recuperação Rua -> Rua da V3 entra antes em
+                    # tracking_reta=True, portanto não cai neste ramo.
+                    tipos_validos = [TipoMovimentoDirecional.MovimentoArco]
+
+                elif _status_in(
+                    status,
+                    [StatusCarroMapa.EntrandoRua, StatusCarroMapa.SaindoRua],
+                ):
+                    # V4: entrada/saída são estados de TRANSIÇÃO. Arco e
+                    # RodasDianteiras são avaliados pelo mesmo MPC.
+                    #
+                    # Colocar o tipo anterior primeiro só estabiliza a divisão
+                    # do orçamento quando sobra um candidato ímpar. A escolha
+                    # final continua sendo pelo custo acumulado.
+                    if tipo_preferido_enum == TipoMovimentoDirecional.RodasDianteiras:
+                        tipos_validos = [
+                            TipoMovimentoDirecional.RodasDianteiras,
+                            TipoMovimentoDirecional.MovimentoArco,
+                        ]
+                    else:
+                        tipos_validos = [
+                            TipoMovimentoDirecional.MovimentoArco,
+                            TipoMovimentoDirecional.RodasDianteiras,
+                        ]
+
+                elif _status_in(status, [StatusCarroMapa.Direcionando]):
+                    # Fora da rua, um heading muito errado deve ser corrigido
+                    # com Arco, que possui raio de giro menor. Quando o rover
+                    # já estiver alinhado, RodasDianteiras finaliza de forma
+                    # mais suave. A faixa 8°..15° é a histerese: se já entrou
+                    # em Arco, permanece nele até cruzar o limiar de saída.
+                    manter_arco = (
+                        tipo_preferido_enum == TipoMovimentoDirecional.MovimentoArco
+                        and erro_heading_alvo_deg > self._direcionando_arco_saida_graus
+                    )
+
+                    if (
+                        erro_heading_alvo_deg >= self._direcionando_arco_entrada_graus
+                        or manter_arco
+                    ):
+                        tipos_validos = [TipoMovimentoDirecional.MovimentoArco]
+                    else:
+                        tipos_validos = [TipoMovimentoDirecional.RodasDianteiras]
+
+                elif _status_in(status, [StatusCarroMapa.RetornandoBase]):
+                    # Guarda defensiva. Em V5 RetornandoBase deve entrar no
+                    # path tracking acima; se algum contexto incompleto impedir
+                    # isso, ainda assim nunca volta ao antigo fallback Front-only.
+                    manter_arco = (
+                        tipo_preferido_enum == TipoMovimentoDirecional.MovimentoArco
+                        and erro_heading_alvo_deg > self._tracking_arco_saida_graus
+                    )
+                    if (
+                        erro_heading_alvo_deg >= self._tracking_arco_entrada_graus
+                        or manter_arco
+                    ):
+                        tipos_validos = [TipoMovimentoDirecional.MovimentoArco]
+                    else:
+                        tipos_validos = [TipoMovimentoDirecional.RodasDianteiras]
 
-                if manter_diagonal or entrar_diagonal:
-                    tipos_validos = [TipoMovimentoDirecional.MovimentoDiagonal]
                 else:
                     tipos_validos = [TipoMovimentoDirecional.RodasDianteiras]
 
+                # Para ranking/telemetria de manobra, preserva erro histórico.
+                _, _, _, _, theta_local, _ = self.cross_track_error_point(
+                    ponto_atual[0], ponto_atual[1], idx_referencia=idx_base
+                )
+                _, e_ori = self._calcula_erro_orientacao(
+                    ponto_atual[0], ponto_atual[1], ponto_atual[2],
+                    ponto_alvo["xy"], theta_local, 0.5, usar_apenas_caminho=False
+                )
+
+            angulo_ideal_rad = float(np.clip(
+                angulo_ideal_rad,
+                -np.radians(self.angulo_max_graus),
+                np.radians(self.angulo_max_graus),
+            ))
+            angulo_ideal_deg = float(np.degrees(angulo_ideal_rad))
+
             # 3) Flags de matriz
             if dados_costmap is None:
                 dados_costmap = self._get_costmap_direcional(contexto)
@@ -2008,6 +2624,11 @@ class ControladorMPC:
             
             # 4) Geração bruta (graus)
             angulos_raw = self._gerar_candidatos_brutos(angulo_ideal_deg, filtrar_matriz)
+            if tipos_validos == [TipoMovimentoDirecional.MovimentoDiagonal]:
+                lim_diag = float(min(self._angulo_diagonal_max_graus, self.angulo_max_graus))
+                angulos_raw = [a for a in angulos_raw if abs(float(a)) <= lim_diag + 1e-9]
+                if not angulos_raw:
+                    angulos_raw = [float(np.clip(angulo_ideal_deg, -lim_diag, lim_diag))]
             if not angulos_raw:
                 # fallback mínimo em torno do ideal
                 base = [angulo_ideal_deg + d for d in (-2.0, -1.0, 0.0, 1.0, 2.0)]
@@ -2449,21 +3070,25 @@ class ControladorMPC:
             omega = float(self.gps_handler.calcular_omega(v, angulo, tipo))
 
             k = np.arange(1, n + 1, dtype=np.float32)
-            if abs(omega) < 1e-6:
-                # reta: mesmo esquema do seu _nova_posicao
-                dx = dist_step * k * math.sin(th0)
-                dy = dist_step * k * math.cos(th0)
+            if tipo in [TipoMovimentoDirecional.MovimentoDiagonal, TipoMovimentoDirecional.MovimentoLateral]:
+                # 4WS em fase: heading do corpo permanece, mas a velocidade
+                # translacional aponta para theta + angulo.
+                direcao = th0 + float(angulo)
+                dx = dist_step * k * math.sin(direcao)
+                dy = dist_step * k * math.cos(direcao)
                 th = th0 + np.zeros_like(k, dtype=np.float32)
                 x = x0 + dx
                 y = y0 + dy
             else:
-                R   = v / omega
+                # Mesma integração semi-implícita usada por _nova_posicao e
+                # pelo simulador C#: primeiro atualiza theta, depois translada.
+                # O cumsum reproduz exatamente a aplicação passo a passo.
                 dth = omega * dt
-                th  = th0 + dth * k
-                s0, c0 = math.sin(th0), math.cos(th0)
-                # solução fechada consistente com x+=sin(θ)*d, y+=cos(θ)*d
-                x = x0 + R * (c0 - np.cos(th))
-                y = y0 + R * (np.sin(th) - s0)
+                th = th0 + dth * k
+                dx_step = dist_step * np.sin(th)
+                dy_step = dist_step * np.cos(th)
+                x = x0 + np.cumsum(dx_step)
+                y = y0 + np.cumsum(dy_step)
 
             return list(zip(x.astype(float).tolist(),
                             y.astype(float).tolist(),
@@ -2841,9 +3466,16 @@ class ControladorMPC:
             sub_dt = dt_total / n_subs
             dist_sub = v_planejado * sub_dt  # custo por metro
 
-            # --- PATCH: estado local p/ amortecimento e anti-chatter
-            e_lat_prev_local = None
-            side_mem_local = 0  # -1 direita, +1 esquerda, 0 centro
+            # Estado inicial real do candidato para derivada/anti-chatter.
+            try:
+                e_lat_inicial, _, _, _, _, _ = self.cross_track_error_point(
+                    x_sim, y_sim, idx_referencia=self._proximo_nao_visitado(pontos_visitados)
+                )
+                e_lat_prev_local = float(e_lat_inicial)
+                side_mem_local = 0 if abs(e_lat_inicial) < max(0.08, self._lateral_deadband_m * 2.0) else (1 if e_lat_inicial > 0 else -1)
+            except Exception:
+                e_lat_prev_local = None
+                side_mem_local = 0
 
             for n_sub in range(n_subs):
                 # IMPORTANTE: _nova_posicao precisa aceitar dt opcional
@@ -2875,90 +3507,109 @@ class ControladorMPC:
                 # --------------------------
                 # ERROS / CUSTOS
                 # --------------------------
-                # posição (mantido)
-                (d_norm, erro_pos) = self._calcula_erro_posicao(x_sim, y_sim, ponto_alvo_sim)
-                
-                # --- PATCH A: usa theta do CAMINHO (segmento projetado) como referência
-                # e usa também margem local pra normalização do custo lateral
-                e_lat, _, _, _, theta_path, margem = self.cross_track_error_point(
-                    x_sim,
-                    y_sim,
-                    idx_referencia=idx_alvo_sim,
+                ref_path_sim = self._referencia_caminho_lookahead(
+                    x_sim, y_sim, idx_alvo_sim, v_planejado
+                )
+                e_lat = float(ref_path_sim["e_lat"])
+                theta_path = float(ref_path_sim["theta_ref"])
+                margem = float(ref_path_sim["margem"])
+                erro_heading_sim = self._erro_heading_caminho(theta_sim, theta_path)
+                tracking_reta_sim = self._tracking_reta_permitido(
+                    contexto, idx_alvo_sim, e_lat, erro_heading_sim
                 )
 
-                # erro de orientação
-                (o_norm, erro_ori_sim) = self._calcula_erro_orientacao(x_sim, y_sim, theta_sim, ponto_alvo_sim, theta_path, 0.5)
+                if tracking_reta_sim:
+                    # Passada: NÃO existe atração para ponto. O custo é path
+                    # tracking puro: heading da centerline + cross-track.
+                    erro_pos = abs(e_lat)
+                    d_norm = 0.0
+                    erro_ori_sim = abs(float(np.degrees(erro_heading_sim)))
+                    o_norm = abs(float(erro_heading_sim)) / np.pi
+                else:
+                    (d_norm, erro_pos) = self._calcula_erro_posicao(
+                        x_sim, y_sim, ponto_alvo_sim
+                    )
+                    (o_norm, erro_ori_sim) = self._calcula_erro_orientacao(
+                        x_sim, y_sim, theta_sim,
+                        ponto_alvo_sim, ref_path_sim["theta_local"], 0.5,
+                        usar_apenas_caminho=False,
+                    )
 
-                # pesos dinâmicos (mantido)
-                pesos = self._calcular_pesos_movimento(contexto, erro_ori_sim, e_lat)
-                (peso_erro_pos, peso_erro_ori, peso_suavidade, peso_fator_re, peso_ideal, peso_lateral, peso_movimento) = pesos
+                pesos = self._calcular_pesos_movimento(
+                    contexto,
+                    erro_ori_sim,
+                    e_lat,
+                    tracking_reta=tracking_reta_sim,
+                    erro_heading_caminho_graus=abs(
+                        float(np.degrees(erro_heading_sim))
+                    ),
+                )
+                (peso_erro_pos, peso_erro_ori, peso_suavidade,
+                 peso_fator_re, peso_ideal, peso_lateral,
+                 peso_movimento) = pesos
 
-                # ângulo relativo ao alvo e suavidade (parcialmente mantido)
-                vetor_alvo = np.array(ponto_alvo_sim) - np.array([x_sim, y_sim])
-                vetor_alvo_norm = vetor_alvo / (np.linalg.norm(vetor_alvo) + 1e-9)
-                # Mesmo referencial de _nova_posicao: x=Leste usa sin(theta)
-                # e y=Norte usa cos(theta).
-                vetor_movel = np.array([np.sin(theta_sim), np.cos(theta_sim)])
-                cos_delta = float(np.dot(vetor_movel, vetor_alvo_norm))
-                delta_angulo = abs(angulo_testado - prev_ang)
+                delta_angulo = abs(self._wrap_pi(float(angulo_testado) - float(prev_ang)))
 
                 # --------------------------
                 # COMPONENTES DE CUSTO
                 # --------------------------
-                # posição (mantido)
-                #custo_pos = erro_pos * peso_erro_pos
-                custo_pos = d_norm * peso_erro_pos
-
-                # --- PATCH B: custo de orientação desacoplado de erro_pos e quadrático
-                #custo_ori = erro_ori_sim * peso_erro_ori
+                custo_pos = 0.0 if tracking_reta_sim else d_norm * peso_erro_pos
                 custo_ori = o_norm * peso_erro_ori
 
-                # --- PATCH C: suavidade SEM depender de erro_pos + piso
-                base_suav = 0.25  # piso de penalização (0.25–0.5)
-                erro_suavidade = ((base_suav + 1.0) * (abs(delta_angulo) / np.pi) * erro_pos)
+                # Suavidade realmente independente do erro de posição.
+                ang_norm = delta_angulo / max(np.radians(self.angulo_max_graus), 1e-6)
+                erro_suavidade = float(ang_norm * ang_norm)
                 custo_suavidade = erro_suavidade * peso_suavidade
 
-                # tipo de movimento (mantido)
                 custo_tipo_movimento = float(peso_movimento[tipo])
+                if prev_tipo != tipo:
+                    custo_tipo_movimento += float(self._penalidade_troca_tipo)
 
-                # fator "re" (mantido como estava)
-                erro_re = -cos_delta                      # já está em [-1, +1]
-                custo_re = erro_re * peso_fator_re        # aplica peso normalmente
+                if tracking_reta_sim:
+                    # Pequena recompensa por manter o corpo paralelo à linha;
+                    # não há vetor carro->ponto neste termo.
+                    cos_delta = math.cos(float(erro_heading_sim))
+                else:
+                    vetor_alvo = np.array(ponto_alvo_sim) - np.array([x_sim, y_sim])
+                    vetor_alvo_norm = vetor_alvo / (np.linalg.norm(vetor_alvo) + 1e-9)
+                    vetor_movel = np.array([np.sin(theta_sim), np.cos(theta_sim)])
+                    cos_delta = float(np.dot(vetor_movel, vetor_alvo_norm))
+                erro_re = -float(cos_delta)
+                custo_re = erro_re * peso_fator_re
 
-                # --- PATCH D: custo lateral com deadband + quadrático + normalização
-                dead = 0.0                            # 5–12 cm
-                half = max(float(margem) / 2.0, 0.5)       # meia-largura mínima
+                # Cross-track sempre ativo, inclusive fora do corredor.
+                dead = self._lateral_deadband_m
+                half = max(float(margem) / 2.0, 0.5)
                 e_eff = max(0.0, abs(e_lat) - dead)
-                r = e_eff/half
+                r = e_eff / half
                 if r <= 1.0:
-                    alpha = 0.25          # 0 = só quadrático; 1 = só linear
+                    alpha = 0.25
                     l_norm = alpha*r + (1-alpha)*r**2
                 else:
-                    gamma = 1.0
-                    l_norm = 1.0 + gamma*(r - 1.0)    # ex.: gamma = 1.0 (ajuste)
+                    l_norm = 1.0 + (r - 1.0)
                 custo_lateral = l_norm * peso_lateral
 
-                # --- PATCH E: amortecimento por derivada do erro lateral (m/s)
+                # Derivada funciona mesmo com n_subporpasso=1 porque o estado
+                # inicial do candidato é usado como memória.
                 if e_lat_prev_local is not None:
-                    de = (e_lat - e_lat_prev_local) / max(sub_dt, 1e-6)  # m/s
-                    peso_lat_der = 0.1 * float(peso_lateral)            # ganho pequeno
+                    de = (e_lat - e_lat_prev_local) / max(sub_dt, 1e-6)
+                    peso_lat_der = 0.08 * float(peso_lateral)
                     custo_lat_der = (de ** 2) * peso_lat_der
-
-                    dlat = (abs(e_lat) - abs(e_lat_prev_local)) / max(sub_dt,1e-6)
-                    pen_away = max(0.0, dlat/half) * (0.2 * peso_lateral)
+                    dlat = (abs(e_lat) - abs(e_lat_prev_local)) / max(sub_dt, 1e-6)
+                    pen_away = max(0.0, dlat / half) * (0.15 * peso_lateral)
                 else:
                     custo_lat_der = 0.0
                     pen_away = 0.0
                 custo_lateral += pen_away
                 e_lat_prev_local = e_lat
 
-                # --- PATCH F: anti-chatter (histerese) perto do centro
-                band = 0.10  # 8–12 cm
+                band = max(0.08, self._lateral_deadband_m * 2.0)
                 side_now = 0 if abs(e_lat) < band else (1 if e_lat > 0 else -1)
                 flip_pen = 0.0
-                if abs(e_lat) < band and side_mem_local != 0 and side_now != side_mem_local:
-                    flip_pen = 0.12 * float(peso_suavidade)  # 0.08–0.20
-                side_mem_local = side_now
+                if side_mem_local != 0 and side_now != 0 and side_now != side_mem_local:
+                    flip_pen = 0.12 * float(peso_suavidade)
+                if side_now != 0:
+                    side_mem_local = side_now
 
                 # soma local (heurística do visual + demais) e NORMALIZA por metro
                 custo_mapa = (
diff --git a/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/visual_worker/camera_manager.py b/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/visual_worker/camera_manager.py
index cb0dea483..ce16e08d8 100644
--- a/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/visual_worker/camera_manager.py
+++ b/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/visual_worker/camera_manager.py
@@ -4,16 +4,16 @@ import datetime
 import os
 import threading
 import time
-from typing import Any, Dict, Optional
+from typing import Optional, Set
 
 import cv2
 import numpy as np
 
 from camera_worker.camera_oak import CameraOak
-from camera_worker.segformer_runner import SegformerNavRunner
+from visual_worker.inferencia.corridor_onnx_runner import CorridorOnnxRunner
 
 from shared.contexto_global_redis import ContextoGlobalRedis, CtxKey
-from shared.enums import StatusModulo, T_Code, TipoFrameCamera
+from shared.enums import StatusModulo, StatusCarroMapa, T_Code, TipoFrameCamera
 from shared.gpu_priority_controller import GpuPriorityController
 from shared.perf_monitor import VisualPerfMonitor
 from shared.utils import get_velocidade_atual_ms
@@ -27,7 +27,6 @@ from visual_worker.processamento.costmap_fuser import CostmapFuser
 from visual_worker.processamento.visual_grid_builder import (
     GridGeometryConfig,
     GridReferenceBuilder,
-    GridConfidenceConfig,
     VisualGridBuilder,
 )
 from visual_worker.processamento.visual_debug_renderer import VisualDebugRenderer
@@ -35,12 +34,12 @@ from visual_worker.processamento.visual_debug_renderer import VisualDebugRendere
 
 class CameraManager:
     """
-    CameraManager v1 do Visual Worker.
+    CameraManager v2 do Visual Worker.
 
     Responsabilidade:
       - Inicializar OAK-D Lite.
-      - Inicializar SegformerNavRunner ONNX/TensorRT.
-      - Inicializar SegmentacaoManager v1.
+      - Inicializar CorridorOnnxRunner ONNX/TensorRT.
+      - Inicializar SegmentacaoManager v2.
       - Inicializar grid builder + CostmapFuser.
       - Orquestrar loops:
           segmentação -> detecção onboard -> grid/costmap -> publicação -> stream/debug.
@@ -83,7 +82,12 @@ class CameraManager:
         self.operante = False
         self.iniciando = False
         self.debug_visual = False
-        self.debug_perf = False
+
+        # Telemetria técnica fica explicitamente controlada pelo config.
+        self.telemetry_enabled = True
+        self.telemetry_console = True
+        self.telemetry_publish_runtime = True
+        self.telemetry_runner_timing = True
 
         self.largura_robo_m = 0.85
 
@@ -92,10 +96,13 @@ class CameraManager:
         self._loop_grid_iniciado = False
         self._loop_publicacao_iniciado = False
         self._loop_stream_iniciado = False
+        self._loop_preview_iniciado = False
         self._loop_analise_iniciado = False
 
         self._vida_lock = threading.RLock()
         self._cache_lock = threading.RLock()
+        self._preview_lock = threading.RLock()
+        self._preview_cond = threading.Condition(self._preview_lock)
         self._pub_lock = threading.RLock()
 
         self._fechando_camera = False
@@ -110,6 +117,13 @@ class CameraManager:
         self._posproc_salvando = False
         self.posproc_intervalo_min_s = 5.0
 
+        # Preview é um fluxo lateral ao runtime crítico. Ele só é produzido
+        # quando algum consumidor solicita um tipo de frame.
+        self.preview_fps = 2.0
+        self.preview_request_timeout_s = 0.35
+        self.preview_cache_fresh_s = 0.75
+        self.replay_jpeg_quality = 70
+
         self._reset_runtime_state()
 
     # ============================================================
@@ -121,7 +135,7 @@ class CameraManager:
         self._ultimo_depth_frame = None
 
         self._ultimo_predictions = None
-        self._ultimo_seg_aux_result = None
+        self._ultimo_status_result = None
         self._ultima_analise_segmentacao = None
 
         self._ultimo_detections = []
@@ -132,24 +146,35 @@ class CameraManager:
         self._startup_grace_until = 0.0
 
         self._seg_cache = {
+            # ts mantido como alias de compatibilidade para result_ts.
             "ts": 0.0,
+            "result_ts": 0.0,
             "frame_ts": 0.0,
             "predictions": None,
-            "aux_result": None,
+            "status_result": None,
             "analise": None,
             "res": None,
             "infer_ms": 0.0,
             "post_ms": 0.0,
+            "runner_timing_ms": {},
         }
 
         self._det_cache = {
+            # ts mantido como alias de compatibilidade para result_ts.
             "ts": 0.0,
+            "result_ts": 0.0,
+            "frame_ts": 0.0,
             "detections": [],
             "res": None,
         }
 
         self._grid_cache = {
+            # ts mantido como alias de compatibilidade para result_ts.
             "ts": 0.0,
+            "result_ts": 0.0,
+            "depth_frame_ts": 0.0,
+            "seg_frame_ts": 0.0,
+            "det_frame_ts": 0.0,
             "snapshot": None,
             "grid_conf": None,
         }
@@ -182,6 +207,7 @@ class CameraManager:
 
         self._ultimo_frame_preview_por_tipo = {}
         self._ultimo_frame_preview_ts_por_tipo = {}
+        self._preview_requests: Set[str] = set()
 
     def inicializar(self, mx_id):
         if self.iniciando:
@@ -206,9 +232,18 @@ class CameraManager:
                 self.det_config = load_det_config()
 
                 self.posproc_intervalo_min_s = float(self.seg_config.get("posproc_intervalo_min_s", 5.0))
+                self.preview_fps = float(self.seg_config.get("preview_fps", 2.0))
+                self.preview_request_timeout_s = float(self.seg_config.get("preview_request_timeout_s", 0.35))
+                self.preview_cache_fresh_s = float(self.seg_config.get("preview_cache_fresh_s", 0.75))
+                self.replay_jpeg_quality = int(self.seg_config.get("replay_jpeg_quality", 70))
 
                 self.debug_visual = bool(self.seg_config.get("debug_visual", False))
-                self.debug_perf = bool(self.seg_config.get("debug_perf", False))
+
+                telemetry_cfg = self.seg_config.get("telemetry", {}) or {}
+                self.telemetry_enabled = bool(telemetry_cfg.get("enabled", True))
+                self.telemetry_console = bool(telemetry_cfg.get("console_performance", True))
+                self.telemetry_publish_runtime = bool(telemetry_cfg.get("publish_runtime_info", True))
+                self.telemetry_runner_timing = bool(telemetry_cfg.get("runner_detailed_timing", True))
 
                 self._inicializar_camera(mx_id)
 
@@ -278,17 +313,37 @@ class CameraManager:
             self.mostrar_log(f"⚠️ Camera com ID {mx_id} não conectada: {e}")
 
     def _inicializar_segmentador(self):
-        self.seg_runner = SegformerNavRunner(self.seg_config)
+        model_path = self.seg_config.get("model_path")
+        if not model_path:
+            raise RuntimeError("Visual Worker v2 sem model_path do ONNX de corredor.")
 
+        self.seg_runner = CorridorOnnxRunner(
+            onnx_path=model_path,
+            runtime_config=self.seg_config.get("model_runtime", {}) or {},
+            telemetry_config=self.seg_config.get("telemetry", {}) or {},
+            mostrar_log=self.mostrar_log,
+        )
+
+        # Contrato ONNX é a fonte da verdade para classes e status.
+        self.seg_runner.contract.validate_status_enum(StatusCarroMapa)
+
+        runtime = self.seg_runner.runtime_info()
         self.mostrar_log(
-            "[SEG_ONNX] Runner iniciado | "
-            f"model={self.seg_config.get('onnx_model_path')} | "
-            f"provider={self.seg_config.get('onnx_provider', 'tensorrt')} | "
-            f"resolution={self.seg_config.get('ia_resolution')}"
+            "[SEG_RUNTIME] pronto | "
+            f"model={runtime['model_path']} | "
+            f"provider={runtime['provider_primary']} | "
+            f"resolution={self.seg_runner.contract.resolution_wh}"
         )
 
     def _inicializar_analisador_segmentacao(self):
-        cfg_dict = self.seg_config.get("segmentacao", {}) or {}
+        cfg_dict = dict(self.seg_config.get("segmentacao", {}) or {})
+
+        # Nunca duplicar estes IDs no config. Eles pertencem ao modelo.
+        cfg_dict["id_navegavel"] = int(self.seg_runner.contract.nav_class_id)
+        cfg_dict["id_nao_navegavel"] = int(self.seg_runner.contract.non_nav_class_id)
+        cfg_dict["include_model_probs"] = bool(
+            (self.seg_config.get("telemetry", {}) or {}).get("include_status_probs", False)
+        )
 
         seg_cfg = SegmentacaoConfig(**{
             k: v
@@ -298,16 +353,20 @@ class CameraManager:
 
         self.segmentacao_manager = SegmentacaoManager(
             config=seg_cfg,
-            color_map=getattr(self.seg_runner, "colormap_rgb", None),
-            classes=getattr(self.seg_runner, "classes", None),
+            color_map=self.seg_runner.colormap_rgb,
+            classes=self.seg_runner.classes,
         )
 
     def _inicializar_grid(self):
         grid_cfg = self.seg_config.get("grid", {}) or {}
         geom_cfg = grid_cfg.get("geometry", {}) or {}
-        conf_cfg = grid_cfg.get("confidence", {}) or {}
+        conf_cfg = dict(grid_cfg.get("confidence", {}) or {})
         fuser_cfg = grid_cfg.get("fuser", {}) or {}
 
+        # IDs semânticos pertencem ao contrato do modelo, não ao config.
+        conf_cfg["id_navegavel"] = int(self.seg_runner.contract.nav_class_id)
+        conf_cfg["id_nao_navegavel"] = int(self.seg_runner.contract.non_nav_class_id)
+
         if "grid_shape" not in geom_cfg:
             geom_cfg["grid_shape"] = tuple(grid_cfg.get("grid_shape", (15, 10)))
 
@@ -388,6 +447,10 @@ class CameraManager:
         freq_inferencia = float(seg_config.get("inferencia_fps", freq_analise))
         freq_deteccao = float(seg_config.get("deteccao_fps", freq_analise))
 
+        if not self._loop_preview_iniciado:
+            self._iniciar_loop_preview(freq=self.preview_fps)
+            self._loop_preview_iniciado = True
+
         if not self._loop_stream_iniciado:
             stream = getattr(self.camera, "stream", None)
             stream_fps = float(getattr(stream, "_op_fps", 2.0) or 2.0)
@@ -460,7 +523,7 @@ class CameraManager:
         depth_frame = None
         depth_ts = None
         predictions = None
-        aux_result = None
+        status_result = None
         analise_seg = None
         dets = []
 
@@ -511,28 +574,32 @@ class CameraManager:
             for i in range(max(0, n_seg)):
                 try:
                     rgb_frame, frame_ts, res = self.get_rgb_frame()
-                    rgb_frame = self._validar_frame_rgb_para_segmentacao(rgb_frame)
+                    rgb_frame = self._validar_frame_camera_para_segmentacao(rgb_frame)
 
                     if rgb_frame is None:
                         continue
 
                     t_inf0 = time.perf_counter()
-                    predictions, ts, roi_resized, roi_info, aux_result = self.seg_runner.infer_ids(rgb_frame)
+                    infer_result = self.seg_runner.infer(rgb_frame)
                     t_inf1 = time.perf_counter()
 
-                    if predictions is None or ts is None:
-                        continue
+                    predictions = infer_result.seg_ids
+                    status_result = infer_result.status
 
                     t_post0 = time.perf_counter()
-                    analise_seg = self.segmentacao_manager.analisar(predictions, aux_result=aux_result)
+                    analise_seg = self.segmentacao_manager.analisar(
+                        predictions,
+                        status_result=status_result,
+                    )
                     t_post1 = time.perf_counter()
 
                     self._store_segmentacao(
                         predictions=predictions,
-                        aux_result=aux_result,
+                        status_result=status_result,
                         analise=analise_seg,
                         frame_ts=frame_ts,
                         res=res,
+                        runner_timing_ms=infer_result.timing_ms,
                         infer_ms=(t_inf1 - t_inf0) * 1000.0,
                         post_ms=(t_post1 - t_post0) * 1000.0,
                     )
@@ -561,7 +628,7 @@ class CameraManager:
                     if depth_frame is None:
                         continue
 
-                    snapshot, grid_conf = self._processar_grid(
+                    snapshot, grid_conf, _grid_timing = self._processar_grid(
                         depth_frame=depth_frame,
                         segmentacao=predictions,
                         dados_visuais_seg=analise_seg.get("dados_visuais", {}),
@@ -572,7 +639,13 @@ class CameraManager:
                     if snapshot is None:
                         continue
 
-                    self._store_grid(snapshot=snapshot, grid_conf=grid_conf)
+                    self._store_grid(
+                        snapshot=snapshot,
+                        grid_conf=grid_conf,
+                        depth_frame_ts=depth_ts or 0.0,
+                        seg_frame_ts=float(self._seg_cache.get("frame_ts", 0.0) or 0.0),
+                        det_frame_ts=float(self._det_cache.get("frame_ts", 0.0) or 0.0),
+                    )
 
                     grid_ok += 1
 
@@ -628,21 +701,25 @@ class CameraManager:
 
         return False
 
-    def _validar_frame_rgb_para_segmentacao(self, rgb_frame):
-        if rgb_frame is None:
+    def _validar_frame_camera_para_segmentacao(self, frame_camera):
+        if frame_camera is None:
             return None
 
-        if not hasattr(rgb_frame, "shape") or rgb_frame.size == 0:
+        if not hasattr(frame_camera, "shape") or frame_camera.size == 0:
             return None
 
-        if rgb_frame.ndim != 3 or rgb_frame.shape[2] != 3:
-            self.mostrar_log(f"[SEG_ONNX] RGB inválido shape={getattr(rgb_frame, 'shape', None)}")
+        if frame_camera.ndim != 3 or frame_camera.shape[2] != 3:
+            self.mostrar_log(f"[SEG_ONNX] Frame câmera inválido shape={getattr(frame_camera, 'shape', None)}")
             return None
 
-        if rgb_frame.dtype != np.uint8:
-            rgb_frame = np.clip(rgb_frame, 0, 255).astype(np.uint8)
+        if frame_camera.dtype != np.uint8:
+            self.mostrar_log(
+                f"[SEG_ONNX] Frame câmera precisa ser uint8; veio dtype={frame_camera.dtype}"
+            )
+            return None
 
-        return np.ascontiguousarray(rgb_frame)
+        # Normalmente o frame da OAK já é C-contiguous. Só copia se a fonte mudar.
+        return frame_camera if frame_camera.flags.c_contiguous else np.ascontiguousarray(frame_camera)
 
     def get_rgb_frame(self):
         cam = self.camera
@@ -651,7 +728,9 @@ class CameraManager:
             return None, None, None
 
         try:
-            frame, res = cam.requisitar_frame_rgb()
+            # Hot path do Visual Worker: referência read-only ao latest frame.
+            # CameraOak substitui o cache, nunca modifica o ndarray antigo.
+            frame, res = cam.requisitar_frame_rgb_ref()
             erro = (res or {}).get("erro")
 
             if self._is_erro_fatal_camera(erro):
@@ -684,7 +763,7 @@ class CameraManager:
             return None, None, None
 
         try:
-            frame, res = cam.requisitar_frame_depth()
+            frame, res = cam.requisitar_frame_depth_ref()
             erro = (res or {}).get("erro")
 
             if self._is_erro_fatal_camera(erro):
@@ -735,50 +814,83 @@ class CameraManager:
     # Store helpers
     # ============================================================
 
-    def _store_segmentacao(self, predictions, aux_result, analise, frame_ts, res, infer_ms=0.0, post_ms=0.0):
-        agora = time.time()
+    def _store_segmentacao(
+        self,
+        predictions,
+        status_result,
+        analise,
+        frame_ts,
+        res,
+        runner_timing_ms=None,
+        infer_ms=0.0,
+        post_ms=0.0,
+    ):
+        result_ts = time.time()
+        frame_ts = float(frame_ts or 0.0)
+        runner_timing_ms = dict(runner_timing_ms or {})
 
         with self._cache_lock:
             self._ultimo_predictions = predictions
-            self._ultimo_seg_aux_result = aux_result
+            self._ultimo_status_result = status_result
             self._ultima_analise_segmentacao = analise
 
             self._seg_cache = {
-                "ts": agora,
-                "frame_ts": frame_ts or 0.0,
+                "ts": result_ts,
+                "result_ts": result_ts,
+                "frame_ts": frame_ts,
                 "predictions": predictions,
-                "aux_result": aux_result,
+                "status_result": status_result,
                 "analise": analise,
                 "res": res,
                 "infer_ms": float(infer_ms or 0.0),
                 "post_ms": float(post_ms or 0.0),
+                "runner_timing_ms": runner_timing_ms,
+                "pipeline_from_frame_ms": (
+                    max(0.0, (result_ts - frame_ts) * 1000.0) if frame_ts > 0 else None
+                ),
             }
 
         self._set_pub_cache("segmentacao", analise.get("dados_visuais", {}))
 
     def _store_detection(self, detections, ts, res):
-        agora = time.time()
+        result_ts = time.time()
+        frame_ts = float(ts or 0.0)
 
         with self._cache_lock:
             self._ultimo_detections = detections or []
             self._det_cache = {
-                "ts": agora,
-                "frame_ts": ts or 0.0,
+                "ts": result_ts,  # compatibilidade
+                "result_ts": result_ts,
+                "frame_ts": frame_ts,
                 "detections": detections or [],
                 "res": res,
+                "pipeline_from_frame_ms": (
+                    max(0.0, (result_ts - frame_ts) * 1000.0) if frame_ts > 0 else None
+                ),
             }
 
         self._set_pub_cache("deteccao", detections or [])
 
-    def _store_grid(self, snapshot, grid_conf):
-        agora = time.time()
+    def _store_grid(
+        self,
+        snapshot,
+        grid_conf,
+        depth_frame_ts=0.0,
+        seg_frame_ts=0.0,
+        det_frame_ts=0.0,
+    ):
+        result_ts = time.time()
 
         with self._cache_lock:
             self._ultimo_snapshot = snapshot
             self._ultima_grid_conf = grid_conf
 
             self._grid_cache = {
-                "ts": agora,
+                "ts": result_ts,  # compatibilidade
+                "result_ts": result_ts,
+                "depth_frame_ts": float(depth_frame_ts or 0.0),
+                "seg_frame_ts": float(seg_frame_ts or 0.0),
+                "det_frame_ts": float(det_frame_ts or 0.0),
                 "snapshot": snapshot,
                 "grid_conf": grid_conf,
             }
@@ -790,7 +902,10 @@ class CameraManager:
     # ============================================================
 
     def _processar_grid(self, depth_frame, segmentacao, dados_visuais_seg, deteccoes, velocidade_ms):
+        t_total0 = time.perf_counter()
+
         try:
+            t_imu0 = time.perf_counter()
             imu_roll = 0.0
 
             try:
@@ -799,8 +914,13 @@ class CameraManager:
             except Exception:
                 imu_roll = 0.0
 
-            grid_ref = self.grid_ref_builder.atualizar_por_pitch(imu_roll)
+            t_imu1 = time.perf_counter()
 
+            t_ref0 = time.perf_counter()
+            grid_ref = self.grid_ref_builder.atualizar_por_pitch(imu_roll)
+            t_ref1 = time.perf_counter()
+
+            t_builder0 = time.perf_counter()
             grid_conf = self.grid_builder.build(
                 depth_mm=depth_frame,
                 seg_ids=segmentacao,
@@ -808,84 +928,222 @@ class CameraManager:
                 grid_shape=self.grid_ref_builder.grid_shape,
                 deteccoes=deteccoes,
             )
+            t_builder1 = time.perf_counter()
 
             if grid_conf is None:
-                return None, None
+                return None, None, {
+                    "imu_ms": (t_imu1 - t_imu0) * 1000.0,
+                    "grid_ref_ms": (t_ref1 - t_ref0) * 1000.0,
+                    "builder_ms": (t_builder1 - t_builder0) * 1000.0,
+                    "fuser_ms": 0.0,
+                    "total_ms": (time.perf_counter() - t_total0) * 1000.0,
+                }
 
+            t_fuser0 = time.perf_counter()
             snapshot = self.costmap_fuser.update(
                 grid_conf,
                 velocidade_ms=velocidade_ms,
                 status_seg=dados_visuais_seg.get("status_corredor"),
             )
+            t_fuser1 = time.perf_counter()
 
-            return snapshot, grid_conf
+            return snapshot, grid_conf, {
+                "imu_ms": (t_imu1 - t_imu0) * 1000.0,
+                "grid_ref_ms": (t_ref1 - t_ref0) * 1000.0,
+                "builder_ms": (t_builder1 - t_builder0) * 1000.0,
+                "fuser_ms": (t_fuser1 - t_fuser0) * 1000.0,
+                "total_ms": (time.perf_counter() - t_total0) * 1000.0,
+            }
 
         except Exception as e:
             self.mostrar_log(f"[visual] erro processando grid: {e}")
-            return None, None
+            return None, None, {
+                "total_ms": (time.perf_counter() - t_total0) * 1000.0,
+            }
 
     # ============================================================
     # Preview / stream
     # ============================================================
 
-    def get_selected_frame(self, frame_type: TipoFrameCamera):
+    def get_selected_frame(
+        self,
+        frame_type: TipoFrameCamera,
+        aguardar_atualizacao: bool = False,
+        timeout_s: Optional[float] = None,
+    ):
+        """
+        Retorna somente frames já produzidos pelo fluxo lateral de preview.
+
+        O runtime crítico nunca renderiza imagem aqui. A chamada apenas solicita
+        ao produtor de preview que atualize aquele tipo e consome o cache pronto.
+
+        PosProcessamento é a única exceção deliberada: representa o RGB limpo em
+        resolução original e é obtido diretamente do último frame para salvamento.
+        """
         try:
             try:
                 tipo_enum = TipoFrameCamera(frame_type)
             except Exception:
                 tipo_enum = frame_type
 
-            # Se visual não está pronto, ainda tenta devolver último frame válido.
-            if not self._visual_disponivel_para_frame():
-                return self._get_cached_preview_frame(tipo_enum)
-
-            with self._cache_lock:
-                rgb = self._copiar_frame(self._ultimo_rgb_frame)
-                depth = self._copiar_frame(self._ultimo_depth_frame)
-                pred = self._ultimo_predictions
-                dets = list(self._ultimo_detections or [])
-                snapshot = self._ultimo_snapshot
-
-            params = getattr(self.camera, "parametros", {}) if self.camera is not None else {}
-
-            # PosProcessamento no Visual Worker = imagem RGB original em melhor qualidade.
             if tipo_enum == TipoFrameCamera.PosProcessamento:
-                if self._frame_valido(rgb):
-                    self._set_cached_preview_frame(tipo_enum, rgb)
-                    return self._copiar_frame(rgb)
+                return self._get_latest_rgb_frame_copy()
 
-                return self._get_cached_preview_frame(tipo_enum)
+            cached, cached_ts = self._get_cached_preview_info(tipo_enum, copiar=False)
 
-            frame = self.renderer.get_selected_frame(
-                frame_type=tipo_enum,
-                rgb_frame=rgb,
-                pred_ids=pred,
-                depth_frame=depth,
-                detections=dets,
-                snapshot=snapshot,
-                camera_params=params,
-                alpha=0.50,
+            if not self._visual_disponivel_para_frame():
+                return self._copiar_frame(cached)
+
+            self._solicitar_preview(tipo_enum)
+
+            agora = time.time()
+            cache_fresco = (
+                self._frame_valido(cached)
+                and cached_ts > 0.0
+                and (agora - cached_ts) <= max(self.preview_cache_fresh_s, 0.0)
             )
 
-            if self._frame_valido(frame):
-                self._set_cached_preview_frame(tipo_enum, frame)
-                return self._copiar_frame(frame)
+            if not aguardar_atualizacao or cache_fresco:
+                return self._copiar_frame(cached)
 
-            # Se o renderer falhou ou faltou algum insumo naquele instante,
-            # devolve o último frame válido desse mesmo tipo.
-            cached = self._get_cached_preview_frame(tipo_enum)
-            if self._frame_valido(cached):
-                return cached
+            timeout = self.preview_request_timeout_s if timeout_s is None else float(timeout_s)
+            deadline = time.time() + max(timeout, 0.0)
 
-            return None
+            with self._preview_cond:
+                while time.time() < deadline:
+                    atual_ts = float(
+                        self._ultimo_frame_preview_ts_por_tipo.get(
+                            self._frame_cache_key(tipo_enum), 0.0
+                        ) or 0.0
+                    )
+                    if atual_ts > cached_ts:
+                        break
+
+                    restante = deadline - time.time()
+                    if restante <= 0:
+                        break
+                    self._preview_cond.wait(timeout=min(restante, 0.05))
+
+            return self._get_cached_preview_frame(tipo_enum)
 
         except Exception as e:
             self.mostrar_log(f"[visual] erro em get_selected_frame({frame_type}): {e}")
+            return self._get_cached_preview_frame(frame_type)
 
+    def _solicitar_preview(self, frame_type) -> bool:
+        try:
+            tipo_enum = TipoFrameCamera(frame_type)
+        except Exception:
+            tipo_enum = frame_type
+
+        if tipo_enum == TipoFrameCamera.PosProcessamento:
+            return False
+
+        if tipo_enum not in self.FRAME_TYPES_PREVIEW:
+            return False
+
+        key = self._frame_cache_key(tipo_enum)
+
+        with self._preview_cond:
+            self._preview_requests.add(key)
+            self._preview_cond.notify_all()
+
+        return True
+
+    def _consumir_solicitacoes_preview(self):
+        with self._preview_cond:
+            if not self._preview_requests:
+                return []
+
+            keys = tuple(self._preview_requests)
+            self._preview_requests.clear()
+
+        tipos = []
+        for key in keys:
             try:
-                return self._get_cached_preview_frame(frame_type)
+                tipos.append(TipoFrameCamera[key])
             except Exception:
-                return None
+                continue
+        return tipos
+
+    def _snapshot_preview_inputs(self, frame_type):
+        """Fotografa apenas referências necessárias, mantendo o lock crítico curto."""
+        try:
+            tipo_enum = TipoFrameCamera(frame_type)
+        except Exception:
+            tipo_enum = frame_type
+
+        rgb = None
+        depth = None
+        pred = None
+        dets = None
+        snapshot = None
+
+        precisa_rgb = tipo_enum in {
+            TipoFrameCamera.Rgb,
+            TipoFrameCamera.Overlay,
+            TipoFrameCamera.MatrizCusto,
+            TipoFrameCamera.Deteccoes,
+            TipoFrameCamera.Debug,
+        }
+        precisa_depth = tipo_enum == TipoFrameCamera.Heatmap
+        precisa_pred = tipo_enum in {
+            TipoFrameCamera.Segmentacao,
+            TipoFrameCamera.Overlay,
+            TipoFrameCamera.Debug,
+        }
+        precisa_dets = tipo_enum == TipoFrameCamera.Deteccoes
+        precisa_snapshot = tipo_enum in {
+            TipoFrameCamera.MatrizCusto,
+            TipoFrameCamera.Debug,
+        }
+
+        with self._cache_lock:
+            if precisa_rgb:
+                rgb = self._ultimo_rgb_frame
+            if precisa_depth:
+                depth = self._ultimo_depth_frame
+            if precisa_pred:
+                pred = self._ultimo_predictions
+            if precisa_dets:
+                dets = list(self._ultimo_detections or [])
+            if precisa_snapshot:
+                snapshot = self._ultimo_snapshot
+
+        params = getattr(self.camera, "parametros", {}) if self.camera is not None else {}
+
+        return {
+            "rgb": rgb,
+            "depth": depth,
+            "pred": pred,
+            "dets": dets,
+            "snapshot": snapshot,
+            "params": params,
+        }
+
+    def _render_preview_frame(self, frame_type):
+        if self.renderer is None:
+            return None
+
+        data = self._snapshot_preview_inputs(frame_type)
+
+        return self.renderer.get_selected_frame(
+            frame_type=frame_type,
+            rgb_frame=data["rgb"],
+            pred_ids=data["pred"],
+            depth_frame=data["depth"],
+            detections=data["dets"],
+            snapshot=data["snapshot"],
+            camera_params=data["params"],
+            alpha=0.50,
+        )
+
+    def _get_latest_rgb_frame_copy(self):
+        with self._cache_lock:
+            frame = self._ultimo_rgb_frame
+
+        # A cópia pesada acontece fora do lock do runtime.
+        return self._copiar_frame(frame)
 
     def _frame_valido(self, frame) -> bool:
         return (
@@ -914,25 +1172,32 @@ class CameraManager:
             return False
 
         key = self._frame_cache_key(frame_type)
-        ts = time.time() if ts is None else ts
-        frame_copy = self._copiar_frame(frame)
+        ts = time.time() if ts is None else float(ts)
 
-        if not self._frame_valido(frame_copy):
-            return False
-
-        with self._cache_lock:
-            self._ultimo_frame_preview_por_tipo[key] = frame_copy
+        # O frame foi criado pelo produtor de preview e não volta a ser mutado.
+        # Portanto o cache pode guardar a própria referência sem uma cópia extra.
+        with self._preview_cond:
+            self._ultimo_frame_preview_por_tipo[key] = frame
             self._ultimo_frame_preview_ts_por_tipo[key] = ts
+            self._preview_cond.notify_all()
 
         return True
 
-    def _get_cached_preview_frame(self, frame_type):
+    def _get_cached_preview_frame(self, frame_type, copiar=True):
+        frame, _ = self._get_cached_preview_info(frame_type, copiar=False)
+        return self._copiar_frame(frame) if copiar else frame
+
+    def _get_cached_preview_info(self, frame_type, copiar=True):
         key = self._frame_cache_key(frame_type)
 
-        with self._cache_lock:
+        with self._preview_lock:
             frame = self._ultimo_frame_preview_por_tipo.get(key)
+            ts = float(self._ultimo_frame_preview_ts_por_tipo.get(key, 0.0) or 0.0)
 
-        return self._copiar_frame(frame)
+        if copiar:
+            frame = self._copiar_frame(frame)
+
+        return frame, ts
 
     def _normalizar_frame_para_uint8(self, frame):
         """
@@ -967,6 +1232,70 @@ class CameraManager:
     # Loops
     # ============================================================
 
+    def _iniciar_loop_preview(self, freq=2.0):
+        """
+        Produtor lateral de preview.
+
+        Dorme sem custo quando não há consumidores. Quando recebe solicitações,
+        renderiza somente os tipos pedidos e sempre fora do lock do runtime.
+        """
+        def loop():
+            ultimo_render_por_tipo = {}
+
+            while True:
+                try:
+                    if self.camera is None or not self.operante:
+                        with self._preview_cond:
+                            self._preview_cond.wait(timeout=0.25)
+                        continue
+
+                    with self._preview_cond:
+                        if not self._preview_requests:
+                            self._preview_cond.wait(timeout=0.50)
+
+                    tipos = self._consumir_solicitacoes_preview()
+                    if not tipos:
+                        continue
+
+                    periodo_min = 1.0 / max(float(self.preview_fps or freq), 0.1)
+
+                    for tipo in tipos:
+                        key = self._frame_cache_key(tipo)
+                        agora = time.time()
+                        ultimo = float(ultimo_render_por_tipo.get(key, 0.0) or 0.0)
+
+                        # Limita o custo máximo do preview mesmo se o operador/UI
+                        # fizer polling acima da frequência configurada.
+                        if ultimo > 0.0 and (agora - ultimo) < periodo_min:
+                            continue
+
+                        t0 = time.perf_counter()
+                        frame = self._render_preview_frame(tipo)
+                        t1 = time.perf_counter()
+
+                        if not self._frame_valido(frame):
+                            self.perf.inc("preview_sem_frame")
+                            continue
+
+                        t_cache0 = time.perf_counter()
+                        self._set_cached_preview_frame(tipo, frame, ts=time.time())
+                        t_cache1 = time.perf_counter()
+
+                        ultimo_render_por_tipo[key] = time.time()
+
+                        self.perf.tick(
+                            "preview",
+                            latencia_ms=(t_cache1 - t0) * 1000.0,
+                            render_ms=(t1 - t0) * 1000.0,
+                            cache_write_ms=(t_cache1 - t_cache0) * 1000.0,
+                        )
+
+                except Exception as e:
+                    self.mostrar_log(f"[visual] erro no produtor de preview: {e}")
+                    time.sleep(0.05)
+
+        threading.Thread(target=loop, daemon=True).start()
+
     def _iniciar_loop_segmentacao(self, freq=8.0):
         def loop():
             ultimo_frame_ts = 0.0
@@ -992,7 +1321,7 @@ class CameraManager:
 
                     t_get0 = time.perf_counter()
                     rgb_frame, frame_ts, res = self.get_rgb_frame()
-                    rgb_frame = self._validar_frame_rgb_para_segmentacao(rgb_frame)
+                    rgb_frame = self._validar_frame_camera_para_segmentacao(rgb_frame)
                     t_get1 = time.perf_counter()
 
                     if rgb_frame is None or frame_ts is None:
@@ -1008,23 +1337,26 @@ class CameraManager:
                     ultimo_frame_ts = frame_ts
 
                     t_inf0 = time.perf_counter()
-                    predictions, ts, roi_resized, roi_info, aux_result = self.seg_runner.infer_ids(rgb_frame)
+                    infer_result = self.seg_runner.infer(rgb_frame)
                     t_inf1 = time.perf_counter()
 
-                    if predictions is None or ts is None:
-                        self.perf.inc("segmentacao_pred_none")
-                        continue
+                    predictions = infer_result.seg_ids
+                    status_result = infer_result.status
 
                     t_post0 = time.perf_counter()
-                    analise = self.segmentacao_manager.analisar(predictions, aux_result=aux_result)
+                    analise = self.segmentacao_manager.analisar(
+                        predictions,
+                        status_result=status_result,
+                    )
                     t_post1 = time.perf_counter()
 
                     self._store_segmentacao(
                         predictions=predictions,
-                        aux_result=aux_result,
+                        status_result=status_result,
                         analise=analise,
                         frame_ts=frame_ts,
                         res=res,
+                        runner_timing_ms=infer_result.timing_ms,
                         infer_ms=(t_inf1 - t_inf0) * 1000.0,
                         post_ms=(t_post1 - t_post0) * 1000.0,
                     )
@@ -1036,6 +1368,13 @@ class CameraManager:
                         latencia_ms=(t_loop1 - t_loop0) * 1000.0,
                         get_rgb_ms=(t_get1 - t_get0) * 1000.0,
                         infer_ms=(t_inf1 - t_inf0) * 1000.0,
+                        geometry_ms=float(infer_result.timing_ms.get("prepare_total_ms", 0.0)),
+                        crop_ms=float(infer_result.timing_ms.get("crop_ms", 0.0)),
+                        resize_ms=float(infer_result.timing_ms.get("resize_ms", 0.0)),
+                        color_ms=float(infer_result.timing_ms.get("color_ms", 0.0)),
+                        session_ms=float(infer_result.timing_ms.get("session_ms", 0.0)),
+                        decode_ms=float(infer_result.timing_ms.get("decode_ms", 0.0)),
+                        runner_total_ms=float(infer_result.timing_ms.get("total_ms", 0.0)),
                         post_ms=(t_post1 - t_post0) * 1000.0,
                         redis_ms=0.0,
                         frame_ts=frame_ts,
@@ -1110,8 +1449,8 @@ class CameraManager:
     def _iniciar_loop_grid(self, freq=5.0):
         def loop():
             ultimo_depth_ts = 0.0
-            ultimo_seg_ts = 0.0
-            ultimo_det_ts = 0.0
+            ultimo_seg_result_ts = 0.0
+            ultimo_det_result_ts = 0.0
 
             while True:
                 t_wall0 = time.time()
@@ -1151,8 +1490,13 @@ class CameraManager:
                     dados_visuais = analise_seg.get("dados_visuais", {})
                     deteccoes = list(det_cache.get("detections", []) or [])
 
-                    seg_ts = float(seg_cache.get("ts", 0.0) or 0.0)
-                    det_ts = float(det_cache.get("ts", 0.0) or 0.0)
+                    # Há dois relógios diferentes e ambos são úteis:
+                    # frame_ts = quando o dado nasceu na câmera;
+                    # result_ts = quando o processamento correspondente terminou.
+                    seg_frame_ts = float(seg_cache.get("frame_ts", 0.0) or 0.0)
+                    det_frame_ts = float(det_cache.get("frame_ts", 0.0) or 0.0)
+                    seg_result_ts = float(seg_cache.get("result_ts", seg_cache.get("ts", 0.0)) or 0.0)
+                    det_result_ts = float(det_cache.get("result_ts", det_cache.get("ts", 0.0)) or 0.0)
                     t_cache1 = time.perf_counter()
 
                     if segmentacao is None:
@@ -1160,14 +1504,18 @@ class CameraManager:
                         time.sleep(0.02)
                         continue
 
-                    if depth_ts == ultimo_depth_ts and seg_ts == ultimo_seg_ts and det_ts == ultimo_det_ts:
+                    if (
+                        depth_ts == ultimo_depth_ts
+                        and seg_result_ts == ultimo_seg_result_ts
+                        and det_result_ts == ultimo_det_result_ts
+                    ):
                         self.perf.inc("grid_pacote_repetido")
                         time.sleep(0.005)
                         continue
 
                     ultimo_depth_ts = depth_ts
-                    ultimo_seg_ts = seg_ts
-                    ultimo_det_ts = det_ts
+                    ultimo_seg_result_ts = seg_result_ts
+                    ultimo_det_result_ts = det_result_ts
 
                     try:
                         vel = get_velocidade_atual_ms()
@@ -1175,7 +1523,7 @@ class CameraManager:
                         vel = 0.0
 
                     t_grid0 = time.perf_counter()
-                    snapshot, grid_conf = self._processar_grid(
+                    snapshot, grid_conf, grid_timing = self._processar_grid(
                         depth_frame=depth_frame,
                         segmentacao=segmentacao,
                         dados_visuais_seg=dados_visuais,
@@ -1189,7 +1537,15 @@ class CameraManager:
                         time.sleep(0.005)
                         continue
 
-                    self._store_grid(snapshot=snapshot, grid_conf=grid_conf)
+                    self._store_grid(
+                        snapshot=snapshot,
+                        grid_conf=grid_conf,
+                        depth_frame_ts=depth_ts,
+                        seg_frame_ts=seg_frame_ts,
+                        det_frame_ts=det_frame_ts,
+                    )
+
+                    agora = time.time()
 
                     self.perf.tick(
                         "grid",
@@ -1197,16 +1553,23 @@ class CameraManager:
                         get_depth_ms=(t_depth1 - t_depth0) * 1000.0,
                         cache_read_ms=(t_cache1 - t_cache0) * 1000.0,
                         build_grid_ms=(t_grid1 - t_grid0) * 1000.0,
-                        fuser_ms=0.0,
+                        imu_ms=float(grid_timing.get("imu_ms", 0.0) or 0.0),
+                        grid_ref_ms=float(grid_timing.get("grid_ref_ms", 0.0) or 0.0),
+                        grid_builder_ms=float(grid_timing.get("builder_ms", 0.0) or 0.0),
+                        fuser_ms=float(grid_timing.get("fuser_ms", 0.0) or 0.0),
                         redis_ms=0.0,
                         depth_ts=depth_ts,
-                        seg_ts=seg_ts,
-                        det_ts=det_ts,
-                        idade_depth_ms=(time.time() - depth_ts) * 1000.0 if depth_ts else None,
-                        idade_seg_ms=(time.time() - seg_ts) * 1000.0 if seg_ts else None,
-                        idade_det_ms=(time.time() - det_ts) * 1000.0 if det_ts else None,
-                        sync_depth_seg_ms=abs(depth_ts - seg_ts) * 1000.0 if depth_ts and seg_ts else None,
-                        sync_depth_det_ms=abs(depth_ts - det_ts) * 1000.0 if depth_ts and det_ts else None,
+                        seg_ts=seg_frame_ts,
+                        det_ts=det_frame_ts,
+                        seg_result_ts=seg_result_ts,
+                        det_result_ts=det_result_ts,
+                        idade_depth_frame_ms=(agora - depth_ts) * 1000.0 if depth_ts else None,
+                        idade_seg_frame_ms=(agora - seg_frame_ts) * 1000.0 if seg_frame_ts else None,
+                        idade_det_frame_ms=(agora - det_frame_ts) * 1000.0 if det_frame_ts else None,
+                        idade_seg_resultado_ms=(agora - seg_result_ts) * 1000.0 if seg_result_ts else None,
+                        idade_det_resultado_ms=(agora - det_result_ts) * 1000.0 if det_result_ts else None,
+                        sync_depth_seg_frame_ms=(abs(depth_ts - seg_frame_ts) * 1000.0) if depth_ts and seg_frame_ts else None,
+                        sync_depth_det_frame_ms=(abs(depth_ts - det_frame_ts) * 1000.0) if depth_ts and det_frame_ts else None,
                         n_dets=len(deteccoes),
                     )
 
@@ -1306,13 +1669,15 @@ class CameraManager:
                             camera_ctx.get("frame_type", TipoFrameCamera.Rgb.value)
                         )
 
-                        frame = self.get_selected_frame(frame_type)
+                        self._solicitar_preview(frame_type)
+                        frame = self._get_cached_preview_frame(frame_type, copiar=False)
 
                         if frame is not None:
                             self.camera.enviar_frame_tcp(frame)
 
                     if self.debug_visual and self._visual_disponivel_para_frame():
-                        dbg = self.get_selected_frame(TipoFrameCamera.Debug)
+                        self._solicitar_preview(TipoFrameCamera.Debug)
+                        dbg = self._get_cached_preview_frame(TipoFrameCamera.Debug, copiar=False)
                         if dbg is not None:
                             cv2.imshow("Visual Worker Debug", dbg)
                             cv2.waitKey(1)
@@ -1366,7 +1731,8 @@ class CameraManager:
                         if ctrl is not None:
                             ctrl.update_from_redis()
 
-                        self._publicar_e_logar_performance()
+                        if self.telemetry_enabled:
+                            self._publicar_e_logar_performance()
 
                 except Exception as e:
                     self.mostrar_log(f"Erro no loop supervisor visual: {e}")
@@ -1417,11 +1783,32 @@ class CameraManager:
             self.seg_config = cfg
 
             self.debug_visual = bool(cfg.get("debug_visual", self.debug_visual))
-            self.debug_perf = bool(cfg.get("debug_perf", self.debug_perf))
+
+            telemetry_cfg = cfg.get("telemetry", {}) or {}
+            self.telemetry_enabled = bool(telemetry_cfg.get("enabled", self.telemetry_enabled))
+            self.telemetry_console = bool(telemetry_cfg.get("console_performance", self.telemetry_console))
+            self.telemetry_publish_runtime = bool(telemetry_cfg.get("publish_runtime_info", self.telemetry_publish_runtime))
+            self.telemetry_runner_timing = bool(telemetry_cfg.get("runner_detailed_timing", self.telemetry_runner_timing))
+
+            include_probs = bool(telemetry_cfg.get("include_status_probs", False))
+            if self.seg_runner is not None:
+                self.seg_runner.include_status_probs = include_probs
+            if self.segmentacao_manager is not None:
+                self.segmentacao_manager.config.include_model_probs = include_probs
 
             self.posproc_intervalo_min_s = float(
                 cfg.get("posproc_intervalo_min_s", self.posproc_intervalo_min_s)
             )
+            self.preview_fps = float(cfg.get("preview_fps", self.preview_fps))
+            self.preview_request_timeout_s = float(
+                cfg.get("preview_request_timeout_s", self.preview_request_timeout_s)
+            )
+            self.preview_cache_fresh_s = float(
+                cfg.get("preview_cache_fresh_s", self.preview_cache_fresh_s)
+            )
+            self.replay_jpeg_quality = int(
+                cfg.get("replay_jpeg_quality", self.replay_jpeg_quality)
+            )
 
             self._ultimo_config_update_ts = agora
 
@@ -1489,19 +1876,38 @@ class CameraManager:
         if self.camera is not None and hasattr(self.camera, "get_cache_stats"):
             resumo["camera_cache"] = self.camera.get_cache_stats()
 
+        agora = time.time()
+        seg_frame_ts = float(self._seg_cache.get("frame_ts", 0.0) or 0.0)
+        seg_result_ts = float(self._seg_cache.get("result_ts", self._seg_cache.get("ts", 0.0)) or 0.0)
+        det_frame_ts = float(self._det_cache.get("frame_ts", 0.0) or 0.0)
+        det_result_ts = float(self._det_cache.get("result_ts", self._det_cache.get("ts", 0.0)) or 0.0)
+        grid_result_ts = float(self._grid_cache.get("result_ts", self._grid_cache.get("ts", 0.0)) or 0.0)
+
         resumo["visual_cache"] = {
-            "seg_ts": self._seg_cache.get("ts", 0.0),
-            "det_ts": self._det_cache.get("ts", 0.0),
-            "grid_ts": self._grid_cache.get("ts", 0.0),
+            "seg_frame_ts": seg_frame_ts,
+            "seg_result_ts": seg_result_ts,
+            "det_frame_ts": det_frame_ts,
+            "det_result_ts": det_result_ts,
+            "grid_result_ts": grid_result_ts,
+            "seg_frame_age_ms": (agora - seg_frame_ts) * 1000.0 if seg_frame_ts else None,
+            "seg_result_age_ms": (agora - seg_result_ts) * 1000.0 if seg_result_ts else None,
+            "det_frame_age_ms": (agora - det_frame_ts) * 1000.0 if det_frame_ts else None,
+            "det_result_age_ms": (agora - det_result_ts) * 1000.0 if det_result_ts else None,
+            "grid_result_age_ms": (agora - grid_result_ts) * 1000.0 if grid_result_ts else None,
+            "seg_pipeline_from_frame_ms": self._seg_cache.get("pipeline_from_frame_ms"),
+            "det_pipeline_from_frame_ms": self._det_cache.get("pipeline_from_frame_ms"),
             "tem_seg": self._ultimo_predictions is not None,
             "tem_grid": self._ultimo_snapshot is not None,
         }
 
         resumo["pub_debug"] = self._ultimo_pub_debug
 
+        if self.telemetry_publish_runtime and self.seg_runner is not None:
+            resumo["seg_runtime"] = self.seg_runner.runtime_info()
+
         self._set_pub_cache("performance_visual", resumo)
 
-        if self.debug_perf:
+        if self.telemetry_console:
             self._logar_performance(resumo)
 
     def _logar_performance(self, resumo):
@@ -1513,13 +1919,14 @@ class CameraManager:
         seg = loops.get("segmentacao", {})
         det = loops.get("deteccao", {})
         grid = loops.get("grid", {})
+        preview = loops.get("preview", {})
         pub = loops.get("publicacao", {})
 
         self.mostrar_log(
             "[PERF_VISUAL] "
             f"fps rgb={self._fps(rgb):.1f} depth={self._fps(depth):.1f} "
             f"seg={self._fps(seg):.1f} det={self._fps(det):.1f} grid={self._fps(grid):.1f} "
-            f"pub={self._fps(pub):.1f} | "
+            f"preview={self._fps(preview):.1f} pub={self._fps(pub):.1f} | "
             f"period seg={self._fmt(self._per(seg))}ms grid={self._fmt(self._per(grid))}ms | "
             f"sync={self._fmt(sync.get('rgb_depth_dt_ms'))}ms"
         )
@@ -1528,17 +1935,46 @@ class CameraManager:
             "[PERF_DETAIL] "
             f"SEG total={self._fmt(self._lat(seg))} "
             f"get={self._fmt(self._m(seg, 'get_rgb_ms'))} "
-            f"infer={self._fmt(self._m(seg, 'infer_ms'))} "
+            f"geom={self._fmt(self._m(seg, 'geometry_ms'))} "
+            f"session={self._fmt(self._m(seg, 'session_ms'))} "
+            f"decode={self._fmt(self._m(seg, 'decode_ms'))} "
             f"post={self._fmt(self._m(seg, 'post_ms'))} | "
             f"GRID total={self._fmt(self._lat(grid))} "
             f"depth={self._fmt(self._m(grid, 'get_depth_ms'))} "
-            f"build={self._fmt(self._m(grid, 'build_grid_ms'))} | "
+            f"imu={self._fmt(self._m(grid, 'imu_ms'))} "
+            f"ref={self._fmt(self._m(grid, 'grid_ref_ms'))} "
+            f"builder={self._fmt(self._m(grid, 'grid_builder_ms'))} "
+            f"fuser={self._fmt(self._m(grid, 'fuser_ms'))} | "
             f"DET total={self._fmt(self._lat(det))} "
             f"read={self._fmt(self._m(det, 'oak_read_ms'))} | "
+            f"PREVIEW total={self._fmt(self._lat(preview))} "
+            f"render={self._fmt(self._m(preview, 'render_ms'))} | "
             f"PUB total={self._fmt(self._lat(pub))} "
             f"redis={self._fmt(self._m(pub, 'redis_ms'))}"
         )
 
+        if self.telemetry_runner_timing:
+            self.mostrar_log(
+                "[PERF_SEG_RUNNER] "
+                f"crop={self._fmt(self._m(seg, 'crop_ms'))}ms "
+                f"resize={self._fmt(self._m(seg, 'resize_ms'))}ms "
+                f"color={self._fmt(self._m(seg, 'color_ms'))}ms "
+                f"session={self._fmt(self._m(seg, 'session_ms'))}ms "
+                f"decode={self._fmt(self._m(seg, 'decode_ms'))}ms "
+                f"runner={self._fmt(self._m(seg, 'runner_total_ms'))}ms"
+            )
+
+        self.mostrar_log(
+            "[PERF_SYNC] "
+            f"depth-seg-frame={self._fmt(self._m(grid, 'sync_depth_seg_frame_ms'))}ms "
+            f"depth-det-frame={self._fmt(self._m(grid, 'sync_depth_det_frame_ms'))}ms | "
+            f"age depth={self._fmt(self._m(grid, 'idade_depth_frame_ms'))}ms "
+            f"seg_frame={self._fmt(self._m(grid, 'idade_seg_frame_ms'))}ms "
+            f"seg_result={self._fmt(self._m(grid, 'idade_seg_resultado_ms'))}ms "
+            f"det_frame={self._fmt(self._m(grid, 'idade_det_frame_ms'))}ms "
+            f"det_result={self._fmt(self._m(grid, 'idade_det_resultado_ms'))}ms"
+        )
+
     def _sleep_loop_adaptativo(self, nome_tarefa, t0_wall, freq_fallback):
         try:
             ctrl = self.gpu_controller
@@ -1557,17 +1993,13 @@ class CameraManager:
 
     def salvar_frames(self, tipos: list, nome: str, pasta="frames_salvos"):
         """
-        Salva frames sob demanda.
+        Salva frames sob demanda sem renderizar no chamador.
 
-        Padrão:
-        - Frames de replay/debug:
-            JPEG leve, qualidade configurada em replay_jpeg_quality.
-        
-        - Tipo PosProcessamento:
-            imagem original em qualidade máxima, sem compressão destrutiva forte.
-            No Visual Worker, isso é apenas a imagem RGB limpa.
+        - Replay/debug: usa somente o preview já produzido pelo Preview Producer
+          e salva JPEG leve.
+        - PosProcessamento: única exceção de alta qualidade, usando RGB original
+          e PNG em thread assíncrona.
         """
-
         if not self._visual_disponivel_para_frame():
             return []
 
@@ -1577,8 +2009,6 @@ class CameraManager:
 
             frame_name = nome.strip() if nome else datetime.datetime.now().strftime("%Y%m%d_%H%M%S_%f")[:-3]
 
-            replay_jpeg_quality = 70
-
             for frame_type in tipos:
                 try:
                     tipo_enum = TipoFrameCamera(frame_type)
@@ -1586,26 +2016,11 @@ class CameraManager:
                     self.mostrar_log(f"[visual] TipoFrameCamera inválido para salvar: {frame_type}")
                     continue
 
-                frame = self.get_selected_frame(tipo_enum)
-
-                if not self._frame_valido(frame):
-                    self.mostrar_log(
-                        f"⚠️ Frame visual não salvo | "
-                        f"tipo={tipo_enum.name} "
-                        f"nome={frame_name}_{tipo_enum.name} "
-                        f"motivo=sem_frame_valido_em_cache"
-                    )
-                    continue
-
-                frame = self._copiar_frame(frame)
-
                 nome_base = f"{frame_name}_{tipo_enum.name}"
 
                 # ====================================================
-                # Pós-processamento
+                # Pós-processamento: RGB original / alta qualidade
                 # ====================================================
-                # No Visual Worker, o dado científico é só a imagem limpa.
-                # Salva com qualidade máxima.
                 if tipo_enum == TipoFrameCamera.PosProcessamento:
                     pode_salvar, motivo = self._pode_salvar_posprocessamento()
 
@@ -1617,66 +2032,76 @@ class CameraManager:
                         )
                         continue
 
-                    frame_pp = frame.copy()
+                    frame_pp = self._get_latest_rgb_frame_copy()
+                    if not self._frame_valido(frame_pp):
+                        self._finalizar_salvamento_posprocessamento()
+                        self.mostrar_log(
+                            f"⚠️ Frame visual não salvo | tipo={tipo_enum.name} "
+                            f"nome={nome_base} motivo=sem_rgb_original"
+                        )
+                        continue
 
                     def salvar_pp_async(nome_base_async, pasta_async, frame_async):
                         try:
                             t_pp0 = time.perf_counter()
-
                             caminho = os.path.join(pasta_async, f"{nome_base_async}.png")
-
                             ok = cv2.imwrite(caminho, frame_async)
                             ok_final = bool(ok and os.path.exists(caminho))
-
                             t_pp1 = time.perf_counter()
 
                             self.mostrar_log(
-                                f"[visual][SAVE_PP] salvo async | "
-                                f"nome={nome_base_async} "
-                                f"ok={ok_final} "
-                                f"tempo_ms={(t_pp1 - t_pp0) * 1000.0:.1f}"
+                                f"[visual][SAVE_PP] salvo async | nome={nome_base_async} "
+                                f"ok={ok_final} tempo_ms={(t_pp1 - t_pp0) * 1000.0:.1f}"
                             )
-
                         except Exception as e:
                             self.mostrar_log(f"❌ Erro no visual SAVE_PP async: {e}")
-
                         finally:
                             self._finalizar_salvamento_posprocessamento()
 
                     threading.Thread(
                         target=salvar_pp_async,
                         args=(nome_base, pasta, frame_pp),
-                        daemon=True
+                        daemon=True,
                     ).start()
 
                     frames_salvos.append(f"ASYNC_STARTED:{nome_base}")
                     continue
 
                 # ====================================================
-                # Replay/debug
+                # Replay/debug: somente preview pronto/cacheado
                 # ====================================================
-                # Salva JPEG leve para não entupir disco durante operação.
-                caminho = os.path.join(pasta, f"{nome_base}.jpg")
+                frame = self.get_selected_frame(
+                    tipo_enum,
+                    aguardar_atualizacao=True,
+                    timeout_s=self.preview_request_timeout_s,
+                )
 
-                if getattr(frame, "dtype", None) != np.uint8:
-                    frame_salvar = self._normalizar_frame_para_uint8(frame)
-                else:
-                    frame_salvar = frame
+                if not self._frame_valido(frame):
+                    self.mostrar_log(
+                        f"⚠️ Frame visual não salvo | tipo={tipo_enum.name} "
+                        f"nome={nome_base} motivo=sem_preview_valido"
+                    )
+                    continue
+
+                caminho = os.path.join(pasta, f"{nome_base}.jpg")
+                frame_salvar = (
+                    self._normalizar_frame_para_uint8(frame)
+                    if getattr(frame, "dtype", None) != np.uint8
+                    else frame
+                )
 
                 ok = cv2.imwrite(
                     caminho,
                     frame_salvar,
-                    [int(cv2.IMWRITE_JPEG_QUALITY), int(replay_jpeg_quality)]
+                    [int(cv2.IMWRITE_JPEG_QUALITY), int(self.replay_jpeg_quality)],
                 )
 
                 if ok and os.path.exists(caminho):
                     frames_salvos.append(caminho)
                 else:
                     self.mostrar_log(
-                        f"❌ Falha ao salvar frame visual | "
-                        f"tipo={tipo_enum.name} "
-                        f"nome={nome_base} "
-                        f"caminho={caminho}"
+                        f"❌ Falha ao salvar frame visual | tipo={tipo_enum.name} "
+                        f"nome={nome_base} caminho={caminho}"
                     )
 
             return frames_salvos
diff --git a/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/visual_worker/config.py b/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/visual_worker/config.py
index cacf2a798..cf9f8f9d7 100644
--- a/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/visual_worker/config.py
+++ b/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/visual_worker/config.py
@@ -50,7 +50,7 @@ _CONFIG_LOCK = threading.Lock()
 
 
 # ============================================================
-# CONFIG V1 - VISUAL WORKER
+# CONFIG V2 - VISUAL WORKER / ONNX FIELD CONTRACT
 # ============================================================
 
 
@@ -117,12 +117,39 @@ VISUAL_DEFAULT_CONFIG = {
     "debug_visual": False,
     "posproc_intervalo_min_s": 5.0,
 
-    # Publica logs de performance no console.
-    "debug_perf": False,
-
     # Tamanho dos frames de preview/debug enviados ao C#.
     "preview_size": (1280, 720),
 
+    # Preview producer lateral: só trabalha quando há consumidor.
+    "preview_fps": 2.0,
+    "preview_request_timeout_s": 0.35,
+    "preview_cache_fresh_s": 0.75,
+
+    # Replay/debug usa JPEG leve. PosProcessamento continua em PNG/RGB original.
+    "replay_jpeg_quality": 70,
+
+    # Telemetria técnica. Tudo que custa payload/log extra fica atrás deste bloco.
+    "telemetry": {
+        # Publica performance_visual no Redis.
+        "enabled": True,
+
+        # Imprime resumo periódico no console.
+        "console_performance": False,
+
+        # Abre o timing da inferência em crop/resize/cor/session/decode.
+        "runner_detailed_timing": True,
+
+        # Inclui o vetor completo de 8 probabilidades do status no debug.
+        # Desligado por padrão porque top1/top2/margin/entropy já bastam no campo.
+        "include_status_probs": False,
+
+        # Anexa contrato/runtime do ONNX em performance_visual.
+        "publish_runtime_info": True,
+
+        # Loga contrato/provider/cache na inicialização.
+        "log_runtime_startup": True,
+    },
+
 
     # ============================================================
     # 2) Frequências dos loops
@@ -144,100 +171,81 @@ VISUAL_DEFAULT_CONFIG = {
 
 
     # ============================================================
-    # 3) Modelo ONNX/TensorRT - segmentação de ruas/corredor
+    # 3) Runtime oficial do modelo de corredor - ONNX FIELD V2
     # ============================================================
-    # Runtime oficial da v1.
-    "runtime_backend": "onnx",
-    "onnx_provider": "tensorrt",
+    # O caminho do ONNX vem do Redis/equipamento: path_ia_model_ruas_seg.
+    # Resolução, ROI, mean/std, layout, nomes de I/O e classes NÃO ficam aqui.
+    # Tudo isso é lido da metadata embutida no próprio ONNX.
+    "model_runtime": {
+        "provider": "tensorrt",
 
-    # Resolução esperada pelo ONNX: [W, H].
-    "ia_resolution": [1024, 576],
+        # Se True: TensorRT -> CUDA -> CPU, sempre expondo degradação na telemetria.
+        # Se False: falha a inicialização quando o provider solicitado não estiver disponível.
+        "allow_provider_fallback": True,
 
-    # ROI vertical da segmentação.
-    # 0.0 + 1.0 = frame inteiro.
-    "ia_roi_begin": 0.0,
-    "ia_roi_size": 1.0,
+        # CameraOak configura ColorCamera.ColorOrder.RGB. O ONNX também exige RGB,
+        # então o hot path evita conversão de cor. Se a fonte mudar no futuro,
+        # BGR continua suportado explicitamente pelo runner.
+        "camera_frame_color": "RGB",
 
-    # Contrato validado do modelo modelseg-2_0.onnx:
-    # entrada:  float32 [1, 3, 576, 1024], RGB 0..1, NCHW
-    # saída 1: semantic_logits
-    # saída 2: label_probs
-    "onnx_has_preprocess": False,
-    "onnx_input_layout": "nchw_float",
-    "onnx_input_scale": 1.0 / 255.0,
-    "onnx_seg_output_format": "logits",
+        # TensorRT. O cache real vira <root>/<sha-do-onnx>/ automaticamente.
+        "trt_fp16_enable": True,
+        "trt_engine_cache_enable": True,
+        "trt_engine_cache_root": "./trt_cache_visual_worker",
+        "trt_max_workspace_size": None,
 
-    # Nomes das entradas/saídas.
-    # input_name None deixa o runner detectar automaticamente.
-    "onnx_input_name": None,
-    "onnx_seg_output_name": "semantic_logits",
-    "onnx_aux_output_name": "label_probs",
-    "onnx_aux_output_format": "probs",
+        # ONNX Runtime.
+        "ort_intra_op_num_threads": 1,
+        "ort_inter_op_num_threads": 1,
 
-    # Auxiliar do status de corredor.
-    # True = retorna label_id, label_name, label_conf.
-    # False = inclui também vetor de probabilidades.
-    "use_compact_aux": True,
+        # Se existir .contract.json, valida o SHA256 do ONNX no startup.
+        "verify_sidecar_sha256": True,
 
-    # Logs internos do runner ONNX.
-    "debug_timing": False,
-    "debug_session": True,
+        # Checagem de IDs únicos a cada inferência. Útil em laboratório, desnecessária no campo.
+        "validate_outputs_each_inference": False,
+    },
 
 
     # ============================================================
-    # 4) TensorRT
+    # 4) Análise geométrica + resolução temporal do status
     # ============================================================
-    "trt_fp16_enable": True,
-    "trt_engine_cache_enable": True,
-    "trt_engine_cache_path": "./trt_cache_visual_worker",
-
-    # Deixe None salvo se não quiser fixar workspace.
-    "trt_max_workspace_size": None,
-
-    # Threads do ONNX Runtime.
-    # Para TensorRT/CUDA, 1 costuma ser suficiente e evita ruído.
-    "onnx_intra_op_num_threads": 1,
-    "onnx_inter_op_num_threads": 1,
-
-
-    # ============================================================
-    # 5) SegmentacaoManager v1
-    # ============================================================
-    # Este bloco controla apenas a análise da máscara:
-    # pred_ids + label_probs -> dados_visuais.
+    # IDs semânticos e nomes de status vêm do contrato ONNX.
     "segmentacao": {
-        # Inclui debug textual no payload de segmentação.
-        # Não gera imagem.
         "include_debug": False,
-
-        # Inclui timing interno do SegmentacaoManager.
         "debug_timing": False,
 
-        # ID das classes no modelo de segmentação.
-        "id_nao_navegavel": 0,
-        "id_navegavel": 1,
-
-        # Frações verticais usadas para estimar centro/ângulo do corredor.
-        # 0.0 = topo, 1.0 = base.
+        # Frações verticais para centro/largura/ângulo. 0=topo, 1=base.
         "scanline_fracs": (0.96, 0.86, 0.74, 0.62, 0.50, 0.38, 0.26),
-
-        # A região próxima ao robô está na base da imagem.
         "near_is_bottom": True,
 
-        # Suavização das saídas usadas pelo controle.
+        # Suavização geométrica.
         "ema_alpha_ang": 0.25,
         "ema_alpha_lat": 0.25,
         "ema_alpha_conf": 0.20,
 
-        # Histerese temporal do status do corredor.
-        "status_window_s": 1.5,
+        # Status v2: voto temporal ponderado + troca rápida quando a cabeça está muito segura.
+        "status_window_s": 0.90,
+        "status_decay_tau_s": 0.35,
         "status_expected_fps": 10.0,
 
-        # Confiança da cabeça auxiliar ONNX.
-        # >= accept: modelo manda.
-        # >= soft: modelo ajuda quando heurística está indefinida.
-        "model_conf_accept": 0.70,
-        "model_conf_soft": 0.45,
+        # Modelo manda quando confiança E margem são boas.
+        "model_conf_accept": 0.72,
+        "model_margin_accept": 0.15,
+
+        # Zona intermediária: só ajuda quando concorda com heurística ou ela está indefinida.
+        "model_conf_soft": 0.55,
+        "model_margin_soft": 0.06,
+
+        # Transições muito fortes podem furar a inércia após N frames consecutivos.
+        "model_conf_fast": 0.90,
+        "model_margin_fast": 0.30,
+        "fast_switch_min_frames": 2,
+
+        # Pesos base do voto temporal.
+        "status_weight_model": 1.35,
+        "status_weight_model_soft": 1.00,
+        "status_weight_heuristic": 0.75,
+        "status_weight_agreement_bonus": 0.40,
 
         # Grid leve interna para centro/fallback do corredor.
         "corridor_grid_rows": 6,
@@ -250,7 +258,7 @@ VISUAL_DEFAULT_CONFIG = {
         "prefer_prev_weight": 0.65,
         "prefer_width_weight": 1.00,
 
-        # Heurística de status quando o modelo auxiliar não está confiante.
+        # Heurística fica como fallback/segundo sensor até termos evidência de campo para removê-la.
         "thr_parado_global": 0.30,
         "thr_direcionando_global": 0.82,
         "thr_caminhando_score": 0.48,
@@ -260,7 +268,7 @@ VISUAL_DEFAULT_CONFIG = {
 
 
     # ============================================================
-    # 6) Grid de confiança/custo
+    # 5) Grid de confiança/custo
     # ============================================================
     "grid": {
         # Formato global da grid: (cols, rows).
@@ -408,7 +416,7 @@ VISUAL_DEFAULT_CONFIG = {
 
 
     # ============================================================
-    # 7) GPU Priority Controller
+    # 6) GPU Priority Controller
     # ============================================================
     # Controla FPS do Visual Worker conforme saúde do Weed Worker.
     "gpu_priority": {
@@ -510,24 +518,16 @@ DET_DEFAULT_CONFIG = {
 def aplicar_overrides_redis_seg(cfg: dict) -> dict:
     equipamento = ContextoGlobalRedis.get_equipamento()
 
-    # Modelo ONNX oficial de ruas/corredor.
-    cfg["ia_onnx_path"] = equipamento.get("path_ia_model_ruas_seg")
-
-    # Labelmap da segmentação.
-    cfg["ia_labelmap_path"] = equipamento.get("path_ia_labelmap_ruas_seg")
+    # Único artefato de IA exigido pelo Visual Worker v2.
+    # Classes, normalização, ROI e I/O vivem dentro do contrato do ONNX.
+    cfg["model_path"] = equipamento.get("path_ia_model_ruas_seg")
 
     return cfg
 
 
 def normalizar_config_runtime_seg(cfg: dict) -> dict:
-    # Alias único interno, para logs e validações.
-    cfg["onnx_model_path"] = cfg.get("ia_onnx_path")
-
-    if not cfg.get("ia_onnx_path"):
-        mostrar_log("[WARN] path do modelo ONNX de ruas não definido no Redis/equipamento.")
-
-    if not cfg.get("ia_labelmap_path"):
-        mostrar_log("[WARN] path do labelmap de ruas não definido no Redis/equipamento.")
+    if not cfg.get("model_path"):
+        mostrar_log("[WARN] path do ONNX de corredor não definido no Redis/equipamento.")
 
     # Garante tuplas onde as dataclasses esperam tuplas.
     cfg["preview_size"] = tuple(cfg.get("preview_size", (1280, 720)))
@@ -556,10 +556,6 @@ def normalizar_config_runtime_seg(cfg: dict) -> dict:
     if "veto_labels" in det and not isinstance(det["veto_labels"], set):
         det["veto_labels"] = set(det["veto_labels"])
 
-    fuser = grid.get("fuser", {}) or {}
-    if "y_range_m" in fuser:
-        fuser["y_range_m"] = tuple(fuser["y_range_m"])
-
     cfg["grid"] = grid
 
     segmentacao = cfg.get("segmentacao", {}) or {}
@@ -615,6 +611,8 @@ def load_seg_config():
 
         # Cópia profunda simples dos blocos aninhados.
         # Evita compartilhar dict interno entre chamadas.
+        cfg["telemetry"] = dict(VISUAL_DEFAULT_CONFIG["telemetry"])
+        cfg["model_runtime"] = dict(VISUAL_DEFAULT_CONFIG["model_runtime"])
         cfg["segmentacao"] = dict(VISUAL_DEFAULT_CONFIG["segmentacao"])
         cfg["grid"] = {
             "grid_shape": VISUAL_DEFAULT_CONFIG["grid"]["grid_shape"],
diff --git a/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/visual_worker/main.py b/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/visual_worker/main.py
index 4ad0d187b..4c1a799da 100644
--- a/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/visual_worker/main.py
+++ b/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/visual_worker/main.py
@@ -20,8 +20,13 @@ def main():
     # Importar e configurar OpenCV antes dos módulos do Visual Worker,
     # pois eles podem carregar OpenCV internamente.
     import cv2
-    cv2.setNumThreads(2)
+    opencv_threads = max(1, int(os.environ.get("VISUAL_OPENCV_THREADS", "4")))
+    cv2.setNumThreads(opencv_threads)
     cv2.ocl.setUseOpenCL(False)
+    print(
+        f"[visual][PERF] OpenCV threads requested={opencv_threads} "
+        f"effective={cv2.getNumThreads()} preview_cache=True"
+    )
 
     from shared.enums import VisualWorkerCommandType, TipoFrameCamera
     from shared.utils import encode_image_base64
@@ -53,6 +58,7 @@ def main():
             time.sleep(0.1)
 
     def redis_callback(dados):
+        acao = None
         try:
             acao = VisualWorkerCommandType(dados.get("cmd", 0))
 
@@ -60,14 +66,14 @@ def main():
                 mx_id = dados.get("params")
                 if mx_id is not None:
                     iniciar_camera_manager(mx_id)
-            elif acao == VisualWorkerCommandType.CalibrarProfundidade:
-                n_frames = dados.get("params", 50)
-                get_camera_manager().gerar_grid_ref(n_frames)
             elif acao == VisualWorkerCommandType.AtualizarSaudeCamera:
                 get_camera_manager().atualizar_saude_camera()
             elif acao == VisualWorkerCommandType.GetCameraFrame:
                 tipo = TipoFrameCamera(dados.get("params", TipoFrameCamera.Rgb.value))
-                frame = get_camera_manager().get_selected_frame(tipo)
+                frame = get_camera_manager().get_selected_frame(
+                    tipo,
+                    aguardar_atualizacao=True,
+                )
                 if frame is not None:
                     base64_img = encode_image_base64(frame)
                     if base64_img is not None:
@@ -82,14 +88,11 @@ def main():
                 pasta = dados.get("params", {}).get("caminho", "frames_salvos")
                 tipos = dados.get("params", {}).get("tipos", [])
                 get_camera_manager().salvar_frames(tipos, nome, pasta)
-            elif acao == VisualWorkerCommandType.EnviarImagemMock:
-                caminho = dados.get("params", "")
-                get_camera_manager().segmentacao_manager.use_mock = caminho != ""
-                get_camera_manager().segmentacao_manager.img_mock = caminho
             else:
                 mostrar_log(f"⚠️ Comando desconhecido: {acao.name}")
         except Exception as e:
-            mostrar_log(f"Erro ao processar comando {acao.name}: {e}")
+            acao_nome = getattr(acao, "name", str(dados.get("cmd", "desconhecido")))
+            mostrar_log(f"Erro ao processar comando {acao_nome}: {e}")
 
     def inicializar():
         mostrar_log("🚀 Iniciando Visual Worker...")
diff --git a/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/visual_worker/processamento/segmentacao_semantica.py b/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/visual_worker/processamento/segmentacao_semantica.py
index a3067e2b4..ac1ed865f 100644
--- a/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/visual_worker/processamento/segmentacao_semantica.py
+++ b/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/visual_worker/processamento/segmentacao_semantica.py
@@ -2,7 +2,6 @@ from __future__ import annotations
 
 import time
 from dataclasses import dataclass, field
-from enum import IntEnum
 from collections import deque
 from typing import Any, Dict, List, Optional, Sequence, Tuple
 
@@ -12,23 +11,19 @@ import numpy as np
 from shared.enums import StatusCarroMapa
 
 
-class ClassesSegmentacao(IntEnum):
-    NAONAVEGAVEL = 0
-    NAVEGAVEL = 1
-
-
 @dataclass
 class SegmentacaoConfig:
     """
-    Configuração v1 do analisador semântico do Visual Worker.
+    Configuração v2 do analisador semântico do Visual Worker.
 
     Esta classe controla apenas análise de máscara.
     Não controla ONNX, câmera, render, preview, stream nem Redis.
     """
 
-    # IDs esperados na máscara de segmentação.
-    id_nao_navegavel: int = int(ClassesSegmentacao.NAONAVEGAVEL)
-    id_navegavel: int = int(ClassesSegmentacao.NAVEGAVEL)
+    # IDs são OBRIGATORIAMENTE injetados do contrato ONNX v2.
+    # -1 faz o componente falhar fechado se alguém tentar inicializá-lo isolado.
+    id_nao_navegavel: int = -1
+    id_navegavel: int = -1
 
     # Scanlines usadas para estimar centro, largura e ângulo do corredor.
     # Frações verticais do frame: 0.0 = topo, 1.0 = base.
@@ -42,13 +37,27 @@ class SegmentacaoConfig:
     ema_alpha_lat: float = 0.25
     ema_alpha_conf: float = 0.20
 
-    # Histórico para estabilizar status do corredor.
-    status_window_s: float = 1.5
+    # Histórico ponderado para estabilizar status do corredor.
+    status_window_s: float = 0.90
+    status_decay_tau_s: float = 0.35
     status_expected_fps: float = 10.0
 
-    # Prioridade do status vindo da cabeça auxiliar ONNX.
-    model_conf_accept: float = 0.70
-    model_conf_soft: float = 0.45
+    # Cabeça de status ONNX v2: confiança + margem top1-top2.
+    model_conf_accept: float = 0.72
+    model_margin_accept: float = 0.15
+    model_conf_soft: float = 0.55
+    model_margin_soft: float = 0.06
+
+    # Troca rápida para transições muito seguras.
+    model_conf_fast: float = 0.90
+    model_margin_fast: float = 0.30
+    fast_switch_min_frames: int = 2
+
+    # Pesos do voto temporal.
+    status_weight_model: float = 1.35
+    status_weight_model_soft: float = 1.00
+    status_weight_heuristic: float = 0.75
+    status_weight_agreement_bonus: float = 0.40
 
     # Grid leve usada para pontuar corredor e fallback de centro.
     corridor_grid_rows: int = 6
@@ -70,15 +79,16 @@ class SegmentacaoConfig:
 
     # Debug leve: inclui métricas extras no retorno, sem criar imagens.
     include_debug: bool = True
+    include_model_probs: bool = False
     debug_timing: bool = False
 
 
 class SegmentacaoManager:
     """
-    Analisador de segmentação semântica v1.
+    Analisador de segmentação semântica v2.
 
     Responsabilidade única:
-      pred_ids + aux_result -> dados_visuais
+      pred_ids + status_result -> dados_visuais
 
     Não renderiza.
     Não colore máscara.
@@ -94,6 +104,14 @@ class SegmentacaoManager:
     ):
         self.config = self._normalizar_config(config)
 
+        if int(self.config.id_navegavel) < 0 or int(self.config.id_nao_navegavel) < 0:
+            raise RuntimeError(
+                "IDs semânticos não foram injetados pelo contrato ONNX v2: "
+                f"nav={self.config.id_navegavel} non_nav={self.config.id_nao_navegavel}"
+            )
+        if int(self.config.id_navegavel) == int(self.config.id_nao_navegavel):
+            raise RuntimeError("IDs navegável/não-navegável são iguais no contrato.")
+
         # Mantidos como metadados úteis, mas não usados no caminho quente.
         self.color_map = color_map
         self.classes = classes
@@ -105,6 +123,8 @@ class SegmentacaoManager:
 
         self._status_hist = deque(maxlen=max_len)
         self._status_final_hist = deque(maxlen=2)
+        self._strong_candidate: Optional[StatusCarroMapa] = None
+        self._strong_candidate_count: int = 0
 
         self._ema_ang: Optional[float] = None
         self._ema_lat: Optional[float] = None
@@ -118,22 +138,22 @@ class SegmentacaoManager:
     # API principal
     # ---------------------------------------------------------------------
 
-    def segmentar(self, predictions: np.ndarray, aux_result: Optional[Dict[str, Any]] = None):
+    def segmentar(self, predictions: np.ndarray, status_result: Optional[Dict[str, Any]] = None):
         """
         Compatibilidade nominal para o CameraManager atual.
 
-        Para a v1, o método preferido é analisar().
+        No contrato v2, o método preferido é analisar().
         Retorna:
             resultado, erro
         """
         try:
-            resultado = self.analisar(predictions, aux_result=aux_result)
+            resultado = self.analisar(predictions, status_result=status_result)
             return resultado, None
         except Exception as e:
             self._ultimo_erro = f"Erro na análise de segmentação: {e}"
             return None, self._ultimo_erro
 
-    def analisar(self, predictions: np.ndarray, aux_result: Optional[Dict[str, Any]] = None) -> Dict[str, Any]:
+    def analisar(self, predictions: np.ndarray, status_result: Optional[Dict[str, Any]] = None) -> Dict[str, Any]:
         t0 = time.perf_counter()
 
         pred = self._validar_predictions(predictions)
@@ -161,7 +181,7 @@ class SegmentacaoManager:
             pred=pred,
             mask_nav=mask_nav,
             corridor_score=corridor_score,
-            aux_result=aux_result,
+            status_result=status_result,
         )
 
         t_status = time.perf_counter()
@@ -182,10 +202,26 @@ class SegmentacaoManager:
             "status_corredor_anterior": int(status_before.value),
             "status_corredor_anterior_nome": str(status_before.name),
 
+            # Telemetria compacta da segunda cabeça. Não depende do debug pesado.
+            "status_fonte": str(status_debug.get("origem_status", "desconhecida")),
+            "status_modelo_id": (status_debug.get("modelo") or {}).get("label_id"),
+            "status_modelo_nome": (status_debug.get("modelo") or {}).get("status_nome"),
+            "status_modelo_conf": (status_debug.get("modelo") or {}).get("conf"),
+            "status_modelo_margin": (status_debug.get("modelo") or {}).get("margin"),
+            "status_modelo_entropy": (status_debug.get("modelo") or {}).get("entropy"),
+            "status_modelo_second_nome": (status_debug.get("modelo") or {}).get("second_name"),
+            "status_modelo_second_conf": (status_debug.get("modelo") or {}).get("second_conf"),
+            "status_fast_switch": bool(status_debug.get("fast_switch", False)),
+
             "centros_corredor": centros,
             "larguras_px": larguras_px,
         }
 
+        if self.config.include_model_probs:
+            probs = (status_debug.get("modelo") or {}).get("probs")
+            if probs is not None:
+                dados_visuais["status_modelo_probs"] = probs
+
         if self.config.include_debug:
             dados_visuais["status_corredor_debug"] = status_debug
             dados_visuais["debug_corredor"] = {
@@ -670,29 +706,100 @@ class SegmentacaoManager:
         pred: np.ndarray,
         mask_nav: np.ndarray,
         corridor_score: float,
-        aux_result: Optional[Dict[str, Any]],
+        status_result: Optional[Dict[str, Any]],
     ):
-        model_status, model_conf, model_debug = self._status_from_aux(aux_result)
+        model_status, model_conf, model_margin, model_debug = self._status_from_model(
+            status_result
+        )
 
-        heur_status, heur_debug = self._status_heuristico(pred, mask_nav, corridor_score)
+        heur_status, heur_debug = self._status_heuristico(
+            pred, mask_nav, corridor_score
+        )
 
         origem = "heuristica"
         status_now = heur_status
+        weight = float(self.config.status_weight_heuristic)
 
-        if model_status is not None and model_conf >= self.config.model_conf_accept:
+        model_fast = (
+            model_status is not None
+            and model_conf >= self.config.model_conf_fast
+            and model_margin >= self.config.model_margin_fast
+        )
+
+        model_accept = (
+            model_status is not None
+            and model_conf >= self.config.model_conf_accept
+            and model_margin >= self.config.model_margin_accept
+        )
+
+        model_soft = (
+            model_status is not None
+            and model_conf >= self.config.model_conf_soft
+            and model_margin >= self.config.model_margin_soft
+        )
+
+        if model_accept:
             status_now = model_status
-            origem = "modelo"
-        elif model_status is not None and model_conf >= self.config.model_conf_soft:
-            # Modelo com média confiança só vence quando a heurística está indefinida.
+            origem = "modelo_fast" if model_fast else "modelo"
+            weight = float(self.config.status_weight_model)
+            # Confiança e margem entram suavemente no peso, sem explodir a janela.
+            weight *= 0.75 + 0.25 * float(np.clip(model_conf, 0.0, 1.0))
+            weight *= 0.85 + 0.15 * float(np.clip(model_margin, 0.0, 1.0))
+
+            if heur_status == model_status:
+                weight += float(self.config.status_weight_agreement_bonus)
+                origem += "_agree"
+
+        elif model_soft:
             if heur_status == StatusCarroMapa.Indefinido:
                 status_now = model_status
                 origem = "modelo_soft"
+                weight = float(self.config.status_weight_model_soft)
+            elif heur_status == model_status:
+                status_now = model_status
+                origem = "modelo_soft_agree"
+                weight = (
+                    float(self.config.status_weight_model_soft)
+                    + float(self.config.status_weight_agreement_bonus)
+                )
             else:
                 status_now = heur_status
-                origem = "heuristica_com_modelo_soft"
+                origem = "heuristica_modelo_soft_discorda"
+
+        now = self._now()
+        self._status_hist.append((status_now, now, float(weight), origem))
+
+        # Transições realmente fortes não precisam aguardar toda a janela temporal.
+        fast_switch = False
+        if model_fast and model_status is not None:
+            if self._strong_candidate == model_status:
+                self._strong_candidate_count += 1
+            else:
+                self._strong_candidate = model_status
+                self._strong_candidate_count = 1
+
+            if self._strong_candidate_count >= max(1, int(self.config.fast_switch_min_frames)):
+                status_final = model_status
+                fast_switch = True
+                # Re-semeia a janela com o estado forte para não voltar no frame seguinte.
+                self._status_hist.clear()
+                self._status_hist.append((
+                    status_final,
+                    now,
+                    float(self.config.status_weight_model)
+                    + float(self.config.status_weight_agreement_bonus),
+                    "fast_switch_seed",
+                ))
+            else:
+                status_final, weighted_scores = self._weighted_status()
+        else:
+            self._strong_candidate = None
+            self._strong_candidate_count = 0
+            status_final, weighted_scores = self._weighted_status()
+
+        if fast_switch:
+            weighted_scores = {status_final.name: 1.0}
 
-        self._status_hist.append((status_now, self._now()))
-        status_final = self._majority_status()
         self._status_final_hist.append(status_final)
 
         status_before = (
@@ -705,34 +812,64 @@ class SegmentacaoManager:
             "origem_status": origem,
             "modelo": model_debug,
             "heuristica": heur_debug,
+            "status_now": status_now.name,
+            "status_final": status_final.name,
+            "candidate_weight": round(float(weight), 4),
+            "weighted_scores": weighted_scores,
+            "fast_switch": bool(fast_switch),
+            "fast_candidate": (
+                self._strong_candidate.name
+                if self._strong_candidate is not None
+                else None
+            ),
+            "fast_candidate_count": int(self._strong_candidate_count),
         }
 
         return status_now, status_final, status_before, debug
 
-    def _status_from_aux(self, aux_result: Optional[Dict[str, Any]]):
-        if not aux_result or aux_result.get("type") != "label":
-            return None, 0.0, {
+    def _status_from_model(self, status_result: Optional[Dict[str, Any]]):
+        if not status_result:
+            return None, 0.0, 0.0, {
                 "disponivel": False,
-                "motivo": "sem_aux_result",
+                "motivo": "sem_status_result",
             }
 
         try:
-            label_id = int(aux_result.get("label_id", StatusCarroMapa.Indefinido.value))
-            conf = float(aux_result.get("label_conf", 0.0))
-            status = StatusCarroMapa(label_id)
+            label_name = str(status_result.get("label_name", "")).strip()
+            conf = float(status_result.get("confidence", 0.0))
+            margin = float(status_result.get("margin", 0.0))
+            entropy = float(status_result.get("entropy", 1.0))
 
-            return status, conf, {
+            if not label_name:
+                raise ValueError("label_name vazio")
+
+            # Mapeamento oficial é por NOME. A ordem dos IDs do treino pode mudar
+            # sem embaralhar a enum do controle.
+            status = StatusCarroMapa[label_name]
+
+            debug = {
                 "disponivel": True,
                 "status": int(status.value),
                 "status_nome": str(status.name),
-                "label_name": aux_result.get("label_name"),
+                "label_id": int(status_result.get("label_id", -1)),
+                "label_name": label_name,
                 "conf": round(conf, 4),
+                "margin": round(margin, 4),
+                "entropy": round(entropy, 4),
+                "second_id": int(status_result.get("second_id", -1)),
+                "second_name": status_result.get("second_name"),
+                "second_conf": round(float(status_result.get("second_confidence", 0.0)), 4),
             }
 
-        except Exception as e:
-            return None, 0.0, {
+            if "probs" in status_result:
+                debug["probs"] = list(status_result["probs"])
+
+            return status, conf, margin, debug
+
+        except Exception as exc:
+            return None, 0.0, 0.0, {
                 "disponivel": False,
-                "motivo": f"aux_invalido: {e}",
+                "motivo": f"status_modelo_invalido: {exc}",
             }
 
     def _status_heuristico(
@@ -827,33 +964,39 @@ class SegmentacaoManager:
     def _now() -> float:
         return time.monotonic()
 
-    def _majority_status(self, janela_s: Optional[float] = None) -> StatusCarroMapa:
-        J = self.config.status_window_s if janela_s is None else float(janela_s)
+    def _weighted_status(self) -> Tuple[StatusCarroMapa, Dict[str, float]]:
+        window_s = max(0.05, float(self.config.status_window_s))
+        tau_s = max(0.05, float(self.config.status_decay_tau_s))
         now = self._now()
 
-        while self._status_hist and (now - self._status_hist[0][1] > J):
+        while self._status_hist and (now - self._status_hist[0][1] > window_s):
             self._status_hist.popleft()
 
         if not self._status_hist:
-            return StatusCarroMapa.Direcionando
+            return StatusCarroMapa.Direcionando, {}
 
-        counts: Dict[StatusCarroMapa, int] = {}
+        scores: Dict[StatusCarroMapa, float] = {}
+        latest_ts: Dict[StatusCarroMapa, float] = {}
 
-        for status, _t in self._status_hist:
-            counts[status] = counts.get(status, 0) + 1
-
-        top = max(counts.values())
-        tied = [status for status, count in counts.items() if count == top]
+        for status, ts, base_weight, _source in self._status_hist:
+            age = max(0.0, now - float(ts))
+            decay = float(np.exp(-age / tau_s))
+            score = max(0.01, float(base_weight)) * decay
+            scores[status] = scores.get(status, 0.0) + score
+            latest_ts[status] = max(latest_ts.get(status, 0.0), float(ts))
 
+        top_score = max(scores.values())
+        tied = [s for s, score in scores.items() if abs(score - top_score) <= 1e-9]
         if len(tied) == 1:
-            return tied[0]
+            winner = tied[0]
+        else:
+            winner = max(tied, key=lambda st: latest_ts.get(st, 0.0))
 
-        # Desempate: status mais recente.
-        for status, _t in reversed(self._status_hist):
-            if status in tied:
-                return status
-
-        return StatusCarroMapa.Direcionando
+        debug_scores = {
+            status.name: round(float(score), 5)
+            for status, score in sorted(scores.items(), key=lambda item: item[1], reverse=True)
+        }
+        return winner, debug_scores
 
     # ---------------------------------------------------------------------
     # Suavização
@@ -879,6 +1022,8 @@ class SegmentacaoManager:
     def reset_estado_temporal(self):
         self._status_hist.clear()
         self._status_final_hist.clear()
+        self._strong_candidate = None
+        self._strong_candidate_count = 0
 
         self._ema_ang = None
         self._ema_lat = None
diff --git a/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/visual_worker/processamento/visual_debug_renderer.py b/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/visual_worker/processamento/visual_debug_renderer.py
index 66b0d0e56..e300257c5 100644
--- a/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/visual_worker/processamento/visual_debug_renderer.py
+++ b/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/visual_worker/processamento/visual_debug_renderer.py
@@ -57,9 +57,12 @@ class VisualDebugRenderer:
             return self.build_detection_overlay(rgb_frame, detections)
 
         if frame_type == TipoFrameCamera.Debug:
-            overlay = self.build_overlay(rgb_frame, pred_ids, alpha=alpha)
+            # Debug prefere costmap. Só monta overlay como fallback, evitando
+            # construir duas imagens completas para descartar uma delas.
             grid = self.build_costmap_debug(rgb_frame, snapshot)
-            return grid if grid is not None else overlay
+            if grid is not None:
+                return grid
+            return self.build_overlay(rgb_frame, pred_ids, alpha=alpha)
 
         return None
 
@@ -67,12 +70,12 @@ class VisualDebugRenderer:
         if rgb_frame is None:
             return None
 
-        img = rgb_frame.copy()
+        if rgb_frame.shape[1::-1] != self.preview_size:
+            # cv2.resize já cria um novo buffer; copiar antes seria tráfego de
+            # memória inútil no caminho de preview.
+            return cv2.resize(rgb_frame, self.preview_size, interpolation=cv2.INTER_AREA)
 
-        if img.shape[1::-1] != self.preview_size:
-            img = cv2.resize(img, self.preview_size, interpolation=cv2.INTER_AREA)
-
-        return img
+        return rgb_frame.copy()
 
     def build_segmentation_preview(self, pred_ids):
         if pred_ids is None or self.color_map is None:
diff --git a/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/visual_worker/processamento/visual_grid_builder.py b/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/visual_worker/processamento/visual_grid_builder.py
index 04db412a4..6a41e0991 100644
--- a/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/visual_worker/processamento/visual_grid_builder.py
+++ b/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/visual_worker/processamento/visual_grid_builder.py
@@ -6,8 +6,6 @@ from typing import Any, Dict, Optional, Sequence, Tuple
 import cv2
 import numpy as np
 
-from visual_worker.processamento.segmentacao_semantica import ClassesSegmentacao
-
 
 @dataclass
 class GridGeometryConfig:
@@ -41,6 +39,10 @@ class DetectionGridConfig:
 
 @dataclass
 class GridConfidenceConfig:
+    # IDs vêm OBRIGATORIAMENTE do contrato ONNX e são injetados pelo CameraManager.
+    id_nao_navegavel: int = -1
+    id_navegavel: int = -1
+
     valid_mm: Tuple[int, int] = (300, 10000)
     min_valid_frac: float = 0.30
     conf_params: Tuple[float, float] = (0.30, 0.80)
@@ -144,6 +146,14 @@ class VisualGridBuilder:
     def __init__(self, config: Optional[GridConfidenceConfig | Dict[str, Any]] = None):
         self.config = self._normalizar_config(config)
 
+        if int(self.config.id_navegavel) < 0 or int(self.config.id_nao_navegavel) < 0:
+            raise RuntimeError(
+                "VisualGridBuilder exige IDs do contrato ONNX v2: "
+                f"nav={self.config.id_navegavel} non_nav={self.config.id_nao_navegavel}"
+            )
+        if int(self.config.id_navegavel) == int(self.config.id_nao_navegavel):
+            raise RuntimeError("Grid recebeu IDs navegável/não-navegável iguais.")
+
     @staticmethod
     def _normalizar_config(config):
         if config is None:
@@ -236,8 +246,8 @@ class VisualGridBuilder:
                     if n == 0:
                         continue
 
-                    n_nav = np.count_nonzero(seg_block == ClassesSegmentacao.NAVEGAVEL.value)
-                    n_naonav = np.count_nonzero(seg_block == ClassesSegmentacao.NAONAVEGAVEL.value)
+                    n_nav = np.count_nonzero(seg_block == cfg.id_navegavel)
+                    n_naonav = np.count_nonzero(seg_block == cfg.id_nao_navegavel)
 
                     pct_navegavel[j, i] = n_nav / n
                     pct_nao_navegavel[j, i] = n_naonav / n
diff --git a/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/visual_worker/utils.py b/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/visual_worker/utils.py
index 6529e4865..457cae6b9 100644
--- a/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/visual_worker/utils.py
+++ b/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/visual_worker/utils.py
@@ -1,144 +1,31 @@
 import cv2
 import numpy as np
-import cupy as cp
-import math
-from scipy import stats
-from shared.enums import T_Code
+
 
 def gerar_heatmap(depth_frame, dist_max_mm=3000.0):
     mask_valid = depth_frame <= dist_max_mm
     depth_normalized = np.zeros(depth_frame.shape, dtype=np.uint8)
-    depth_normalized[mask_valid] = (255 - ((depth_frame[mask_valid] / dist_max_mm) * 255)).astype(np.uint8)
+    depth_normalized[mask_valid] = (
+        255 - ((depth_frame[mask_valid] / dist_max_mm) * 255)
+    ).astype(np.uint8)
     heatmap = cv2.applyColorMap(depth_normalized, cv2.COLORMAP_JET)
     heatmap[~mask_valid] = [0, 0, 0]
     return heatmap
 
-def calcular_threshold_anomalias(velocidade_ms=None, threshold_base=0.15, fator_sensibilidade=100):
-    """
-    Calcula o threshold dinamicamente baseado na velocidade do robô.
-    - velocidade_ms: Velocidade em metros por segundo (m/s).
-    """
-    try:
-        if velocidade_ms is None:
-            from shared.contexto_global_redis import ContextoGlobalRedis
-            velocidade_ms = ContextoGlobalRedis.get_contexto().get("Gerais", {}).get("velocidade_ms", 0.0)
-        threshold = threshold_base + (velocidade_ms * fator_sensibilidade)
-        return round(threshold, 3)
-    except Exception as e:
-        print(f"❌ Erro ao calcular thresold de anomalias: {e}")
-        return threshold_base
-
-def obter_regiao_solo(depth_frame, percentual_altura):
-    depth_frame = cp.asarray(depth_frame)  # 🔥 Garante que está na GPU
-
-    altura, largura = depth_frame.shape
-    linhas_solo = int((percentual_altura / 100) * altura)
-    y_inicio = altura - linhas_solo
-
-    depth_solo = depth_frame[y_inicio:altura, :]  # 🔥 Slice direto na GPU
-
-    return depth_solo, y_inicio, linhas_solo, altura, largura
-
-def calcular_referencia(frames, metodo="moda"):
-    stack = np.stack(frames)
-
-    if metodo == "moda":
-        ref = stats.mode(stack, axis=0, keepdims=True)[0][0]
-    elif metodo == "media":
-        ref = np.mean(stack, axis=0)
-    elif metodo == "mediana":
-        ref = np.median(stack, axis=0)
-    else:
-        raise ValueError("Método inválido. Use 'moda', 'media' ou 'mediana'.")
-
-    return ref
-
-def calcular_inclinacao_solo(perfil, fov_h):
-    """
-    Calcula o ângulo de inclinação do solo em graus.
-
-    :param perfil: Lista de profundidades dos setores [z1, z2, ..., zn]
-    :param largura_total_metros: Largura total em metros do campo de visão
-    :return: Ângulo de inclinação em graus
-    """
-    if len(perfil) < 2:
-        return 0  # Não dá pra calcular
-
-    distancia_media = np.mean(perfil)
-    largura = 2 * math.tan(fov_h / 2) * distancia_media
-
-    delta_z = perfil[-1] - perfil[0]
-    delta_x = largura
-
-    angulo_rad = math.atan2(delta_z, delta_x)
-    angulo_graus = math.degrees(angulo_rad)
-
-    return round(angulo_graus, 3)
-
-def calcular_profundidades_setores(depth_solo, n_setores, largura, distancia_max_mm):
-    depth_solo = cp.asarray(depth_solo)  # 🔥 Garante que está na GPU
-
-    largura_setor = largura // n_setores
-    profundidades = []
-
-    for i in range(n_setores):
-        x_inicio = i * largura_setor
-        x_fim = (i + 1) * largura_setor if (i < n_setores - 1) else largura
-
-        setor = depth_solo[:, x_inicio:x_fim]
-
-        # 🔥 Filtra profundidade válida
-        setor_valido = setor[(setor > 0) & (setor < distancia_max_mm)]
-
-        z = cp.median(setor_valido).item() if setor_valido.size > 0 else 0  # 🔥 .item() p/ float
-        profundidades.append(z)
-
-    return cp.asnumpy(cp.array(profundidades)), largura_setor
-
-def histogram2d_gpu(x, y, bins, range):
-    try:
-        if x.size == 0 or y.size == 0:
-            return cp.zeros(bins, dtype=cp.int32)
-        
-        x_bins, y_bins = bins
-        (x_min, x_max), (y_min, y_max) = range
-
-        x_idx = cp.floor((x - x_min) / (x_max - x_min) * x_bins).astype(cp.int32)
-        y_idx = cp.floor((y - y_min) / (y_max - y_min) * y_bins).astype(cp.int32)
-
-        mask = (x_idx >= 0) & (x_idx < x_bins) & (y_idx >= 0) & (y_idx < y_bins)
-        x_idx = x_idx[mask]
-        y_idx = y_idx[mask]
-
-        if x_idx.size == 0:
-            return cp.zeros((x_bins, y_bins), dtype=cp.int32)
-
-        linear_idx = x_idx * y_bins + y_idx
-
-        if linear_idx.size == 0:
-            return cp.zeros((x_bins, y_bins), dtype=cp.int32)
-
-        hist_flat = cp.bincount(linear_idx, minlength=x_bins * y_bins)
-        hist = hist_flat.reshape((x_bins, y_bins))
-        return hist
-    except Exception as e:
-        print(f"Erro ao gerar histograma 2d: {e}")
-        return cp.zeros((x_bins, y_bins), dtype=cp.int32)
 
 def converter_valores_numpy(obj):
     if isinstance(obj, dict):
         return {k: converter_valores_numpy(v) for k, v in obj.items()}
-    elif isinstance(obj, list):
+    if isinstance(obj, list):
         return [converter_valores_numpy(v) for v in obj]
-    elif isinstance(obj, tuple):
+    if isinstance(obj, tuple):
         return tuple(converter_valores_numpy(v) for v in obj)
-    elif isinstance(obj, (np.integer, np.int32, np.int64)):
+    if isinstance(obj, (np.integer, np.int32, np.int64)):
         return int(obj)
-    elif isinstance(obj, (np.floating, np.float32, np.float64)):
+    if isinstance(obj, (np.floating, np.float32, np.float64)):
         return float(obj)
-    elif isinstance(obj, (np.bool_)):
+    if isinstance(obj, np.bool_):
         return bool(obj)
-    elif isinstance(obj, np.ndarray):
+    if isinstance(obj, np.ndarray):
         return obj.tolist()
-    else:
-        return obj
+    return obj
diff --git a/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/weed_worker/camera_manager.py b/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/weed_worker/camera_manager.py
index 83de8c489..c233ea788 100644
--- a/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/weed_worker/camera_manager.py
+++ b/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/weed_worker/camera_manager.py
@@ -28,7 +28,7 @@ from visual_worker.utils import converter_valores_numpy
 from shared.perf_monitor import VisualPerfMonitor
 
 
-CAMERA_MANAGER_VERSION = "production_v1_2026_09_12_runtime_priority_v3"
+CAMERA_MANAGER_VERSION = "production_v1_2026_09_15_debug_geometry_sync_v4"
 
 PRODUCT_ASSEMBLY_SCHEMA = "multispec_module_params_assembly_v1"
 PRODUCT_MODULE_SCHEMA = "multispec_module_params_v3"
@@ -2044,6 +2044,135 @@ class CameraManager:
 
         return True
 
+    @staticmethod
+    def _clamp01_debug(value, default=0.0):
+        try:
+            return max(0.0, min(1.0, float(value)))
+        except Exception:
+            return max(0.0, min(1.0, float(default)))
+
+    def _resolver_geometria_debug_bicos(self, config, analise_debug=None):
+        """
+        Resolve a geometria que o Debug deve desenhar.
+
+        Autoridade, nesta ordem:
+          1) estatísticas produzidas pelo WeedDetector na última decisão real;
+          2) o próprio helper do WeedDetector, quando disponível;
+          3) fallback local com o MESMO contrato: 0=topo, 1=pé.
+
+        Assim o preview não mantém uma segunda implementação da lógica de atuação.
+        """
+        cfg = config or {}
+        analise_debug = analise_debug or {}
+
+        dados_visuais = analise_debug.get("dados_visuais", {}) or {}
+        estatisticas = dados_visuais.get("estatisticas", {}) or {}
+        estat_bicos = estatisticas.get("bicos", {}) or {}
+        estat_atuacao = estat_bicos.get("atuacao", {}) or {}
+
+        ordem_invertida = bool(
+            estat_bicos.get(
+                "ordem_bicos_invertida",
+                cfg.get("inverter_ordem_bicos", False),
+            )
+        )
+        sentido_vertical_invertido = bool(
+            estat_bicos.get(
+                "sentido_vertical_invertido",
+                estat_atuacao.get(
+                    "sentido_vertical_invertido",
+                    cfg.get("inverter_sentido_vertical_bicos", False),
+                ),
+            )
+        )
+
+        zona_ini = estat_atuacao.get("zona_atuacao_inicio_frac")
+        zona_fim = estat_atuacao.get("zona_atuacao_fim_frac")
+        eval_ini = estat_atuacao.get("zona_eval_inicio_frac")
+        eval_fim = estat_atuacao.get("zona_eval_fim_frac")
+
+        # Se ainda não existe uma decisão real (startup, reset etc.), pede ao
+        # detector a mesma função usada pelo controle. Evita duplicar semântica.
+        if zona_ini is None or zona_fim is None:
+            detector = self.weed_detector
+            helper = getattr(detector, "_calcular_zona_atuacao_frac_from_config", None)
+            if callable(helper):
+                try:
+                    zona_ini, zona_fim = helper(cfg)
+                except Exception:
+                    zona_ini = None
+                    zona_fim = None
+
+        # Último fallback, mantido fail-soft para o Debug. Esta é a convenção
+        # oficial atual do WeedDetector: faixa é medida de cima para baixo.
+        if zona_ini is None or zona_fim is None:
+            faixa = self._clamp01_debug(cfg.get("faixa_atuacao_bicos", 0.70), 0.70)
+            area = self._clamp01_debug(cfg.get("area_atuacao_bicos", 0.10), 0.10)
+            area = max(0.001, area)
+
+            zona_ini = faixa
+            zona_fim = min(1.0, faixa + area)
+
+            if sentido_vertical_invertido:
+                zona_ini, zona_fim = 1.0 - zona_fim, 1.0 - zona_ini
+
+        zona_ini = self._clamp01_debug(zona_ini, 0.0)
+        zona_fim = self._clamp01_debug(zona_fim, zona_ini)
+        if zona_fim < zona_ini:
+            zona_ini, zona_fim = zona_fim, zona_ini
+
+        # Sem compensação de latência disponível, a zona avaliada coincide com
+        # a zona física. Quando o detector informar outra zona, mostramos ambas.
+        if eval_ini is None or eval_fim is None:
+            eval_ini, eval_fim = zona_ini, zona_fim
+
+        eval_ini = self._clamp01_debug(eval_ini, zona_ini)
+        eval_fim = self._clamp01_debug(eval_fim, zona_fim)
+        if eval_fim < eval_ini:
+            eval_ini, eval_fim = eval_fim, eval_ini
+
+        pulverizacao = dados_visuais.get("pulverizacao", {}) or {}
+
+        return {
+            "zona_inicio_frac": float(zona_ini),
+            "zona_fim_frac": float(zona_fim),
+            "zona_eval_inicio_frac": float(eval_ini),
+            "zona_eval_fim_frac": float(eval_fim),
+            "ordem_bicos_invertida": bool(ordem_invertida),
+            "sentido_vertical_invertido": bool(sentido_vertical_invertido),
+            "pulverizacao": pulverizacao,
+            "fonte_detector": bool(estat_atuacao),
+        }
+
+    @staticmethod
+    def _frac_vertical_para_pixels(inicio_frac, fim_frac, altura):
+        if altura <= 1:
+            return 0, 0
+
+        inicio = max(0.0, min(1.0, float(inicio_frac)))
+        fim = max(0.0, min(1.0, float(fim_frac)))
+        if fim < inicio:
+            inicio, fim = fim, inicio
+
+        y0 = int(round(inicio * (altura - 1)))
+        y1 = int(round(fim * (altura - 1)))
+        y0 = max(0, min(altura - 1, y0))
+        y1 = max(0, min(altura - 1, y1))
+        return min(y0, y1), max(y0, y1)
+
+    @staticmethod
+    def _intervalo_visual_bico(indice_bico, qtd_bicos, largura, ordem_invertida):
+        """Mapeia índice FÍSICO do bico para a coluna correspondente da imagem."""
+        qtd = max(1, int(qtd_bicos))
+        idx = int(indice_bico)
+        idx_visual = (qtd - 1 - idx) if ordem_invertida else idx
+
+        x0 = int(round(idx_visual * largura / float(qtd)))
+        x1 = int(round((idx_visual + 1) * largura / float(qtd))) - 1
+        x0 = max(0, min(largura - 1, x0))
+        x1 = max(x0, min(largura - 1, x1))
+        return x0, x1
+
     def get_debug_frame(self, mostrar=False, overlay_bgr=None):
         overlay = overlay_bgr if overlay_bgr is not None else self._ultimo_preview_overlay
 
@@ -2057,12 +2186,21 @@ class CameraManager:
             "fps_loop": self._ultimo_loop_analise_fps,
         }
 
+        # A detecção publica controle final + análise sob _vida_lock. Tiramos o
+        # snapshot sob o mesmo lock para o Debug não misturar estados de ciclos
+        # diferentes enquanto a thread de detecção está atualizando os caches.
+        with self._vida_lock:
+            atuacao_bicos = dict(self._ultimo_controle or {})
+            analise_debug = dict(self._ultima_analise or {})
+            config_debug = dict(self.seg_config or {})
+
         # Não publica diretamente no cache: o preview producer faz a
         # publicação atômica depois de reduzir/comprimir o bundle inteiro.
         return self._montar_debug_overlay(
             overlay_bgr=overlay,
-            atuacao_bicos=self._ultimo_controle or {},
-            config=self.seg_config or {},
+            atuacao_bicos=atuacao_bicos,
+            config=config_debug,
+            analise_debug=analise_debug,
             metricas_perf=metricas_perf,
             mostrar=mostrar,
         )
@@ -2072,6 +2210,7 @@ class CameraManager:
         overlay_bgr,
         atuacao_bicos,
         config,
+        analise_debug=None,
         metricas_perf=None,
         mostrar=False,
     ):
@@ -2080,6 +2219,10 @@ class CameraManager:
                 return None
 
             metricas_perf = metricas_perf or {}
+            geometria = self._resolver_geometria_debug_bicos(
+                config=config,
+                analise_debug=analise_debug,
+            )
 
             W, H = self._dbg_shape
 
@@ -2104,25 +2247,68 @@ class CameraManager:
                     self._dbg_img[...] = overlay_bgr
 
             qtd_bicos = int(config.get("qtd_bicos", self.qtd_bicos or 1) or 1)
-            zona_inicio = float(config.get("faixa_atuacao_bicos", 0.7))
-            faixa_atuacao = float(config.get("area_atuacao_bicos", 0.1))
+            qtd_bicos = max(1, qtd_bicos)
 
-            y_inicio = int((1.0 - zona_inicio) * H)
-            y_fim = int((1.0 - (zona_inicio + faixa_atuacao)) * H)
-            y0, y1 = min(y_inicio, y_fim), max(y_inicio, y_fim)
+            y0, y1 = self._frac_vertical_para_pixels(
+                geometria["zona_inicio_frac"],
+                geometria["zona_fim_frac"],
+                H,
+            )
+            ey0, ey1 = self._frac_vertical_para_pixels(
+                geometria["zona_eval_inicio_frac"],
+                geometria["zona_eval_fim_frac"],
+                H,
+            )
 
+            # ------------------------------------------------------------
+            # Zona FÍSICA dos bicos. É a faixa configurada pelo operador.
+            # ------------------------------------------------------------
             self._dbg_layer.fill(0)
             cv2.rectangle(
                 self._dbg_layer,
                 (0, y0),
-                (W, y1),
+                (W - 1, y1),
                 (220, 220, 100),
                 thickness=-1,
             )
-            cv2.addWeighted(self._dbg_layer, 0.18, self._dbg_img, 0.82, 0, dst=self._dbg_img)
-            cv2.rectangle(self._dbg_img, (0, y0), (W, y1), (180, 180, 80), 2)
+            cv2.addWeighted(
+                self._dbg_layer,
+                0.18,
+                self._dbg_img,
+                0.82,
+                0,
+                dst=self._dbg_img,
+            )
+            cv2.rectangle(
+                self._dbg_img,
+                (0, y0),
+                (W - 1, y1),
+                (180, 180, 80),
+                2,
+            )
+
+            # Zona REALMENTE avaliada pelo detector quando há antecipação por
+            # latência/velocidade. Se coincidir com a zona física, não duplica.
+            eval_diferente = abs(ey0 - y0) > 1 or abs(ey1 - y1) > 1
+            if eval_diferente:
+                cv2.rectangle(
+                    self._dbg_img,
+                    (0, ey0),
+                    (W - 1, ey1),
+                    (255, 255, 0),
+                    2,
+                )
+                cv2.putText(
+                    self._dbg_img,
+                    "EVAL",
+                    (max(5, W - 60), max(16, min(H - 5, ey0 + 16))),
+                    cv2.FONT_HERSHEY_SIMPLEX,
+                    0.45,
+                    (255, 255, 0),
+                    1,
+                    cv2.LINE_AA,
+                )
 
-            largura_bico = W / float(max(qtd_bicos, 1))
             cores = [
                 (0, 255, 0),
                 (255, 0, 0),
@@ -2134,33 +2320,64 @@ class CameraManager:
                 (0, 0, 255),
             ]
 
+            ordem_invertida = bool(geometria["ordem_bicos_invertida"])
+
+            # ------------------------------------------------------------
+            # Estado ON/OFF FINAL, já depois do gate de pulverização.
+            # Cada índice é o bico FÍSICO. A posição visual respeita
+            # inverter_ordem_bicos exatamente como o WeedDetector.
+            # ------------------------------------------------------------
             self._dbg_layer.fill(0)
-
             for i in range(qtd_bicos):
-                x0 = int(i * largura_bico)
-                x1 = int((i + 1) * largura_bico)
+                x0, x1 = self._intervalo_visual_bico(
+                    i,
+                    qtd_bicos,
+                    W,
+                    ordem_invertida,
+                )
                 cor = cores[i % len(cores)]
 
-                if atuacao_bicos.get(i, False):
-                    cv2.rectangle(self._dbg_layer, (x0, y0), (x1, y1), cor, thickness=-1)
+                if bool(atuacao_bicos.get(i, False)):
+                    cv2.rectangle(
+                        self._dbg_layer,
+                        (x0, y0),
+                        (x1, y1),
+                        cor,
+                        thickness=-1,
+                    )
 
-            cv2.addWeighted(self._dbg_layer, 0.15, self._dbg_img, 0.85, 0, dst=self._dbg_img)
+            cv2.addWeighted(
+                self._dbg_layer,
+                0.15,
+                self._dbg_img,
+                0.85,
+                0,
+                dst=self._dbg_img,
+            )
 
             for i in range(qtd_bicos):
-                x0 = int(i * largura_bico)
-                x1 = int((i + 1) * largura_bico)
+                x0, x1 = self._intervalo_visual_bico(
+                    i,
+                    qtd_bicos,
+                    W,
+                    ordem_invertida,
+                )
                 cor = cores[i % len(cores)]
-                status = "ON" if atuacao_bicos.get(i, False) else "OFF"
+                status = "ON" if bool(atuacao_bicos.get(i, False)) else "OFF"
 
                 cv2.rectangle(self._dbg_img, (x0, y0), (x1, y1), cor, 1)
+
+                texto = f"Bico {i} {status}"
+                text_y = max(18, min(H - 6, y0 + 20))
                 cv2.putText(
                     self._dbg_img,
-                    f"Bico {i} {status}",
-                    (x0 + 5, min(H - 5, y1 + 20)),
+                    texto,
+                    (x0 + 4, text_y),
                     cv2.FONT_HERSHEY_SIMPLEX,
-                    0.55,
+                    0.48,
                     cor,
-                    2,
+                    1,
+                    cv2.LINE_AA,
                 )
 
             fps_infer = float(metricas_perf.get("fps_infer") or 0.0)
@@ -2188,6 +2405,51 @@ class CameraManager:
                 2,
             )
 
+            zona_txt = (
+                f"Zona={geometria['zona_inicio_frac'] * 100:.0f}-"
+                f"{geometria['zona_fim_frac'] * 100:.0f}%"
+            )
+            if eval_diferente:
+                zona_txt += (
+                    f" | Eval={geometria['zona_eval_inicio_frac'] * 100:.0f}-"
+                    f"{geometria['zona_eval_fim_frac'] * 100:.0f}%"
+                )
+
+            zona_txt += (
+                f" | V={'INV' if geometria['sentido_vertical_invertido'] else 'NORMAL'}"
+                f" H={'INV' if ordem_invertida else 'NORMAL'}"
+            )
+
+            cv2.putText(
+                self._dbg_img,
+                zona_txt,
+                (10, 88),
+                cv2.FONT_HERSHEY_SIMPLEX,
+                0.48,
+                (230, 230, 230),
+                1,
+                cv2.LINE_AA,
+            )
+
+            pulverizacao = geometria.get("pulverizacao", {}) or {}
+            if pulverizacao:
+                permitida = bool(pulverizacao.get("permitida", False))
+                motivo = str(pulverizacao.get("motivo", "") or "")
+                gate_txt = f"Pulv: {'ON' if permitida else 'OFF'}"
+                if motivo and motivo != "ok":
+                    gate_txt += f" ({motivo[:48]})"
+
+                cv2.putText(
+                    self._dbg_img,
+                    gate_txt,
+                    (10, 108),
+                    cv2.FONT_HERSHEY_SIMPLEX,
+                    0.45,
+                    (200, 255, 200) if permitida else (180, 180, 255),
+                    1,
+                    cv2.LINE_AA,
+                )
+
             if mostrar:
                 cv2.imshow("Debug Weed Worker", self._dbg_img)
                 cv2.waitKey(1)
diff --git a/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/weed_worker/config.py b/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/weed_worker/config.py
index 3cbb17f1e..b76c82bc8 100644
--- a/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/weed_worker/config.py
+++ b/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/weed_worker/config.py
@@ -256,7 +256,7 @@ WEED_DEFAULT_CONFIG = {
 
     # True no rover atual:
     # a aproximação física do alvo ocorre do pé para o topo da imagem.
-    "inverter_sentido_vertical_bicos": True,
+    "inverter_sentido_vertical_bicos": False,
 
     # Zona física de atuação no frame.
     # Ambos podem ser sobrescritos pela configuração da operação.
diff --git a/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/weed_worker/weed_detector.py b/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/weed_worker/weed_detector.py
index 166718d68..f5c48bbed 100644
--- a/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/weed_worker/weed_detector.py
+++ b/AgroBase/AgroBase/bin/x64/Debug/Python/Scripts/workers/weed_worker/weed_detector.py
@@ -25,11 +25,14 @@ class WeedDetector:
       - Desloca a memória pela distância real percorrida pelo rover.
       - Atua quando a evidência chega na zona física de atuação.
 
-    Convenção espacial:
-      - cell index 0 = região mais próxima da linha dos bicos.
-      - cell index maior = região mais distante, vista antes pela câmera.
-      - Conforme o rover anda, a memória se desloca de índices maiores
-        para índices menores.
+    Convenção espacial oficial:
+      - eixo vertical sempre usa coordenada de imagem: 0.0 = topo, 1.0 = pé.
+      - cell index 0 = topo da imagem.
+      - cell index N-1 = pé da imagem.
+      - inverter_sentido_vertical_bicos=False (rover atual):
+          alvo caminha topo -> pé; faixa=0.85, area=0.15 atua em 85%..100%.
+      - inverter_sentido_vertical_bicos=True:
+          alvo caminha pé -> topo; a mesma faixa é espelhada para 0%..15%.
 
     Este arquivo NÃO:
       - gera imagem;
@@ -41,6 +44,7 @@ class WeedDetector:
 
     CONTRATO_OFICIAL = "target_binary"
     TARGET_ID = 1
+    VERSION = "weed_detector_v2_2026_09_15_vertical_contract_fix"
 
     def __init__(self, config: Optional[dict] = None):
         # O CameraManager já mantém um snapshot de configuração. Quando ele é
@@ -76,8 +80,14 @@ class WeedDetector:
             self.config.get("inverter_ordem_bicos", False)
         )
 
+        # Convenção oficial:
+        #   False = fluxo normal topo -> pé (rover atual).
+        #   True  = fluxo invertido pé -> topo.
+        #
+        # IMPORTANTE: esta flag altera tanto o sentido da memória espacial
+        # quanto a posição vertical da zona física de atuação.
         self.inverter_sentido_vertical_bicos = bool(
-            self.config.get("inverter_sentido_vertical_bicos", True)
+            self.config.get("inverter_sentido_vertical_bicos", False)
         )
 
         self.detector_debug_perf = bool(
@@ -168,11 +178,14 @@ class WeedDetector:
         0.0 = topo da imagem
         1.0 = pé da imagem
 
-        Normal:
+        Normal (inverter_sentido_vertical_bicos=False):
         faixa=0.90, area=0.10 -> 90% até 100%, embaixo.
 
-        Com inverter_sentido_vertical_bicos=True:
+        Invertido (inverter_sentido_vertical_bicos=True):
         faixa=0.90, area=0.10 -> 0% até 10%, em cima.
+
+        Portanto, para o rover atual, em que o alvo se aproxima do topo para
+        o pé da imagem, a configuração correta é False.
         """
 
         faixa = float(cfg.get("faixa_atuacao_bicos", 0.70) or 0.70)
@@ -918,11 +931,12 @@ class WeedDetector:
         k_shift = max(0.0, k_shift)
 
         # A zona configurada representa a posição física fixa dos bicos.
-        # Com faixa=0,85, área=0,10 e inversão vertical, 85%..95% da imagem
-        # original vira 5%..15% na convenção interna. Essa zona NÃO muda de
-        # parado até a velocidade de referência. Somente a parcela de
-        # velocidade ACIMA da referência antecipa a avaliação para compensar
-        # o tempo de resposta do sistema.
+        # Exemplo no rover atual (sentido normal / False):
+        #   faixa=0,85, área=0,15 -> 85%..100%, no pé da imagem.
+        # No sentido invertido / True a mesma faixa é espelhada para 0%..15%.
+        # Essa zona NÃO muda de parado até a velocidade de referência. Somente
+        # a parcela de velocidade ACIMA da referência antecipa a avaliação para
+        # compensar o tempo de resposta do sistema.
         vel_referencia = max(
             0.0,
             float(cfg.get("velocidade_referencia_atuacao_mps", 0.65) or 0.65),
@@ -1282,19 +1296,46 @@ class WeedDetector:
         vel_norm: float,
         cfg: dict,
     ):
-        # Convenção legada:
-        # zona_inicio e zona_altura são frações medidas a partir da parte inferior da imagem.
-        y_inicio = int((1.0 - zona_inicio) * h)
-        y_fim = int((1.0 - (zona_inicio + zona_altura)) * h)
+        """
+        ROI do fallback legado usando a MESMA convenção do caminho oficial.
 
-        y_top = min(y_inicio, y_fim)
-        y_bot = max(y_inicio, y_fim)
+        Coordenada vertical:
+          0.0 = topo
+          1.0 = pé
 
+        False: alvo topo -> pé; 0.85 + 0.15 => faixa inferior.
+        True : alvo pé -> topo; a mesma faixa é espelhada para o topo.
+
+        Os argumentos zona_inicio/zona_altura são mantidos na assinatura por
+        compatibilidade, mas o cálculo usa os valores de cfg para garantir que
+        exista uma única fonte de verdade.
+        """
+        _ = zona_inicio, zona_altura
+
+        zona_ini_frac, zona_fim_frac = self._calcular_zona_atuacao_frac_from_config(cfg)
+
+        y_top = int(np.floor(zona_ini_frac * h))
+        y_bot = int(np.ceil(zona_fim_frac * h))
+
+        y_top = max(0, min(h, y_top))
+        y_bot = max(y_top, min(h, y_bot))
+
+        # Antecipação legada em pixels. A avaliação deve se deslocar para a
+        # região que o alvo ocupa ANTES de chegar aos bicos:
+        #   fluxo topo -> pé : antecipa para cima (subtrai y)
+        #   fluxo pé -> topo : antecipa para baixo (soma y)
         roi_shift_per_v = float(cfg.get("k_roi_shift_px_per_vnorm", 24.0))
-        shift = int(roi_shift_per_v * self._clamp01(vel_norm))
+        shift = int(max(0.0, roi_shift_per_v) * self._clamp01(vel_norm))
 
-        y_top = max(0, y_top - shift)
-        y_bot = max(y_top, min(h, y_bot - shift))
+        if getattr(self, "inverter_sentido_vertical_bicos", False):
+            y_top = min(h, y_top + shift)
+            y_bot = min(h, y_bot + shift)
+        else:
+            y_top = max(0, y_top - shift)
+            y_bot = max(0, y_bot - shift)
+
+        if y_bot < y_top:
+            y_top, y_bot = y_bot, y_top
 
         return int(y_top), int(y_bot), int(shift)
 
diff --git a/Python/OAK/datasets/oak-d/_0_capture_corridor.py b/Python/OAK/datasets/oak-d/_0_capture_corridor.py
new file mode 100644
index 000000000..6cf9206c6
--- /dev/null
+++ b/Python/OAK/datasets/oak-d/_0_capture_corridor.py
@@ -0,0 +1,851 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+"""
+Capture OAK-D Lite - Agri Corridor Dataset
+==========================================
+
+Captura RGB para o dataset da camera frontal.
+
+Contrato desta etapa:
+    - roda de dentro da pasta oak-d/
+    - captura o mesmo dominio de video usado no runtime:
+        OAK-D Lite -> ColorCamera 1080p -> cam.video -> host
+    - salva SOMENTE a imagem RGB original em PNG, sem resize e sem normalizacao
+    - destino padrao:
+        dataset/brutas/
+    - resize existe apenas para PREVIEW da interface e nunca toca a imagem salva
+
+Teclas:
+    SPACE / S : salvar frame atual
+    A         : ligar/desligar auto-save
+    E         : alternar exposicao AUTO/MANUAL
+    + / =     : aumentar ISO no modo manual
+    -         : diminuir ISO no modo manual
+    M         : aumentar exposicao no modo manual
+    N         : diminuir exposicao no modo manual
+    F         : alternar foco AUTO/MANUAL
+    ]         : aumentar foco manual
+    [         : diminuir foco manual
+    V         : alternar fullscreen
+    Q / ESC   : sair
+
+Exemplo:
+    python _0_capture.py
+
+Auto-save a cada 1 segundo:
+    python _0_capture.py --auto --interval 1.0
+
+Outra pasta:
+    python _0_capture.py --out dataset/brutas
+"""
+
+from __future__ import annotations
+
+import argparse
+import time
+from concurrent.futures import ThreadPoolExecutor
+from datetime import datetime
+from pathlib import Path
+from typing import Optional, Tuple
+
+import cv2
+import depthai as dai
+import numpy as np
+
+
+CAPTURE_W = 1920
+CAPTURE_H = 1080
+CAPTURE_FPS = 30.0
+
+DEFAULT_OUT = Path("dataset") / "brutas"
+
+DEFAULT_EXPOSURE_US = 6000
+DEFAULT_ISO = 400
+DEFAULT_FOCUS = 130
+
+EXPOSURE_MIN_US = 100
+EXPOSURE_MAX_US = 30000
+EXPOSURE_STEP_US = 500
+
+ISO_MIN = 100
+ISO_MAX = 1600
+ISO_STEP = 50
+
+FOCUS_MIN = 0
+FOCUS_MAX = 255
+FOCUS_STEP = 5
+
+WINDOW_NAME = "OAK-D Lite | Dataset Corredor Agricola"
+CAPTURE_SESSION_ID = datetime.now().strftime("%Y%m%d_%H%M%S")
+
+
+def timestamp_name() -> str:
+    """
+    Nome novo:
+        img_sYYYYMMDD_HHMMSS__YYYYMMDD_HHMMSS_micro.png
+
+    O primeiro timestamp identifica a SESSAO de captura (uma execução do script).
+    O segundo identifica o frame.
+
+    Isso permite ao split manter uma mesma sessão inteira em train OU val.
+    """
+    frame_ts = datetime.now().strftime("%Y%m%d_%H%M%S_%f")
+    return f"img_s{CAPTURE_SESSION_ID}__{frame_ts}"
+
+
+def save_png(path: Path, frame_bgr: np.ndarray) -> Tuple[bool, str]:
+    path.parent.mkdir(parents=True, exist_ok=True)
+
+    ok = cv2.imwrite(
+        str(path),
+        frame_bgr,
+        [cv2.IMWRITE_PNG_COMPRESSION, 3],
+    )
+
+    return bool(ok), str(path)
+
+
+def fit_inside(image: np.ndarray, max_w: int, max_h: int) -> np.ndarray:
+    h, w = image.shape[:2]
+
+    if w <= 0 or h <= 0:
+        return image
+
+    scale = min(
+        float(max_w) / float(w),
+        float(max_h) / float(h),
+    )
+    scale = max(scale, 1e-6)
+
+    new_w = max(1, int(round(w * scale)))
+    new_h = max(1, int(round(h * scale)))
+
+    interp = cv2.INTER_AREA if scale < 1.0 else cv2.INTER_LINEAR
+    return cv2.resize(image, (new_w, new_h), interpolation=interp)
+
+
+def live_quality(frame_bgr: np.ndarray) -> dict:
+    """
+    QA apenas visual.
+    Nunca rejeita nem altera a imagem salva.
+    """
+    small = fit_inside(frame_bgr, 640, 360)
+    gray = cv2.cvtColor(small, cv2.COLOR_BGR2GRAY)
+
+    mean = float(gray.mean())
+    dark_pct = float((gray <= 5).mean() * 100.0)
+    bright_pct = float((gray >= 250).mean() * 100.0)
+    sharpness = float(cv2.Laplacian(gray, cv2.CV_64F).var())
+
+    warnings = []
+
+    if mean < 35:
+        warnings.append("MUITO ESCURA")
+    elif mean > 225:
+        warnings.append("MUITO CLARA")
+
+    if dark_pct > 20.0:
+        warnings.append("CLIP PRETO")
+
+    if bright_pct > 8.0:
+        warnings.append("CLIP BRANCO")
+
+    if sharpness < 35.0:
+        warnings.append("FOCO/BLUR?")
+
+    return {
+        "mean": mean,
+        "dark_pct": dark_pct,
+        "bright_pct": bright_pct,
+        "sharpness": sharpness,
+        "warnings": warnings,
+    }
+
+
+def put_line(
+    canvas: np.ndarray,
+    text: str,
+    xy: Tuple[int, int],
+    scale: float = 0.65,
+    thickness: int = 1,
+    color=(235, 235, 235),
+):
+    cv2.putText(
+        canvas,
+        text,
+        xy,
+        cv2.FONT_HERSHEY_SIMPLEX,
+        scale,
+        color,
+        thickness,
+        cv2.LINE_AA,
+    )
+
+
+class CameraControls:
+    def __init__(
+        self,
+        queue,
+        exposure_us: int,
+        iso: int,
+        focus: int,
+    ):
+        self.queue = queue
+
+        self.exposure_auto = True
+        self.focus_auto = True
+
+        self.exposure_us = int(exposure_us)
+        self.iso = int(iso)
+        self.focus = int(focus)
+
+    def send_initial_auto(self):
+        ctrl = dai.CameraControl()
+
+        try:
+            ctrl.setAutoExposureEnable()
+        except Exception:
+            pass
+
+        try:
+            ctrl.setAutoWhiteBalanceLock(False)
+        except Exception:
+            pass
+
+        try:
+            ctrl.setAutoFocusMode(
+                dai.CameraControl.AutoFocusMode.CONTINUOUS_VIDEO
+            )
+            ctrl.setAutoFocusTrigger()
+        except Exception:
+            pass
+
+        self.queue.send(ctrl)
+
+    def set_exposure_auto(self):
+        ctrl = dai.CameraControl()
+
+        try:
+            ctrl.setAutoExposureEnable()
+        except Exception as exc:
+            print(f"[WARN] Nao consegui ativar AE: {exc}")
+            return
+
+        self.queue.send(ctrl)
+        self.exposure_auto = True
+        print("[CAM] Exposicao: AUTO")
+
+    def set_exposure_manual(self):
+        self.exposure_us = int(
+            np.clip(
+                self.exposure_us,
+                EXPOSURE_MIN_US,
+                EXPOSURE_MAX_US,
+            )
+        )
+        self.iso = int(
+            np.clip(
+                self.iso,
+                ISO_MIN,
+                ISO_MAX,
+            )
+        )
+
+        ctrl = dai.CameraControl()
+        ctrl.setManualExposure(
+            self.exposure_us,
+            self.iso,
+        )
+        self.queue.send(ctrl)
+
+        self.exposure_auto = False
+
+        print(
+            f"[CAM] Exposicao: MANUAL | "
+            f"{self.exposure_us} us | ISO {self.iso}"
+        )
+
+    def toggle_exposure(self):
+        if self.exposure_auto:
+            self.set_exposure_manual()
+        else:
+            self.set_exposure_auto()
+
+    def adjust_exposure(self, delta_us: int):
+        if self.exposure_auto:
+            print("[CAM] Ajuste de exposicao ignorado: pressione E para modo MANUAL.")
+            return
+
+        self.exposure_us += int(delta_us)
+        self.set_exposure_manual()
+
+    def adjust_iso(self, delta_iso: int):
+        if self.exposure_auto:
+            print("[CAM] Ajuste de ISO ignorado: pressione E para modo MANUAL.")
+            return
+
+        self.iso += int(delta_iso)
+        self.set_exposure_manual()
+
+    def set_focus_auto(self):
+        ctrl = dai.CameraControl()
+
+        try:
+            ctrl.setAutoFocusMode(
+                dai.CameraControl.AutoFocusMode.CONTINUOUS_VIDEO
+            )
+            ctrl.setAutoFocusTrigger()
+        except Exception as exc:
+            print(f"[WARN] Autofocus nao disponivel: {exc}")
+            return
+
+        self.queue.send(ctrl)
+        self.focus_auto = True
+        print("[CAM] Foco: AUTO CONTINUOUS")
+
+    def set_focus_manual(self):
+        self.focus = int(
+            np.clip(
+                self.focus,
+                FOCUS_MIN,
+                FOCUS_MAX,
+            )
+        )
+
+        ctrl = dai.CameraControl()
+
+        try:
+            ctrl.setManualFocus(self.focus)
+        except Exception as exc:
+            print(f"[WARN] Foco manual nao disponivel: {exc}")
+            return
+
+        self.queue.send(ctrl)
+        self.focus_auto = False
+        print(f"[CAM] Foco: MANUAL | lens={self.focus}")
+
+    def toggle_focus(self):
+        if self.focus_auto:
+            self.set_focus_manual()
+        else:
+            self.set_focus_auto()
+
+    def adjust_focus(self, delta: int):
+        if self.focus_auto:
+            print("[CAM] Ajuste de foco ignorado: pressione F para modo MANUAL.")
+            return
+
+        self.focus += int(delta)
+        self.set_focus_manual()
+
+
+def create_pipeline(fps: float) -> dai.Pipeline:
+    pipeline = dai.Pipeline()
+
+    cam_rgb = pipeline.createColorCamera()
+    cam_rgb.setBoardSocket(dai.CameraBoardSocket.CAM_A)
+    cam_rgb.setResolution(
+        dai.ColorCameraProperties.SensorResolution.THE_1080_P
+    )
+    cam_rgb.setFps(float(fps))
+    cam_rgb.setInterleaved(False)
+
+    # Mesmo contrato do runtime atual.
+    cam_rgb.setColorOrder(
+        dai.ColorCameraProperties.ColorOrder.RGB
+    )
+
+    xout = pipeline.createXLinkOut()
+    xout.setStreamName("rgb")
+
+    # IMPORTANTE: VIDEO, nao preview.
+    cam_rgb.video.link(xout.input)
+
+    control_in = pipeline.createXLinkIn()
+    control_in.setStreamName("control")
+    control_in.out.link(cam_rgb.inputControl)
+
+    return pipeline
+
+
+def build_canvas(
+    frame_bgr: np.ndarray,
+    last_saved: Optional[np.ndarray],
+    *,
+    controls: CameraControls,
+    auto_save: bool,
+    interval_s: float,
+    saved_count: int,
+    out_dir: Path,
+    fps_view: float,
+    quality: dict,
+) -> np.ndarray:
+    canvas_w = 1600
+    canvas_h = 900
+
+    canvas = np.zeros(
+        (canvas_h, canvas_w, 3),
+        dtype=np.uint8,
+    )
+
+    live = fit_inside(frame_bgr, 1150, 780)
+    lh, lw = live.shape[:2]
+
+    live_x = 20
+    live_y = 60
+
+    canvas[
+        live_y:live_y + lh,
+        live_x:live_x + lw,
+    ] = live
+
+    panel_x = 1200
+
+    put_line(
+        canvas,
+        "OAK-D LITE / CORREDOR",
+        (panel_x, 60),
+        scale=0.78,
+        thickness=2,
+    )
+
+    put_line(
+        canvas,
+        "Fonte: VIDEO 1920x1080",
+        (panel_x, 100),
+    )
+
+    put_line(
+        canvas,
+        f"View FPS: {fps_view:.1f}",
+        (panel_x, 130),
+    )
+
+    put_line(
+        canvas,
+        f"Salvas: {saved_count}",
+        (panel_x, 160),
+    )
+
+    exp_text = (
+        "AUTO"
+        if controls.exposure_auto
+        else f"MANUAL {controls.exposure_us}us ISO{controls.iso}"
+    )
+
+    focus_text = (
+        "AUTO"
+        if controls.focus_auto
+        else f"MANUAL {controls.focus}"
+    )
+
+    put_line(
+        canvas,
+        f"Exposure: {exp_text}",
+        (panel_x, 205),
+    )
+
+    put_line(
+        canvas,
+        f"Focus: {focus_text}",
+        (panel_x, 235),
+    )
+
+    auto_text = (
+        f"ON ({interval_s:.2f}s)"
+        if auto_save
+        else "OFF"
+    )
+
+    put_line(
+        canvas,
+        f"Auto-save: {auto_text}",
+        (panel_x, 265),
+    )
+
+    put_line(
+        canvas,
+        f"Luma mean: {quality['mean']:.1f}",
+        (panel_x, 315),
+    )
+    put_line(
+        canvas,
+        f"Dark clip: {quality['dark_pct']:.1f}%",
+        (panel_x, 345),
+    )
+    put_line(
+        canvas,
+        f"White clip: {quality['bright_pct']:.1f}%",
+        (panel_x, 375),
+    )
+    put_line(
+        canvas,
+        f"Sharpness: {quality['sharpness']:.1f}",
+        (panel_x, 405),
+    )
+
+    if quality["warnings"]:
+        put_line(
+            canvas,
+            "QA: " + " | ".join(quality["warnings"]),
+            (panel_x, 440),
+            scale=0.55,
+            thickness=2,
+            color=(0, 180, 255),
+        )
+    else:
+        put_line(
+            canvas,
+            "QA: OK",
+            (panel_x, 440),
+            scale=0.62,
+            thickness=2,
+            color=(80, 230, 80),
+        )
+
+    put_line(canvas, "SPACE/S  salvar", (panel_x, 510))
+    put_line(canvas, "A        auto-save", (panel_x, 540))
+    put_line(canvas, "E        auto/manual exp", (panel_x, 570))
+    put_line(canvas, "+/-      ISO manual", (panel_x, 600))
+    put_line(canvas, "M/N      exposicao manual", (panel_x, 630))
+    put_line(canvas, "F        auto/manual foco", (panel_x, 660))
+    put_line(canvas, "[/]      foco manual", (panel_x, 690))
+    put_line(canvas, "V        fullscreen", (panel_x, 720))
+    put_line(canvas, "Q/ESC    sair", (panel_x, 750))
+
+    if last_saved is not None:
+        thumb = fit_inside(last_saved, 360, 100)
+        th, tw = thumb.shape[:2]
+
+        tx = panel_x
+        ty = 780
+
+        if ty + th <= canvas_h and tx + tw <= canvas_w:
+            canvas[
+                ty:ty + th,
+                tx:tx + tw,
+            ] = thumb
+
+    put_line(
+        canvas,
+        f"Saida: {out_dir}",
+        (20, 30),
+        scale=0.62,
+        color=(200, 220, 255),
+    )
+
+    return canvas
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description=(
+            "Captura PNG lossless da OAK-D Lite para dataset "
+            "de corredor agricola."
+        )
+    )
+
+    parser.add_argument(
+        "--out",
+        default=str(DEFAULT_OUT),
+        help="Fila de imagens brutas (somente PNGs).",
+    )
+
+    parser.add_argument(
+        "--fps",
+        type=float,
+        default=CAPTURE_FPS,
+        help="FPS da camera.",
+    )
+
+    parser.add_argument(
+        "--auto",
+        action="store_true",
+        help="Inicia com auto-save ligado.",
+    )
+
+    parser.add_argument(
+        "--interval",
+        type=float,
+        default=1.0,
+        help="Intervalo do auto-save em segundos.",
+    )
+
+    parser.add_argument(
+        "--exposure_us",
+        type=int,
+        default=DEFAULT_EXPOSURE_US,
+        help="Exposicao inicial quando entrar em modo manual.",
+    )
+
+    parser.add_argument(
+        "--iso",
+        type=int,
+        default=DEFAULT_ISO,
+        help="ISO inicial quando entrar em modo manual.",
+    )
+
+    parser.add_argument(
+        "--focus",
+        type=int,
+        default=DEFAULT_FOCUS,
+        help="Lens position inicial quando entrar em foco manual.",
+    )
+
+    args = parser.parse_args()
+
+    out_dir = Path(args.out)
+    out_dir.mkdir(parents=True, exist_ok=True)
+
+    saved_count = len(list(out_dir.glob("*.png")))
+
+    print("=" * 72)
+    print("OAK-D Lite | Captura Dataset Corredor Agricola")
+    print("=" * 72)
+    print(f"Saida       : {out_dir.resolve()}")
+    print("Formato     : PNG lossless")
+    print(f"Resolucao   : {CAPTURE_W}x{CAPTURE_H}")
+    print("Fonte       : ColorCamera.video")
+    print(f"FPS camera  : {args.fps}")
+    print(f"Ja existem  : {saved_count} PNGs")
+    print(f"Sessao       : {CAPTURE_SESSION_ID}")
+    print("=" * 72)
+    print("SPACE/S salvar | A auto-save | E exposure | F foco | Q sair")
+
+    pipeline = create_pipeline(args.fps)
+
+    saver = ThreadPoolExecutor(
+        max_workers=1,
+        thread_name_prefix="png_saver",
+    )
+
+    pending = []
+    last_saved: Optional[np.ndarray] = None
+
+    auto_save = bool(args.auto)
+    interval_s = max(0.10, float(args.interval))
+    last_auto_save_t = 0.0
+
+    fps_view = 0.0
+    fps_frames = 0
+    fps_t0 = time.perf_counter()
+
+    fullscreen = False
+
+    try:
+        with dai.Device(pipeline) as device:
+            control_queue = device.getInputQueue("control")
+
+            rgb_queue = device.getOutputQueue(
+                "rgb",
+                maxSize=2,
+                blocking=False,
+            )
+
+            controls = CameraControls(
+                control_queue,
+                exposure_us=args.exposure_us,
+                iso=args.iso,
+                focus=args.focus,
+            )
+
+            controls.send_initial_auto()
+
+            cv2.namedWindow(
+                WINDOW_NAME,
+                cv2.WINDOW_NORMAL,
+            )
+            cv2.resizeWindow(
+                WINDOW_NAME,
+                1600,
+                900,
+            )
+
+            while True:
+                packet = rgb_queue.get()
+                frame = packet.getCvFrame()
+
+                if (
+                    frame is None
+                    or frame.ndim != 3
+                    or frame.shape[2] != 3
+                ):
+                    print(
+                        f"[WARN] Frame invalido: "
+                        f"{getattr(frame, 'shape', None)}"
+                    )
+                    continue
+
+                h, w = frame.shape[:2]
+                if (w, h) != (CAPTURE_W, CAPTURE_H):
+                    print(
+                        f"[WARN] Frame veio {w}x{h}, "
+                        f"esperado {CAPTURE_W}x{CAPTURE_H}. "
+                        "Nao sera redimensionado silenciosamente."
+                    )
+
+                now_perf = time.perf_counter()
+
+                fps_frames += 1
+                dt_fps = now_perf - fps_t0
+                if dt_fps >= 1.0:
+                    fps_view = fps_frames / dt_fps
+                    fps_frames = 0
+                    fps_t0 = now_perf
+
+                quality = live_quality(frame)
+
+                def request_save(current_frame: np.ndarray):
+                    nonlocal saved_count, last_saved
+
+                    name = timestamp_name() + ".png"
+                    path = out_dir / name
+
+                    snapshot = current_frame.copy()
+
+                    future = saver.submit(
+                        save_png,
+                        path,
+                        snapshot,
+                    )
+                    pending.append((future, path))
+
+                    last_saved = snapshot
+                    saved_count += 1
+
+                    print(
+                        f"[CAPTURE] #{saved_count} -> {path}"
+                    )
+
+                if auto_save:
+                    if (
+                        last_auto_save_t <= 0.0
+                        or now_perf - last_auto_save_t >= interval_s
+                    ):
+                        request_save(frame)
+                        last_auto_save_t = now_perf
+
+                still_pending = []
+                for future, path in pending:
+                    if not future.done():
+                        still_pending.append((future, path))
+                        continue
+
+                    try:
+                        ok, saved_path = future.result()
+                    except Exception as exc:
+                        print(
+                            f"[ERRO] Falha salvando {path}: {exc}"
+                        )
+                        continue
+
+                    if not ok:
+                        print(
+                            f"[ERRO] cv2.imwrite retornou False: "
+                            f"{saved_path}"
+                        )
+
+                pending = still_pending
+
+                canvas = build_canvas(
+                    frame,
+                    last_saved,
+                    controls=controls,
+                    auto_save=auto_save,
+                    interval_s=interval_s,
+                    saved_count=saved_count,
+                    out_dir=out_dir,
+                    fps_view=fps_view,
+                    quality=quality,
+                )
+
+                cv2.imshow(
+                    WINDOW_NAME,
+                    canvas,
+                )
+
+                key = cv2.waitKey(1) & 0xFF
+
+                if key in (ord("q"), 27):
+                    print("[INFO] Encerrando captura...")
+                    break
+
+                if key in (ord("s"), ord(" ")):
+                    request_save(frame)
+
+                elif key == ord("a"):
+                    auto_save = not auto_save
+                    last_auto_save_t = 0.0
+                    print(
+                        f"[CAPTURE] Auto-save "
+                        f"{'ON' if auto_save else 'OFF'} "
+                        f"| interval={interval_s:.2f}s"
+                    )
+
+                elif key == ord("e"):
+                    controls.toggle_exposure()
+
+                elif key in (ord("+"), ord("=")):
+                    controls.adjust_iso(ISO_STEP)
+
+                elif key == ord("-"):
+                    controls.adjust_iso(-ISO_STEP)
+
+                elif key == ord("m"):
+                    controls.adjust_exposure(EXPOSURE_STEP_US)
+
+                elif key == ord("n"):
+                    controls.adjust_exposure(-EXPOSURE_STEP_US)
+
+                elif key == ord("f"):
+                    controls.toggle_focus()
+
+                elif key == ord("]"):
+                    controls.adjust_focus(FOCUS_STEP)
+
+                elif key == ord("["):
+                    controls.adjust_focus(-FOCUS_STEP)
+
+                elif key == ord("v"):
+                    fullscreen = not fullscreen
+                    cv2.setWindowProperty(
+                        WINDOW_NAME,
+                        cv2.WND_PROP_FULLSCREEN,
+                        (
+                            cv2.WINDOW_FULLSCREEN
+                            if fullscreen
+                            else cv2.WINDOW_NORMAL
+                        ),
+                    )
+
+    finally:
+        cv2.destroyAllWindows()
+
+        print(
+            f"[INFO] Aguardando {len(pending)} PNG(s) "
+            "que ja foram solicitados..."
+        )
+
+        saver.shutdown(wait=True)
+
+        failures = 0
+        for future, path in pending:
+            try:
+                ok, _ = future.result()
+                if not ok:
+                    failures += 1
+            except Exception as exc:
+                failures += 1
+                print(f"[ERRO] {path}: {exc}")
+
+        print("=" * 72)
+        print("Captura encerrada.")
+        print(f"Total contabilizado : {saved_count}")
+        print(f"Falhas finais       : {failures}")
+        print(f"Saida               : {out_dir.resolve()}")
+        print("=" * 72)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/Python/OAK/datasets/oak-d/_1_annotate_corridor.py b/Python/OAK/datasets/oak-d/_1_annotate_corridor.py
new file mode 100644
index 000000000..8fc03a1a8
--- /dev/null
+++ b/Python/OAK/datasets/oak-d/_1_annotate_corridor.py
@@ -0,0 +1,2376 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+"""
+_1_annotate_corridor.py
+=======================
+
+Editor definitivo da etapa BRUTAS para a camera frontal OAK-D Lite.
+
+Objetivo por imagem:
+    1) criar/editar mascara binaria do corredor:
+       - nao navegavel
+       - navegavel
+    2) definir o estado global do corredor:
+       - classes vindas de config.json -> label_classes
+    3) derivar automaticamente o grupo pela mascara final:
+       - naonavegavel
+       - navegavel
+       - naonavegavel_navegavel
+    4) arquivar a amostra pronta em dataset/original/group/<grupo>/:
+       images/<base>.png
+       masks/<base>.png       # RGB colorida conforme labelmap
+       labels/<base>.json
+
+A mascara nova nasce 100% NAO NAVEGAVEL.
+O operador normalmente desenha apenas o corredor NAVEGAVEL.
+
+Ferramentas:
+    P = poligono
+        clique esquerdo -> adiciona ponto
+        clique direito  -> remove ultimo ponto
+        ENTER            -> fecha/preenche
+
+    B = pincel
+        arrastar botao esquerdo -> pinta classe selecionada
+        arrastar botao direito  -> pinta classe oposta
+
+    F = balde / flood fill
+        clique esquerdo -> preenche regiao conectada com classe selecionada
+
+    C = troca classe da mascara
+    [ / ] = diminui/aumenta pincel
+    Z = undo
+    Y = redo
+    R = reset para 100% nao navegavel
+    O = alterna visualizacao do overlay
+
+Estado global:
+    1..9 / 0 = seleciona label conforme config["label_classes"]
+    A label NAO salva automaticamente.
+
+Fluxo:
+    S ou SPACE = SALVAR mascara + label e ir para a proxima
+    A / seta esquerda = anterior
+    D / seta direita = proxima sem salvar
+    Q / ESC = sair
+
+Execucao:
+    Rode de dentro da pasta oak-d:
+
+        python _1_annotate_corridor.py
+
+Estrutura esperada:
+    oak-d/
+      config.json
+      dataset/
+        labelmap.txt
+        brutas/                 # fila: somente PNGs ainda nao mapeadas
+          img_....png
+        original/
+          group/
+            navegavel/
+            naonavegavel/
+            naonavegavel_navegavel/
+
+Observacoes:
+    - mascara BRUTA e salva em PNG RGB colorida, conforme labelmap;
+    - a imagem original nunca e recomprimida; ela e copiada/movida byte a byte;
+    - modo padrao MOVE: somente remove a PNG de brutas depois que image+mask+label finais existem;
+    - --mode copy preserva a PNG em brutas, mas amostras ja arquivadas sao puladas por padrao;
+    - esta etapa nao escolhe train/val. Ela apenas organiza o dataset original por grupo semantico.
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import re
+import shutil
+import time
+from dataclasses import dataclass
+from datetime import datetime
+from pathlib import Path
+from typing import Dict, List, Optional, Sequence, Tuple
+
+import cv2
+import numpy as np
+
+
+# =============================================================================
+# Defaults
+# =============================================================================
+
+WINDOW_NAME = "Agrobot | Editor Corredor OAK-D Lite"
+
+IMG_EXTS = {
+    ".png",
+    ".jpg",
+    ".jpeg",
+    ".bmp",
+    ".webp",
+}
+
+DEFAULT_BRUTAS_DIR = Path("dataset") / "brutas"
+DEFAULT_ORIGINAL_GROUP_ROOT = Path("dataset") / "original" / "group"
+DEFAULT_LABELMAP = Path("dataset") / "labelmap.txt"
+DEFAULT_CONFIG = Path("config.json")
+
+MAX_WINDOW_W = 1800
+MAX_WINDOW_H = 1000
+PANEL_W = 430
+HEADER_H = 72
+FOOTER_H = 92
+
+DEFAULT_OVERLAY_ALPHA = 0.42
+DEFAULT_BRUSH_RADIUS = 24
+
+MASK_SCHEMA = "agrobot_corridor_mask_rgb_v1"
+LABEL_SCHEMA = "agrobot_corridor_state_v2"
+
+# Cores de fallback apenas para visualizacao.
+FALLBACK_COLORS_RGB = {
+    0: (210, 55, 55),
+    1: (45, 215, 70),
+}
+
+
+# =============================================================================
+# Estruturas
+# =============================================================================
+
+@dataclass
+class SampleItem:
+    image_path: Path
+    mask_path: Path
+    label_path: Path
+    base: str
+
+
+@dataclass
+class ViewTransform:
+    x0: int = 0
+    y0: int = 0
+    width: int = 1
+    height: int = 1
+    scale: float = 1.0
+
+
+# =============================================================================
+# Utilitarios
+# =============================================================================
+
+def natural_key(text: str):
+    parts = re.split(r"(\d+)", str(text))
+    return [
+        int(p) if p.isdigit() else p.lower()
+        for p in parts
+    ]
+
+
+def now_iso() -> str:
+    return datetime.now().isoformat(timespec="seconds")
+
+
+def atomic_write_json(path: Path, data: dict) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    tmp = path.with_suffix(path.suffix + ".tmp")
+
+    with tmp.open("w", encoding="utf-8") as f:
+        json.dump(
+            data,
+            f,
+            ensure_ascii=False,
+            indent=2,
+        )
+
+    os.replace(tmp, path)
+
+
+def atomic_write_png_color(path: Path, mask_bgr: np.ndarray) -> None:
+    """Salva a máscara BRUTA colorida, lossless, conforme as cores exatas do labelmap."""
+    path.parent.mkdir(parents=True, exist_ok=True)
+
+    if mask_bgr.dtype != np.uint8:
+        mask_bgr = mask_bgr.astype(np.uint8)
+
+    if mask_bgr.ndim != 3 or mask_bgr.shape[2] != 3:
+        raise ValueError(
+            f"Máscara colorida deve ser HxWx3 uint8; recebido {mask_bgr.shape}"
+        )
+
+    tmp = path.with_name(path.stem + ".__tmp__.png")
+
+    ok = cv2.imwrite(
+        str(tmp),
+        mask_bgr,
+        [cv2.IMWRITE_PNG_COMPRESSION, 3],
+    )
+
+    if not ok:
+        try:
+            if tmp.exists():
+                tmp.unlink()
+        except Exception:
+            pass
+        raise IOError(f"cv2.imwrite retornou False para {tmp}")
+
+    os.replace(tmp, path)
+
+
+def safe_relative(path: Path, root: Path) -> str:
+    try:
+        return str(
+            path.resolve().relative_to(root.resolve())
+        ).replace("\\", "/")
+    except Exception:
+        return str(path).replace("\\", "/")
+
+
+def read_json_safe(path: Path) -> Optional[dict]:
+    if not path.is_file():
+        return None
+
+    try:
+        with path.open("r", encoding="utf-8") as f:
+            value = json.load(f)
+        return value if isinstance(value, dict) else None
+    except Exception:
+        return None
+
+
+def list_images(folder: Path) -> List[Path]:
+    if not folder.is_dir():
+        raise FileNotFoundError(f"Pasta de imagens nao encontrada: {folder}")
+
+    files = [
+        p
+        for p in folder.iterdir()
+        if p.is_file()
+        and p.suffix.lower() in IMG_EXTS
+    ]
+
+    files.sort(key=lambda p: natural_key(p.name))
+    return files
+
+
+# =============================================================================
+# Labelmap
+# =============================================================================
+
+def load_labelmap_classes(
+    path: Path,
+) -> Dict[int, Tuple[str, Tuple[int, int, int]]]:
+    """
+    Retorna:
+        id -> (nome, cor RGB)
+
+    Formatos aceitos:
+        naonavegavel: 220,40,40
+        navegavel: 40,220,80
+
+        0 naonavegavel 220 40 40
+        1 navegavel 40 220 80
+
+        0:naonavegavel
+        1:navegavel
+    """
+    if not path.is_file():
+        raise FileNotFoundError(
+            f"labelmap obrigatorio nao encontrado: {path}"
+        )
+
+    classes: Dict[int, Tuple[str, Tuple[int, int, int]]] = {}
+    next_id = 0
+
+    with path.open("r", encoding="utf-8") as f:
+        for raw_line in f:
+            s = raw_line.strip()
+
+            if not s or s.startswith("#"):
+                continue
+
+            cid: Optional[int] = None
+            name: Optional[str] = None
+            color: Optional[Tuple[int, int, int]] = None
+
+            left = s.split(":", 1)[0].strip()
+
+            # Formato nome: R,G,B :: ...
+            if ":" in s and not left.isdigit():
+                name_part, rest = s.split(":", 1)
+                name = name_part.strip()
+
+                color_txt = (
+                    rest.split("::", 1)[0]
+                    .strip()
+                    .strip(":")
+                )
+
+                parts = [
+                    p.strip()
+                    for p in color_txt.split(",")
+                    if p.strip()
+                ]
+
+                if len(parts) >= 3:
+                    try:
+                        color = tuple(
+                            int(float(x))
+                            for x in parts[:3]
+                        )
+                    except Exception:
+                        color = None
+
+            else:
+                parts = (
+                    s.replace(",", " ")
+                    .replace(":", " ")
+                    .split()
+                )
+
+                if len(parts) >= 2 and parts[0].isdigit():
+                    cid = int(parts[0])
+                    name = parts[1]
+
+                    if len(parts) >= 5:
+                        try:
+                            color = tuple(
+                                int(float(x))
+                                for x in parts[2:5]
+                            )
+                        except Exception:
+                            color = None
+
+                elif parts:
+                    name = parts[0]
+
+            if not name:
+                continue
+
+            lname = name.strip().lower()
+            if lname in {
+                "ignore",
+                "void",
+                "background_ignore",
+            }:
+                continue
+
+            if cid is None:
+                cid = next_id
+
+            next_id = max(next_id, cid + 1)
+
+            if color is None:
+                color = FALLBACK_COLORS_RGB.get(
+                    int(cid),
+                    (255, 255, 255),
+                )
+
+            classes[int(cid)] = (
+                str(name),
+                tuple(map(int, color)),
+            )
+
+    if len(classes) != 2:
+        raise RuntimeError(
+            "Este editor e especializado em segmentacao binaria "
+            f"navegavel/nao navegavel. labelmap={classes}"
+        )
+
+    return classes
+
+
+def class_color_bgr(
+    classes: Dict[int, Tuple[str, Tuple[int, int, int]]],
+    cid: int,
+) -> Tuple[int, int, int]:
+    _name, rgb = classes[int(cid)]
+    return (
+        int(rgb[2]),
+        int(rgb[1]),
+        int(rgb[0]),
+    )
+
+
+def resolve_nav_classes(
+    classes: Dict[int, Tuple[str, Tuple[int, int, int]]],
+    main_class_name: str,
+) -> Tuple[int, int]:
+    target = str(main_class_name).strip().lower()
+
+    exact = [
+        cid
+        for cid, (name, _rgb) in classes.items()
+        if str(name).strip().lower() == target
+    ]
+
+    if len(exact) != 1:
+        # Compatibilidade historica do projeto.
+        aliases = {
+            "navegavel",
+            "navegável",
+            "nav",
+            "navigable",
+        }
+
+        exact = [
+            cid
+            for cid, (name, _rgb) in classes.items()
+            if str(name).strip().lower() in aliases
+        ]
+
+    if len(exact) != 1:
+        raise RuntimeError(
+            f"Nao consegui resolver classe navegavel. "
+            f"main_class_name={main_class_name!r} classes={classes}"
+        )
+
+    nav_id = int(exact[0])
+    non_nav = [
+        int(cid)
+        for cid in classes
+        if int(cid) != nav_id
+    ]
+
+    if len(non_nav) != 1:
+        raise RuntimeError(
+            f"Nao consegui resolver classe nao navegavel: {classes}"
+        )
+
+    return nav_id, non_nav[0]
+
+
+# =============================================================================
+# Mask IO
+# =============================================================================
+
+def colored_mask_to_ids(
+    mask_bgr: np.ndarray,
+    classes: Dict[int, Tuple[str, Tuple[int, int, int]]],
+) -> np.ndarray:
+    h, w = mask_bgr.shape[:2]
+    ids = np.full(
+        (h, w),
+        255,
+        dtype=np.uint8,
+    )
+
+    for cid in classes:
+        bgr = np.array(
+            class_color_bgr(classes, cid),
+            dtype=np.uint8,
+        )
+        match = np.all(
+            mask_bgr == bgr[None, None, :],
+            axis=2,
+        )
+        ids[match] = int(cid)
+
+    if np.any(ids == 255):
+        unknown = int((ids == 255).sum())
+        raise ValueError(
+            f"Mascara colorida possui {unknown} pixels "
+            "que nao casam exatamente com o labelmap."
+        )
+
+    return ids
+
+
+def load_mask_ids(
+    path: Path,
+    expected_hw: Tuple[int, int],
+    classes: Dict[int, Tuple[str, Tuple[int, int, int]]],
+) -> np.ndarray:
+    raw = cv2.imread(
+        str(path),
+        cv2.IMREAD_UNCHANGED,
+    )
+
+    if raw is None:
+        raise FileNotFoundError(f"Nao consegui abrir mask: {path}")
+
+    if raw.ndim == 2:
+        ids = raw.astype(
+            np.uint8,
+            copy=False,
+        )
+    elif raw.ndim == 3 and raw.shape[2] >= 3:
+        ids = colored_mask_to_ids(
+            raw[:, :, :3],
+            classes,
+        )
+    else:
+        raise ValueError(
+            f"Formato de mask invalido: shape={raw.shape}"
+        )
+
+    if ids.shape != expected_hw:
+        raise ValueError(
+            f"Mask {path.name} shape={ids.shape} "
+            f"mas imagem shape={expected_hw}. "
+            "Nao farei resize silencioso."
+        )
+
+    valid_ids = set(int(x) for x in np.unique(ids))
+    allowed = set(int(cid) for cid in classes)
+
+    if not valid_ids.issubset(allowed):
+        raise ValueError(
+            f"Mask {path.name} possui IDs invalidos "
+            f"{sorted(valid_ids - allowed)}; permitidos={sorted(allowed)}"
+        )
+
+    return ids.copy()
+
+
+def ids_to_color(
+    ids: np.ndarray,
+    classes: Dict[int, Tuple[str, Tuple[int, int, int]]],
+) -> np.ndarray:
+    out = np.zeros(
+        (ids.shape[0], ids.shape[1], 3),
+        dtype=np.uint8,
+    )
+
+    for cid in classes:
+        out[ids == int(cid)] = class_color_bgr(
+            classes,
+            int(cid),
+        )
+
+    return out
+
+
+# =============================================================================
+# Editor
+# =============================================================================
+
+class CorridorAnnotationEditor:
+    TOOL_POLYGON = "POLIGONO"
+    TOOL_BRUSH = "PINCEL"
+    TOOL_FILL = "BALDE"
+
+    OVERLAY_NAV_ONLY = 0
+    OVERLAY_BOTH = 1
+    OVERLAY_OFF = 2
+
+    def __init__(
+        self,
+        *,
+        dataset_root: Path,
+        samples: Sequence[SampleItem],
+        states: Sequence[str],
+        classes: Dict[int, Tuple[str, Tuple[int, int, int]]],
+        nav_id: int,
+        non_nav_id: int,
+        original_group_root: Path,
+        archive_mode: str,
+        archived_before: int,
+        overlay_alpha: float,
+        brush_radius: int,
+        start_index: Optional[int],
+    ):
+        self.dataset_root = dataset_root
+        self.samples = list(samples)
+        self.states = [str(x) for x in states]
+        self.classes = classes
+
+        self.nav_id = int(nav_id)
+        self.non_nav_id = int(non_nav_id)
+        self.original_group_root = Path(original_group_root)
+        self.archive_mode = str(archive_mode).lower()
+        if self.archive_mode not in {"move", "copy"}:
+            raise ValueError(f"archive_mode invalido: {self.archive_mode}")
+        self.archived_before = int(archived_before)
+        self.saved_this_session = 0
+        self.finished = False
+        self.initial_queue_size = len(self.samples)
+
+        self.resume = False
+        self.overlay_alpha = float(
+            np.clip(
+                overlay_alpha,
+                0.0,
+                1.0,
+            )
+        )
+        self.brush_radius = max(
+            2,
+            int(brush_radius),
+        )
+
+        self.idx = 0
+
+        self.image: Optional[np.ndarray] = None
+        self.mask: Optional[np.ndarray] = None
+        self.original_loaded_mask: Optional[np.ndarray] = None
+
+        self.selected_state: Optional[int] = None
+        self.loaded_state: Optional[int] = None
+
+        self.selected_class = self.nav_id
+        self.tool = self.TOOL_POLYGON
+        self.overlay_mode = self.OVERLAY_NAV_ONLY
+
+        self.polygon_points: List[Tuple[int, int]] = []
+
+        self.undo_stack: List[np.ndarray] = []
+        self.redo_stack: List[np.ndarray] = []
+
+        self.mask_dirty = False
+        self.label_dirty = False
+
+        self.view = ViewTransform()
+        self.last_canvas_shape = (
+            MAX_WINDOW_H,
+            MAX_WINDOW_W,
+        )
+
+        self.mouse_down = False
+        self.mouse_button: Optional[int] = None
+        self.brush_action_started = False
+        self.mouse_src = (0, 0)
+
+        self.message = ""
+        self.message_until = 0.0
+
+        self._completed_cache: Dict[str, bool] = {}
+
+        if start_index is not None:
+            self.idx = max(
+                0,
+                min(
+                    len(self.samples) - 1,
+                    int(start_index),
+                ),
+            )
+
+        cv2.namedWindow(
+            WINDOW_NAME,
+            cv2.WINDOW_NORMAL,
+        )
+        cv2.resizeWindow(
+            WINDOW_NAME,
+            MAX_WINDOW_W,
+            MAX_WINDOW_H,
+        )
+        cv2.setMouseCallback(
+            WINDOW_NAME,
+            self.on_mouse,
+        )
+
+        self.load_current()
+
+    # -------------------------------------------------------------------------
+    # Completion / navigation
+    # -------------------------------------------------------------------------
+
+    def sample_complete(
+        self,
+        item: SampleItem,
+        refresh: bool = False,
+    ) -> bool:
+        if not refresh and item.base in self._completed_cache:
+            return self._completed_cache[item.base]
+
+        ok = False
+
+        if (
+            item.mask_path.is_file()
+            and item.label_path.is_file()
+        ):
+            try:
+                img = cv2.imread(
+                    str(item.image_path),
+                    cv2.IMREAD_COLOR,
+                )
+
+                if img is not None:
+                    _ = load_mask_ids(
+                        item.mask_path,
+                        img.shape[:2],
+                        self.classes,
+                    )
+
+                    lab = read_json_safe(
+                        item.label_path,
+                    )
+
+                    lid = (
+                        int(lab["label_id"])
+                        if lab is not None
+                        and "label_id" in lab
+                        else -1
+                    )
+
+                    ok = (
+                        0 <= lid < len(self.states)
+                    )
+            except Exception:
+                ok = False
+
+        self._completed_cache[item.base] = ok
+        return ok
+
+    def find_first_incomplete(self) -> Optional[int]:
+        for i, item in enumerate(self.samples):
+            if not self.sample_complete(item):
+                return i
+        return None
+
+    def completed_count(self) -> int:
+        return self.archived_before + self.saved_this_session
+
+    def can_discard_current(self) -> bool:
+        return not (
+            self.mask_dirty
+            or self.label_dirty
+            or self.polygon_points
+        )
+
+    def navigate(
+        self,
+        delta: int,
+    ):
+        delta = -1 if int(delta) < 0 else 1
+
+        if (
+            self.mask_dirty
+            or self.label_dirty
+            or self.polygon_points
+        ):
+            now = time.monotonic()
+            pending = getattr(
+                self,
+                "_pending_nav",
+                None,
+            )
+
+            same_request = (
+                isinstance(pending, tuple)
+                and len(pending) == 2
+                and int(pending[0]) == delta
+                and now - float(pending[1]) <= 2.0
+            )
+
+            if not same_request:
+                self._pending_nav = (
+                    delta,
+                    now,
+                )
+                direction = (
+                    "anterior"
+                    if delta < 0
+                    else "proxima"
+                )
+                self.flash(
+                    f"Alteracoes nao salvas. "
+                    f"Repita o comando para ir para {direction} "
+                    "DESCARTANDO, ou S/SPACE para salvar.",
+                    seconds=2.2,
+                )
+                return
+
+            self._pending_nav = None
+            self.flash(
+                "Alteracoes descartadas.",
+                seconds=1.0,
+            )
+
+        else:
+            self._pending_nav = None
+
+        self.idx = max(
+            0,
+            min(
+                len(self.samples) - 1,
+                self.idx + delta,
+            ),
+        )
+
+        self.load_current()
+
+    # -------------------------------------------------------------------------
+    # Load / save
+    # -------------------------------------------------------------------------
+
+    def load_current(self):
+        if not self.samples:
+            raise RuntimeError("Nenhuma imagem em dataset/brutas")
+
+        item = self.samples[self.idx]
+
+        image = cv2.imread(
+            str(item.image_path),
+            cv2.IMREAD_COLOR,
+        )
+
+        if image is None:
+            raise FileNotFoundError(
+                f"Nao consegui abrir imagem: {item.image_path}"
+            )
+
+        self.image = image
+
+        if item.mask_path.is_file():
+            self.mask = load_mask_ids(
+                item.mask_path,
+                image.shape[:2],
+                self.classes,
+            )
+        else:
+            self.mask = np.full(
+                image.shape[:2],
+                self.non_nav_id,
+                dtype=np.uint8,
+            )
+
+        self.original_loaded_mask = self.mask.copy()
+
+        existing = read_json_safe(
+            item.label_path,
+        )
+
+        self.selected_state = None
+        self.loaded_state = None
+
+        if existing is not None:
+            try:
+                lid = int(existing.get("label_id"))
+            except Exception:
+                lid = -1
+
+            if 0 <= lid < len(self.states):
+                self.selected_state = lid
+                self.loaded_state = lid
+
+        self.selected_class = self.nav_id
+        self.tool = self.TOOL_POLYGON
+
+        self.polygon_points = []
+
+        self.undo_stack = []
+        self.redo_stack = []
+
+        self.mask_dirty = False
+        self.label_dirty = False
+
+        self.mouse_down = False
+        self.mouse_button = None
+        self.brush_action_started = False
+
+        status = (
+            "existente carregado"
+            if self.sample_complete(item, refresh=True)
+            else "novo/incompleto"
+        )
+
+        self.flash(
+            f"{item.image_path.name} | {status}",
+            seconds=1.4,
+        )
+
+    def derive_group(self) -> str:
+        assert self.mask is not None
+
+        present = set(int(x) for x in np.unique(self.mask))
+        allowed = {self.nav_id, self.non_nav_id}
+        if not present.issubset(allowed):
+            raise ValueError(
+                f"Mascara possui IDs inesperados: {sorted(present - allowed)}"
+            )
+
+        if present == {self.non_nav_id}:
+            return "naonavegavel"
+        if present == {self.nav_id}:
+            return "navegavel"
+        if present == {self.nav_id, self.non_nav_id}:
+            return "naonavegavel_navegavel"
+
+        raise ValueError(f"Nao consegui derivar grupo da mascara: IDs={sorted(present)}")
+
+    def final_paths(self, item: SampleItem, group: str):
+        root = self.original_group_root / group
+        image_dir = root / "images"
+        mask_dir = root / "masks"
+        label_dir = root / "labels"
+        for folder in (image_dir, mask_dir, label_dir):
+            folder.mkdir(parents=True, exist_ok=True)
+
+        return (
+            image_dir / item.image_path.name,
+            mask_dir / f"{item.base}.png",
+            label_dir / f"{item.base}.json",
+        )
+
+    def build_label_record(
+        self,
+        item: SampleItem,
+        group: str,
+        final_image: Path,
+        final_mask: Path,
+        final_label: Path,
+    ) -> dict:
+        assert self.selected_state is not None
+        assert self.mask is not None
+        assert self.image is not None
+
+        nav_pct = float(
+            (self.mask == self.nav_id).mean()
+            * 100.0
+        )
+
+        return {
+            "schema": LABEL_SCHEMA,
+            "estado_corredor": self.states[self.selected_state],
+            "label_id": int(self.selected_state),
+            "states": list(self.states),
+
+            "base": item.base,
+            "group": group,
+            "image": safe_relative(
+                final_image,
+                self.dataset_root,
+            ),
+            "mask": safe_relative(
+                final_mask,
+                self.dataset_root,
+            ),
+            "label": safe_relative(
+                final_label,
+                self.dataset_root,
+            ),
+            "source_bruta": safe_relative(
+                item.image_path,
+                self.dataset_root,
+            ),
+
+            "mask_schema": MASK_SCHEMA,
+            "mask_format": "rgb_color_png",
+            "mask_dtype": "uint8",
+            "mask_shape_hw": [
+                int(self.mask.shape[0]),
+                int(self.mask.shape[1]),
+            ],
+            "mask_classes": {
+                str(cid): {
+                    "name": self.classes[cid][0],
+                    "rgb": list(self.classes[cid][1]),
+                }
+                for cid in sorted(self.classes)
+            },
+            "nav_class_id": int(self.nav_id),
+            "non_nav_class_id": int(self.non_nav_id),
+            "navegavel_pct": nav_pct,
+
+            "annotated_at": now_iso(),
+            "source": "agri_corridor_editor",
+        }
+
+    def save_and_next(self):
+        item = self.samples[self.idx]
+
+        if self.image is None or self.mask is None:
+            return
+
+        if self.polygon_points:
+            self.flash(
+                "Ha poligono aberto. ENTER confirma ou BACKSPACE cancela."
+            )
+            return
+
+        if self.selected_state is None:
+            self.flash(
+                "Selecione o estado global com 1..9/0 antes de salvar."
+            )
+            return
+
+        valid_ids = set(int(x) for x in np.unique(self.mask))
+        allowed = set(int(cid) for cid in self.classes)
+
+        if not valid_ids.issubset(allowed):
+            self.flash(
+                f"ERRO mask IDs invalidos: {sorted(valid_ids - allowed)}",
+                seconds=3.0,
+            )
+            return
+
+        nav_pct = float((self.mask == self.nav_id).mean() * 100.0)
+
+        try:
+            group = self.derive_group()
+            final_image, final_mask, final_label = self.final_paths(item, group)
+
+            # Nunca sobrescreve silenciosamente uma amostra já arquivada.
+            collisions = [
+                str(p) for p in (final_image, final_mask, final_label) if p.exists()
+            ]
+            if collisions:
+                raise FileExistsError(
+                    "Amostra ja existe no destino final: " + ", ".join(collisions)
+                )
+
+            mask_bgr = ids_to_color(self.mask, self.classes)
+            record = self.build_label_record(
+                item,
+                group,
+                final_image,
+                final_mask,
+                final_label,
+            )
+
+            created = []
+            try:
+                # Copia a imagem sem recomprimir. O source só é removido no fim.
+                final_image.parent.mkdir(parents=True, exist_ok=True)
+                shutil.copy2(item.image_path, final_image)
+                created.append(final_image)
+
+                atomic_write_png_color(final_mask, mask_bgr)
+                created.append(final_mask)
+
+                atomic_write_json(final_label, record)
+                created.append(final_label)
+
+                if not all(p.is_file() for p in (final_image, final_mask, final_label)):
+                    raise IOError("Commit incompleto: image/mask/label final nao confirmado")
+
+                if self.archive_mode == "move":
+                    item.image_path.unlink()
+
+            except Exception:
+                # Como colisões são bloqueadas antes, tudo em created pertence a este commit.
+                for path in reversed(created):
+                    try:
+                        if path.exists():
+                            path.unlink()
+                    except Exception:
+                        pass
+                raise
+
+        except Exception as exc:
+            self.flash(
+                f"ERRO salvando: {exc}",
+                seconds=4.0,
+            )
+            return
+
+        print(
+            f"[SAVE] {item.base} | group={group} | "
+            f"label={self.states[self.selected_state]} | "
+            f"nav={nav_pct:.2f}% | mode={self.archive_mode}"
+        )
+
+        self.saved_this_session += 1
+        self.mask_dirty = False
+        self.label_dirty = False
+
+        # A fila é consumida durante a sessão em MOVE e também em COPY.
+        # No COPY o arquivo físico permanece em brutas, mas na próxima execução
+        # collect_samples pula bases já arquivadas por padrão.
+        self.samples.pop(self.idx)
+
+        if not self.samples:
+            self.finished = True
+            self.flash(
+                "Fila de brutas concluida. Todas as amostras desta sessao foram arquivadas.",
+                seconds=4.0,
+            )
+            return
+
+        self.idx = min(self.idx, len(self.samples) - 1)
+        self.load_current()
+
+    # -------------------------------------------------------------------------
+    # History
+    # -------------------------------------------------------------------------
+
+    def push_undo(self):
+        assert self.mask is not None
+
+        self.undo_stack.append(
+            self.mask.copy()
+        )
+
+        if len(self.undo_stack) > 40:
+            self.undo_stack.pop(0)
+
+        self.redo_stack.clear()
+
+    def undo(self):
+        if not self.undo_stack:
+            self.flash("Nada para desfazer.")
+            return
+
+        assert self.mask is not None
+
+        self.redo_stack.append(
+            self.mask.copy()
+        )
+        self.mask = self.undo_stack.pop()
+        self.mask_dirty = True
+        self.polygon_points = []
+
+    def redo(self):
+        if not self.redo_stack:
+            self.flash("Nada para refazer.")
+            return
+
+        assert self.mask is not None
+
+        self.undo_stack.append(
+            self.mask.copy()
+        )
+        self.mask = self.redo_stack.pop()
+        self.mask_dirty = True
+        self.polygon_points = []
+
+    def reset_non_nav(self):
+        assert self.mask is not None
+
+        self.push_undo()
+        self.mask.fill(
+            self.non_nav_id
+        )
+        self.mask_dirty = True
+        self.polygon_points = []
+
+        self.flash(
+            "Mascara resetada para 100% nao navegavel."
+        )
+
+    # -------------------------------------------------------------------------
+    # Tools
+    # -------------------------------------------------------------------------
+
+    def opposite_class(
+        self,
+        cid: int,
+    ) -> int:
+        return (
+            self.non_nav_id
+            if int(cid) == self.nav_id
+            else self.nav_id
+        )
+
+    def toggle_class(self):
+        self.selected_class = self.opposite_class(
+            self.selected_class
+        )
+        self.flash(
+            f"Classe de pintura: {self.classes[self.selected_class][0]}"
+        )
+
+    def set_tool(
+        self,
+        tool: str,
+    ):
+        self.tool = tool
+        self.polygon_points = []
+        self.mouse_down = False
+        self.brush_action_started = False
+        self.flash(f"Ferramenta: {tool}")
+
+    def commit_polygon(self):
+        if len(self.polygon_points) < 3:
+            self.flash(
+                "Poligono precisa de pelo menos 3 pontos."
+            )
+            return
+
+        assert self.mask is not None
+
+        self.push_undo()
+
+        poly = np.array(
+            self.polygon_points,
+            dtype=np.int32,
+        )
+
+        cv2.fillPoly(
+            self.mask,
+            [poly],
+            int(self.selected_class),
+        )
+
+        self.polygon_points = []
+        self.mask_dirty = True
+
+    def cancel_polygon(self):
+        self.polygon_points = []
+        self.flash(
+            "Poligono aberto cancelado."
+        )
+
+    def paint_brush(
+        self,
+        src_x: int,
+        src_y: int,
+        cid: int,
+    ):
+        assert self.mask is not None
+
+        cv2.circle(
+            self.mask,
+            (int(src_x), int(src_y)),
+            int(self.brush_radius),
+            int(cid),
+            -1,
+            cv2.LINE_8,
+        )
+
+        self.mask_dirty = True
+
+    def flood_fill(
+        self,
+        src_x: int,
+        src_y: int,
+    ):
+        assert self.mask is not None
+
+        y = int(src_y)
+        x = int(src_x)
+
+        old = int(self.mask[y, x])
+        new = int(self.selected_class)
+
+        if old == new:
+            self.flash(
+                "Balde: regiao ja possui essa classe."
+            )
+            return
+
+        self.push_undo()
+
+        # floodFill trabalha diretamente em uint8.
+        tmp = self.mask.copy()
+
+        cv2.floodFill(
+            tmp,
+            None,
+            (x, y),
+            newVal=new,
+            loDiff=0,
+            upDiff=0,
+            flags=4,
+        )
+
+        self.mask = tmp
+        self.mask_dirty = True
+
+    # -------------------------------------------------------------------------
+    # Mouse mapping
+    # -------------------------------------------------------------------------
+
+    def window_to_source(
+        self,
+        x: int,
+        y: int,
+    ) -> Optional[Tuple[int, int]]:
+        if self.image is None:
+            return None
+
+        v = self.view
+
+        if (
+            x < v.x0
+            or y < v.y0
+            or x >= v.x0 + v.width
+            or y >= v.y0 + v.height
+        ):
+            return None
+
+        sx = int(
+            round(
+                (x - v.x0)
+                / max(v.scale, 1e-8)
+            )
+        )
+        sy = int(
+            round(
+                (y - v.y0)
+                / max(v.scale, 1e-8)
+            )
+        )
+
+        h, w = self.image.shape[:2]
+
+        sx = int(
+            np.clip(
+                sx,
+                0,
+                w - 1,
+            )
+        )
+        sy = int(
+            np.clip(
+                sy,
+                0,
+                h - 1,
+            )
+        )
+
+        return sx, sy
+
+    def source_to_window(
+        self,
+        x: int,
+        y: int,
+    ) -> Tuple[int, int]:
+        v = self.view
+
+        return (
+            int(round(v.x0 + x * v.scale)),
+            int(round(v.y0 + y * v.scale)),
+        )
+
+    def on_mouse(
+        self,
+        event,
+        x,
+        y,
+        flags,
+        param,
+    ):
+        src = self.window_to_source(
+            int(x),
+            int(y),
+        )
+
+        if src is None:
+            if event in {
+                cv2.EVENT_LBUTTONUP,
+                cv2.EVENT_RBUTTONUP,
+            }:
+                self.mouse_down = False
+                self.mouse_button = None
+                self.brush_action_started = False
+            return
+
+        sx, sy = src
+        self.mouse_src = (
+            sx,
+            sy,
+        )
+
+        # -------------------------------------------------------------
+        # Polygon
+        # -------------------------------------------------------------
+
+        if self.tool == self.TOOL_POLYGON:
+            if event == cv2.EVENT_LBUTTONDOWN:
+                self.polygon_points.append(
+                    (sx, sy)
+                )
+
+            elif event == cv2.EVENT_RBUTTONDOWN:
+                if self.polygon_points:
+                    self.polygon_points.pop()
+
+            return
+
+        # -------------------------------------------------------------
+        # Fill
+        # -------------------------------------------------------------
+
+        if self.tool == self.TOOL_FILL:
+            if event == cv2.EVENT_LBUTTONDOWN:
+                self.flood_fill(
+                    sx,
+                    sy,
+                )
+            return
+
+        # -------------------------------------------------------------
+        # Brush
+        # -------------------------------------------------------------
+
+        if self.tool == self.TOOL_BRUSH:
+            if event in {
+                cv2.EVENT_LBUTTONDOWN,
+                cv2.EVENT_RBUTTONDOWN,
+            }:
+                self.mouse_down = True
+                self.mouse_button = event
+                self.brush_action_started = True
+
+                self.push_undo()
+
+                cid = (
+                    self.selected_class
+                    if event == cv2.EVENT_LBUTTONDOWN
+                    else self.opposite_class(
+                        self.selected_class
+                    )
+                )
+
+                self.paint_brush(
+                    sx,
+                    sy,
+                    cid,
+                )
+
+            elif (
+                event == cv2.EVENT_MOUSEMOVE
+                and self.mouse_down
+            ):
+                cid = (
+                    self.selected_class
+                    if self.mouse_button
+                    == cv2.EVENT_LBUTTONDOWN
+                    else self.opposite_class(
+                        self.selected_class
+                    )
+                )
+
+                self.paint_brush(
+                    sx,
+                    sy,
+                    cid,
+                )
+
+            elif event in {
+                cv2.EVENT_LBUTTONUP,
+                cv2.EVENT_RBUTTONUP,
+            }:
+                self.mouse_down = False
+                self.mouse_button = None
+                self.brush_action_started = False
+
+    # -------------------------------------------------------------------------
+    # Render
+    # -------------------------------------------------------------------------
+
+    def overlay_image(self) -> np.ndarray:
+        assert self.image is not None
+        assert self.mask is not None
+
+        if self.overlay_mode == self.OVERLAY_OFF:
+            return self.image.copy()
+
+        mask_color = ids_to_color(
+            self.mask,
+            self.classes,
+        )
+
+        blended = cv2.addWeighted(
+            self.image,
+            1.0 - self.overlay_alpha,
+            mask_color,
+            self.overlay_alpha,
+            0.0,
+        )
+
+        if self.overlay_mode == self.OVERLAY_BOTH:
+            return blended
+
+        # Default: mostra apenas navegavel.
+        out = self.image.copy()
+        nav_region = (
+            self.mask == self.nav_id
+        )
+        out[nav_region] = blended[nav_region]
+        return out
+
+    def build_canvas(self) -> np.ndarray:
+        assert self.image is not None
+        assert self.mask is not None
+
+        canvas = np.zeros(
+            (
+                MAX_WINDOW_H,
+                MAX_WINDOW_W,
+                3,
+            ),
+            dtype=np.uint8,
+        )
+
+        left_w = MAX_WINDOW_W - PANEL_W
+        image_area_h = (
+            MAX_WINDOW_H
+            - HEADER_H
+            - FOOTER_H
+        )
+
+        src_h, src_w = self.image.shape[:2]
+
+        scale = min(
+            left_w / float(src_w),
+            image_area_h / float(src_h),
+        )
+
+        disp_w = max(
+            1,
+            int(round(src_w * scale)),
+        )
+        disp_h = max(
+            1,
+            int(round(src_h * scale)),
+        )
+
+        x0 = max(
+            0,
+            (left_w - disp_w) // 2,
+        )
+        y0 = HEADER_H + max(
+            0,
+            (image_area_h - disp_h) // 2,
+        )
+
+        self.view = ViewTransform(
+            x0=x0,
+            y0=y0,
+            width=disp_w,
+            height=disp_h,
+            scale=scale,
+        )
+
+        overlay = self.overlay_image()
+
+        display = cv2.resize(
+            overlay,
+            (disp_w, disp_h),
+            interpolation=cv2.INTER_AREA,
+        )
+
+        canvas[
+            y0:y0 + disp_h,
+            x0:x0 + disp_w,
+        ] = display
+
+        # -------------------------------------------------------------
+        # Polygon in progress
+        # -------------------------------------------------------------
+
+        if self.polygon_points:
+            color = class_color_bgr(
+                self.classes,
+                self.selected_class,
+            )
+
+            pts_win = np.array(
+                [
+                    self.source_to_window(
+                        px,
+                        py,
+                    )
+                    for px, py in self.polygon_points
+                ],
+                dtype=np.int32,
+            )
+
+            for px, py in pts_win:
+                cv2.circle(
+                    canvas,
+                    (int(px), int(py)),
+                    4,
+                    color,
+                    -1,
+                    cv2.LINE_AA,
+                )
+                cv2.circle(
+                    canvas,
+                    (int(px), int(py)),
+                    6,
+                    (255, 255, 255),
+                    1,
+                    cv2.LINE_AA,
+                )
+
+            if len(pts_win) >= 2:
+                cv2.polylines(
+                    canvas,
+                    [pts_win],
+                    isClosed=False,
+                    color=color,
+                    thickness=2,
+                    lineType=cv2.LINE_AA,
+                )
+
+        # -------------------------------------------------------------
+        # Brush cursor
+        # -------------------------------------------------------------
+
+        if self.tool == self.TOOL_BRUSH:
+            mouse_win = self.source_to_window(
+                self.mouse_src[0],
+                self.mouse_src[1],
+            )
+
+            rr = max(
+                2,
+                int(
+                    round(
+                        self.brush_radius
+                        * self.view.scale
+                    )
+                ),
+            )
+
+            cv2.circle(
+                canvas,
+                mouse_win,
+                rr,
+                (255, 255, 255),
+                1,
+                cv2.LINE_AA,
+            )
+
+        # -------------------------------------------------------------
+        # Header
+        # -------------------------------------------------------------
+
+        item = self.samples[self.idx]
+        complete = self.sample_complete(
+            item,
+        )
+
+        cv2.putText(
+            canvas,
+            (
+                f"[{self.idx + 1}/{len(self.samples)}] "
+                f"{item.image_path.name}"
+            ),
+            (16, 28),
+            cv2.FONT_HERSHEY_SIMPLEX,
+            0.66,
+            (245, 245, 245),
+            2,
+            cv2.LINE_AA,
+        )
+
+        cv2.putText(
+            canvas,
+            (
+                f"{'PRONTO' if complete else 'INCOMPLETO'} | "
+                f"mask={'editada' if self.mask_dirty else 'ok'} | "
+                f"label={'editado' if self.label_dirty else 'ok'}"
+            ),
+            (16, 56),
+            cv2.FONT_HERSHEY_SIMPLEX,
+            0.50,
+            (
+                (70, 220, 80)
+                if complete
+                else (70, 180, 255)
+            ),
+            1,
+            cv2.LINE_AA,
+        )
+
+        # -------------------------------------------------------------
+        # Right panel
+        # -------------------------------------------------------------
+
+        px = MAX_WINDOW_W - PANEL_W + 18
+
+        cv2.putText(
+            canvas,
+            "MASCARA",
+            (px, 38),
+            cv2.FONT_HERSHEY_SIMPLEX,
+            0.72,
+            (255, 255, 255),
+            2,
+            cv2.LINE_AA,
+        )
+
+        cname = self.classes[
+            self.selected_class
+        ][0]
+        ccolor = class_color_bgr(
+            self.classes,
+            self.selected_class,
+        )
+
+        cv2.rectangle(
+            canvas,
+            (px, 55),
+            (px + 28, 83),
+            ccolor,
+            -1,
+        )
+        cv2.rectangle(
+            canvas,
+            (px, 55),
+            (px + 28, 83),
+            (255, 255, 255),
+            1,
+        )
+
+        cv2.putText(
+            canvas,
+            f"Classe: {self.selected_class} {cname}",
+            (px + 40, 78),
+            cv2.FONT_HERSHEY_SIMPLEX,
+            0.51,
+            (225, 225, 225),
+            1,
+            cv2.LINE_AA,
+        )
+
+        cv2.putText(
+            canvas,
+            f"Ferramenta: {self.tool}",
+            (px, 112),
+            cv2.FONT_HERSHEY_SIMPLEX,
+            0.52,
+            (220, 220, 220),
+            1,
+            cv2.LINE_AA,
+        )
+
+        nav_pct = float(
+            (self.mask == self.nav_id).mean()
+            * 100.0
+        )
+
+        cv2.putText(
+            canvas,
+            f"Navegavel: {nav_pct:.2f}%",
+            (px, 142),
+            cv2.FONT_HERSHEY_SIMPLEX,
+            0.52,
+            (220, 220, 220),
+            1,
+            cv2.LINE_AA,
+        )
+
+        try:
+            group_preview = self.derive_group()
+        except Exception:
+            group_preview = "?"
+
+        cv2.putText(
+            canvas,
+            f"Grupo: {group_preview}",
+            (px, 172),
+            cv2.FONT_HERSHEY_SIMPLEX,
+            0.48,
+            (210, 235, 255),
+            1,
+            cv2.LINE_AA,
+        )
+
+        cv2.putText(
+            canvas,
+            f"Pincel: {self.brush_radius}px",
+            (px, 198),
+            cv2.FONT_HERSHEY_SIMPLEX,
+            0.52,
+            (220, 220, 220),
+            1,
+            cv2.LINE_AA,
+        )
+
+        overlay_names = {
+            self.OVERLAY_NAV_ONLY: "so navegavel",
+            self.OVERLAY_BOTH: "duas classes",
+            self.OVERLAY_OFF: "desligado",
+        }
+
+        cv2.putText(
+            canvas,
+            f"Overlay: {overlay_names[self.overlay_mode]}",
+            (px, 226),
+            cv2.FONT_HERSHEY_SIMPLEX,
+            0.52,
+            (220, 220, 220),
+            1,
+            cv2.LINE_AA,
+        )
+
+        cv2.putText(
+            canvas,
+            "P poligono | B pincel | F balde",
+            (px, 266),
+            cv2.FONT_HERSHEY_SIMPLEX,
+            0.47,
+            (190, 220, 255),
+            1,
+            cv2.LINE_AA,
+        )
+
+        cv2.putText(
+            canvas,
+            "C classe | [ ] pincel | O overlay",
+            (px, 294),
+            cv2.FONT_HERSHEY_SIMPLEX,
+            0.47,
+            (190, 220, 255),
+            1,
+            cv2.LINE_AA,
+        )
+
+        cv2.putText(
+            canvas,
+            "Z undo | Y redo | R reset",
+            (px, 322),
+            cv2.FONT_HERSHEY_SIMPLEX,
+            0.47,
+            (190, 220, 255),
+            1,
+            cv2.LINE_AA,
+        )
+
+        # Status labels.
+        cv2.putText(
+            canvas,
+            "ESTADO DO CORREDOR",
+            (px, 374),
+            cv2.FONT_HERSHEY_SIMPLEX,
+            0.70,
+            (255, 255, 255),
+            2,
+            cv2.LINE_AA,
+        )
+
+        y = 410
+
+        for i, state in enumerate(
+            self.states
+        ):
+            key = (
+                str(i + 1)
+                if i < 9
+                else "0"
+            )
+
+            selected = (
+                self.selected_state == i
+            )
+
+            color = (
+                (80, 255, 120)
+                if selected
+                else (205, 205, 205)
+            )
+
+            prefix = ">" if selected else " "
+
+            cv2.putText(
+                canvas,
+                f"{prefix} [{key}] {state}",
+                (px, y),
+                cv2.FONT_HERSHEY_SIMPLEX,
+                0.49,
+                color,
+                2 if selected else 1,
+                cv2.LINE_AA,
+            )
+
+            y += 28
+
+        if self.selected_state is None:
+            state_txt = "NAO SELECIONADO"
+            state_color = (
+                80,
+                180,
+                255,
+            )
+        else:
+            state_txt = self.states[
+                self.selected_state
+            ]
+            state_color = (
+                80,
+                255,
+                120,
+            )
+
+        cv2.putText(
+            canvas,
+            f"Atual: {state_txt}",
+            (px, y + 15),
+            cv2.FONT_HERSHEY_SIMPLEX,
+            0.52,
+            state_color,
+            2,
+            cv2.LINE_AA,
+        )
+
+        done = self.completed_count()
+
+        cv2.putText(
+            canvas,
+            f"Concluidas: {done}/{len(self.samples)}",
+            (px, min(MAX_WINDOW_H - 122, y + 58)),
+            cv2.FONT_HERSHEY_SIMPLEX,
+            0.50,
+            (220, 220, 220),
+            1,
+            cv2.LINE_AA,
+        )
+
+        # -------------------------------------------------------------
+        # Footer
+        # -------------------------------------------------------------
+
+        footer_y = MAX_WINDOW_H - FOOTER_H + 28
+
+        cv2.putText(
+            canvas,
+            (
+                "S/SPACE=salvar+proxima | A/D ou setas=navegar | "
+                "ENTER=fechar poligono | BACKSPACE=cancelar poligono | Q=sair"
+            ),
+            (16, footer_y),
+            cv2.FONT_HERSHEY_SIMPLEX,
+            0.49,
+            (225, 225, 225),
+            1,
+            cv2.LINE_AA,
+        )
+
+        cv2.putText(
+            canvas,
+            (
+                "Pincel: esquerdo=classe selecionada, direito=classe oposta | "
+                "Mascara bruta salva colorida conforme labelmap no tamanho original"
+            ),
+            (16, footer_y + 30),
+            cv2.FONT_HERSHEY_SIMPLEX,
+            0.47,
+            (165, 205, 255),
+            1,
+            cv2.LINE_AA,
+        )
+
+        if (
+            self.message
+            and time.monotonic()
+            < self.message_until
+        ):
+            cv2.putText(
+                canvas,
+                self.message[:150],
+                (16, footer_y + 60),
+                cv2.FONT_HERSHEY_SIMPLEX,
+                0.49,
+                (80, 255, 160),
+                2,
+                cv2.LINE_AA,
+            )
+
+        return canvas
+
+    # -------------------------------------------------------------------------
+    # Labels / messages
+    # -------------------------------------------------------------------------
+
+    def select_state(
+        self,
+        idx: int,
+    ):
+        if not (
+            0 <= idx < len(self.states)
+        ):
+            return
+
+        self.selected_state = int(idx)
+
+        self.label_dirty = (
+            self.loaded_state
+            != self.selected_state
+        )
+
+        self.flash(
+            f"Estado: {self.states[idx]}"
+        )
+
+    def flash(
+        self,
+        message: str,
+        seconds: float = 2.0,
+    ):
+        self.message = str(message)
+        self.message_until = (
+            time.monotonic()
+            + float(seconds)
+        )
+        print(f"[INFO] {message}")
+
+    # -------------------------------------------------------------------------
+    # Keys / run
+    # -------------------------------------------------------------------------
+
+    def on_key(
+        self,
+        k: int,
+    ) -> bool:
+        # Quit.
+        if k in {
+            27,
+            ord("q"),
+            ord("Q"),
+        }:
+            if (
+                self.mask_dirty
+                or self.label_dirty
+                or self.polygon_points
+            ):
+                self.flash(
+                    "Ha alteracoes nao salvas. "
+                    "Pressione Q novamente em ate 2s para sair mesmo assim."
+                )
+
+                now = time.monotonic()
+
+                last_q = getattr(
+                    self,
+                    "_last_q",
+                    0.0,
+                )
+
+                if (
+                    last_q > 0
+                    and now - last_q <= 2.0
+                ):
+                    return False
+
+                self._last_q = now
+                return True
+
+            return False
+
+        self._last_q = 0.0
+
+        # Save.
+        if k in {
+            ord("s"),
+            ord("S"),
+            ord(" "),
+        }:
+            self.save_and_next()
+            return True
+
+        # Tools.
+        if k in {
+            ord("p"),
+            ord("P"),
+        }:
+            self.set_tool(
+                self.TOOL_POLYGON
+            )
+            return True
+
+        if k in {
+            ord("b"),
+            ord("B"),
+        }:
+            self.set_tool(
+                self.TOOL_BRUSH
+            )
+            return True
+
+        if k in {
+            ord("f"),
+            ord("F"),
+        }:
+            self.set_tool(
+                self.TOOL_FILL
+            )
+            return True
+
+        if k in {
+            ord("c"),
+            ord("C"),
+        }:
+            self.toggle_class()
+            return True
+
+        if k == ord("["):
+            self.brush_radius = max(
+                2,
+                self.brush_radius - 4,
+            )
+            self.flash(
+                f"Pincel={self.brush_radius}px"
+            )
+            return True
+
+        if k == ord("]"):
+            self.brush_radius = min(
+                300,
+                self.brush_radius + 4,
+            )
+            self.flash(
+                f"Pincel={self.brush_radius}px"
+            )
+            return True
+
+        if k in {
+            ord("o"),
+            ord("O"),
+        }:
+            self.overlay_mode = (
+                self.overlay_mode + 1
+            ) % 3
+            return True
+
+        if k in {
+            ord("z"),
+            ord("Z"),
+        }:
+            self.undo()
+            return True
+
+        if k in {
+            ord("y"),
+            ord("Y"),
+        }:
+            self.redo()
+            return True
+
+        if k in {
+            ord("r"),
+            ord("R"),
+        }:
+            self.reset_non_nav()
+            return True
+
+        # Polygon commit.
+        if k in {
+            10,
+            13,
+        }:
+            if (
+                self.tool
+                == self.TOOL_POLYGON
+            ):
+                self.commit_polygon()
+            return True
+
+        # Backspace / Delete.
+        if k in {
+            8,
+            127,
+            3014656,
+        }:
+            if self.polygon_points:
+                self.cancel_polygon()
+            return True
+
+        # Global labels.
+        if (
+            ord("1")
+            <= k
+            <= ord("9")
+        ):
+            idx = int(chr(k)) - 1
+            self.select_state(idx)
+            return True
+
+        if k == ord("0"):
+            self.select_state(9)
+            return True
+
+        # Navigation ASCII.
+        if k in {
+            ord("a"),
+            ord("A"),
+        }:
+            self.navigate(-1)
+            return True
+
+        if k in {
+            ord("d"),
+            ord("D"),
+        }:
+            self.navigate(+1)
+            return True
+
+        # OpenCV Windows arrows via waitKeyEx.
+        # Common values:
+        # left=2424832, right=2555904
+        if k == 2424832:
+            self.navigate(-1)
+            return True
+
+        if k == 2555904:
+            self.navigate(+1)
+            return True
+
+        return True
+
+    def run(self):
+        try:
+            while not self.finished:
+                canvas = self.build_canvas()
+
+                cv2.imshow(
+                    WINDOW_NAME,
+                    canvas,
+                )
+
+                k = cv2.waitKeyEx(20)
+
+                if k == -1:
+                    continue
+
+                keep = self.on_key(
+                    int(k),
+                )
+
+                if not keep:
+                    break
+
+        finally:
+            cv2.destroyAllWindows()
+
+
+# =============================================================================
+# Main
+# =============================================================================
+
+def find_archived_bases(original_group_root: Path) -> set[str]:
+    out: set[str] = set()
+    if not original_group_root.is_dir():
+        return out
+
+    for group_dir in original_group_root.iterdir():
+        if not group_dir.is_dir():
+            continue
+        image_dir = group_dir / "images"
+        if not image_dir.is_dir():
+            continue
+        for path in image_dir.iterdir():
+            if path.is_file() and path.suffix.lower() in IMG_EXTS:
+                out.add(path.stem)
+    return out
+
+
+def collect_samples(
+    brutas_dir: Path,
+    original_group_root: Path,
+    skip_archived: bool = True,
+) -> Tuple[List[SampleItem], int]:
+    images = list_images(brutas_dir)
+    archived = find_archived_bases(original_group_root)
+
+    samples: List[SampleItem] = []
+    for img in images:
+        if skip_archived and img.stem in archived:
+            continue
+
+        # Paths pendentes só existem para manter compatibilidade interna do editor.
+        # O destino real é resolvido no SAVE depois que o grupo é derivado da máscara.
+        samples.append(
+            SampleItem(
+                image_path=img,
+                mask_path=Path("__pending__") / "masks" / f"{img.stem}.png",
+                label_path=Path("__pending__") / "labels" / f"{img.stem}.json",
+                base=img.stem,
+            )
+        )
+
+    return samples, len(archived)
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description=(
+            "Editor integrado de máscara + estado. Consome dataset/brutas/ "
+            "e arquiva em dataset/original/group/<grupo>/."
+        )
+    )
+
+    parser.add_argument("--config", default=str(DEFAULT_CONFIG))
+    parser.add_argument("--brutas", default=str(DEFAULT_BRUTAS_DIR))
+    parser.add_argument(
+        "--original-root",
+        default=str(DEFAULT_ORIGINAL_GROUP_ROOT),
+        help="Raiz dos grupos finais originais.",
+    )
+    parser.add_argument("--labelmap", default=str(DEFAULT_LABELMAP))
+    parser.add_argument(
+        "--mode",
+        choices=["move", "copy"],
+        default="move",
+        help="move=remove de brutas após commit; copy=preserva a PNG bruta.",
+    )
+    parser.add_argument(
+        "--include-archived",
+        action="store_true",
+        help="Também mostra bases que já existem em original/group (normalmente não use).",
+    )
+    parser.add_argument("--start-index", type=int, default=None)
+    parser.add_argument("--alpha", type=float, default=DEFAULT_OVERLAY_ALPHA)
+    parser.add_argument("--brush-radius", type=int, default=DEFAULT_BRUSH_RADIUS)
+
+    args = parser.parse_args()
+
+    config_path = Path(args.config)
+    if not config_path.is_file():
+        raise FileNotFoundError(f"config.json nao encontrado: {config_path}")
+
+    with config_path.open("r", encoding="utf-8") as f:
+        config = json.load(f)
+
+    states = config.get("label_classes")
+    if not isinstance(states, list) or not states:
+        raise RuntimeError("config['label_classes'] obrigatorio.")
+    if len(states) > 10:
+        raise RuntimeError("Editor suporta ate 10 estados nas teclas 1..9/0.")
+
+    main_class_name = str(config.get("main_class_name", "navegavel"))
+
+    brutas_dir = Path(args.brutas)
+    original_group_root = Path(args.original_root)
+    labelmap_path = Path(args.labelmap)
+
+    classes = load_labelmap_classes(labelmap_path)
+    nav_id, non_nav_id = resolve_nav_classes(classes, main_class_name)
+
+    samples, archived_before = collect_samples(
+        brutas_dir,
+        original_group_root,
+        skip_archived=not args.include_archived,
+    )
+
+    if not samples:
+        print("[INFO] Nenhuma PNG pendente em brutas. Fila vazia.")
+        return
+
+    dataset_root = Path("dataset")
+
+    print("=" * 78)
+    print("Agrobot | Editor de Corredor OAK-D Lite")
+    print("=" * 78)
+    print(f"Fila brutas       : {brutas_dir.resolve()}")
+    print(f"Destino original  : {original_group_root.resolve()}")
+    print(f"Modo              : {args.mode.upper()}")
+    print(f"Pendentes         : {len(samples)}")
+    print(f"Já arquivadas     : {archived_before}")
+    print(f"Labelmap          : {labelmap_path.resolve()}")
+    print(f"Navegavel         : id={nav_id} name={classes[nav_id][0]}")
+    print(f"Nao navegavel     : id={non_nav_id} name={classes[non_nav_id][0]}")
+    print("Grupos automáticos: navegavel | naonavegavel | naonavegavel_navegavel")
+    print("Estados:")
+    for i, state in enumerate(states):
+        key = str(i + 1) if i < 9 else "0"
+        print(f"  [{key}] {state}")
+    print("=" * 78)
+
+    editor = CorridorAnnotationEditor(
+        dataset_root=dataset_root,
+        samples=samples,
+        states=states,
+        classes=classes,
+        nav_id=nav_id,
+        non_nav_id=non_nav_id,
+        original_group_root=original_group_root,
+        archive_mode=args.mode,
+        archived_before=archived_before,
+        overlay_alpha=args.alpha,
+        brush_radius=args.brush_radius,
+        start_index=args.start_index,
+    )
+
+    editor.run()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/Python/OAK/datasets/oak-d/_2_normalize_corridor.py b/Python/OAK/datasets/oak-d/_2_normalize_corridor.py
new file mode 100644
index 000000000..86eb8ebc6
--- /dev/null
+++ b/Python/OAK/datasets/oak-d/_2_normalize_corridor.py
@@ -0,0 +1,1938 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+"""
+_2_normalize_corridor.py
+========================
+
+Normaliza o dataset ORIGINAL agrupado da OAK-D Lite para a resolução do modelo.
+
+Executar de dentro da pasta oak-d:
+
+    python _2_normalize_corridor.py
+
+Entrada:
+    dataset/original/group/
+      navegavel/
+        images/<base>.png
+        masks/<base>.png       # RGB colorida conforme labelmap
+        labels/<base>.json
+      naonavegavel/
+        images/
+        masks/
+        labels/
+      naonavegavel_navegavel/
+        images/
+        masks/
+        labels/
+
+Saída, preservando grupos:
+    dataset/<W>x<H>/group/
+      navegavel/{images,masks,labels}
+      naonavegavel/{images,masks,labels}
+      naonavegavel_navegavel/{images,masks,labels}
+      normalize_report.json
+
+Contrato da saída:
+    images: PNG RGB/BGR convencional, somente resize INTER_AREA
+    masks : PNG 1 canal uint8 com IDs de classe, resize INTER_NEAREST
+    labels: JSON global atualizado e rastreável
+
+Este script NÃO:
+    - seleciona amostras;
+    - faz augmentation offline;
+    - calcula mean/std;
+    - cria train/val.
+
+Pareamento:
+    - dados novos: image/mask/label usam o mesmo stem;
+    - legado: sufixos como _Rgb, _mask, _label são reconciliados;
+    - apenas tripletas completas entram na saída;
+    - incompletos não abortam por padrão: são registrados no normalize_report;
+    - poucos pixels RGB fora do labelmap tentam reparo conservador;
+    - falha de conteúdo vai para dataset/normalize_quarantine e o lote continua;
+    - por padrão invalida split/norm_stats/report/manifest antigos da mesma resolução.
+
+mean/std serão calculados DEPOIS do split, somente sobre train.
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import re
+import shutil
+from collections import Counter, defaultdict
+from dataclasses import dataclass
+from datetime import datetime
+from pathlib import Path
+from typing import Dict, List, Optional, Sequence, Tuple
+
+import cv2
+import numpy as np
+
+
+# =============================================================================
+# Defaults
+# =============================================================================
+
+CONFIG_PATH = Path("config.json")
+DATASET_ROOT = Path("dataset")
+DEFAULT_ORIGINAL_GROUP_ROOT = DATASET_ROOT / "original" / "group"
+DEFAULT_LABELMAP = DATASET_ROOT / "labelmap.txt"
+DEFAULT_QUARANTINE_ROOT = DATASET_ROOT / "normalize_quarantine"
+
+IMG_EXTS = {".png", ".jpg", ".jpeg", ".bmp", ".webp"}
+MASK_EXTS = {".png", ".jpg", ".jpeg", ".bmp", ".webp"}
+LABEL_EXTS = {".json"}
+
+VALID_GROUPS = {
+    "navegavel",
+    "naonavegavel",
+    "naonavegavel_navegavel",
+}
+
+REPORT_SCHEMA = "agrobot_corridor_normalize_grouped_v2"
+NORMALIZED_LABEL_SCHEMA = "agrobot_corridor_state_normalized_v2"
+
+FALLBACK_COLORS_RGB = {
+    0: (210, 55, 55),
+    1: (45, 215, 70),
+}
+
+
+# =============================================================================
+# Data
+# =============================================================================
+
+@dataclass(frozen=True)
+class Sample:
+    group: str
+    base: str
+    image: Path
+    mask: Path
+    label: Path
+
+
+# =============================================================================
+# Generic helpers
+# =============================================================================
+
+def now_iso() -> str:
+    return datetime.now().isoformat(timespec="seconds")
+
+
+def natural_key(text: str):
+    parts = re.split(r"(\d+)", str(text))
+    return [int(p) if p.isdigit() else p.lower() for p in parts]
+
+
+def safe_rel(path: Path, root: Path) -> str:
+    try:
+        return str(path.resolve().relative_to(root.resolve())).replace("\\", "/")
+    except Exception:
+        return str(path).replace("\\", "/")
+
+
+def atomic_write_json(path: Path, data: dict) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    tmp = path.with_suffix(path.suffix + ".tmp")
+    with tmp.open("w", encoding="utf-8") as f:
+        json.dump(data, f, ensure_ascii=False, indent=2)
+    os.replace(tmp, path)
+
+
+def atomic_write_png(path: Path, arr: np.ndarray) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    tmp = path.with_name(path.stem + ".__tmp__.png")
+    ok = cv2.imwrite(str(tmp), arr, [cv2.IMWRITE_PNG_COMPRESSION, 3])
+    if not ok:
+        try:
+            tmp.unlink(missing_ok=True)
+        except Exception:
+            pass
+        raise IOError(f"cv2.imwrite retornou False: {tmp}")
+    os.replace(tmp, path)
+
+
+def map_by_stem(folder: Path, exts: set[str]) -> Dict[str, Path]:
+    if not folder.is_dir():
+        return {}
+
+    result: Dict[str, Path] = {}
+    for path in sorted(folder.iterdir(), key=lambda p: natural_key(p.name)):
+        if not path.is_file() or path.suffix.lower() not in exts:
+            continue
+        if path.stem in result:
+            raise RuntimeError(
+                f"Stem duplicado '{path.stem}' em {folder}: "
+                f"{result[path.stem].name} e {path.name}"
+            )
+        result[path.stem] = path
+    return result
+
+
+LEGACY_BASE_SUFFIXES = (
+    "_rgb",
+    "_image",
+    "_img",
+    "_frame",
+    "_mask",
+    "_masks",
+    "_seg",
+    "_segment",
+    "_segmentacao",
+    "_label",
+    "_labels",
+)
+
+
+def canonical_base(stem: str) -> str:
+    """
+    Normaliza nomes históricos para casar a mesma amostra.
+
+    Exemplos:
+        500_Rgb.png   -> 500
+        500_mask.png  -> 500
+        500_label.json -> 500
+
+    O novo pipeline já produz o mesmo stem nas três pastas, então para dados
+    novos esta função é essencialmente identidade.
+    """
+    out = str(stem)
+
+    changed = True
+    while changed:
+        changed = False
+        low = out.lower()
+
+        for suffix in LEGACY_BASE_SUFFIXES:
+            if low.endswith(suffix) and len(out) > len(suffix):
+                out = out[:-len(suffix)]
+                changed = True
+                break
+
+    return out
+
+
+def map_by_canonical_base(
+    folder: Path,
+    exts: set[str],
+) -> Tuple[Dict[str, Path], List[dict], List[dict]]:
+    """
+    Retorna:
+        canonical -> arquivo
+        aliases    -> casos em que stem original != canonical
+        collisions -> dois arquivos da MESMA pasta colidiram no mesmo canonical
+
+    Colisão não é resolvida silenciosamente. O canonical fica fora do mapa para
+    impedir que o normalize escolha uma versão arbitrária.
+    """
+    if not folder.is_dir():
+        return {}, [], []
+
+    buckets: Dict[str, List[Path]] = {}
+
+    for p in sorted(folder.iterdir(), key=lambda x: natural_key(x.name)):
+        if not p.is_file() or p.suffix.lower() not in exts:
+            continue
+
+        base = canonical_base(p.stem)
+        buckets.setdefault(base, []).append(p)
+
+    result: Dict[str, Path] = {}
+    aliases: List[dict] = []
+    collisions: List[dict] = []
+
+    for base, candidates in sorted(
+        buckets.items(),
+        key=lambda kv: natural_key(kv[0]),
+    ):
+        if len(candidates) != 1:
+            collisions.append({
+                "base": base,
+                "files": [p.name for p in candidates],
+                "folder": str(folder),
+            })
+            continue
+
+        p = candidates[0]
+        result[base] = p
+
+        if p.stem != base:
+            aliases.append({
+                "base": base,
+                "original_stem": p.stem,
+                "file": p.name,
+                "folder": str(folder),
+            })
+
+    return result, aliases, collisions
+
+
+# =============================================================================
+# Labelmap
+# =============================================================================
+
+def load_labelmap_classes(
+    path: Path,
+) -> Dict[int, Tuple[str, Tuple[int, int, int]]]:
+    if not path.is_file():
+        raise FileNotFoundError(f"labelmap não encontrado: {path}")
+
+    classes: Dict[int, Tuple[str, Tuple[int, int, int]]] = {}
+    next_id = 0
+
+    with path.open("r", encoding="utf-8") as f:
+        for raw in f:
+            s = raw.strip()
+            if not s or s.startswith("#"):
+                continue
+
+            cid: Optional[int] = None
+            name: Optional[str] = None
+            color: Optional[Tuple[int, int, int]] = None
+            left = s.split(":", 1)[0].strip()
+
+            if ":" in s and not left.isdigit():
+                name_part, rest = s.split(":", 1)
+                name = name_part.strip()
+                color_text = rest.split("::", 1)[0].strip().strip(":")
+                parts = [p.strip() for p in color_text.split(",") if p.strip()]
+                if len(parts) >= 3:
+                    try:
+                        color = tuple(int(float(x)) for x in parts[:3])
+                    except Exception:
+                        color = None
+            else:
+                parts = s.replace(",", " ").replace(":", " ").split()
+                if len(parts) >= 2 and parts[0].isdigit():
+                    cid = int(parts[0])
+                    name = parts[1]
+                    if len(parts) >= 5:
+                        try:
+                            color = tuple(int(float(x)) for x in parts[2:5])
+                        except Exception:
+                            color = None
+                elif parts:
+                    name = parts[0]
+
+            if not name:
+                continue
+
+            if name.strip().lower() in {"ignore", "void", "background_ignore"}:
+                continue
+
+            if cid is None:
+                cid = next_id
+            next_id = max(next_id, cid + 1)
+
+            if color is None:
+                color = FALLBACK_COLORS_RGB.get(cid, (255, 255, 255))
+
+            classes[int(cid)] = (str(name), tuple(map(int, color)))
+
+    if len(classes) != 2:
+        raise RuntimeError(
+            "Normalize especializado em duas classes. "
+            f"labelmap carregado={classes}"
+        )
+
+    return classes
+
+
+def class_color_bgr(classes, cid: int) -> Tuple[int, int, int]:
+    _name, rgb = classes[int(cid)]
+    return int(rgb[2]), int(rgb[1]), int(rgb[0])
+
+
+def resolve_nav_classes(classes, main_class_name: str) -> Tuple[int, int]:
+    target = str(main_class_name).strip().lower()
+
+    nav = [
+        int(cid)
+        for cid, (name, _rgb) in classes.items()
+        if str(name).strip().lower() == target
+    ]
+
+    if len(nav) != 1:
+        aliases = {"navegavel", "navegável", "nav", "navigable"}
+        nav = [
+            int(cid)
+            for cid, (name, _rgb) in classes.items()
+            if str(name).strip().lower() in aliases
+        ]
+
+    if len(nav) != 1:
+        raise RuntimeError(
+            f"Não consegui resolver navegável. main_class_name={main_class_name!r} "
+            f"classes={classes}"
+        )
+
+    nav_id = nav[0]
+    other = [int(cid) for cid in classes if int(cid) != nav_id]
+    if len(other) != 1:
+        raise RuntimeError(f"Classe não navegável ambígua: {classes}")
+    return nav_id, other[0]
+
+
+# =============================================================================
+# Mask contract
+# =============================================================================
+
+def _exact_color_decode(
+    mask_bgr: np.ndarray,
+    classes,
+) -> Tuple[np.ndarray, np.ndarray]:
+    """Converte somente cores EXATAS do labelmap. Retorna ids + unknown mask."""
+    h, w = mask_bgr.shape[:2]
+    ids = np.full((h, w), 255, dtype=np.uint8)
+
+    for cid in sorted(classes):
+        bgr = np.array(
+            class_color_bgr(classes, cid),
+            dtype=np.uint8,
+        )
+        match = np.all(
+            mask_bgr[:, :, :3] == bgr[None, None, :],
+            axis=2,
+        )
+        ids[match] = int(cid)
+
+    return ids, ids == 255
+
+
+def _repair_unknown_spatial(
+    ids: np.ndarray,
+    classes,
+    max_iters: int = 6,
+) -> Tuple[np.ndarray, int]:
+    """
+    Recupera pixels desconhecidos quando a vizinhança é fortemente consistente.
+
+    Não usa interpolação na máscara. O pixel só herda uma classe se houver
+    maioria local clara entre os 8 vizinhos já conhecidos.
+    """
+    out = ids.copy()
+    repaired = 0
+    class_ids = [int(cid) for cid in sorted(classes)]
+
+    for _ in range(max(1, int(max_iters))):
+        unknown = out == 255
+        if not np.any(unknown):
+            break
+
+        padded = np.pad(
+            out,
+            1,
+            mode="constant",
+            constant_values=255,
+        )
+
+        neighbors = [
+            padded[0:-2, 0:-2],
+            padded[0:-2, 1:-1],
+            padded[0:-2, 2:],
+            padded[1:-1, 0:-2],
+            padded[1:-1, 2:],
+            padded[2:, 0:-2],
+            padded[2:, 1:-1],
+            padded[2:, 2:],
+        ]
+
+        votes = np.stack(
+            [
+                sum((n == cid).astype(np.uint8) for n in neighbors)
+                for cid in class_ids
+            ],
+            axis=0,
+        )
+
+        order = np.argsort(votes, axis=0)
+        top_idx = order[-1]
+        top_votes = np.take_along_axis(
+            votes,
+            top_idx[None, ...],
+            axis=0,
+        )[0]
+
+        if len(class_ids) > 1:
+            second_idx = order[-2]
+            second_votes = np.take_along_axis(
+                votes,
+                second_idx[None, ...],
+                axis=0,
+            )[0]
+        else:
+            second_votes = np.zeros_like(top_votes)
+
+        margin = top_votes.astype(np.int16) - second_votes.astype(np.int16)
+
+        # Conservador:
+        # pelo menos 4 vizinhos concordam e há margem >= 3 sobre a outra classe.
+        accept = (
+            unknown
+            & (top_votes >= 4)
+            & (margin >= 3)
+        )
+
+        if not np.any(accept):
+            break
+
+        top_class = np.take(
+            np.asarray(class_ids, dtype=np.uint8),
+            top_idx,
+        )
+
+        out[accept] = top_class[accept]
+        repaired += int(accept.sum())
+
+    return out, repaired
+
+
+def _repair_unknown_color(
+    ids: np.ndarray,
+    mask_bgr: np.ndarray,
+    classes,
+    max_distance: float = 55.0,
+    min_margin: float = 20.0,
+) -> Tuple[np.ndarray, int]:
+    """
+    Segunda chance por distância de cor, mas somente quando a melhor classe
+    está claramente mais próxima que a segunda.
+
+    Isso NÃO é nearest-color irrestrito.
+    """
+    out = ids.copy()
+    unknown = out == 255
+
+    if not np.any(unknown):
+        return out, 0
+
+    class_ids = [int(cid) for cid in sorted(classes)]
+    colors = np.asarray(
+        [
+            class_color_bgr(classes, cid)
+            for cid in class_ids
+        ],
+        dtype=np.float32,
+    )
+
+    pixels = mask_bgr[:, :, :3][unknown].astype(np.float32)
+
+    # [N, K]
+    dist = np.linalg.norm(
+        pixels[:, None, :] - colors[None, :, :],
+        axis=2,
+    )
+
+    nearest_idx = np.argmin(dist, axis=1)
+    nearest_dist = dist[
+        np.arange(dist.shape[0]),
+        nearest_idx,
+    ]
+
+    if dist.shape[1] >= 2:
+        sorted_dist = np.sort(dist, axis=1)
+        margin = sorted_dist[:, 1] - sorted_dist[:, 0]
+    else:
+        margin = np.full_like(
+            nearest_dist,
+            np.inf,
+        )
+
+    accept = (
+        (nearest_dist <= float(max_distance))
+        & (margin >= float(min_margin))
+    )
+
+    if not np.any(accept):
+        return out, 0
+
+    unknown_yx = np.argwhere(unknown)
+    accepted_yx = unknown_yx[accept]
+    accepted_ids = np.asarray(
+        class_ids,
+        dtype=np.uint8,
+    )[nearest_idx[accept]]
+
+    out[
+        accepted_yx[:, 0],
+        accepted_yx[:, 1],
+    ] = accepted_ids
+
+    return out, int(accept.sum())
+
+
+def _repair_unknown_nearest_class(
+    ids: np.ndarray,
+    classes,
+    non_nav_id: Optional[int] = None,
+) -> Tuple[np.ndarray, int, int]:
+    """
+    Fallback geométrico para máscaras categóricas legadas.
+
+    Depois das tentativas conservadoras, pixels ainda desconhecidos são
+    atribuídos à classe VÁLIDA espacialmente mais próxima.
+
+    Caso duas classes estejam à mesma distância, o empate vai para
+    `non_nav_id`, quando disponível. Isso é deliberadamente conservador:
+    uma fronteira antiga/anti-aliased de 1 px não deve virar falso navegável.
+
+    Retorna:
+        ids_reparados
+        quantidade_reparada
+        quantidade_de_empates_resolvidos_para_non_nav
+    """
+    out = ids.copy()
+    unknown = out == 255
+
+    n_unknown = int(unknown.sum())
+
+    if n_unknown == 0:
+        return out, 0, 0
+
+    present_class_ids = [
+        int(cid)
+        for cid in sorted(classes)
+        if np.any(out == int(cid))
+    ]
+
+    if not present_class_ids:
+        return out, 0, 0
+
+    # Máscara pura com poucos artefatos: só existe uma classe válida.
+    if len(present_class_ids) == 1:
+        out[unknown] = int(present_class_ids[0])
+        return out, n_unknown, 0
+
+    distance_maps = []
+
+    for cid in present_class_ids:
+        # distanceTransform mede distância ao zero mais próximo.
+        # Portanto pixels da classe atual viram zero.
+        src = np.ones(
+            out.shape,
+            dtype=np.uint8,
+        )
+        src[out == int(cid)] = 0
+
+        dist = cv2.distanceTransform(
+            src,
+            cv2.DIST_L2,
+            3,
+        )
+
+        distance_maps.append(
+            dist
+        )
+
+    dists = np.stack(
+        distance_maps,
+        axis=0,
+    )
+
+    ys, xs = np.where(
+        unknown
+    )
+
+    d_unknown = dists[
+        :,
+        ys,
+        xs,
+    ]
+
+    min_dist = np.min(
+        d_unknown,
+        axis=0,
+    )
+
+    # Tolerância pequena porque distanceTransform usa aproximação.
+    tie_eps = 1e-3
+
+    assignments = np.empty(
+        n_unknown,
+        dtype=np.uint8,
+    )
+
+    tie_to_non_nav = 0
+
+    non_nav_available = (
+        non_nav_id is not None
+        and int(non_nav_id)
+        in present_class_ids
+    )
+
+    class_ids_arr = np.asarray(
+        present_class_ids,
+        dtype=np.uint8,
+    )
+
+    for i in range(n_unknown):
+        tied = np.where(
+            np.abs(
+                d_unknown[:, i]
+                - min_dist[i]
+            )
+            <= tie_eps
+        )[0]
+
+        if len(tied) == 1:
+            assignments[i] = class_ids_arr[
+                tied[0]
+            ]
+            continue
+
+        if non_nav_available:
+            non_nav_pos = present_class_ids.index(
+                int(non_nav_id)
+            )
+
+            if non_nav_pos in tied:
+                assignments[i] = int(
+                    non_nav_id
+                )
+                tie_to_non_nav += 1
+                continue
+
+        # Fallback determinístico se non_nav não estiver entre os empatados.
+        tied_ids = class_ids_arr[
+            tied
+        ]
+        assignments[i] = int(
+            np.min(
+                tied_ids
+            )
+        )
+
+    out[
+        ys,
+        xs,
+    ] = assignments
+
+    return (
+        out,
+        n_unknown,
+        tie_to_non_nav,
+    )
+
+
+def colored_mask_to_ids(
+    mask_bgr: np.ndarray,
+    classes,
+    *,
+    non_nav_id: Optional[int] = None,
+    repair_enabled: bool = True,
+    max_unknown_pixels: int = 4096,
+    max_unknown_fraction: float = 0.002,
+    color_max_distance: float = 55.0,
+    color_min_margin: float = 20.0,
+) -> Tuple[np.ndarray, dict]:
+    """
+    RGB/BGR -> IDs com recuperação CONSERVADORA.
+
+    Ordem:
+        1) cores exatas do labelmap;
+        2) consenso espacial;
+        3) distância de cor com margem forte;
+        4) fallback geométrico para a classe válida espacialmente mais próxima;
+           empate -> não navegável;
+        5) quarentena somente se o volume de pixels inválidos já exceder
+           os limites de reparo ou se algo realmente inconsistente permanecer.
+    """
+    ids, unknown = _exact_color_decode(
+        mask_bgr,
+        classes,
+    )
+
+    unknown_initial = int(unknown.sum())
+    total_pixels = int(ids.size)
+
+    info = {
+        "enabled": bool(repair_enabled),
+        "unknown_initial": unknown_initial,
+        "unknown_fraction_initial": (
+            float(unknown_initial / max(1, total_pixels))
+        ),
+        "repaired_spatial": 0,
+        "repaired_color": 0,
+        "repaired_nearest": 0,
+        "nearest_ties_to_non_nav": 0,
+        "unknown_final": unknown_initial,
+        "repaired": False,
+    }
+
+    if unknown_initial == 0:
+        return ids, info
+
+    pixels = mask_bgr[:, :, :3][unknown]
+    uniq = np.unique(
+        pixels.reshape(-1, 3),
+        axis=0,
+    )
+
+    if not repair_enabled:
+        raise ValueError(
+            f"Máscara possui {unknown_initial} pixels com cores fora do labelmap. "
+            f"Primeiras cores BGR={uniq[:10].tolist()}"
+        )
+
+    fraction = unknown_initial / max(1, total_pixels)
+
+    if (
+        unknown_initial > int(max_unknown_pixels)
+        or fraction > float(max_unknown_fraction)
+    ):
+        raise ValueError(
+            f"Máscara possui {unknown_initial} pixels desconhecidos "
+            f"({fraction * 100.0:.5f}%), acima do limite de reparo "
+            f"(pixels<={int(max_unknown_pixels)}, "
+            f"fração<={float(max_unknown_fraction) * 100.0:.5f}%). "
+            f"Primeiras cores BGR={uniq[:10].tolist()}"
+        )
+
+    ids, n_spatial = _repair_unknown_spatial(
+        ids,
+        classes,
+    )
+    info["repaired_spatial"] = int(n_spatial)
+
+    ids, n_color = _repair_unknown_color(
+        ids,
+        mask_bgr,
+        classes,
+        max_distance=float(color_max_distance),
+        min_margin=float(color_min_margin),
+    )
+    info["repaired_color"] = int(n_color)
+
+    # Legado: anti-alias/ruído de 1 px frequentemente cai EXATAMENTE
+    # na fronteira das duas classes. Nesse caso maioria local e distância
+    # de cor podem ficar empatadas. Resolve pela geometria da máscara.
+    ids, n_nearest, n_ties_non_nav = _repair_unknown_nearest_class(
+        ids,
+        classes,
+        non_nav_id=non_nav_id,
+    )
+    info["repaired_nearest"] = int(n_nearest)
+    info["nearest_ties_to_non_nav"] = int(n_ties_non_nav)
+
+    remaining = ids == 255
+    unknown_final = int(remaining.sum())
+
+    info["unknown_final"] = unknown_final
+    info["repaired"] = (
+        unknown_initial > 0
+        and unknown_final == 0
+    )
+
+    if unknown_final:
+        pixels_final = mask_bgr[:, :, :3][remaining]
+        uniq_final = np.unique(
+            pixels_final.reshape(-1, 3),
+            axis=0,
+        )
+
+        raise ValueError(
+            f"Máscara tinha {unknown_initial} pixels fora do labelmap; "
+            f"reparo resolveu "
+            f"{int(n_spatial) + int(n_color) + int(n_nearest)}, "
+            f"mas restaram {unknown_final} "
+            f"pixels ambíguos. Primeiras cores BGR restantes="
+            f"{uniq_final[:10].tolist()}"
+        )
+
+    return ids, info
+
+
+def load_mask_ids(
+    path: Path,
+    classes,
+    expected_hw: Tuple[int, int],
+    *,
+    non_nav_id: Optional[int] = None,
+    repair_enabled: bool = True,
+    max_unknown_pixels: int = 4096,
+    max_unknown_fraction: float = 0.002,
+    color_max_distance: float = 55.0,
+    color_min_margin: float = 20.0,
+):
+    raw = cv2.imread(
+        str(path),
+        cv2.IMREAD_UNCHANGED,
+    )
+
+    if raw is None:
+        raise FileNotFoundError(
+            f"Não consegui abrir mask: {path}"
+        )
+
+    if raw.shape[:2] != expected_hw:
+        raise ValueError(
+            f"Mask {path.name} shape={raw.shape[:2]} != imagem {expected_hw}"
+        )
+
+    repair_info = {
+        "enabled": bool(repair_enabled),
+        "unknown_initial": 0,
+        "unknown_fraction_initial": 0.0,
+        "repaired_spatial": 0,
+        "repaired_color": 0,
+        "repaired_nearest": 0,
+        "nearest_ties_to_non_nav": 0,
+        "unknown_final": 0,
+        "repaired": False,
+    }
+
+    if raw.ndim == 2:
+        ids = raw.astype(
+            np.uint8,
+            copy=False,
+        )
+        contract = "ids"
+
+    elif raw.ndim == 3 and raw.shape[2] >= 3:
+        ids, repair_info = colored_mask_to_ids(
+            raw[:, :, :3],
+            classes,
+            non_nav_id=non_nav_id,
+            repair_enabled=repair_enabled,
+            max_unknown_pixels=max_unknown_pixels,
+            max_unknown_fraction=max_unknown_fraction,
+            color_max_distance=color_max_distance,
+            color_min_margin=color_min_margin,
+        )
+
+        contract = (
+            "rgb_color_repaired"
+            if repair_info["repaired"]
+            else "rgb_color"
+        )
+
+    else:
+        raise ValueError(
+            f"Mask inválida {path.name}: shape={raw.shape}"
+        )
+
+    present = set(
+        int(x)
+        for x in np.unique(ids)
+    )
+    allowed = set(
+        int(cid)
+        for cid in classes
+    )
+
+    if not present.issubset(allowed):
+        raise ValueError(
+            f"Mask {path.name} IDs inválidos="
+            f"{sorted(present - allowed)}"
+        )
+
+    return ids.copy(), contract, repair_info
+
+
+def build_unknown_mask_diagnostic(
+    mask_path: Path,
+    classes,
+) -> Optional[np.ndarray]:
+    """
+    Gera uma cópia visual da máscara com pixels de cor desconhecida em magenta.
+    Útil para revisão manual da quarentena.
+    """
+    raw = cv2.imread(
+        str(mask_path),
+        cv2.IMREAD_UNCHANGED,
+    )
+
+    if (
+        raw is None
+        or raw.ndim != 3
+        or raw.shape[2] < 3
+    ):
+        return None
+
+    vis = raw[:, :, :3].copy()
+    _ids, unknown = _exact_color_decode(
+        vis,
+        classes,
+    )
+
+    if not np.any(unknown):
+        return None
+
+    # BGR magenta.
+    vis[unknown] = np.array(
+        [255, 0, 255],
+        dtype=np.uint8,
+    )
+
+    return vis
+
+
+def derive_group(mask_ids: np.ndarray, nav_id: int, non_nav_id: int) -> str:
+    present = set(int(x) for x in np.unique(mask_ids))
+
+    if present == {non_nav_id}:
+        return "naonavegavel"
+    if present == {nav_id}:
+        return "navegavel"
+    if present == {nav_id, non_nav_id}:
+        return "naonavegavel_navegavel"
+
+    raise ValueError(f"Composição inesperada da mask: IDs={sorted(present)}")
+
+
+# =============================================================================
+# Dataset grouped
+# =============================================================================
+
+def collect_group_samples(
+    original_group_root: Path,
+    strict: bool = False,
+) -> Tuple[List[Sample], dict]:
+    """
+    Coleta SOMENTE tripletas completas.
+
+    Importante:
+        - nomes legados são reconciliados por canonical_base();
+        - item incompleto é ignorado por padrão e registrado no relatório;
+        - colisão ambígua de nomes também é ignorada/registrada;
+        - --strict-input transforma esses casos em erro.
+
+    Portanto:
+        500_Rgb.png + 500.png + 500.json
+    é uma tripleta válida com base canônica "500".
+    """
+    if not original_group_root.is_dir():
+        raise FileNotFoundError(
+            f"Raiz original agrupada não encontrada: {original_group_root}"
+        )
+
+    samples: List[Sample] = []
+    group_report = {}
+
+    all_incomplete: List[dict] = []
+    all_collisions: List[dict] = []
+    all_aliases: List[dict] = []
+
+    group_dirs = sorted(
+        [p for p in original_group_root.iterdir() if p.is_dir()],
+        key=lambda p: natural_key(p.name),
+    )
+
+    for group_dir in group_dirs:
+        group = group_dir.name
+
+        images, image_aliases, image_collisions = map_by_canonical_base(
+            group_dir / "images",
+            IMG_EXTS,
+        )
+        masks, mask_aliases, mask_collisions = map_by_canonical_base(
+            group_dir / "masks",
+            MASK_EXTS,
+        )
+        labels, label_aliases, label_collisions = map_by_canonical_base(
+            group_dir / "labels",
+            LABEL_EXTS,
+        )
+
+        aliases = image_aliases + mask_aliases + label_aliases
+
+        collisions = []
+        for modality, rows in (
+            ("image", image_collisions),
+            ("mask", mask_collisions),
+            ("label", label_collisions),
+        ):
+            for row in rows:
+                row = dict(row)
+                row["group"] = group
+                row["modality"] = modality
+                collisions.append(row)
+
+        for row in aliases:
+            row["group"] = group
+
+        all_aliases.extend(aliases)
+        all_collisions.extend(collisions)
+
+        # Canonicals com colisão são conhecidos, mas foram removidos do map.
+        collision_bases_by_modality = {
+            "image": {r["base"] for r in image_collisions},
+            "mask": {r["base"] for r in mask_collisions},
+            "label": {r["base"] for r in label_collisions},
+        }
+
+        all_bases = sorted(
+            set(images)
+            | set(masks)
+            | set(labels)
+            | collision_bases_by_modality["image"]
+            | collision_bases_by_modality["mask"]
+            | collision_bases_by_modality["label"],
+            key=natural_key,
+        )
+
+        complete = 0
+        incomplete = []
+
+        for base in all_bases:
+            missing = []
+            ambiguous = []
+
+            if base in collision_bases_by_modality["image"]:
+                ambiguous.append("image")
+            elif base not in images:
+                missing.append("image")
+
+            if base in collision_bases_by_modality["mask"]:
+                ambiguous.append("mask")
+            elif base not in masks:
+                missing.append("mask")
+
+            if base in collision_bases_by_modality["label"]:
+                ambiguous.append("label")
+            elif base not in labels:
+                missing.append("label")
+
+            if missing or ambiguous:
+                row = {
+                    "group": group,
+                    "base": base,
+                    "missing": missing,
+                    "ambiguous": ambiguous,
+                }
+                incomplete.append(row)
+                all_incomplete.append(row)
+                continue
+
+            samples.append(
+                Sample(
+                    group=group,
+                    base=base,
+                    image=images[base],
+                    mask=masks[base],
+                    label=labels[base],
+                )
+            )
+            complete += 1
+
+        group_report[group] = {
+            "images_canonical": len(images),
+            "masks_canonical": len(masks),
+            "labels_canonical": len(labels),
+            "complete": complete,
+            "incomplete_count": len(incomplete),
+            "incomplete": incomplete,
+            "legacy_alias_count": len(aliases),
+            "legacy_aliases": aliases,
+            "collision_count": len(collisions),
+            "collisions": collisions,
+        }
+
+    if strict and (all_incomplete or all_collisions):
+        lines = []
+
+        for x in all_incomplete[:20]:
+            details = []
+            if x.get("missing"):
+                details.append(
+                    "faltando " + ",".join(x["missing"])
+                )
+            if x.get("ambiguous"):
+                details.append(
+                    "ambíguo " + ",".join(x["ambiguous"])
+                )
+
+            lines.append(
+                f"{x['group']}/{x['base']}: " + "; ".join(details)
+            )
+
+        raise RuntimeError(
+            "Dataset original possui itens não utilizáveis no pareamento "
+            f"(incompletos={len(all_incomplete)}, colisões={len(all_collisions)}).\n"
+            + "\n".join(lines)
+        )
+
+    return samples, {
+        "groups": group_report,
+        "complete_count": len(samples),
+        "incomplete_count": len(all_incomplete),
+        "collision_count": len(all_collisions),
+        "legacy_alias_count": len(all_aliases),
+        "incomplete": all_incomplete,
+        "collisions": all_collisions,
+        "legacy_aliases": all_aliases,
+    }
+
+
+def validate_label(label_path: Path, states: Sequence[str], expected_group: str) -> dict:
+    try:
+        with label_path.open("r", encoding="utf-8") as f:
+            data = json.load(f)
+    except Exception as exc:
+        raise ValueError(f"Label JSON inválido {label_path}: {exc}") from exc
+
+    if not isinstance(data, dict):
+        raise ValueError(f"Label não é objeto JSON: {label_path}")
+
+    if "label_id" not in data:
+        raise ValueError(f"Label sem label_id: {label_path}")
+
+    lid = int(data["label_id"])
+    if not 0 <= lid < len(states):
+        raise ValueError(
+            f"label_id={lid} fora de [0,{len(states)-1}] em {label_path}"
+        )
+
+    expected_state = str(states[lid])
+    got_state = data.get("estado_corredor")
+    if got_state is not None and str(got_state) != expected_state:
+        raise ValueError(
+            f"Label inconsistente {label_path}: id={lid} => {expected_state!r}, "
+            f"estado_corredor={got_state!r}"
+        )
+
+    got_group = data.get("group")
+    if got_group is not None and str(got_group) != expected_group:
+        raise ValueError(
+            f"Label {label_path.name} diz group={got_group!r}, "
+            f"mas está em {expected_group!r}"
+        )
+
+    return data
+
+
+# =============================================================================
+# Downstream invalidation
+# =============================================================================
+
+def clear_downstream_artifacts(
+    resolution_root: Path,
+) -> List[str]:
+    """
+    Reexecutar NORMALIZE invalida tudo que depende dele.
+
+    Remove:
+        dataset/<W>x<H>/split/
+        dataset/<W>x<H>/norm_stats.json
+        dataset/<W>x<H>/split_report.json
+        dataset/<W>x<H>/split_manifest.csv
+
+    A fonte normalizada group/ será limpa separadamente pelo próprio normalize.
+    """
+    removed: List[str] = []
+
+    split_root = resolution_root / "split"
+    if split_root.exists():
+        shutil.rmtree(split_root)
+        removed.append(str(split_root))
+
+    for name in (
+        "norm_stats.json",
+        "split_report.json",
+        "split_manifest.csv",
+    ):
+        p = resolution_root / name
+        if p.exists():
+            p.unlink()
+            removed.append(str(p))
+
+    return removed
+
+
+# =============================================================================
+# Quarantine
+# =============================================================================
+
+def prepare_quarantine_root(
+    root: Path,
+    clean: bool,
+) -> None:
+    if clean and root.exists():
+        shutil.rmtree(root)
+
+    root.mkdir(
+        parents=True,
+        exist_ok=True,
+    )
+
+
+def quarantine_sample(
+    sample: Sample,
+    *,
+    quarantine_root: Path,
+    classes,
+    error: Exception,
+) -> dict:
+    """
+    Copia a tripleta ORIGINAL para revisão manual.
+
+    Estrutura:
+        normalize_quarantine/<grupo>/
+          images/
+          masks/
+          labels/
+          diagnostics/
+          errors/
+    """
+    group_root = (
+        quarantine_root
+        / sample.group
+    )
+
+    dirs = {
+        "images": group_root / "images",
+        "masks": group_root / "masks",
+        "labels": group_root / "labels",
+        "diagnostics": group_root / "diagnostics",
+        "errors": group_root / "errors",
+    }
+
+    for folder in dirs.values():
+        folder.mkdir(
+            parents=True,
+            exist_ok=True,
+        )
+
+    out_image = (
+        dirs["images"]
+        / f"{sample.base}{sample.image.suffix.lower()}"
+    )
+    out_mask = (
+        dirs["masks"]
+        / f"{sample.base}{sample.mask.suffix.lower()}"
+    )
+    out_label = (
+        dirs["labels"]
+        / f"{sample.base}{sample.label.suffix.lower()}"
+    )
+
+    shutil.copy2(
+        sample.image,
+        out_image,
+    )
+    shutil.copy2(
+        sample.mask,
+        out_mask,
+    )
+    shutil.copy2(
+        sample.label,
+        out_label,
+    )
+
+    diagnostic_path = None
+
+    try:
+        diagnostic = build_unknown_mask_diagnostic(
+            sample.mask,
+            classes,
+        )
+
+        if diagnostic is not None:
+            diagnostic_path = (
+                dirs["diagnostics"]
+                / f"{sample.base}_unknown.png"
+            )
+            atomic_write_png(
+                diagnostic_path,
+                diagnostic,
+            )
+    except Exception:
+        diagnostic_path = None
+
+    error_path = (
+        dirs["errors"]
+        / f"{sample.base}.json"
+    )
+
+    payload = {
+        "schema": "agrobot_normalize_quarantine_v1",
+        "created_at": now_iso(),
+        "group": sample.group,
+        "base": sample.base,
+        "error_type": type(error).__name__,
+        "error": str(error),
+        "source_image": str(sample.image),
+        "source_mask": str(sample.mask),
+        "source_label": str(sample.label),
+        "quarantine_image": str(out_image),
+        "quarantine_mask": str(out_mask),
+        "quarantine_label": str(out_label),
+        "diagnostic_unknown_pixels": (
+            str(diagnostic_path)
+            if diagnostic_path is not None
+            else None
+        ),
+    }
+
+    atomic_write_json(
+        error_path,
+        payload,
+    )
+
+    return payload
+
+
+# =============================================================================
+# Processing
+# =============================================================================
+
+def check_aspect_ratio(
+    src_wh: Tuple[int, int],
+    dst_wh: Tuple[int, int],
+    tolerance: float = 1e-3,
+):
+    sw, sh = src_wh
+    dw, dh = dst_wh
+    src_ar = sw / float(sh)
+    dst_ar = dw / float(dh)
+    rel = abs(src_ar - dst_ar) / max(dst_ar, 1e-9)
+    if rel > tolerance:
+        raise ValueError(
+            f"Aspect ratio incompatível: origem={sw}x{sh} ({src_ar:.6f}) vs "
+            f"destino={dw}x{dh} ({dst_ar:.6f}). "
+            "Não será feito crop/letterbox silencioso."
+        )
+
+
+def prepare_output_root(output_group_root: Path, clean: bool):
+    if clean and output_group_root.exists():
+        shutil.rmtree(output_group_root)
+    output_group_root.mkdir(parents=True, exist_ok=True)
+
+
+def process_sample(
+    sample: Sample,
+    *,
+    target_wh: Tuple[int, int],
+    classes,
+    states: Sequence[str],
+    nav_id: int,
+    non_nav_id: int,
+    dataset_root: Path,
+    output_root: Path,
+    repair_enabled: bool,
+    repair_max_unknown_pixels: int,
+    repair_max_unknown_fraction: float,
+    repair_color_max_distance: float,
+    repair_color_min_margin: float,
+) -> dict:
+    target_w, target_h = target_wh
+
+    img = cv2.imread(str(sample.image), cv2.IMREAD_COLOR)
+    if img is None:
+        raise FileNotFoundError(f"Não consegui abrir imagem: {sample.image}")
+
+    src_h, src_w = img.shape[:2]
+    check_aspect_ratio((src_w, src_h), target_wh)
+
+    mask_ids, source_mask_contract, mask_repair = load_mask_ids(
+        sample.mask,
+        classes,
+        expected_hw=(src_h, src_w),
+        non_nav_id=non_nav_id,
+        repair_enabled=repair_enabled,
+        max_unknown_pixels=repair_max_unknown_pixels,
+        max_unknown_fraction=repair_max_unknown_fraction,
+        color_max_distance=repair_color_max_distance,
+        color_min_margin=repair_color_min_margin,
+    )
+
+    derived_group = derive_group(mask_ids, nav_id, non_nav_id)
+    if sample.group != derived_group:
+        raise ValueError(
+            f"Grupo divergente em {sample.base}: pasta={sample.group!r}, "
+            f"mask implica={derived_group!r}"
+        )
+
+    label = validate_label(sample.label, states, expected_group=sample.group)
+    label_id = int(label["label_id"])
+    estado = str(states[label_id])
+
+    group_root = output_root / "group" / sample.group
+    out_images = group_root / "images"
+    out_masks = group_root / "masks"
+    out_labels = group_root / "labels"
+
+    out_img = out_images / f"{sample.base}.png"
+    out_mask = out_masks / f"{sample.base}.png"
+    out_label = out_labels / f"{sample.base}.json"
+
+    img_resized = cv2.resize(
+        img,
+        (target_w, target_h),
+        interpolation=cv2.INTER_AREA,
+    )
+    atomic_write_png(out_img, img_resized)
+
+    mask_resized = cv2.resize(
+        mask_ids,
+        (target_w, target_h),
+        interpolation=cv2.INTER_NEAREST,
+    ).astype(np.uint8, copy=False)
+
+    resized_group = derive_group(mask_resized, nav_id, non_nav_id)
+    if resized_group != sample.group:
+        # Em teoria nearest pode eliminar uma classe presente em pouquíssimos pixels.
+        # Isso é informação importante, não deve passar silenciosamente.
+        raise ValueError(
+            f"Após resize, composição mudou de grupo: {sample.group} -> {resized_group}. "
+            "Revise esta máscara; a classe minoritária pode ser pequena demais."
+        )
+
+    atomic_write_png(out_mask, mask_resized)
+
+    nav_pct = float((mask_resized == nav_id).mean() * 100.0)
+
+    normalized = dict(label)
+    normalized.update({
+        "schema": NORMALIZED_LABEL_SCHEMA,
+        "group": sample.group,
+        "estado_corredor": estado,
+        "label_id": label_id,
+        "states": list(states),
+        "base": sample.base,
+        "normalized": True,
+        "normalized_at": now_iso(),
+        "normalized_resolution_wh": [target_w, target_h],
+        "image": safe_rel(out_img, output_root),
+        "mask": safe_rel(out_mask, output_root),
+        "label": safe_rel(out_label, output_root),
+        "source_image": safe_rel(sample.image, dataset_root),
+        "source_mask": safe_rel(sample.mask, dataset_root),
+        "source_label": safe_rel(sample.label, dataset_root),
+        "source_mask_contract": source_mask_contract,
+        "mask_repair": mask_repair,
+        "normalized_mask_contract": "uint8_class_ids",
+        "mask_classes": {
+            str(cid): {
+                "name": classes[cid][0],
+                "rgb": list(classes[cid][1]),
+            }
+            for cid in sorted(classes)
+        },
+        "nav_class_id": int(nav_id),
+        "non_nav_class_id": int(non_nav_id),
+        "navegavel_pct_normalized": nav_pct,
+    })
+
+    atomic_write_json(out_label, normalized)
+
+    return {
+        "group": sample.group,
+        "base": sample.base,
+        "source_wh": [src_w, src_h],
+        "target_wh": [target_w, target_h],
+        "source_mask_contract": source_mask_contract,
+        "mask_repair": mask_repair,
+        "label_id": label_id,
+        "estado_corredor": estado,
+        "navegavel_pct": nav_pct,
+    }
+
+
+# =============================================================================
+# Main
+# =============================================================================
+
+def main():
+    ap = argparse.ArgumentParser(
+        description=(
+            "Normaliza dataset/original/group da OAK-D Lite preservando grupos."
+        )
+    )
+    ap.add_argument("--config", default=str(CONFIG_PATH))
+    ap.add_argument("--original-root", default=str(DEFAULT_ORIGINAL_GROUP_ROOT))
+    ap.add_argument("--labelmap", default=str(DEFAULT_LABELMAP))
+    ap.add_argument(
+        "--strict-input",
+        action="store_true",
+        help=(
+            "Abortar se houver item incompleto ou colisão de nomes. "
+            "Default: processa apenas tripletas completas e registra o restante."
+        ),
+    )
+    ap.add_argument(
+        "--allow-incomplete",
+        action="store_true",
+        help=(
+            "Compatibilidade com versões anteriores. Hoje já é o comportamento "
+            "default; mantido para não quebrar comandos antigos."
+        ),
+    )
+    ap.add_argument(
+        "--continue-on-error",
+        action="store_true",
+        help=(
+            "Compatibilidade: hoje continuar após erro de conteúdo já é o default."
+        ),
+    )
+    ap.add_argument(
+        "--fail-fast-content",
+        action="store_true",
+        help=(
+            "Abortar no primeiro erro de conteúdo de uma tripleta completa. "
+            "Default: envia a tripleta para quarentena e continua."
+        ),
+    )
+    ap.add_argument(
+        "--quarantine-root",
+        default=str(DEFAULT_QUARANTINE_ROOT),
+        help=(
+            "Pasta para tripletas que falharem durante normalização."
+        ),
+    )
+    ap.add_argument(
+        "--no-mask-repair",
+        action="store_true",
+        help=(
+            "Desativa recuperação conservadora de poucos pixels RGB fora do labelmap."
+        ),
+    )
+    ap.add_argument(
+        "--repair-max-unknown-pixels",
+        type=int,
+        default=4096,
+        help="Máximo absoluto de pixels desconhecidos elegíveis a reparo.",
+    )
+    ap.add_argument(
+        "--repair-max-unknown-fraction",
+        type=float,
+        default=0.002,
+        help=(
+            "Máxima fração de pixels desconhecidos elegível a reparo. "
+            "0.002 = 0.2%% da máscara."
+        ),
+    )
+    ap.add_argument(
+        "--repair-color-max-distance",
+        type=float,
+        default=55.0,
+        help="Distância BGR máxima para segunda chance por cor.",
+    )
+    ap.add_argument(
+        "--repair-color-min-margin",
+        type=float,
+        default=20.0,
+        help="Margem mínima entre melhor e segunda melhor cor para aceitar reparo.",
+    )
+    ap.add_argument(
+        "--no-clean",
+        action="store_true",
+        help=(
+            "Não limpa a saída normalizada anterior e não invalida artefatos "
+            "downstream. Use somente quando souber exatamente o que está fazendo."
+        ),
+    )
+    args = ap.parse_args()
+
+    config_path = Path(args.config)
+    if not config_path.is_file():
+        raise FileNotFoundError(f"config não encontrado: {config_path}")
+
+    with config_path.open("r", encoding="utf-8") as f:
+        config = json.load(f)
+
+    resolution = config.get("resolucao")
+    if not isinstance(resolution, list) or len(resolution) != 2:
+        raise RuntimeError("config['resolucao'] deve ser [W,H].")
+
+    target_w, target_h = int(resolution[0]), int(resolution[1])
+    if target_w <= 0 or target_h <= 0:
+        raise RuntimeError(f"Resolução inválida: {resolution}")
+
+    states = config.get("label_classes")
+    if not isinstance(states, list) or not states:
+        raise RuntimeError("config['label_classes'] obrigatório.")
+
+    classes = load_labelmap_classes(Path(args.labelmap))
+    nav_id, non_nav_id = resolve_nav_classes(
+        classes,
+        str(config.get("main_class_name", "navegavel")),
+    )
+
+    original_group_root = Path(args.original_root)
+
+    strict_input = bool(args.strict_input) and not bool(args.allow_incomplete)
+
+    # Pipeline legado grande: por padrão erro de conteúdo vai para quarentena
+    # e o normalize continua. --fail-fast-content restaura o modo rígido.
+    fail_fast_processing = bool(args.fail_fast_content)
+
+    repair_enabled = not bool(args.no_mask_repair)
+
+    if args.repair_max_unknown_pixels < 0:
+        raise ValueError("--repair-max-unknown-pixels deve ser >= 0")
+
+    if not (0.0 <= args.repair_max_unknown_fraction <= 1.0):
+        raise ValueError("--repair-max-unknown-fraction deve ficar entre 0 e 1")
+
+    quarantine_root = Path(args.quarantine_root)
+
+    samples, collect_report = collect_group_samples(
+        original_group_root,
+        strict=strict_input,
+    )
+
+    if not samples:
+        raise RuntimeError(f"Nenhuma amostra completa em {original_group_root}")
+
+    output_root = DATASET_ROOT / f"{target_w}x{target_h}"
+    output_group_root = output_root / "group"
+
+    removed_downstream = []
+
+    if not args.no_clean:
+        removed_downstream = clear_downstream_artifacts(
+            output_root
+        )
+
+    prepare_output_root(
+        output_group_root,
+        clean=not args.no_clean,
+    )
+    prepare_quarantine_root(
+        quarantine_root,
+        clean=not args.no_clean,
+    )
+
+    print("=" * 82)
+    print("OAK-D Lite | Normalize Corridor Dataset por grupos")
+    print("=" * 82)
+    print(f"Entrada            : {original_group_root.resolve()}")
+    print(f"Saída              : {output_root.resolve()}")
+    print(f"Resolução          : {target_w}x{target_h}")
+    print(f"Tripletas válidas  : {len(samples)}")
+    print(f"Incompletas skip   : {collect_report['incomplete_count']}")
+    print(f"Aliases legados    : {collect_report['legacy_alias_count']}")
+    print(f"Colisões skip      : {collect_report['collision_count']}")
+    print(f"Navegável          : id={nav_id} {classes[nav_id][0]}")
+    print(f"Não navegável      : id={non_nav_id} {classes[non_nav_id][0]}")
+    print("Augmentation       : NÃO, fica no trainer")
+    print("Mean/std           : NÃO, será calculado depois do split apenas no TRAIN")
+    print(
+        f"Mask repair        : {'ON' if repair_enabled else 'OFF'} "
+        f"(max_px={args.repair_max_unknown_pixels}, "
+        f"max_frac={args.repair_max_unknown_fraction:.6f})"
+    )
+    print(f"Quarentena         : {quarantine_root.resolve()}")
+    print(
+        f"Erro conteúdo      : "
+        f"{'ABORTA' if fail_fast_processing else 'QUARENTENA + CONTINUA'}"
+    )
+    print(
+        f"Clear downstream   : "
+        f"{'OFF (--no-clean)' if args.no_clean else 'ON'}"
+    )
+
+    if removed_downstream:
+        print("[CLEAN] Artefatos downstream antigos removidos:")
+        for removed_path in removed_downstream:
+            print(f"  - {removed_path}")
+    print("=" * 82)
+
+    if collect_report["incomplete_count"]:
+        print("[PAIRING] Itens incompletos ignorados (primeiros 12):")
+        for row in collect_report["incomplete"][:12]:
+            details = []
+            if row.get("missing"):
+                details.append("faltando=" + ",".join(row["missing"]))
+            if row.get("ambiguous"):
+                details.append("ambíguo=" + ",".join(row["ambiguous"]))
+            print(
+                f"  - {row['group']}/{row['base']} | "
+                + " | ".join(details)
+            )
+
+    if collect_report["collision_count"]:
+        print("[PAIRING] Colisões de nome ignoradas (primeiras 8):")
+        for row in collect_report["collisions"][:8]:
+            print(
+                f"  - {row['group']}/{row['base']} "
+                f"[{row['modality']}] -> {row['files']}"
+            )
+
+    if collect_report["legacy_alias_count"]:
+        print(
+            f"[PAIRING] {collect_report['legacy_alias_count']} nome(s) legado(s) "
+            "foram reconciliados por base canônica."
+        )
+
+    print()
+
+    processed = []
+    failed = []
+    repaired = []
+    quarantined = []
+    count_by_group = Counter()
+    count_by_state = Counter()
+    cross = defaultdict(Counter)
+
+    for i, sample in enumerate(samples, start=1):
+        try:
+            row = process_sample(
+                sample,
+                target_wh=(target_w, target_h),
+                classes=classes,
+                states=states,
+                nav_id=nav_id,
+                non_nav_id=non_nav_id,
+                dataset_root=DATASET_ROOT,
+                output_root=output_root,
+                repair_enabled=repair_enabled,
+                repair_max_unknown_pixels=int(args.repair_max_unknown_pixels),
+                repair_max_unknown_fraction=float(args.repair_max_unknown_fraction),
+                repair_color_max_distance=float(args.repair_color_max_distance),
+                repair_color_min_margin=float(args.repair_color_min_margin),
+            )
+            processed.append(row)
+
+            if row.get("mask_repair", {}).get("repaired"):
+                repaired.append({
+                    "group": sample.group,
+                    "base": sample.base,
+                    **row["mask_repair"],
+                })
+
+            count_by_group[row["group"]] += 1
+            count_by_state[row["estado_corredor"]] += 1
+            cross[row["group"]][row["estado_corredor"]] += 1
+
+            repair_tag = ""
+
+            if row.get("mask_repair", {}).get("repaired"):
+                mr = row["mask_repair"]
+                repair_tag = (
+                    f" | REPAIR {mr['unknown_initial']}px "
+                    f"(local={mr['repaired_spatial']}, "
+                    f"color={mr['repaired_color']}, "
+                    f"nearest={mr.get('repaired_nearest', 0)}, "
+                    f"tieNonNav={mr.get('nearest_ties_to_non_nav', 0)})"
+                )
+
+            print(
+                f"[{i:5d}/{len(samples):5d}] OK "
+                f"{sample.group}/{sample.base} | "
+                f"mask={row['source_mask_contract']} -> ids | "
+                f"nav={row['navegavel_pct']:.1f}% | "
+                f"label={row['estado_corredor']}"
+                f"{repair_tag}"
+            )
+
+        except Exception as exc:
+            failure = {
+                "group": sample.group,
+                "base": sample.base,
+                "error_type": type(exc).__name__,
+                "error": str(exc),
+            }
+            failed.append(failure)
+
+            quarantine_info = quarantine_sample(
+                sample,
+                quarantine_root=quarantine_root,
+                classes=classes,
+                error=exc,
+            )
+            quarantined.append(quarantine_info)
+
+            print(
+                f"[{i:5d}/{len(samples):5d}] QUARENTENA "
+                f"{sample.group}/{sample.base}: {exc}"
+            )
+
+            if fail_fast_processing:
+                raise
+
+    report = {
+        "schema": REPORT_SCHEMA,
+        "created_at": now_iso(),
+        "config_path": str(config_path),
+        "labelmap_path": str(Path(args.labelmap)),
+        "input_root": str(original_group_root),
+        "output_root": str(output_root),
+        "target_resolution_wh": [target_w, target_h],
+        "classes": {
+            str(cid): {
+                "name": classes[cid][0],
+                "rgb": list(classes[cid][1]),
+            }
+            for cid in sorted(classes)
+        },
+        "nav_class_id": nav_id,
+        "non_nav_class_id": non_nav_id,
+        "states": list(states),
+        "collection": collect_report,
+        "processed_count": len(processed),
+        "repaired_count": len(repaired),
+        "repaired": repaired,
+        "failed_count": len(failed),
+        "failed": failed,
+        "quarantined_count": len(quarantined),
+        "quarantine_root": str(quarantine_root),
+        "quarantined": quarantined,
+        "counts": {
+            "by_group": dict(count_by_group),
+            "by_state": dict(count_by_state),
+            "group_x_state": {
+                group: dict(counter)
+                for group, counter in cross.items()
+            },
+        },
+        "policy": {
+            "image_resize": "cv2.INTER_AREA",
+            "mask_input": "rgb_color_exact_or_legacy_ids",
+            "mask_repair_enabled": repair_enabled,
+            "mask_repair_policy": (
+                "exact -> local_consensus -> strong_color_margin -> "
+                "nearest_valid_class(tie=non_nav) -> quarantine_if_over_limits"
+            ),
+            "repair_max_unknown_pixels": int(args.repair_max_unknown_pixels),
+            "repair_max_unknown_fraction": float(args.repair_max_unknown_fraction),
+            "repair_color_max_distance": float(args.repair_color_max_distance),
+            "repair_color_min_margin": float(args.repair_color_min_margin),
+            "mask_output": "uint8_class_ids",
+            "mask_resize": "cv2.INTER_NEAREST",
+            "preserve_groups": True,
+            "clear_downstream_on_clean": True,
+            "downstream_removed": removed_downstream,
+            "pairing": "canonical_legacy_stem + complete_triplets_only",
+            "incomplete_policy": "skip_and_report",
+            "collision_policy": "skip_and_report",
+            "complete_triplet_processing_error_policy": (
+                "fail_fast"
+                if fail_fast_processing
+                else "quarantine_and_continue"
+            ),
+            "validate_folder_group_against_mask": True,
+            "augmentation_offline": False,
+            "mean_std_computed_here": False,
+        },
+    }
+
+    atomic_write_json(output_root / "normalize_report.json", report)
+
+    print()
+    print("=" * 82)
+    print("NORMALIZE CONCLUÍDO")
+    print("=" * 82)
+    print(f"Processadas        : {len(processed)}")
+    print(f"Reparadas mask     : {len(repaired)}")
+    print(f"Quarentena         : {len(quarantined)}")
+    print(f"Falhas conteúdo    : {len(failed)}")
+    print(f"Incompletas skip   : {collect_report['incomplete_count']}")
+    print(f"Aliases reconciliados: {collect_report['legacy_alias_count']}")
+    print(f"Colisões skip      : {collect_report['collision_count']}")
+    print(f"Por grupo   : {dict(count_by_group)}")
+    print(f"Por estado  : {dict(count_by_state)}")
+    print(f"Saída       : {output_root.resolve()}")
+    print(f"Revisão     : {quarantine_root.resolve()}")
+    print("Próximo     : split estratificado por grupo + estado, depois norm_stats do TRAIN")
+    print("=" * 82)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/Python/OAK/datasets/oak-d/_3_split_corridor.py b/Python/OAK/datasets/oak-d/_3_split_corridor.py
new file mode 100644
index 000000000..d09bc1e3a
--- /dev/null
+++ b/Python/OAK/datasets/oak-d/_3_split_corridor.py
@@ -0,0 +1,1682 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+"""
+_3_split_corridor.py
+====================
+
+Split inteligente do dataset normalizado da OAK-D Lite.
+
+Executar DE DENTRO da pasta oak-d:
+
+    python _3_split_corridor.py
+
+Entrada:
+    config.json
+
+    dataset/<W>x<H>/group/
+      navegavel/
+        images/
+        masks/
+        labels/
+      naonavegavel/
+        images/
+        masks/
+        labels/
+      naonavegavel_navegavel/
+        images/
+        masks/
+        labels/
+
+Saída:
+    dataset/<W>x<H>/split/
+      train/
+        images/
+        masks/
+        labels/
+      val/
+        images/
+        masks/
+        labels/
+
+    dataset/<W>x<H>/
+      norm_stats.json
+      split_report.json
+
+Objetivos:
+    1) preservar a distribuição conjunta:
+         grupo da máscara x estado global do corredor
+
+    2) reduzir vazamento temporal:
+         uma sessão inteira de captura vai para train OU para val
+
+    3) calcular mean/std SOMENTE sobre train/images
+
+    4) não alterar dataset/<W>x<H>/group, que continua sendo a fonte
+       normalizada completa.
+
+Importante:
+    - o split é determinístico para a mesma seed;
+    - não usa sklearn;
+    - classes extremamente raras podem não conseguir aparecer nos dois lados;
+      isso é relatado explicitamente;
+    - se houver poucas sessões, a proporção train/val pode se afastar de 80/20;
+      isso é preferível a vazar frames quase idênticos entre os splits;
+    - labels copiados para split recebem metadados split/split_seed.
+"""
+
+from __future__ import annotations
+
+import argparse
+import csv
+import json
+import math
+import os
+import random
+import re
+import shutil
+from collections import Counter, defaultdict
+from dataclasses import dataclass, field
+from datetime import datetime
+from pathlib import Path
+from typing import Dict, Iterable, List, Optional, Sequence, Tuple
+
+import cv2
+import numpy as np
+
+
+# =============================================================================
+# Defaults
+# =============================================================================
+
+CONFIG_PATH = Path("config.json")
+DATASET_ROOT = Path("dataset")
+
+DEFAULT_VAL_RATIO = 0.20
+DEFAULT_SEED = 42
+
+# Blocos temporais:
+# - abre bloco novo se a captura ficou muito tempo sem imagem;
+# - mesmo numa sequência contínua, limita o span do bloco para evitar um
+#   bloco gigantesco que inviabilize o balanceamento.
+DEFAULT_SESSION_GAP_S = 8.0
+DEFAULT_FALLBACK_BLOCK_SIZE = 12
+
+REPORT_SCHEMA = "agrobot_corridor_split_report_v1"
+NORM_SCHEMA = "agrobot_rgb_norm_stats_v2"
+
+IMG_EXTS = {".png", ".jpg", ".jpeg", ".bmp", ".webp"}
+MASK_EXTS = {".png"}
+LABEL_EXTS = {".json"}
+
+
+# =============================================================================
+# Data
+# =============================================================================
+
+@dataclass(frozen=True)
+class Sample:
+    base: str
+    group: str
+    image: Path
+    mask: Path
+    label: Path
+    label_id: int
+    label_name: str
+    timestamp_s: Optional[float]
+    capture_session_id: Optional[str]
+
+    @property
+    def stratum(self) -> Tuple[str, int]:
+        return self.group, self.label_id
+
+
+@dataclass
+class Block:
+    block_id: int
+    samples: List[Sample] = field(default_factory=list)
+
+    @property
+    def size(self) -> int:
+        return len(self.samples)
+
+    @property
+    def strata(self) -> Counter:
+        return Counter(s.stratum for s in self.samples)
+
+    @property
+    def first_timestamp(self) -> Optional[float]:
+        vals = [s.timestamp_s for s in self.samples if s.timestamp_s is not None]
+        return min(vals) if vals else None
+
+    @property
+    def last_timestamp(self) -> Optional[float]:
+        vals = [s.timestamp_s for s in self.samples if s.timestamp_s is not None]
+        return max(vals) if vals else None
+
+
+# =============================================================================
+# General helpers
+# =============================================================================
+
+def now_iso() -> str:
+    return datetime.now().isoformat(timespec="seconds")
+
+
+def natural_key(text: str):
+    parts = re.split(r"(\d+)", str(text))
+    return [int(p) if p.isdigit() else p.lower() for p in parts]
+
+
+def safe_rel(path: Path, root: Path) -> str:
+    try:
+        return str(path.resolve().relative_to(root.resolve())).replace("\\", "/")
+    except Exception:
+        return str(path).replace("\\", "/")
+
+
+def atomic_write_json(path: Path, data: dict) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    tmp = path.with_suffix(path.suffix + ".tmp")
+
+    with tmp.open("w", encoding="utf-8") as f:
+        json.dump(data, f, ensure_ascii=False, indent=2)
+
+    os.replace(tmp, path)
+
+
+def atomic_write_csv(path: Path, rows: Sequence[dict], fieldnames: Sequence[str]) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    tmp = path.with_suffix(path.suffix + ".tmp")
+
+    with tmp.open("w", newline="", encoding="utf-8-sig") as f:
+        writer = csv.DictWriter(
+            f,
+            fieldnames=list(fieldnames),
+            extrasaction="ignore",
+        )
+        writer.writeheader()
+        writer.writerows(rows)
+
+    os.replace(tmp, path)
+
+
+def map_by_stem(folder: Path, exts: set[str]) -> Dict[str, Path]:
+    if not folder.is_dir():
+        return {}
+
+    out: Dict[str, Path] = {}
+
+    for p in sorted(folder.iterdir(), key=lambda x: natural_key(x.name)):
+        if not p.is_file() or p.suffix.lower() not in exts:
+            continue
+
+        if p.stem in out:
+            raise RuntimeError(
+                f"Stem duplicado '{p.stem}' em {folder}: "
+                f"{out[p.stem].name} e {p.name}"
+            )
+
+        out[p.stem] = p
+
+    return out
+
+
+def parse_capture_session_id(base: str) -> Optional[str]:
+    """
+    Novo padrão do capture:
+        img_sYYYYMMDD_HHMMSS__YYYYMMDD_HHMMSS_micro
+
+    Retorna a identidade da execução do capture.
+
+    Arquivos antigos não possuem essa informação e retornam None.
+    """
+    m = re.search(
+        r"(?:^|_)s(\d{8}_\d{6})__",
+        str(base),
+    )
+    if not m:
+        return None
+    return m.group(1)
+
+
+def parse_capture_timestamp(base: str) -> Optional[float]:
+    """
+    Suporta:
+        img_YYYYMMDD_HHMMSS_microseconds                 # legado
+        img_sSESSION__YYYYMMDD_HHMMSS_microseconds       # atual
+
+    Retorna POSIX timestamp local apenas para ordenar/calcular diferenças.
+    """
+    matches = list(
+        re.finditer(
+            r"(\d{8})_(\d{6})(?:_(\d{1,6}))?",
+            str(base),
+        )
+    )
+
+    if not matches:
+        return None
+
+    # No formato novo, a primeira data é a sessão e a última é o frame.
+    m = matches[-1]
+
+    date_txt = m.group(1)
+    time_txt = m.group(2)
+    frac_txt = m.group(3) or "0"
+
+    frac_txt = (frac_txt + "000000")[:6]
+
+    try:
+        dt = datetime.strptime(
+            date_txt + time_txt + frac_txt,
+            "%Y%m%d%H%M%S%f",
+        )
+        return dt.timestamp()
+    except Exception:
+        return None
+
+
+# =============================================================================
+# Read source dataset
+# =============================================================================
+
+def read_label(path: Path, states: Sequence[str]) -> Tuple[int, str, dict]:
+    try:
+        with path.open("r", encoding="utf-8") as f:
+            data = json.load(f)
+    except Exception as exc:
+        raise ValueError(f"Label JSON inválido {path}: {exc}") from exc
+
+    if not isinstance(data, dict):
+        raise ValueError(f"Label não é objeto JSON: {path}")
+
+    if "label_id" not in data:
+        raise ValueError(f"Label sem label_id: {path}")
+
+    label_id = int(data["label_id"])
+
+    if not (0 <= label_id < len(states)):
+        raise ValueError(
+            f"label_id={label_id} fora de [0,{len(states)-1}] em {path}"
+        )
+
+    expected = str(states[label_id])
+    got = data.get("estado_corredor")
+
+    if got is not None and str(got) != expected:
+        raise ValueError(
+            f"estado_corredor inconsistente em {path}: "
+            f"id={label_id} => {expected!r}, arquivo={got!r}"
+        )
+
+    return label_id, expected, data
+
+
+def collect_samples(
+    group_root: Path,
+    states: Sequence[str],
+) -> List[Sample]:
+    if not group_root.is_dir():
+        raise FileNotFoundError(f"group root não encontrado: {group_root}")
+
+    result: List[Sample] = []
+    seen_bases: Dict[str, str] = {}
+
+    group_dirs = sorted(
+        [p for p in group_root.iterdir() if p.is_dir()],
+        key=lambda p: natural_key(p.name),
+    )
+
+    for group_dir in group_dirs:
+        group = group_dir.name
+
+        images = map_by_stem(group_dir / "images", IMG_EXTS)
+        masks = map_by_stem(group_dir / "masks", MASK_EXTS)
+        labels = map_by_stem(group_dir / "labels", LABEL_EXTS)
+
+        all_bases = set(images) | set(masks) | set(labels)
+
+        incomplete = []
+
+        for base in sorted(all_bases, key=natural_key):
+            missing = []
+
+            if base not in images:
+                missing.append("image")
+            if base not in masks:
+                missing.append("mask")
+            if base not in labels:
+                missing.append("label")
+
+            if missing:
+                incomplete.append((base, missing))
+                continue
+
+            if base in seen_bases:
+                raise RuntimeError(
+                    f"Base duplicada entre grupos: {base!r} aparece em "
+                    f"{seen_bases[base]!r} e {group!r}. "
+                    "O split é plano e exige stems globalmente únicos."
+                )
+
+            seen_bases[base] = group
+
+            label_id, label_name, _data = read_label(
+                labels[base],
+                states,
+            )
+
+            result.append(
+                Sample(
+                    base=base,
+                    group=group,
+                    image=images[base],
+                    mask=masks[base],
+                    label=labels[base],
+                    label_id=label_id,
+                    label_name=label_name,
+                    timestamp_s=parse_capture_timestamp(base),
+                    capture_session_id=parse_capture_session_id(base),
+                )
+            )
+
+        if incomplete:
+            preview = "\n".join(
+                f"  {group}/{base}: faltando {','.join(missing)}"
+                for base, missing in incomplete[:20]
+            )
+
+            raise RuntimeError(
+                f"Grupo {group!r} possui {len(incomplete)} amostra(s) incompleta(s):\n"
+                + preview
+            )
+
+    if not result:
+        raise RuntimeError(f"Nenhuma amostra válida encontrada em {group_root}")
+
+    return result
+
+
+# =============================================================================
+# Temporal blocks
+# =============================================================================
+
+def build_capture_sessions(
+    samples: Sequence[Sample],
+    session_gap_s: float,
+    fallback_block_size: int,
+) -> Tuple[List[Block], dict]:
+    """
+    Constrói unidades indivisíveis de split.
+
+    Prioridade:
+        1) arquivos novos: usa capture_session_id explícito no nome;
+        2) arquivos legados com timestamp: infere sessão por gap temporal;
+        3) arquivos sem timestamp: fallback em blocos naturais.
+
+    Regra central:
+        uma sessão inteira vai para TRAIN ou VAL.
+
+    Isso é mais importante do que acertar exatamente 80/20.
+    Se existirem poucas sessões, o script aceita uma proporção menos perfeita
+    em vez de permitir vazamento temporal.
+    """
+    explicit: Dict[str, List[Sample]] = defaultdict(list)
+    legacy_with_ts: List[Sample] = []
+    without_ts: List[Sample] = []
+
+    for sample in samples:
+        if sample.capture_session_id:
+            explicit[str(sample.capture_session_id)].append(sample)
+        elif sample.timestamp_s is not None:
+            legacy_with_ts.append(sample)
+        else:
+            without_ts.append(sample)
+
+    blocks: List[Block] = []
+    next_id = 0
+
+    # ---------------------------------------------------------
+    # 1) Sessões explícitas do capture novo
+    # ---------------------------------------------------------
+    explicit_session_count = 0
+
+    for session_id in sorted(explicit, key=natural_key):
+        ss = sorted(
+            explicit[session_id],
+            key=lambda s: (
+                float("inf") if s.timestamp_s is None else float(s.timestamp_s),
+                natural_key(s.base),
+            ),
+        )
+
+        blocks.append(
+            Block(
+                block_id=next_id,
+                samples=ss,
+            )
+        )
+        next_id += 1
+        explicit_session_count += 1
+
+    # ---------------------------------------------------------
+    # 2) Dataset legado: sessão inferida por gap
+    # ---------------------------------------------------------
+    inferred_session_count = 0
+
+    if legacy_with_ts:
+        ordered = sorted(
+            legacy_with_ts,
+            key=lambda s: (float(s.timestamp_s), natural_key(s.base)),
+        )
+
+        current: List[Sample] = []
+        prev_ts: Optional[float] = None
+
+        for sample in ordered:
+            ts = float(sample.timestamp_s)
+
+            if (
+                current
+                and prev_ts is not None
+                and ts - prev_ts > session_gap_s
+            ):
+                blocks.append(
+                    Block(
+                        block_id=next_id,
+                        samples=current,
+                    )
+                )
+                next_id += 1
+                inferred_session_count += 1
+                current = []
+
+            current.append(sample)
+            prev_ts = ts
+
+        if current:
+            blocks.append(
+                Block(
+                    block_id=next_id,
+                    samples=current,
+                )
+            )
+            next_id += 1
+            inferred_session_count += 1
+
+    # ---------------------------------------------------------
+    # 3) Sem timestamp: fallback por blocos de ordenação natural
+    # ---------------------------------------------------------
+    fallback_block_count = 0
+
+    if without_ts:
+        ordered = sorted(
+            without_ts,
+            key=lambda s: natural_key(s.base),
+        )
+
+        bs = max(1, int(fallback_block_size))
+
+        for start_i in range(0, len(ordered), bs):
+            chunk = ordered[start_i:start_i + bs]
+
+            blocks.append(
+                Block(
+                    block_id=next_id,
+                    samples=list(chunk),
+                )
+            )
+            next_id += 1
+            fallback_block_count += 1
+
+    # Ordem cronológica aproximada só para legibilidade.
+    blocks.sort(
+        key=lambda b: (
+            float("inf") if b.first_timestamp is None else b.first_timestamp,
+            b.block_id,
+        )
+    )
+
+    blocks = [
+        Block(
+            block_id=i,
+            samples=b.samples,
+        )
+        for i, b in enumerate(blocks)
+    ]
+
+    sizes = [b.size for b in blocks]
+
+    report = {
+        "policy": "whole_capture_session_per_split",
+        "explicit_session_count": explicit_session_count,
+        "legacy_inferred_session_count": inferred_session_count,
+        "fallback_block_count": fallback_block_count,
+        "session_count_total": len(blocks),
+
+        "explicit_session_samples": sum(len(v) for v in explicit.values()),
+        "legacy_timestamp_samples": len(legacy_with_ts),
+        "timestamp_missing_samples": len(without_ts),
+
+        "session_size_min": min(sizes) if sizes else 0,
+        "session_size_max": max(sizes) if sizes else 0,
+        "session_size_mean": float(np.mean(sizes)) if sizes else 0.0,
+
+        "legacy_session_gap_s": float(session_gap_s),
+        "fallback_block_size": int(fallback_block_size),
+    }
+
+    return blocks, report
+
+
+# =============================================================================
+# Intelligent split
+# =============================================================================
+
+def stratum_key(group: str, label_id: int) -> str:
+    return f"{group}__label_{label_id}"
+
+
+def target_val_counts(
+    samples: Sequence[Sample],
+    val_ratio: float,
+) -> Tuple[int, Counter, Counter]:
+    total = len(samples)
+
+    target_total = int(round(total * val_ratio))
+
+    if total >= 2:
+        target_total = max(1, min(total - 1, target_total))
+    else:
+        target_total = 0
+
+    total_strata = Counter(s.stratum for s in samples)
+
+    target_strata: Counter = Counter()
+
+    for stratum, count in total_strata.items():
+        raw = count * val_ratio
+
+        if count <= 1:
+            target = 0
+        else:
+            target = int(round(raw))
+
+            # Sempre que possível, tenta manter ao menos 1 de cada lado.
+            target = max(1, target)
+            target = min(count - 1, target)
+
+        target_strata[stratum] = target
+
+    return target_total, total_strata, target_strata
+
+
+def split_loss(
+    val_count: int,
+    val_strata: Counter,
+    target_total: int,
+    target_strata: Counter,
+    total_strata: Counter,
+) -> float:
+    """
+    Função de custo.
+
+    - erro por stratum recebe peso maior quando a classe é rara;
+    - erro de tamanho total ajuda a segurar a razão global;
+    - overfill é permitido, mas custa.
+    """
+    if target_total <= 0:
+        total_term = float(val_count > 0)
+    else:
+        total_term = abs(val_count - target_total) / max(1.0, float(target_total))
+
+    strata_term = 0.0
+    weight_sum = 0.0
+
+    for st, total_n in total_strata.items():
+        target_n = int(target_strata.get(st, 0))
+        current_n = int(val_strata.get(st, 0))
+
+        denom = max(1.0, float(max(target_n, 1)))
+        err = abs(current_n - target_n) / denom
+
+        rarity_weight = 1.0 / math.sqrt(max(1.0, float(total_n)))
+        # Multiplica por sqrt(N_total) para não deixar todos os pesos minúsculos.
+        rarity_weight *= math.sqrt(max(1.0, sum(total_strata.values())))
+
+        strata_term += err * rarity_weight
+        weight_sum += rarity_weight
+
+    if weight_sum > 0:
+        strata_term /= weight_sum
+
+    return (
+        0.72 * strata_term
+        + 0.28 * total_term
+    )
+
+
+def choose_val_blocks(
+    blocks: Sequence[Block],
+    samples: Sequence[Sample],
+    val_ratio: float,
+    seed: int,
+) -> Tuple[set[int], dict]:
+    """
+    Seleciona sessões inteiras para VAL.
+
+    Prioridades, nesta ordem:
+        1) quando um stratum (grupo × label) aparece em >=2 sessões,
+           tentar manter pelo menos uma ocorrência em TRAIN e VAL;
+        2) aproximar a distribuição conjunta grupo × label;
+        3) aproximar a razão global de validação.
+
+    A proporção 80/20 é um ALVO, não uma regra mais importante que a
+    independência temporal ou a cobertura de casos.
+    """
+    rng = random.Random(seed)
+
+    target_total, total_strata, target_strata = target_val_counts(
+        samples,
+        val_ratio,
+    )
+
+    if target_total <= 0:
+        return set(), {
+            "target_total": target_total,
+            "target_strata": {},
+            "coverage_feasible_strata": [],
+            "coverage_missing": [],
+            "loss": 0.0,
+        }
+
+    # Em quantas sessões independentes cada stratum aparece?
+    stratum_blocks: Dict[Tuple[str, int], set[int]] = defaultdict(set)
+
+    for block in blocks:
+        for st in block.strata:
+            stratum_blocks[st].add(block.block_id)
+
+    feasible_both = {
+        st
+        for st, block_ids in stratum_blocks.items()
+        if len(block_ids) >= 2
+        and total_strata.get(st, 0) >= 2
+    }
+
+    def coverage_missing(
+        val_strata: Counter,
+    ) -> List[Tuple[Tuple[str, int], str]]:
+        missing = []
+
+        for st in sorted(feasible_both):
+            val_n = int(val_strata.get(st, 0))
+            total_n = int(total_strata.get(st, 0))
+
+            if val_n <= 0:
+                missing.append((st, "val"))
+            elif val_n >= total_n:
+                missing.append((st, "train"))
+
+        return missing
+
+    def objective(
+        val_count: int,
+        val_strata: Counter,
+    ) -> Tuple[int, float, int]:
+        missing = coverage_missing(val_strata)
+
+        base_loss = split_loss(
+            val_count,
+            val_strata,
+            target_total,
+            target_strata,
+            total_strata,
+        )
+
+        size_error = abs(
+            int(val_count)
+            - int(target_total)
+        )
+
+        # Lexicográfico:
+        # cobertura primeiro, depois qualidade da estratificação,
+        # depois proximidade do tamanho.
+        return (
+            len(missing),
+            float(base_loss),
+            int(size_error),
+        )
+
+    def calc_for_selected(
+        selected: set[int],
+    ) -> Tuple[int, Counter, Tuple[int, float, int]]:
+        vc = 0
+        vs = Counter()
+
+        for block in blocks:
+            if block.block_id in selected:
+                vc += block.size
+                vs += block.strata
+
+        return vc, vs, objective(vc, vs)
+
+    # -------------------------------------------------------------
+    # Forward greedy
+    # -------------------------------------------------------------
+
+    remaining = list(blocks)
+    rng.shuffle(remaining)
+
+    selected: set[int] = set()
+    val_count = 0
+    val_strata: Counter = Counter()
+    current_obj = objective(
+        val_count,
+        val_strata,
+    )
+
+    while remaining:
+        candidates = []
+
+        for block in remaining:
+            new_count = val_count + block.size
+            new_strata = val_strata + block.strata
+            obj = objective(
+                new_count,
+                new_strata,
+            )
+
+            # O tamanho do bloco entra apenas como desempate final.
+            candidates.append(
+                (
+                    obj,
+                    block.size,
+                    block.block_id,
+                    block,
+                    new_count,
+                    new_strata,
+                )
+            )
+
+        candidates.sort(
+            key=lambda x: (
+                x[0][0],
+                x[0][1],
+                x[0][2],
+                x[1],
+                x[2],
+            )
+        )
+
+        (
+            best_obj,
+            _best_size,
+            _best_id,
+            best_block,
+            new_count,
+            new_strata,
+        ) = candidates[0]
+
+        # Continua enquanto:
+        # - ainda faltam casos cobríveis, OU
+        # - não atingimos o alvo total, OU
+        # - a inclusão melhora o objetivo.
+        need_coverage = current_obj[0] > 0
+        need_size = val_count < target_total
+        improves = best_obj < current_obj
+
+        if (
+            val_count > 0
+            and not need_coverage
+            and not need_size
+            and not improves
+        ):
+            break
+
+        # Nunca joga todas as sessões em val.
+        if len(selected) + 1 >= len(blocks):
+            break
+
+        selected.add(
+            best_block.block_id
+        )
+        remaining = [
+            b
+            for b in remaining
+            if b.block_id
+            != best_block.block_id
+        ]
+
+        val_count = new_count
+        val_strata = new_strata
+        current_obj = best_obj
+
+        if (
+            val_count >= target_total
+            and current_obj[0] == 0
+        ):
+            # Já temos cobertura e tamanho.
+            # Deixa o hill-climb encontrar ajustes melhores.
+            break
+
+    # -------------------------------------------------------------
+    # Hill-climb add/remove/swap
+    # -------------------------------------------------------------
+
+    if not selected:
+        best = min(
+            blocks,
+            key=lambda b: abs(
+                b.size - target_total
+            ),
+        )
+        selected = {
+            best.block_id
+        }
+
+    val_count, val_strata, current_obj = calc_for_selected(
+        selected
+    )
+
+    for _ in range(12):
+        best_sel = set(selected)
+        best_obj = current_obj
+
+        selected_ids = sorted(selected)
+        unselected_ids = sorted(
+            b.block_id
+            for b in blocks
+            if b.block_id not in selected
+        )
+
+        # Remove.
+        for rid in selected_ids:
+            candidate = set(selected)
+            candidate.remove(rid)
+
+            if not candidate:
+                continue
+
+            _vc, _vs, obj = calc_for_selected(
+                candidate
+            )
+
+            if obj < best_obj:
+                best_obj = obj
+                best_sel = candidate
+
+        # Add.
+        for aid in unselected_ids:
+            candidate = set(selected)
+            candidate.add(aid)
+
+            if len(candidate) >= len(blocks):
+                continue
+
+            _vc, _vs, obj = calc_for_selected(
+                candidate
+            )
+
+            if obj < best_obj:
+                best_obj = obj
+                best_sel = candidate
+
+        # Swap.
+        # Para datasets enormes limita a vizinhança de busca.
+        max_side = 100
+
+        for rid in selected_ids[:max_side]:
+            for aid in unselected_ids[:max_side]:
+                candidate = set(selected)
+                candidate.remove(rid)
+                candidate.add(aid)
+
+                if (
+                    not candidate
+                    or len(candidate) >= len(blocks)
+                ):
+                    continue
+
+                _vc, _vs, obj = calc_for_selected(
+                    candidate
+                )
+
+                if obj < best_obj:
+                    best_obj = obj
+                    best_sel = candidate
+
+        if best_sel == selected:
+            break
+
+        selected = best_sel
+        val_count, val_strata, current_obj = calc_for_selected(
+            selected
+        )
+
+    # Garante os dois lados mesmo em situação patológica.
+    if len(selected) >= len(blocks):
+        smallest = min(
+            (
+                b
+                for b in blocks
+                if b.block_id in selected
+            ),
+            key=lambda b: b.size,
+        )
+        selected.remove(
+            smallest.block_id
+        )
+
+    val_count, val_strata, current_obj = calc_for_selected(
+        selected
+    )
+
+    missing_final = coverage_missing(
+        val_strata
+    )
+
+    info = {
+        "target_total": int(target_total),
+        "actual_total": int(val_count),
+
+        "target_strata": {
+            stratum_key(g, lid): int(v)
+            for (g, lid), v in sorted(
+                target_strata.items()
+            )
+        },
+
+        "actual_strata": {
+            stratum_key(g, lid): int(v)
+            for (g, lid), v in sorted(
+                val_strata.items()
+            )
+        },
+
+        "coverage_feasible_strata": [
+            stratum_key(g, lid)
+            for (g, lid) in sorted(
+                feasible_both
+            )
+        ],
+
+        "coverage_missing": [
+            {
+                "stratum": stratum_key(
+                    st[0],
+                    st[1],
+                ),
+                "missing_side": side,
+            }
+            for st, side in missing_final
+        ],
+
+        "objective": {
+            "coverage_missing_count": int(
+                current_obj[0]
+            ),
+            "distribution_loss": float(
+                current_obj[1]
+            ),
+            "size_error": int(
+                current_obj[2]
+            ),
+        },
+    }
+
+    return selected, info
+
+
+# =============================================================================
+# Distribution audit
+# =============================================================================
+
+def distribution(samples: Sequence[Sample], states: Sequence[str]) -> dict:
+    by_group = Counter(s.group for s in samples)
+    by_label = Counter(s.label_id for s in samples)
+    by_joint = Counter(s.stratum for s in samples)
+
+    total = len(samples)
+
+    return {
+        "total": total,
+
+        "by_group": {
+            group: {
+                "count": int(count),
+                "pct": float(count * 100.0 / max(1, total)),
+            }
+            for group, count in sorted(by_group.items())
+        },
+
+        "by_label": {
+            str(label_id): {
+                "name": str(states[label_id]),
+                "count": int(count),
+                "pct": float(count * 100.0 / max(1, total)),
+            }
+            for label_id, count in sorted(by_label.items())
+        },
+
+        "by_group_label": {
+            stratum_key(group, label_id): {
+                "group": group,
+                "label_id": int(label_id),
+                "label_name": str(states[label_id]),
+                "count": int(count),
+                "pct": float(count * 100.0 / max(1, total)),
+            }
+            for (group, label_id), count in sorted(by_joint.items())
+        },
+    }
+
+
+def rare_strata_warnings(
+    all_samples: Sequence[Sample],
+    train_samples: Sequence[Sample],
+    val_samples: Sequence[Sample],
+    states: Sequence[str],
+) -> List[str]:
+    total = Counter(s.stratum for s in all_samples)
+    train = Counter(s.stratum for s in train_samples)
+    val = Counter(s.stratum for s in val_samples)
+
+    warnings = []
+
+    for (group, label_id), count in sorted(total.items()):
+        t = int(train.get((group, label_id), 0))
+        v = int(val.get((group, label_id), 0))
+
+        if count == 1:
+            warnings.append(
+                f"{group} × {states[label_id]} possui somente 1 amostra; "
+                f"ficou em {'train' if t else 'val'}."
+            )
+        elif t == 0 or v == 0:
+            warnings.append(
+                f"{group} × {states[label_id]} possui {count} amostras, "
+                f"mas distribuição ficou train={t}, val={v}. "
+                "Blocos temporais podem impedir divisão sem vazamento."
+            )
+
+    return warnings
+
+
+# =============================================================================
+# Copy split
+# =============================================================================
+
+def prepare_split_dirs(split_root: Path, clean: bool) -> Dict[str, Dict[str, Path]]:
+    if clean and split_root.exists():
+        shutil.rmtree(split_root)
+
+    result: Dict[str, Dict[str, Path]] = {}
+
+    for split in ("train", "val"):
+        root = split_root / split
+
+        result[split] = {
+            "root": root,
+            "images": root / "images",
+            "masks": root / "masks",
+            "labels": root / "labels",
+        }
+
+        for key in ("images", "masks", "labels"):
+            result[split][key].mkdir(parents=True, exist_ok=True)
+
+    return result
+
+
+def copy_sample_to_split(
+    sample: Sample,
+    split: str,
+    dirs: Dict[str, Dict[str, Path]],
+    seed: int,
+    val_ratio: float,
+    source_root: Path,
+) -> dict:
+    target = dirs[split]
+
+    out_image = target["images"] / sample.image.name
+    out_mask = target["masks"] / f"{sample.base}.png"
+    out_label = target["labels"] / f"{sample.base}.json"
+
+    shutil.copy2(sample.image, out_image)
+    shutil.copy2(sample.mask, out_mask)
+
+    with sample.label.open("r", encoding="utf-8") as f:
+        label = json.load(f)
+
+    label = dict(label)
+    label.update({
+        "split": split,
+        "split_at": now_iso(),
+        "split_seed": int(seed),
+        "split_val_ratio": float(val_ratio),
+        "split_source_group": sample.group,
+
+        "image": f"images/{out_image.name}",
+        "mask": f"masks/{out_mask.name}",
+        "label": f"labels/{out_label.name}",
+
+        "source_normalized_image": safe_rel(sample.image, source_root),
+        "source_normalized_mask": safe_rel(sample.mask, source_root),
+        "source_normalized_label": safe_rel(sample.label, source_root),
+    })
+
+    atomic_write_json(out_label, label)
+
+    return {
+        "base": sample.base,
+        "split": split,
+        "group": sample.group,
+        "label_id": sample.label_id,
+        "label_name": sample.label_name,
+        "timestamp_s": sample.timestamp_s,
+        "image": str(out_image),
+        "mask": str(out_mask),
+        "label": str(out_label),
+    }
+
+
+# =============================================================================
+# Norm stats, TRAIN ONLY
+# =============================================================================
+
+def compute_train_norm_stats(
+    train_samples: Sequence[Sample],
+) -> dict:
+    """
+    Calcula estatísticas RGB em [0,1] sobre TODAS as imagens de train.
+
+    Acumulação em float64 para evitar perda numérica.
+    """
+    channel_sum = np.zeros(3, dtype=np.float64)
+    channel_sumsq = np.zeros(3, dtype=np.float64)
+    pixels = 0
+
+    shape_counts = Counter()
+
+    for i, sample in enumerate(train_samples, start=1):
+        img_bgr = cv2.imread(str(sample.image), cv2.IMREAD_COLOR)
+
+        if img_bgr is None:
+            raise FileNotFoundError(
+                f"Não consegui abrir imagem de train para norm_stats: {sample.image}"
+            )
+
+        img_rgb = cv2.cvtColor(img_bgr, cv2.COLOR_BGR2RGB)
+        h, w, c = img_rgb.shape
+
+        if c != 3:
+            raise RuntimeError(
+                f"Imagem não RGB 3 canais: {sample.image} shape={img_rgb.shape}"
+            )
+
+        shape_counts[(w, h)] += 1
+
+        arr = img_rgb.astype(np.float64) / 255.0
+
+        channel_sum += arr.sum(axis=(0, 1))
+        channel_sumsq += np.square(arr).sum(axis=(0, 1))
+        pixels += h * w
+
+    if pixels <= 0:
+        raise RuntimeError("Train vazio; não é possível calcular norm_stats.")
+
+    mean = channel_sum / pixels
+    var = channel_sumsq / pixels - np.square(mean)
+    std = np.sqrt(np.maximum(var, 1e-12))
+
+    return {
+        "schema": NORM_SCHEMA,
+        "created_at": now_iso(),
+
+        "channels": ["R", "G", "B"],
+        "value_domain": "[0,1]",
+
+        "mean": mean.tolist(),
+        "std": std.tolist(),
+
+        "image_count": len(train_samples),
+        "pixels_per_channel": int(pixels),
+
+        "source": "split/train/images_only",
+        "split_scope": "train_only",
+
+        "image_shapes_wh": {
+            f"{w}x{h}": int(count)
+            for (w, h), count in sorted(shape_counts.items())
+        },
+    }
+
+
+# =============================================================================
+# Temporal audit
+# =============================================================================
+
+def nearest_cross_split_time_gap(
+    train_samples: Sequence[Sample],
+    val_samples: Sequence[Sample],
+) -> Optional[float]:
+    train_ts = sorted(
+        float(s.timestamp_s)
+        for s in train_samples
+        if s.timestamp_s is not None
+    )
+    val_ts = sorted(
+        float(s.timestamp_s)
+        for s in val_samples
+        if s.timestamp_s is not None
+    )
+
+    if not train_ts or not val_ts:
+        return None
+
+    i = 0
+    j = 0
+    best = float("inf")
+
+    while i < len(train_ts) and j < len(val_ts):
+        best = min(best, abs(train_ts[i] - val_ts[j]))
+
+        if train_ts[i] < val_ts[j]:
+            i += 1
+        else:
+            j += 1
+
+    return float(best)
+
+
+# =============================================================================
+# Main
+# =============================================================================
+
+def main():
+    parser = argparse.ArgumentParser(
+        description=(
+            "Split inteligente grupo × label com proteção temporal "
+            "e norm_stats calculado somente no train."
+        )
+    )
+
+    parser.add_argument(
+        "--config",
+        default=str(CONFIG_PATH),
+    )
+
+    parser.add_argument(
+        "--val-ratio",
+        type=float,
+        default=DEFAULT_VAL_RATIO,
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=DEFAULT_SEED,
+    )
+
+    parser.add_argument(
+        "--session-gap-s",
+        type=float,
+        default=DEFAULT_SESSION_GAP_S,
+        help=(
+            "Somente para arquivos legados sem session id: "
+            "gap que define nova sessão de captura."
+        ),
+    )
+
+    parser.add_argument(
+        "--fallback-block-size",
+        type=int,
+        default=DEFAULT_FALLBACK_BLOCK_SIZE,
+        help="Tamanho do bloco para nomes sem timestamp reconhecível.",
+    )
+
+    parser.add_argument(
+        "--no-clean",
+        action="store_true",
+        help="Não apaga split anterior antes de gerar.",
+    )
+
+    args = parser.parse_args()
+
+    if not (0.0 < args.val_ratio < 0.5):
+        raise ValueError("--val-ratio deve ficar entre 0 e 0.5")
+
+    if args.session_gap_s <= 0:
+        raise ValueError("--session-gap-s deve ser > 0")
+
+    config_path = Path(args.config)
+
+    if not config_path.is_file():
+        raise FileNotFoundError(f"config não encontrado: {config_path}")
+
+    with config_path.open("r", encoding="utf-8") as f:
+        config = json.load(f)
+
+    resolution = config.get("resolucao")
+
+    if (
+        not isinstance(resolution, list)
+        or len(resolution) != 2
+    ):
+        raise RuntimeError("config['resolucao'] deve ser [W,H]")
+
+    W = int(resolution[0])
+    H = int(resolution[1])
+
+    states = config.get("label_classes")
+
+    if not isinstance(states, list) or not states:
+        raise RuntimeError("config['label_classes'] obrigatório.")
+
+    normalized_root = DATASET_ROOT / f"{W}x{H}"
+    group_root = normalized_root / "group"
+    split_root = normalized_root / "split"
+
+    samples = collect_samples(
+        group_root,
+        states,
+    )
+
+    blocks, temporal_report = build_capture_sessions(
+        samples,
+        session_gap_s=float(args.session_gap_s),
+        fallback_block_size=max(1, int(args.fallback_block_size)),
+    )
+
+    if len(blocks) < 2:
+        raise RuntimeError(
+            "O dataset possui menos de 2 sessões independentes de captura. "
+            "Não vou separar frames da mesma sessão entre train e val, pois isso "
+            "contaminaria a validação. Colete pelo menos mais uma sessão."
+        )
+
+    val_block_ids, split_algorithm = choose_val_blocks(
+        blocks,
+        samples,
+        val_ratio=float(args.val_ratio),
+        seed=int(args.seed),
+    )
+
+    train_samples: List[Sample] = []
+    val_samples: List[Sample] = []
+
+    block_by_base = {}
+
+    for block in blocks:
+        target = (
+            val_samples
+            if block.block_id in val_block_ids
+            else train_samples
+        )
+
+        target.extend(block.samples)
+
+        for sample in block.samples:
+            block_by_base[sample.base] = block.block_id
+
+    # Ordem estável.
+    train_samples.sort(key=lambda s: natural_key(s.base))
+    val_samples.sort(key=lambda s: natural_key(s.base))
+
+    if not train_samples or not val_samples:
+        raise RuntimeError(
+            f"Split inválido: train={len(train_samples)} val={len(val_samples)}"
+        )
+
+    dirs = prepare_split_dirs(
+        split_root,
+        clean=not args.no_clean,
+    )
+
+    print("=" * 88)
+    print("OAK-D Lite | Intelligent Corridor Split")
+    print("=" * 88)
+    print(f"Fonte normalizada : {group_root.resolve()}")
+    print(f"Saída split       : {split_root.resolve()}")
+    print(f"Total             : {len(samples)}")
+    print(f"Train             : {len(train_samples)} ({len(train_samples)/len(samples)*100:.1f}%)")
+    print(f"Val               : {len(val_samples)} ({len(val_samples)/len(samples)*100:.1f}%)")
+    print(f"Val alvo          : {args.val_ratio*100:.1f}%")
+    print(f"Seed              : {args.seed}")
+    print(f"Sessões de captura: {len(blocks)}")
+    print(
+        f"Política temporal : sessão inteira em apenas um split "
+        f"(legado gap>{args.session_gap_s:.1f}s)"
+    )
+    print("=" * 88)
+
+    train_set = {s.base for s in train_samples}
+    val_set = {s.base for s in val_samples}
+
+    manifest_rows = []
+
+    for sample in samples:
+        split = "train" if sample.base in train_set else "val"
+
+        row = copy_sample_to_split(
+            sample,
+            split=split,
+            dirs=dirs,
+            seed=int(args.seed),
+            val_ratio=float(args.val_ratio),
+            source_root=normalized_root,
+        )
+
+        row["capture_session_block"] = int(block_by_base[sample.base])
+        row["capture_session_id"] = sample.capture_session_id or ""
+        manifest_rows.append(row)
+
+    # Norm stats usa imagens de origem normalizada dos samples TRAIN.
+    norm_stats = compute_train_norm_stats(
+        train_samples
+    )
+
+    norm_stats.update({
+        "resolution_wh": [W, H],
+        "split_seed": int(args.seed),
+        "split_val_ratio": float(args.val_ratio),
+        "train_count": len(train_samples),
+        "val_count": len(val_samples),
+    })
+
+    norm_path = normalized_root / "norm_stats.json"
+    atomic_write_json(
+        norm_path,
+        norm_stats,
+    )
+
+    dist_all = distribution(samples, states)
+    dist_train = distribution(train_samples, states)
+    dist_val = distribution(val_samples, states)
+
+    warnings = rare_strata_warnings(
+        samples,
+        train_samples,
+        val_samples,
+        states,
+    )
+
+    nearest_gap = nearest_cross_split_time_gap(
+        train_samples,
+        val_samples,
+    )
+
+    if (
+        nearest_gap is not None
+        and nearest_gap < args.session_gap_s
+    ):
+        warnings.append(
+            f"Menor distância temporal entre um frame train e val = "
+            f"{nearest_gap:.3f}s. Isso é menor que session_gap_s="
+            f"{args.session_gap_s:.3f}s. Se os arquivos usam session id explícito, "
+            "isso significa que duas execuções distintas do capture ocorreram muito "
+            "próximas no tempo; revise se são realmente sessões independentes."
+        )
+
+    report = {
+        "schema": REPORT_SCHEMA,
+        "created_at": now_iso(),
+
+        "config_path": str(config_path),
+        "normalized_root": str(normalized_root),
+        "group_root": str(group_root),
+        "split_root": str(split_root),
+
+        "resolution_wh": [W, H],
+
+        "seed": int(args.seed),
+        "val_ratio_requested": float(args.val_ratio),
+        "val_ratio_actual": float(len(val_samples) / len(samples)),
+
+        "total_count": len(samples),
+        "train_count": len(train_samples),
+        "val_count": len(val_samples),
+
+        "temporal_blocks": temporal_report,
+        "split_algorithm": split_algorithm,
+
+        "distribution": {
+            "all": dist_all,
+            "train": dist_train,
+            "val": dist_val,
+        },
+
+        "nearest_train_val_timestamp_gap_s": nearest_gap,
+
+        "warnings": warnings,
+
+        "norm_stats": {
+            "path": str(norm_path),
+            "scope": "train_only",
+            "channels": norm_stats["channels"],
+            "mean": norm_stats["mean"],
+            "std": norm_stats["std"],
+            "pixels_per_channel": norm_stats["pixels_per_channel"],
+        },
+    }
+
+    report_path = normalized_root / "split_report.json"
+
+    atomic_write_json(
+        report_path,
+        report,
+    )
+
+    manifest_path = normalized_root / "split_manifest.csv"
+
+    atomic_write_csv(
+        manifest_path,
+        sorted(manifest_rows, key=lambda r: natural_key(r["base"])),
+        fieldnames=[
+            "base",
+            "split",
+            "group",
+            "label_id",
+            "label_name",
+            "capture_session_block",
+            "capture_session_id",
+            "timestamp_s",
+            "image",
+            "mask",
+            "label",
+        ],
+    )
+
+    # -------------------------------------------------------------
+    # Human-readable console audit
+    # -------------------------------------------------------------
+
+    print()
+    print("DISTRIBUIÇÃO POR GRUPO")
+    print("-" * 88)
+
+    all_groups = sorted(
+        set(dist_all["by_group"])
+        | set(dist_train["by_group"])
+        | set(dist_val["by_group"])
+    )
+
+    for group in all_groups:
+        a = dist_all["by_group"].get(group, {"count": 0})["count"]
+        t = dist_train["by_group"].get(group, {"count": 0})["count"]
+        v = dist_val["by_group"].get(group, {"count": 0})["count"]
+
+        print(
+            f"{group:32s} total={a:5d} | train={t:5d} | val={v:5d}"
+        )
+
+    print()
+    print("DISTRIBUIÇÃO POR ESTADO")
+    print("-" * 88)
+
+    for label_id, state in enumerate(states):
+        key = str(label_id)
+
+        a = dist_all["by_label"].get(key, {"count": 0})["count"]
+        t = dist_train["by_label"].get(key, {"count": 0})["count"]
+        v = dist_val["by_label"].get(key, {"count": 0})["count"]
+
+        print(
+            f"{label_id:2d} {str(state):28s} "
+            f"total={a:5d} | train={t:5d} | val={v:5d}"
+        )
+
+    print()
+    print("DISTRIBUIÇÃO CONJUNTA grupo × estado")
+    print("-" * 88)
+
+    joint_keys = sorted(dist_all["by_group_label"])
+
+    for key in joint_keys:
+        info = dist_all["by_group_label"][key]
+        group = info["group"]
+        label_id = int(info["label_id"])
+        state = info["label_name"]
+
+        a = info["count"]
+        t = dist_train["by_group_label"].get(key, {"count": 0})["count"]
+        v = dist_val["by_group_label"].get(key, {"count": 0})["count"]
+
+        print(
+            f"{group:28s} × {state:24s} "
+            f"total={a:4d} train={t:4d} val={v:4d}"
+        )
+
+    print()
+    print("NORM STATS - TRAIN ONLY")
+    print("-" * 88)
+    print(f"Arquivo : {norm_path.resolve()}")
+    print(f"Mean RGB: {[round(x, 8) for x in norm_stats['mean']]}")
+    print(f"Std  RGB: {[round(x, 8) for x in norm_stats['std']]}")
+
+    if nearest_gap is not None:
+        print(f"Menor gap temporal train↔val: {nearest_gap:.3f}s")
+
+    if warnings:
+        print()
+        print("AVISOS")
+        print("-" * 88)
+        for warning in warnings:
+            print(f"[WARN] {warning}")
+
+    print()
+    print("=" * 88)
+    print("SPLIT CONCLUÍDO")
+    print("=" * 88)
+    print(f"Train          : {dirs['train']['root'].resolve()}")
+    print(f"Val            : {dirs['val']['root'].resolve()}")
+    print(f"norm_stats     : {norm_path.resolve()}")
+    print(f"split_report   : {report_path.resolve()}")
+    print(f"split_manifest : {manifest_path.resolve()}")
+    print("=" * 88)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/Python/OAK/datasets/oak-d/_4_train_corridor.py b/Python/OAK/datasets/oak-d/_4_train_corridor.py
new file mode 100644
index 000000000..22cf1d95e
--- /dev/null
+++ b/Python/OAK/datasets/oak-d/_4_train_corridor.py
@@ -0,0 +1,4198 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+"""
+Agri Corridor Teacher v2
+========================
+
+Trainer especializado para a camera frontal OAK-D Lite do robo agricola.
+
+Tarefa operacional:
+    1) Segmentacao binaria:
+       - area nao navegavel
+       - area navegavel
+
+    2) Classificacao global do estado do corredor:
+       - classes declaradas em config["label_classes"]
+
+Filosofia:
+    - nao tenta ser um trainer generico;
+    - prioriza robustez em campo;
+    - augmentation preserva a semantica global do corredor;
+    - loss de segmentacao combina CE + Dice + borda + penalidade de falso navegavel;
+    - status head trabalha em features nativas e resumo espacial da segmentacao,
+      sem fazer upsample gigante de features profundas;
+    - sampler ajuda a balancear estados raros sem explodir repeticoes;
+    - LR discriminativo entre encoder, decoder e status head;
+    - warmup + polynomial decay;
+    - AMP, grad accumulation e gradient clipping;
+    - checkpoints separados para navegacao, status e score operacional;
+    - best_operational so nasce quando a epoca passa pelos gates de campo;
+    - consome SOMENTE o split final em dataset/<W>x<H>/split;
+    - norm_stats deve vir do split e ser train-only;
+    - sampler conjunto grupo-da-mascara x estado global;
+    - preflight audita paths, shapes, IDs, labels e vazamento de stems;
+    - metricas de validacao tambem sao quebradas por grupo da mascara;
+    - checkpoint salva o contrato necessario para exportacao ONNX posterior;
+    - persiste historico completo por epoca em JSONL + JSON + CSV.
+
+Dependencias:
+    numpy
+    pillow
+    torch
+    transformers
+    roi_seg_dataset.ROISegDataset
+
+Exemplo:
+    python train_corridor_agri_v2.py ^
+        --config config.json ^
+        --epochs 140 ^
+        --batch 4 ^
+        --num_workers 4 ^
+        --amp ^
+        --amp_val ^
+        --grad_accum 2
+"""
+
+from __future__ import annotations
+
+import argparse
+import copy
+import csv
+import json
+import math
+import os
+import random
+import time
+from collections import Counter
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Any, Dict, List, Optional, Sequence, Tuple
+
+import cv2
+import numpy as np
+from PIL import Image
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.cuda.amp import GradScaler
+from torch.utils.data import DataLoader, WeightedRandomSampler
+from transformers import SegformerForSemanticSegmentation
+
+from roi_seg_dataset import ROISegDataset
+
+
+# =============================================================================
+# Contrato / defaults
+# =============================================================================
+
+TRAINER_VERSION = "agri_corridor_teacher_v2.1_operational_gate"
+IGNORE_INDEX = 255
+LABEL_IGNORE_INDEX = -100
+
+DEFAULT_TRAINING = {
+    "data": {
+        # Campo: label de status ausente normalmente significa dataset inconsistente.
+        "strict_labels": True,
+
+        # O pipeline novo exige split/ e norm_stats produzidos pela etapa _3_split.
+        "strict_pipeline_contract": True,
+
+        # Quantas amostras de cada split auditar profundamente antes do treino.
+        # 0 = todas.
+        "preflight_max_samples": 0,
+    },
+
+    "augmentation": {
+        "enabled": True,
+
+        # Geometria segura para corredor.
+        # Horizontal flip e pequenas perturbacoes preservam o significado global.
+        "horizontal_flip_p": 0.50,
+        "affine_p": 0.60,
+        "rotate_deg": 4.0,
+        "scale_min": 0.92,
+        "scale_max": 1.08,
+        "translate_frac": 0.03,
+
+        # Fotometria pensada para sol, sombra e autoexposure no campo.
+        "global_gain_p": 0.45,
+        "global_gain_min": 0.82,
+        "global_gain_max": 1.18,
+
+        "contrast_p": 0.35,
+        "contrast_min": 0.85,
+        "contrast_max": 1.15,
+
+        "gamma_p": 0.30,
+        "gamma_min": 0.82,
+        "gamma_max": 1.18,
+
+        "rgb_gain_p": 0.25,
+        "rgb_gain_min": 0.94,
+        "rgb_gain_max": 1.06,
+
+        "shadow_p": 0.35,
+        "shadow_strength_min": 0.12,
+        "shadow_strength_max": 0.42,
+
+        "noise_p": 0.18,
+        "noise_sigma_min": 0.003,
+        "noise_sigma_max": 0.018,
+
+        "blur_p": 0.12,
+        "blur_kernel": 3,
+
+        # Nao habilitado por padrao. Pode ser usado futuramente para lente suja/folha.
+        "occlusion_p": 0.00,
+        "occlusion_min_frac": 0.04,
+        "occlusion_max_frac": 0.15,
+
+        # Comeca leve e chega a 100% da intensidade.
+        "ramp_epochs": 6,
+    },
+
+    "sampler": {
+        # O dataset possui dois eixos pedagogicos:
+        #   1) composicao da mascara: navegavel / nao navegavel / misto
+        #   2) estado global: EntrandoRua / CaminhandoRua / ...
+        #
+        # "joint" balanceia a combinacao grupo x estado.
+        # Se metadado de grupo faltar, cai automaticamente para status-only.
+        "mode": "joint",
+        "joint_power": 0.30,
+        "status_power": 0.12,
+        "group_power": 0.10,
+        "max_weight_ratio": 5.0,
+        "samples_per_epoch": 0,
+    },
+
+    "class_weighting": {
+        # Segmentacao.
+        "seg_method": "log_inverse",
+        "seg_log_offset": 1.02,
+        "seg_min_weight": 0.35,
+        "seg_max_weight": 3.0,
+        "seg_max_samples": 1200,
+
+        # Status.
+        "status_enabled": True,
+        "status_power": 0.30,
+        "status_min_weight": 0.50,
+        "status_max_weight": 3.0,
+    },
+
+    "loss": {
+        # Segmentacao.
+        "seg_ce_weight": 0.65,
+        "seg_dice_weight": 0.35,
+
+        # Aumenta a atencao na fronteira corredor / nao corredor.
+        "boundary_boost": 0.15,
+
+        # Penaliza confianca navegavel sobre GT nao navegavel.
+        # Mantido pequeno para nao ensinar o modelo a "ter medo de tudo".
+        "unsafe_nav_weight": 0.06,
+
+        # Status global.
+        "status_weight": 0.40,
+        "status_ramp_epochs": 8,
+        "status_label_smoothing": 0.03,
+    },
+
+    "status_head": {
+        # Resumo espacial compacto. Preserva geometria global sem upsample de feature.
+        "pool_h": 3,
+        "pool_w": 4,
+        "hidden": 256,
+        "dropout": 0.20,
+
+        # Evita que a loss de status force o decoder semantico a "codificar status".
+        # A label head ainda ensina o encoder atraves de feat.
+        "detach_seg_summary": True,
+
+        # Usa probabilidade, nao logits crus, como resumo da geometria semantica.
+        "seg_summary": "probabilities",
+    },
+
+    "optimizer": {
+        "encoder_lr_mult": 0.50,
+        "decoder_lr_mult": 1.00,
+        "status_lr_mult": 2.00,
+        "weight_decay": None,  # None => usa --wd
+        "no_decay_bias": True,
+        "no_decay_norm": True,
+        "betas": [0.9, 0.999],
+        "eps": 1e-8,
+
+        # Primeiro deixa heads se organizarem sem mexer no backbone inteiro.
+        "freeze_encoder_epochs": 2,
+    },
+
+    "scheduler": {
+        "mode": "poly",
+        "warmup_ratio": 0.05,
+        "warmup_start_factor": 0.10,
+        "poly_power": 1.0,
+        "min_lr_ratio": 0.02,
+    },
+
+    "optimization": {
+        "grad_clip_norm": 1.0,
+        "matmul_precision": "high",
+        "cudnn_benchmark": True,
+        "persistent_workers": True,
+        "prefetch_factor": 2,
+    },
+
+    "score": {
+        # O FIELD_SCORE RANQUEIA somente os modelos que passaram pelo gate.
+        "nav_iou": 0.40,
+        "nav_f1": 0.20,
+        "status_macro_f1": 0.20,
+        "safety": 0.20,
+    },
+
+    "operational_gate": {
+        # 1) elegibilidade para campo; 2) FIELD_SCORE escolhe o melhor elegivel.
+        "enabled": True,
+        "min_epoch": 5,
+        "min_nav_iou": 0.70,
+        "min_nav_f1": 0.80,
+        "min_non_nav_iou": 0.60,
+        "min_status_macro_f1": 0.60,
+        "max_unsafe_nav_rate": 0.03,
+
+        "group_unsafe": {
+            "enabled": True,
+            "min_non_nav_pixels": 5000,
+            "max_rate": 0.05,
+            "max_rate_by_group": {
+                "naonavegavel": 0.04,
+                "navegavel_naonavegavel": 0.05
+            }
+        },
+
+        "status_recall": {
+            "enabled": False,
+            "min_support": 5,
+            "minimums": {}
+        }
+    },
+
+    "checkpoint": {
+        "early_stop_patience": 18,
+        "early_stop_min_delta": 3e-4,
+    },
+}
+
+
+# =============================================================================
+# Utilitarios
+# =============================================================================
+
+def deep_merge(base: dict, override: Optional[dict]) -> dict:
+    out = copy.deepcopy(base)
+    if not isinstance(override, dict):
+        return out
+
+    def rec(dst: dict, src: dict):
+        for k, v in src.items():
+            if isinstance(v, dict) and isinstance(dst.get(k), dict):
+                rec(dst[k], v)
+            else:
+                dst[k] = copy.deepcopy(v)
+
+    rec(out, override)
+    return out
+
+
+def seed_everything(seed: int = 42):
+    random.seed(seed)
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    torch.cuda.manual_seed_all(seed)
+
+
+def seed_worker(worker_id: int):
+    seed = torch.initial_seed() % (2 ** 32)
+    np.random.seed(seed)
+    random.seed(seed)
+
+
+def capture_rng_state() -> dict:
+    state = {
+        "python": random.getstate(),
+        "numpy": np.random.get_state(),
+        "torch": torch.get_rng_state(),
+    }
+    if torch.cuda.is_available():
+        state["cuda"] = torch.cuda.get_rng_state_all()
+    return state
+
+
+def restore_rng_state(state: Optional[dict]):
+    if not state:
+        return
+    if "python" in state:
+        random.setstate(state["python"])
+    if "numpy" in state:
+        np.random.set_state(state["numpy"])
+    if "torch" in state:
+        torch.set_rng_state(state["torch"])
+    if torch.cuda.is_available() and "cuda" in state:
+        torch.cuda.set_rng_state_all(state["cuda"])
+
+
+def ensure_img_tensor(img: Any) -> torch.Tensor:
+    if isinstance(img, Image.Image):
+        img = np.asarray(img)
+
+    if isinstance(img, np.ndarray):
+        img = torch.from_numpy(img)
+
+    if not isinstance(img, torch.Tensor):
+        raise TypeError(f"Imagem com tipo inesperado: {type(img)}")
+
+    if img.ndim == 3 and img.shape[-1] == 3:
+        img = img.permute(2, 0, 1)
+
+    if img.ndim != 3 or img.shape[0] != 3:
+        raise RuntimeError(f"Imagem deve ser CHW RGB, recebido shape={tuple(img.shape)}")
+
+    img = img.float()
+    if img.numel() > 0 and float(img.max()) > 1.5:
+        img = img / 255.0
+
+    return img.contiguous().clamp(0.0, 1.0)
+
+
+def ensure_mask_tensor(mask: Any) -> torch.Tensor:
+    if isinstance(mask, Image.Image):
+        mask = np.asarray(mask)
+
+    if isinstance(mask, np.ndarray):
+        mask = torch.from_numpy(mask)
+
+    if not isinstance(mask, torch.Tensor):
+        raise TypeError(f"Mask com tipo inesperado: {type(mask)}")
+
+    if mask.ndim == 3:
+        if mask.shape[0] == 1:
+            mask = mask[0]
+        elif mask.shape[-1] == 1:
+            mask = mask[..., 0]
+        else:
+            mask = mask[..., 0]
+
+    return mask.long().contiguous()
+
+
+def read_labelmap(path: Path) -> Tuple[Dict[int, str], Dict[str, int]]:
+    if not path.is_file():
+        raise FileNotFoundError(f"labelmap nao encontrado: {path}")
+
+    id2label: Dict[int, str] = {}
+
+    with path.open("r", encoding="utf-8") as f:
+        for raw in f:
+            line = raw.strip()
+            if not line or line.startswith("#"):
+                continue
+
+            parts = [p.strip() for p in line.replace("=", ":").split(":") if p.strip()]
+
+            if len(parts) >= 2:
+                try:
+                    cid = int(parts[0])
+                    name = parts[1]
+                except ValueError:
+                    cid = len(id2label)
+                    name = parts[0]
+            else:
+                cid = len(id2label)
+                name = parts[0]
+
+            if name.lower() in ("ignore", "void", "background_ignore"):
+                continue
+
+            id2label[int(cid)] = str(name)
+
+    label2id = {name.lower(): cid for cid, name in id2label.items()}
+    return id2label, label2id
+
+
+def resolve_project_root(
+    config_path: Path,
+    camera_value: Optional[str] = None,
+) -> Path:
+    """
+    O pipeline 2026-09 roda de DENTRO da pasta da camera (oak-d).
+
+    Regra principal:
+        config.json.parent precisa conter dataset/.
+
+    Fallback apenas para compatibilidade:
+        tenta camera_value se o config for chamado de outro local.
+
+    Nao cria caminhos do tipo oak-d/oak-d.
+    """
+    primary = config_path.parent.resolve()
+
+    if (primary / "dataset").is_dir():
+        return primary
+
+    candidates: List[Path] = []
+
+    if camera_value:
+        raw = Path(str(camera_value))
+
+        if raw.is_absolute():
+            candidates.append(raw.resolve())
+        else:
+            candidates.extend([
+                (Path.cwd() / raw).resolve(),
+                (config_path.parent / raw).resolve(),
+            ])
+
+    # cwd tambem pode ser a raiz, mesmo se config foi fornecido com caminho absoluto.
+    candidates.append(Path.cwd().resolve())
+
+    seen = set()
+
+    for cand in candidates:
+        key = str(cand)
+        if key in seen:
+            continue
+        seen.add(key)
+
+        if (cand / "dataset").is_dir():
+            return cand
+
+    pretty = "\n".join(f"  - {p}" for p in [primary] + candidates)
+
+    raise FileNotFoundError(
+        "Nao consegui resolver a raiz do projeto/camera. "
+        "O trainer novo deve ser executado dentro de oak-d, com dataset/ ao lado "
+        f"de config.json.\nCaminhos testados:\n{pretty}"
+    )
+
+
+def load_norm_stats(
+    path: Path,
+    device: torch.device,
+    expected_resolution_wh: Tuple[int, int],
+    strict_contract: bool = True,
+) -> Tuple[torch.Tensor, torch.Tensor, Path, dict]:
+    """
+    Carrega o norm_stats produzido pela etapa _3_split_corridor.py.
+
+    Contrato novo:
+        dataset/<W>x<H>/norm_stats.json
+        split_scope == "train_only"
+        channels == ["R","G","B"]
+        resolution_wh == [W,H]
+    """
+    path = Path(path)
+
+    if not path.is_file():
+        raise FileNotFoundError(
+            f"norm_stats.json obrigatorio nao encontrado: {path}\n"
+            "Execute _3_split_corridor.py antes do treino."
+        )
+
+    with path.open("r", encoding="utf-8") as f:
+        stats = json.load(f)
+
+    channels = [str(x).upper() for x in stats.get("channels", [])]
+    mean = list(stats.get("mean", []))
+    std = list(stats.get("std", []))
+
+    if channels != ["R", "G", "B"]:
+        raise ValueError(
+            f"norm_stats precisa ter channels=['R','G','B']; recebido={channels}"
+        )
+
+    if len(mean) != 3 or len(std) != 3:
+        raise ValueError(f"mean/std invalidos: mean={mean} std={std}")
+
+    if any((not np.isfinite(float(x))) for x in mean + std):
+        raise ValueError("norm_stats contem NaN/Inf.")
+
+    if any(float(x) <= 0 for x in std):
+        raise ValueError(f"std precisa ser > 0; recebido={std}")
+
+    split_scope = str(stats.get("split_scope", "")).strip().lower()
+
+    if strict_contract and split_scope != "train_only":
+        raise RuntimeError(
+            f"norm_stats fora do contrato novo: split_scope={split_scope!r}. "
+            "Esperado 'train_only'. Reexecute _3_split_corridor.py."
+        )
+
+    expected_wh = [int(expected_resolution_wh[0]), int(expected_resolution_wh[1])]
+    stats_wh = stats.get("resolution_wh")
+
+    if strict_contract and stats_wh is not None:
+        got_wh = [int(stats_wh[0]), int(stats_wh[1])]
+
+        if got_wh != expected_wh:
+            raise RuntimeError(
+                f"norm_stats resolution_wh={got_wh}, mas config.resolucao={expected_wh}."
+            )
+
+    value_domain = str(stats.get("value_domain", "[0,1]")).replace(" ", "")
+
+    if strict_contract and value_domain not in {"[0,1]", "0..1", "0-1"}:
+        raise RuntimeError(
+            f"norm_stats value_domain inesperado: {stats.get('value_domain')!r}"
+        )
+
+    mean_t = torch.tensor(
+        mean,
+        dtype=torch.float32,
+        device=device,
+    ).view(1, 3, 1, 1)
+
+    std_t = torch.tensor(
+        std,
+        dtype=torch.float32,
+        device=device,
+    ).view(1, 3, 1, 1).clamp_min(1e-6)
+
+    return mean_t, std_t, path, stats
+
+
+# =============================================================================
+# Dataset de corredor + status
+# =============================================================================
+
+class CorridorDataset(torch.utils.data.Dataset):
+    """
+    Wrapper do ROISegDataset para o contrato final do split.
+
+    Esperado:
+        split_root/
+          images/<base>.png
+          masks/<base>.png
+          labels/<base>.json
+
+    O JSON do split contem:
+        label_id
+        estado_corredor
+        split_source_group
+
+    Mantemos o grupo porque ele e um eixo pedagogico importante tanto para
+    sampler quanto para metricas de validacao.
+    """
+
+    def __init__(
+        self,
+        base_ds: ROISegDataset,
+        split_root: Path,
+        label_classes: Sequence[str],
+        strict_labels: bool = True,
+    ):
+        self.base_ds = base_ds
+        self.split_root = Path(split_root)
+        self.label_classes = [str(x) for x in label_classes]
+        self.num_label_classes = len(self.label_classes)
+        self.strict_labels = bool(strict_labels)
+
+        self.label_name_to_id = {
+            str(name).strip().lower(): i
+            for i, name in enumerate(self.label_classes)
+        }
+
+        self.label_ids: List[int] = []
+        self.label_paths: List[Optional[str]] = []
+        self.group_names: List[str] = []
+        self.group_ids: List[int] = []
+        self.missing_labels: List[str] = []
+
+        raw_groups: List[str] = []
+
+        for idx in range(len(self.base_ds)):
+            mask_path = self._get_mask_path(idx)
+            label_path = self._label_path_from_mask_path(mask_path)
+
+            label_id = LABEL_IGNORE_INDEX
+            label_name: Optional[str] = None
+            group_name = "unknown"
+
+            if label_path is None or not os.path.exists(label_path):
+                self.missing_labels.append(mask_path)
+            else:
+                label_id_opt, label_name, group_name = self._read_label(label_path)
+
+                if label_id_opt is None and label_name:
+                    label_id_opt = self.label_name_to_id.get(
+                        str(label_name).strip().lower()
+                    )
+
+                if label_id_opt is None:
+                    self.missing_labels.append(mask_path)
+                else:
+                    label_id = int(label_id_opt)
+
+                    if label_id < 0 or label_id >= self.num_label_classes:
+                        raise ValueError(
+                            f"label_id={label_id} fora do intervalo "
+                            f"[0,{self.num_label_classes - 1}] em {label_path}"
+                        )
+
+                    expected_name = self.label_classes[label_id]
+
+                    if (
+                        label_name is not None
+                        and str(label_name) != str(expected_name)
+                    ):
+                        raise ValueError(
+                            f"Label inconsistente em {label_path}: "
+                            f"label_id={label_id} implica {expected_name!r}, "
+                            f"mas nome={label_name!r}."
+                        )
+
+            self.label_ids.append(int(label_id))
+            self.label_paths.append(label_path)
+            raw_groups.append(str(group_name or "unknown"))
+
+        if self.strict_labels and self.missing_labels:
+            preview = "\n".join(
+                f"  - {p}"
+                for p in self.missing_labels[:10]
+            )
+
+            raise RuntimeError(
+                f"{len(self.missing_labels)} amostras sem label de status "
+                f"no split {self.split_root}.\nPrimeiras:\n{preview}\n"
+                "Corrija o split/dataset ou use --allow_missing_labels."
+            )
+
+        unique_groups = sorted(set(raw_groups))
+        self.group_name_by_id = {
+            i: name
+            for i, name in enumerate(unique_groups)
+        }
+        self.group_id_by_name = {
+            name: i
+            for i, name in self.group_name_by_id.items()
+        }
+
+        self.group_names = raw_groups
+        self.group_ids = [
+            self.group_id_by_name[g]
+            for g in raw_groups
+        ]
+
+    def __len__(self):
+        return len(self.base_ds)
+
+    def _get_mask_path(self, idx: int) -> str:
+        if hasattr(self.base_ds, "msk_paths"):
+            return str(self.base_ds.msk_paths[idx])
+
+        item = self.base_ds[idx]
+
+        if isinstance(item, dict):
+            for key in ("mask_path", "mask_file", "maskname", "mask_name"):
+                if item.get(key):
+                    return str(item[key])
+
+        raise RuntimeError("ROISegDataset nao expoe caminho da mask.")
+
+    def _label_path_from_mask_path(
+        self,
+        mask_path: str,
+    ) -> Optional[str]:
+        norm = os.path.normpath(mask_path)
+        parts = norm.split(os.sep)
+        stem = os.path.splitext(os.path.basename(mask_path))[0]
+
+        try:
+            i = parts.index("masks")
+            root = os.sep.join(parts[:i] + ["labels"])
+        except ValueError:
+            root = str(self.split_root / "labels")
+
+        # Pipeline novo usa JSON. NPY/TXT ficam apenas como compatibilidade.
+        for ext in (".json", ".npy", ".txt"):
+            candidate = os.path.join(root, stem + ext)
+            if os.path.exists(candidate):
+                return candidate
+
+        return os.path.join(root, stem + ".json")
+
+    @staticmethod
+    def _read_label(
+        path: str,
+    ) -> Tuple[Optional[int], Optional[str], str]:
+        ext = os.path.splitext(path)[1].lower()
+
+        if ext == ".npy":
+            value = np.load(path)
+            return int(np.asarray(value).reshape(-1)[0]), None, "unknown"
+
+        if ext == ".json":
+            with open(path, "r", encoding="utf-8") as f:
+                data = json.load(f)
+
+            label_id = data.get("label_id")
+
+            label_name = (
+                data.get("estado_corredor")
+                or data.get("label")
+                or data.get("state")
+            )
+
+            group_name = (
+                data.get("split_source_group")
+                or data.get("group")
+                or data.get("source_group")
+                or "unknown"
+            )
+
+            return (
+                int(label_id) if label_id is not None else None,
+                str(label_name) if label_name is not None else None,
+                str(group_name),
+            )
+
+        if ext == ".txt":
+            txt = Path(path).read_text(
+                encoding="utf-8"
+            ).strip()
+
+            try:
+                return int(txt), None, "unknown"
+            except ValueError:
+                return None, txt, "unknown"
+
+        return None, None, "unknown"
+
+    def __getitem__(self, idx: int):
+        item = self.base_ds[idx]
+
+        if (
+            isinstance(item, (tuple, list))
+            and len(item) == 1
+            and isinstance(item[0], dict)
+        ):
+            item = item[0]
+
+        if isinstance(item, dict):
+            img = item["image"]
+            mask = item["mask"]
+        elif isinstance(item, (tuple, list)) and len(item) >= 2:
+            img, mask = item[:2]
+        else:
+            raise RuntimeError(
+                f"Formato inesperado de item: {type(item)}"
+            )
+
+        return {
+            "image": ensure_img_tensor(img),
+            "mask": ensure_mask_tensor(mask),
+            "label": int(self.label_ids[idx]),
+            "group_id": int(self.group_ids[idx]),
+            "group_name": self.group_names[idx],
+            "label_path": self.label_paths[idx],
+            "mask_path": self._get_mask_path(idx),
+        }
+
+
+def collate_corridor(batch: Sequence[dict]):
+    imgs = torch.stack(
+        [x["image"] for x in batch],
+        dim=0,
+    )
+    masks = torch.stack(
+        [x["mask"] for x in batch],
+        dim=0,
+    )
+    labels = torch.tensor(
+        [int(x["label"]) for x in batch],
+        dtype=torch.long,
+    )
+    group_ids = torch.tensor(
+        [int(x["group_id"]) for x in batch],
+        dtype=torch.long,
+    )
+
+    return imgs, masks, labels, group_ids
+
+
+# =============================================================================
+# Augmentation de campo
+# =============================================================================
+
+class FieldCorridorAugmenter:
+    """
+    Augmentation sincronizado imagem/mask.
+
+    Importante:
+    - nao faz vertical flip;
+    - nao faz random crop por padrao;
+    - perturbacoes geometricas sao pequenas;
+    - status global permanece semanticamente valido.
+    """
+
+    def __init__(self, cfg: dict, ignore_index: int = IGNORE_INDEX):
+        self.cfg = cfg or {}
+        self.ignore_index = int(ignore_index)
+        self.strength = 1.0
+
+    def set_epoch(self, epoch: int):
+        ramp = int(self.cfg.get("ramp_epochs", 0))
+        if ramp <= 0:
+            self.strength = 1.0
+        else:
+            self.strength = min(1.0, max(0.20, float(epoch) / float(ramp)))
+
+    def __call__(
+        self,
+        imgs: torch.Tensor,
+        masks: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        if not bool(self.cfg.get("enabled", True)):
+            return imgs, masks
+
+        imgs = imgs.clone()
+        masks = masks.clone()
+
+        imgs, masks = self._geometry(imgs, masks)
+        imgs = self._photometric(imgs)
+
+        return imgs.clamp(0.0, 1.0), masks
+
+    def _geometry(
+        self,
+        imgs: torch.Tensor,
+        masks: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        B, _, H, W = imgs.shape
+        device = imgs.device
+
+        # Horizontal flip.
+        p_flip = float(self.cfg.get("horizontal_flip_p", 0.5)) * self.strength
+        do_flip = torch.rand(B, device=device) < p_flip
+
+        if bool(do_flip.any()):
+            imgs[do_flip] = torch.flip(imgs[do_flip], dims=[3])
+            masks[do_flip] = torch.flip(masks[do_flip], dims=[2])
+
+        # Affine leve.
+        p_aff = float(self.cfg.get("affine_p", 0.6)) * self.strength
+        do_aff = torch.rand(B, device=device) < p_aff
+
+        if not bool(do_aff.any()):
+            return imgs, masks
+
+        rotate_deg = float(self.cfg.get("rotate_deg", 4.0)) * self.strength
+
+        smin = float(self.cfg.get("scale_min", 0.92))
+        smax = float(self.cfg.get("scale_max", 1.08))
+        smin = 1.0 - (1.0 - smin) * self.strength
+        smax = 1.0 + (smax - 1.0) * self.strength
+
+        trans = float(self.cfg.get("translate_frac", 0.03)) * self.strength
+
+        n = int(do_aff.sum().item())
+        angle = (torch.rand(n, device=device) * 2.0 - 1.0) * math.radians(rotate_deg)
+        scale = torch.empty(n, device=device).uniform_(smin, smax)
+        tx = (torch.rand(n, device=device) * 2.0 - 1.0) * (2.0 * trans)
+        ty = (torch.rand(n, device=device) * 2.0 - 1.0) * (2.0 * trans)
+
+        cos = torch.cos(angle) * scale
+        sin = torch.sin(angle) * scale
+
+        theta = torch.zeros((n, 2, 3), dtype=imgs.dtype, device=device)
+        theta[:, 0, 0] = cos
+        theta[:, 0, 1] = -sin
+        theta[:, 1, 0] = sin
+        theta[:, 1, 1] = cos
+        theta[:, 0, 2] = tx
+        theta[:, 1, 2] = ty
+
+        img_sel = imgs[do_aff]
+        mask_sel = masks[do_aff].float().unsqueeze(1)
+
+        grid = F.affine_grid(
+            theta,
+            size=img_sel.size(),
+            align_corners=False,
+        )
+
+        img_aug = F.grid_sample(
+            img_sel,
+            grid,
+            mode="bilinear",
+            padding_mode="border",
+            align_corners=False,
+        )
+
+        mask_aug = F.grid_sample(
+            mask_sel,
+            grid,
+            mode="nearest",
+            padding_mode="zeros",
+            align_corners=False,
+        ).squeeze(1).long()
+
+        # Marca pixels fora da imagem como ignore, em vez de ensinar classe 0 falsa.
+        valid_src = torch.ones(
+            (n, 1, H, W),
+            dtype=imgs.dtype,
+            device=device,
+        )
+        valid_aug = F.grid_sample(
+            valid_src,
+            grid,
+            mode="nearest",
+            padding_mode="zeros",
+            align_corners=False,
+        ).squeeze(1) > 0.5
+
+        mask_aug[~valid_aug] = self.ignore_index
+
+        imgs[do_aff] = img_aug
+        masks[do_aff] = mask_aug
+
+        return imgs, masks
+
+    def _photometric(self, imgs: torch.Tensor) -> torch.Tensor:
+        B, C, H, W = imgs.shape
+        device = imgs.device
+        dtype = imgs.dtype
+
+        def choose(p: float) -> torch.Tensor:
+            return torch.rand(B, device=device) < (float(p) * self.strength)
+
+        # Ganho global / exposicao.
+        sel = choose(self.cfg.get("global_gain_p", 0.45))
+        if bool(sel.any()):
+            lo = float(self.cfg.get("global_gain_min", 0.82))
+            hi = float(self.cfg.get("global_gain_max", 1.18))
+            lo = 1.0 - (1.0 - lo) * self.strength
+            hi = 1.0 + (hi - 1.0) * self.strength
+            gain = torch.empty((int(sel.sum()), 1, 1, 1), device=device, dtype=dtype).uniform_(lo, hi)
+            imgs[sel] = imgs[sel] * gain
+
+        # Contraste.
+        sel = choose(self.cfg.get("contrast_p", 0.35))
+        if bool(sel.any()):
+            lo = float(self.cfg.get("contrast_min", 0.85))
+            hi = float(self.cfg.get("contrast_max", 1.15))
+            lo = 1.0 - (1.0 - lo) * self.strength
+            hi = 1.0 + (hi - 1.0) * self.strength
+            contrast = torch.empty((int(sel.sum()), 1, 1, 1), device=device, dtype=dtype).uniform_(lo, hi)
+            mean = imgs[sel].mean(dim=(2, 3), keepdim=True)
+            imgs[sel] = (imgs[sel] - mean) * contrast + mean
+
+        # Gamma.
+        sel = choose(self.cfg.get("gamma_p", 0.30))
+        if bool(sel.any()):
+            lo = float(self.cfg.get("gamma_min", 0.82))
+            hi = float(self.cfg.get("gamma_max", 1.18))
+            lo = 1.0 - (1.0 - lo) * self.strength
+            hi = 1.0 + (hi - 1.0) * self.strength
+            gamma = torch.empty((int(sel.sum()), 1, 1, 1), device=device, dtype=dtype).uniform_(lo, hi)
+            imgs[sel] = imgs[sel].clamp(1e-5, 1.0).pow(gamma)
+
+        # Pequeno desbalanceamento RGB.
+        sel = choose(self.cfg.get("rgb_gain_p", 0.25))
+        if bool(sel.any()):
+            lo = float(self.cfg.get("rgb_gain_min", 0.94))
+            hi = float(self.cfg.get("rgb_gain_max", 1.06))
+            lo = 1.0 - (1.0 - lo) * self.strength
+            hi = 1.0 + (hi - 1.0) * self.strength
+            gains = torch.empty((int(sel.sum()), C, 1, 1), device=device, dtype=dtype).uniform_(lo, hi)
+            imgs[sel] = imgs[sel] * gains
+
+        # Sombra direcional suave.
+        sel = choose(self.cfg.get("shadow_p", 0.35))
+        if bool(sel.any()):
+            yy = torch.linspace(-1.0, 1.0, H, device=device, dtype=dtype).view(H, 1)
+            xx = torch.linspace(-1.0, 1.0, W, device=device, dtype=dtype).view(1, W)
+
+            for bi in torch.where(sel)[0].tolist():
+                angle = random.uniform(0.0, math.pi * 2.0)
+                proj = math.cos(angle) * xx + math.sin(angle) * yy
+                proj = (proj - proj.min()) / (proj.max() - proj.min() + 1e-6)
+
+                lo = float(self.cfg.get("shadow_strength_min", 0.12))
+                hi = float(self.cfg.get("shadow_strength_max", 0.42))
+                strength = random.uniform(lo, hi) * self.strength
+
+                # Metade do gradiente recebe sombra; borda fica suave.
+                center = random.uniform(0.30, 0.70)
+                softness = random.uniform(0.12, 0.28)
+                shade = torch.sigmoid((proj - center) / max(softness, 1e-3))
+                factor = 1.0 - strength * shade
+                imgs[bi] = imgs[bi] * factor.unsqueeze(0)
+
+        # Ruido.
+        sel = choose(self.cfg.get("noise_p", 0.18))
+        if bool(sel.any()):
+            lo = float(self.cfg.get("noise_sigma_min", 0.003))
+            hi = float(self.cfg.get("noise_sigma_max", 0.018))
+            sigmas = torch.empty((int(sel.sum()), 1, 1, 1), device=device, dtype=dtype).uniform_(lo, hi)
+            sigmas = sigmas * self.strength
+            noise = torch.randn_like(imgs[sel]) * sigmas
+            imgs[sel] = imgs[sel] + noise
+
+        # Blur leve.
+        sel = choose(self.cfg.get("blur_p", 0.12))
+        if bool(sel.any()):
+            k = int(self.cfg.get("blur_kernel", 3))
+            if k % 2 == 0:
+                k += 1
+            k = max(3, k)
+            blurred = F.avg_pool2d(imgs[sel], kernel_size=k, stride=1, padding=k // 2)
+            imgs[sel] = blurred
+
+        # Oclusao opcional, desabilitada por padrao.
+        sel = choose(self.cfg.get("occlusion_p", 0.0))
+        if bool(sel.any()):
+            fmin = float(self.cfg.get("occlusion_min_frac", 0.04))
+            fmax = float(self.cfg.get("occlusion_max_frac", 0.15))
+
+            for bi in torch.where(sel)[0].tolist():
+                frac = random.uniform(fmin, fmax) * self.strength
+                rh = max(2, int(H * frac))
+                rw = max(2, int(W * frac))
+                y0 = random.randint(0, max(0, H - rh))
+                x0 = random.randint(0, max(0, W - rw))
+                factor = random.uniform(0.35, 0.75)
+                imgs[bi, :, y0:y0 + rh, x0:x0 + rw] *= factor
+
+        return imgs.clamp(0.0, 1.0)
+
+
+# =============================================================================
+# Status head otimizada
+# =============================================================================
+
+class CorridorStatusHead(nn.Module):
+    """
+    Classificacao global do estado do corredor.
+
+    Diferenca importante para a versao antiga:
+      - NAO faz upsample da feature profunda ate a resolucao da segmentacao;
+      - preserva estrutura espacial por pooling 3x4 (configuravel);
+      - resume separadamente feature profunda e probabilidade semantica;
+      - concatena vetores compactos antes do MLP.
+
+    Isso e muito mais barato e preserva informacao espacial util para distinguir:
+      EntrandoRua / CaminhandoRua / SaindoRua / Manobrando / etc.
+    """
+
+    def __init__(
+        self,
+        feat_ch: int,
+        num_seg_classes: int,
+        num_label_classes: int,
+        pool_hw: Tuple[int, int] = (3, 4),
+        hidden: int = 256,
+        dropout: float = 0.20,
+        detach_seg_summary: bool = True,
+        seg_summary: str = "probabilities",
+    ):
+        super().__init__()
+
+        self.feat_ch = int(feat_ch)
+        self.num_seg_classes = int(num_seg_classes)
+        self.num_label_classes = int(num_label_classes)
+        self.pool_hw = (int(pool_hw[0]), int(pool_hw[1]))
+        self.detach_seg_summary = bool(detach_seg_summary)
+        self.seg_summary = str(seg_summary).lower()
+
+        pooled_cells = self.pool_hw[0] * self.pool_hw[1]
+        in_dim = (self.feat_ch + self.num_seg_classes) * pooled_cells
+
+        self.feat_pool = nn.AdaptiveAvgPool2d(self.pool_hw)
+        self.seg_pool = nn.AdaptiveAvgPool2d(self.pool_hw)
+
+        self.net = nn.Sequential(
+            nn.Linear(in_dim, int(hidden)),
+            nn.ReLU(inplace=True),
+            nn.Dropout(float(dropout)),
+            nn.Linear(int(hidden), self.num_label_classes),
+        )
+
+    def forward(
+        self,
+        feat: torch.Tensor,
+        seg_logits_native: torch.Tensor,
+    ) -> torch.Tensor:
+        feat_grid = self.feat_pool(feat).flatten(1)
+
+        seg_src = seg_logits_native.detach() if self.detach_seg_summary else seg_logits_native
+
+        if self.seg_summary == "probabilities":
+            seg_src = torch.softmax(seg_src, dim=1)
+        elif self.seg_summary != "logits":
+            raise RuntimeError(
+                f"status_head.seg_summary invalido: {self.seg_summary}. "
+                "Use 'probabilities' ou 'logits'."
+            )
+
+        seg_grid = self.seg_pool(seg_src).flatten(1)
+        x = torch.cat([feat_grid, seg_grid], dim=1)
+        return self.net(x)
+
+    def export_config(self) -> dict:
+        return {
+            "feat_ch": self.feat_ch,
+            "num_seg_classes": self.num_seg_classes,
+            "num_label_classes": self.num_label_classes,
+            "pool_hw": list(self.pool_hw),
+            "detach_seg_summary": self.detach_seg_summary,
+            "seg_summary": self.seg_summary,
+        }
+
+
+# =============================================================================
+# Pesos de classes e sampler
+# =============================================================================
+
+def estimate_seg_class_weights(
+    ds: CorridorDataset,
+    num_classes: int,
+    ignore_index: int,
+    cfg: dict,
+) -> torch.Tensor:
+    """
+    Estima frequencia de classes lendo SOMENTE masks.
+
+    A versao v1 fazia ds[idx], carregando imagem inteira para contar pixels
+    da máscara. Aqui evitamos esse I/O desnecessario.
+    """
+    max_samples = int(cfg.get("seg_max_samples", 1200))
+    n = min(len(ds), max_samples)
+
+    if n <= 0:
+        return torch.ones(
+            num_classes,
+            dtype=torch.float32,
+        )
+
+    idxs = np.linspace(
+        0,
+        len(ds) - 1,
+        n,
+    ).astype(int)
+
+    counts = np.zeros(
+        num_classes,
+        dtype=np.float64,
+    )
+
+    for idx in idxs:
+        mask_path = ds._get_mask_path(int(idx))
+
+        arr = np.asarray(
+            Image.open(mask_path)
+        )
+
+        if arr.ndim == 3:
+            arr = arr[..., 0]
+
+        mask = arr.reshape(-1).astype(np.int64, copy=False)
+        mask = mask[mask != ignore_index]
+
+        if mask.size == 0:
+            continue
+
+        valid = (
+            (mask >= 0)
+            & (mask < num_classes)
+        )
+
+        invalid_count = int((~valid).sum())
+
+        if invalid_count:
+            bad = np.unique(mask[~valid])[:10].tolist()
+            raise RuntimeError(
+                f"Mask normalizada invalida {mask_path}: "
+                f"{invalid_count} pixels fora das classes; ids={bad}"
+            )
+
+        counts += np.bincount(
+            mask[valid],
+            minlength=num_classes,
+        )[:num_classes]
+
+    if counts.sum() <= 0:
+        return torch.ones(
+            num_classes,
+            dtype=torch.float32,
+        )
+
+    freq = counts / counts.sum()
+    freq = np.clip(freq, 1e-12, 1.0)
+
+    method = str(
+        cfg.get(
+            "seg_method",
+            "log_inverse",
+        )
+    ).lower()
+
+    if method == "log_inverse":
+        offset = float(
+            cfg.get(
+                "seg_log_offset",
+                1.02,
+            )
+        )
+        weights = 1.0 / np.log(
+            offset + freq
+        )
+    elif method == "sqrt_inverse":
+        weights = 1.0 / np.sqrt(freq)
+    else:
+        raise ValueError(
+            f"seg_method desconhecido: {method}"
+        )
+
+    weights = weights / max(
+        weights.mean(),
+        1e-12,
+    )
+
+    weights = np.clip(
+        weights,
+        float(
+            cfg.get(
+                "seg_min_weight",
+                0.35,
+            )
+        ),
+        float(
+            cfg.get(
+                "seg_max_weight",
+                3.0,
+            )
+        ),
+    )
+
+    print(
+        f"[LOSS][SEG] counts={counts.astype(np.int64).tolist()}"
+    )
+    print(
+        f"[LOSS][SEG] freq={np.round(freq, 6).tolist()}"
+    )
+    print(
+        f"[LOSS][SEG] weights={np.round(weights, 4).tolist()}"
+    )
+
+    return torch.tensor(
+        weights,
+        dtype=torch.float32,
+    )
+
+
+def status_counts(ds: CorridorDataset, num_label_classes: int) -> np.ndarray:
+    counts = np.zeros(num_label_classes, dtype=np.int64)
+    for label_id in ds.label_ids:
+        if 0 <= int(label_id) < num_label_classes:
+            counts[int(label_id)] += 1
+    return counts
+
+
+def estimate_status_class_weights(
+    ds: CorridorDataset,
+    num_label_classes: int,
+    cfg: dict,
+) -> torch.Tensor:
+    counts = status_counts(ds, num_label_classes).astype(np.float64)
+
+    if counts.sum() <= 0:
+        return torch.ones(num_label_classes, dtype=torch.float32)
+
+    freq = counts / counts.sum()
+    freq = np.clip(freq, 1e-12, 1.0)
+
+    power = float(cfg.get("status_power", 0.30))
+    weights = (1.0 / freq) ** power
+    weights = weights / max(weights.mean(), 1e-12)
+
+    weights = np.clip(
+        weights,
+        float(cfg.get("status_min_weight", 0.50)),
+        float(cfg.get("status_max_weight", 3.0)),
+    )
+
+    print(f"[LOSS][STATUS] counts={counts.astype(np.int64).tolist()}")
+    print(f"[LOSS][STATUS] weights={np.round(weights, 4).tolist()}")
+
+    return torch.tensor(weights, dtype=torch.float32)
+
+
+def build_training_sampler(
+    ds: CorridorDataset,
+    num_label_classes: int,
+    cfg: dict,
+) -> Optional[WeightedRandomSampler]:
+    """
+    Sampler pedagogico.
+
+    mode="joint":
+        peso considera grupo x estado, estado isolado e grupo isolado.
+        Potencias baixas impedem que raridades extremas dominem a epoca.
+
+    mode="status":
+        compatibilidade com a filosofia v1.
+
+    mode="none":
+        shuffle normal.
+    """
+    mode = str(
+        cfg.get(
+            "mode",
+            "joint",
+        )
+    ).strip().lower()
+
+    if mode in {"none", "off", "false"}:
+        print("[SAMPLER] disabled")
+        return None
+
+    labels = [
+        int(x)
+        for x in ds.label_ids
+    ]
+    groups = [
+        str(x)
+        for x in ds.group_names
+    ]
+
+    valid_indices = [
+        i
+        for i, label_id in enumerate(labels)
+        if 0 <= label_id < num_label_classes
+    ]
+
+    if not valid_indices:
+        print("[SAMPLER] nenhum label valido; usando shuffle.")
+        return None
+
+    status_counts_map = Counter(
+        labels[i]
+        for i in valid_indices
+    )
+
+    group_counts_map = Counter(
+        groups[i]
+        for i in valid_indices
+    )
+
+    joint_counts_map = Counter(
+        (groups[i], labels[i])
+        for i in valid_indices
+    )
+
+    joint_power = float(
+        cfg.get(
+            "joint_power",
+            0.30,
+        )
+    )
+    status_power = float(
+        cfg.get(
+            "status_power",
+            0.12,
+        )
+    )
+    group_power = float(
+        cfg.get(
+            "group_power",
+            0.10,
+        )
+    )
+
+    sample_weights: List[float] = []
+
+    for i, label_id in enumerate(labels):
+        if not (0 <= label_id < num_label_classes):
+            sample_weights.append(1.0)
+            continue
+
+        group = groups[i]
+
+        if mode == "status":
+            w = (
+                1.0
+                / max(
+                    1,
+                    status_counts_map[label_id],
+                )
+            ) ** max(
+                status_power,
+                joint_power,
+            )
+        elif mode == "joint":
+            n_joint = max(
+                1,
+                joint_counts_map[
+                    (
+                        group,
+                        label_id,
+                    )
+                ],
+            )
+            n_status = max(
+                1,
+                status_counts_map[label_id],
+            )
+            n_group = max(
+                1,
+                group_counts_map[group],
+            )
+
+            w = (
+                (1.0 / n_joint) ** joint_power
+                * (1.0 / n_status) ** status_power
+                * (1.0 / n_group) ** group_power
+            )
+        else:
+            raise ValueError(
+                f"sampler.mode desconhecido: {mode!r}"
+            )
+
+        sample_weights.append(
+            float(w)
+        )
+
+    weights_np = np.asarray(
+        sample_weights,
+        dtype=np.float64,
+    )
+
+    positive = weights_np[
+        weights_np > 0
+    ]
+
+    if positive.size == 0:
+        return None
+
+    weights_np /= positive.mean()
+
+    max_ratio = float(
+        cfg.get(
+            "max_weight_ratio",
+            5.0,
+        )
+    )
+
+    positive = weights_np[
+        weights_np > 0
+    ]
+
+    lo = float(
+        positive.min()
+    )
+    hi = lo * max(
+        1.0,
+        max_ratio,
+    )
+
+    weights_np = np.where(
+        weights_np > 0,
+        np.clip(
+            weights_np,
+            lo,
+            hi,
+        ),
+        0.0,
+    )
+
+    num_samples = int(
+        cfg.get(
+            "samples_per_epoch",
+            0,
+        )
+    )
+
+    if num_samples <= 0:
+        num_samples = len(ds)
+
+    print(f"[SAMPLER] mode={mode}")
+    print(
+        "[SAMPLER] groups="
+        + str(
+            dict(
+                sorted(
+                    group_counts_map.items()
+                )
+            )
+        )
+    )
+    print(
+        "[SAMPLER] status="
+        + str(
+            dict(
+                sorted(
+                    status_counts_map.items()
+                )
+            )
+        )
+    )
+
+    if mode == "joint":
+        print(
+            "[SAMPLER] joint="
+            + str(
+                {
+                    f"{g}×{lid}": n
+                    for (
+                        g,
+                        lid,
+                    ), n in sorted(
+                        joint_counts_map.items()
+                    )
+                }
+            )
+        )
+
+    print(
+        f"[SAMPLER] sample_weight range="
+        f"{weights_np.min():.4f}..{weights_np.max():.4f} "
+        f"samples_per_epoch={num_samples}"
+    )
+
+    return WeightedRandomSampler(
+        weights=torch.tensor(
+            weights_np,
+            dtype=torch.double,
+        ),
+        num_samples=num_samples,
+        replacement=True,
+    )
+
+
+# =============================================================================
+# Losses
+# =============================================================================
+
+def boundary_map(
+    targets: torch.Tensor,
+    nav_class_id: int,
+    ignore_index: int,
+) -> torch.Tensor:
+    valid = targets != ignore_index
+    nav = (targets == nav_class_id).float().unsqueeze(1)
+
+    dil = F.max_pool2d(nav, kernel_size=3, stride=1, padding=1)
+    ero = -F.max_pool2d(-nav, kernel_size=3, stride=1, padding=1)
+
+    bnd = (dil - ero).squeeze(1) > 1e-6
+    return bnd & valid
+
+
+def multiclass_dice_loss(
+    logits: torch.Tensor,
+    targets: torch.Tensor,
+    num_classes: int,
+    ignore_index: int,
+    smooth: float = 1.0,
+) -> torch.Tensor:
+    probs = torch.softmax(logits, dim=1)
+    valid = targets != ignore_index
+
+    losses = []
+
+    for cid in range(num_classes):
+        p = probs[:, cid]
+        t = (targets == cid).float()
+
+        p = p * valid
+        t = t * valid
+
+        inter = (p * t).sum(dim=(1, 2))
+        denom = p.sum(dim=(1, 2)) + t.sum(dim=(1, 2))
+
+        dice = (2.0 * inter + smooth) / (denom + smooth)
+        losses.append(1.0 - dice)
+
+    return torch.stack(losses, dim=0).mean()
+
+
+def corridor_seg_loss(
+    logits: torch.Tensor,
+    targets: torch.Tensor,
+    class_weights: Optional[torch.Tensor],
+    nav_class_id: int,
+    non_nav_class_id: int,
+    num_classes: int,
+    ignore_index: int,
+    cfg: dict,
+) -> Tuple[torch.Tensor, Dict[str, float]]:
+    ce_map = F.cross_entropy(
+        logits,
+        targets,
+        weight=class_weights,
+        ignore_index=ignore_index,
+        reduction="none",
+    )
+
+    valid = targets != ignore_index
+    boundary_boost = float(cfg.get("boundary_boost", 0.15))
+
+    if boundary_boost > 0:
+        bnd = boundary_map(targets, nav_class_id, ignore_index)
+        ce_map = ce_map * (1.0 + boundary_boost * bnd.float())
+
+    if bool(valid.any()):
+        ce = ce_map[valid].mean()
+    else:
+        ce = logits.sum() * 0.0
+
+    dice = multiclass_dice_loss(
+        logits,
+        targets,
+        num_classes=num_classes,
+        ignore_index=ignore_index,
+    )
+
+    probs = torch.softmax(logits, dim=1)
+    p_nav = probs[:, nav_class_id]
+    gt_non_nav = targets == non_nav_class_id
+
+    if bool(gt_non_nav.any()):
+        # Quadratico: baixa probabilidade falsa quase nao pesa; alta confianca falsa pesa.
+        unsafe = (p_nav[gt_non_nav] ** 2).mean()
+    else:
+        unsafe = logits.sum() * 0.0
+
+    w_ce = float(cfg.get("seg_ce_weight", 0.65))
+    w_dice = float(cfg.get("seg_dice_weight", 0.35))
+    w_unsafe = float(cfg.get("unsafe_nav_weight", 0.06))
+
+    total = w_ce * ce + w_dice * dice + w_unsafe * unsafe
+
+    return total, {
+        "ce": float(ce.detach().item()),
+        "dice": float(dice.detach().item()),
+        "unsafe": float(unsafe.detach().item()),
+    }
+
+
+def corridor_status_loss(
+    logits: torch.Tensor,
+    labels: torch.Tensor,
+    class_weights: Optional[torch.Tensor],
+    label_smoothing: float,
+) -> torch.Tensor:
+    valid = labels != LABEL_IGNORE_INDEX
+    if not bool(valid.any()):
+        return logits.sum() * 0.0
+
+    return F.cross_entropy(
+        logits[valid],
+        labels[valid],
+        weight=class_weights,
+        label_smoothing=float(label_smoothing),
+    )
+
+
+# =============================================================================
+# Metricas
+# =============================================================================
+
+@torch.no_grad()
+def update_confusion_matrix(
+    cm: torch.Tensor,
+    preds: torch.Tensor,
+    labels: torch.Tensor,
+    num_classes: int,
+    ignore_index: int,
+):
+    preds = preds.reshape(-1)
+    labels = labels.reshape(-1)
+
+    valid = labels != ignore_index
+    preds = preds[valid]
+    labels = labels[valid]
+
+    if labels.numel() == 0:
+        return
+
+    idx = labels * num_classes + preds
+    bins = torch.bincount(idx, minlength=num_classes * num_classes)
+    cm += bins.view(num_classes, num_classes)
+
+
+@torch.no_grad()
+def update_label_confusion(
+    cm: torch.Tensor,
+    logits: torch.Tensor,
+    labels: torch.Tensor,
+    num_classes: int,
+):
+    valid = labels != LABEL_IGNORE_INDEX
+    if not bool(valid.any()):
+        return
+
+    preds = torch.argmax(logits[valid], dim=1)
+    gt = labels[valid]
+
+    idx = gt * num_classes + preds
+    bins = torch.bincount(idx, minlength=num_classes * num_classes)
+    cm += bins.view(num_classes, num_classes)
+
+
+def safe_div(a: torch.Tensor, b: torch.Tensor, eps: float = 1e-8) -> torch.Tensor:
+    return a / (b + eps)
+
+
+def segmentation_metrics(
+    cm: torch.Tensor,
+    nav_class_id: int,
+    non_nav_class_id: int,
+) -> Dict[str, Any]:
+    cmf = cm.float()
+    tp = torch.diag(cmf)
+    fp = cmf.sum(0) - tp
+    fn = cmf.sum(1) - tp
+
+    iou = safe_div(tp, tp + fp + fn)
+    precision = safe_div(tp, tp + fp)
+    recall = safe_div(tp, tp + fn)
+    f1 = safe_div(2 * precision * recall, precision + recall)
+
+    total = cmf.sum()
+    acc = safe_div(tp.sum(), total)
+
+    # Safety: entre pixels realmente nao navegaveis, quantos foram liberados como navegaveis?
+    unsafe_nav = cmf[non_nav_class_id, nav_class_id]
+    gt_non_nav = cmf[non_nav_class_id].sum()
+    unsafe_nav_rate = safe_div(unsafe_nav, gt_non_nav)
+    safety = 1.0 - unsafe_nav_rate
+
+    return {
+        "acc": float(acc.item()),
+        "miou": float(iou.mean().item()),
+        "iou_per_class": iou.cpu().tolist(),
+        "precision_per_class": precision.cpu().tolist(),
+        "recall_per_class": recall.cpu().tolist(),
+        "f1_per_class": f1.cpu().tolist(),
+        "nav_iou": float(iou[nav_class_id].item()),
+        "nav_precision": float(precision[nav_class_id].item()),
+        "nav_recall": float(recall[nav_class_id].item()),
+        "nav_f1": float(f1[nav_class_id].item()),
+        "non_nav_iou": float(iou[non_nav_class_id].item()),
+        "unsafe_nav_rate": float(unsafe_nav_rate.item()),
+        "safety": float(safety.item()),
+        "support_per_class": cmf.sum(1).cpu().tolist(),
+        "gt_nav_pixels": int(cmf[nav_class_id].sum().item()),
+        "gt_non_nav_pixels": int(gt_non_nav.item()),
+    }
+
+
+def status_metrics(cm: torch.Tensor) -> Dict[str, Any]:
+    cmf = cm.float()
+    tp = torch.diag(cmf)
+    fp = cmf.sum(0) - tp
+    fn = cmf.sum(1) - tp
+    support = cmf.sum(1)
+
+    precision = safe_div(tp, tp + fp)
+    recall = safe_div(tp, tp + fn)
+    f1 = safe_div(2 * precision * recall, precision + recall)
+
+    acc = safe_div(tp.sum(), cmf.sum())
+
+    present = support > 0
+    if bool(present.any()):
+        macro_f1 = f1[present].mean()
+        macro_recall = recall[present].mean()
+    else:
+        macro_f1 = torch.tensor(0.0, device=cm.device)
+        macro_recall = torch.tensor(0.0, device=cm.device)
+
+    return {
+        "acc": float(acc.item()),
+        "macro_f1": float(macro_f1.item()),
+        "macro_recall": float(macro_recall.item()),
+        "precision_per_class": precision.cpu().tolist(),
+        "recall_per_class": recall.cpu().tolist(),
+        "f1_per_class": f1.cpu().tolist(),
+        "support_per_class": support.cpu().tolist(),
+    }
+
+
+def field_score(seg: dict, status: dict, cfg: dict) -> float:
+    weights = {
+        "nav_iou": float(cfg.get("nav_iou", 0.40)),
+        "nav_f1": float(cfg.get("nav_f1", 0.20)),
+        "status_macro_f1": float(cfg.get("status_macro_f1", 0.20)),
+        "safety": float(cfg.get("safety", 0.20)),
+    }
+
+    total_w = sum(weights.values())
+    if total_w <= 0:
+        raise ValueError("score weights precisam somar > 0")
+
+    score = (
+        weights["nav_iou"] * seg["nav_iou"]
+        + weights["nav_f1"] * seg["nav_f1"]
+        + weights["status_macro_f1"] * status["macro_f1"]
+        + weights["safety"] * seg["safety"]
+    ) / total_w
+
+    return float(score)
+
+
+
+def evaluate_operational_gate(
+    *,
+    epoch: int,
+    val_metrics: dict,
+    cfg: dict,
+    label_name_by_id: Dict[int, str],
+) -> Dict[str, Any]:
+    """Avalia se uma epoca e elegivel para concorrer a best_operational."""
+    enabled = bool(cfg.get("enabled", True))
+
+    result: Dict[str, Any] = {
+        "schema": "agrobot.corridor.operational_gate.v1",
+        "enabled": enabled,
+        "eligible": True,
+        "checks": {},
+        "failures": [],
+        "warnings": [],
+    }
+
+    if not enabled:
+        result["warnings"].append("GATE_DISABLED")
+        return result
+
+    seg = val_metrics["seg"]
+    status = val_metrics["status"]
+    seg_by_group = val_metrics.get("seg_by_group", {}) or {}
+
+    def add_min(name: str, value: float, threshold_value) -> None:
+        if threshold_value is None:
+            return
+        threshold = float(threshold_value)
+        value_f = float(value)
+        passed = value_f >= threshold
+        result["checks"][name] = {
+            "passed": bool(passed),
+            "value": value_f,
+            "operator": ">=",
+            "threshold": threshold,
+        }
+        if not passed:
+            result["failures"].append(
+                f"{name}:{value_f:.6f}<{threshold:.6f}"
+            )
+
+    def add_max(name: str, value: float, threshold_value) -> None:
+        if threshold_value is None:
+            return
+        threshold = float(threshold_value)
+        value_f = float(value)
+        passed = value_f <= threshold
+        result["checks"][name] = {
+            "passed": bool(passed),
+            "value": value_f,
+            "operator": "<=",
+            "threshold": threshold,
+        }
+        if not passed:
+            result["failures"].append(
+                f"{name}:{value_f:.6f}>{threshold:.6f}"
+            )
+
+    min_epoch = int(cfg.get("min_epoch", 1))
+    epoch_passed = int(epoch) >= min_epoch
+    result["checks"]["min_epoch"] = {
+        "passed": bool(epoch_passed),
+        "value": int(epoch),
+        "operator": ">=",
+        "threshold": int(min_epoch),
+    }
+    if not epoch_passed:
+        result["failures"].append(
+            f"min_epoch:{int(epoch)}<{int(min_epoch)}"
+        )
+
+    add_min("min_nav_iou", seg["nav_iou"], cfg.get("min_nav_iou"))
+    add_min("min_nav_f1", seg["nav_f1"], cfg.get("min_nav_f1"))
+    add_min("min_non_nav_iou", seg["non_nav_iou"], cfg.get("min_non_nav_iou"))
+    add_min(
+        "min_status_macro_f1",
+        status["macro_f1"],
+        cfg.get("min_status_macro_f1"),
+    )
+    add_max(
+        "max_unsafe_nav_rate",
+        seg["unsafe_nav_rate"],
+        cfg.get("max_unsafe_nav_rate"),
+    )
+
+    group_cfg = cfg.get("group_unsafe", {}) or {}
+    if bool(group_cfg.get("enabled", True)):
+        min_non_nav_pixels = int(group_cfg.get("min_non_nav_pixels", 0))
+        default_max_rate = float(group_cfg.get("max_rate", 1.0))
+        by_group = group_cfg.get("max_rate_by_group", {}) or {}
+        group_checks = {}
+
+        for group_name in sorted(seg_by_group):
+            gm = seg_by_group[group_name]
+            support = int(gm.get("gt_non_nav_pixels", 0))
+            rate = float(gm.get("unsafe_nav_rate", 0.0))
+            threshold = float(by_group.get(group_name, default_max_rate))
+            applicable = support >= min_non_nav_pixels
+            passed = (rate <= threshold) if applicable else True
+
+            group_checks[group_name] = {
+                "applicable": bool(applicable),
+                "passed": bool(passed),
+                "gt_non_nav_pixels": int(support),
+                "min_non_nav_pixels": int(min_non_nav_pixels),
+                "unsafe_nav_rate": rate,
+                "max_unsafe_nav_rate": threshold,
+            }
+
+            if applicable and not passed:
+                result["failures"].append(
+                    f"group_unsafe:{group_name}:{rate:.6f}>{threshold:.6f}"
+                )
+
+        result["checks"]["group_unsafe"] = group_checks
+
+    status_cfg = cfg.get("status_recall", {}) or {}
+    if bool(status_cfg.get("enabled", False)):
+        minimums = status_cfg.get("minimums", {}) or {}
+        min_support = int(status_cfg.get("min_support", 1))
+        recalls = status.get("recall_per_class", [])
+        supports = status.get("support_per_class", [])
+        status_checks = {}
+        name_to_id = {
+            str(name): int(cid)
+            for cid, name in label_name_by_id.items()
+        }
+
+        for class_name, threshold_raw in minimums.items():
+            if class_name not in name_to_id:
+                status_checks[class_name] = {
+                    "applicable": False,
+                    "passed": False,
+                    "reason": "UNKNOWN_STATUS_CLASS",
+                }
+                result["failures"].append(
+                    f"status_recall:{class_name}:UNKNOWN_STATUS_CLASS"
+                )
+                continue
+
+            cid = name_to_id[class_name]
+            support = int(supports[cid] if cid < len(supports) else 0)
+            recall = float(recalls[cid] if cid < len(recalls) else 0.0)
+            threshold = float(threshold_raw)
+            applicable = support >= min_support
+            passed = (recall >= threshold) if applicable else True
+
+            status_checks[class_name] = {
+                "applicable": bool(applicable),
+                "passed": bool(passed),
+                "support": int(support),
+                "min_support": int(min_support),
+                "recall": recall,
+                "min_recall": threshold,
+            }
+
+            if applicable and not passed:
+                result["failures"].append(
+                    f"status_recall:{class_name}:{recall:.6f}<{threshold:.6f}"
+                )
+            if not applicable:
+                result["warnings"].append(
+                    f"status_recall:{class_name}:support={support}<min_support={min_support}"
+                )
+
+        result["checks"]["status_recall"] = status_checks
+
+    result["eligible"] = len(result["failures"]) == 0
+    return result
+
+
+def format_operational_gate(gate: dict, max_failures: int = 6) -> str:
+    if not bool(gate.get("enabled", True)):
+        return "DISABLED"
+    if bool(gate.get("eligible", False)):
+        return "ELIGIBLE"
+
+    failures = list(gate.get("failures", []))
+    shown = failures[:max_failures]
+    suffix = (
+        ""
+        if len(failures) <= max_failures
+        else f" (+{len(failures) - max_failures})"
+    )
+    return "FAIL | " + " | ".join(shown) + suffix
+
+
+def pretty_status(values: Sequence[float], names: Dict[int, str]) -> str:
+    return " | ".join(
+        f"{names.get(i, str(i))}:{float(v):.3f}"
+        for i, v in enumerate(values)
+    )
+
+
+# =============================================================================
+# Optimizer / scheduler
+# =============================================================================
+
+def is_no_decay_param(name: str, p: torch.Tensor, cfg: dict) -> bool:
+    lname = name.lower()
+
+    if bool(cfg.get("no_decay_bias", True)) and lname.endswith(".bias"):
+        return True
+
+    if bool(cfg.get("no_decay_norm", True)):
+        if p.ndim <= 1:
+            return True
+        if any(token in lname for token in ("norm", "bn", "batchnorm", "layernorm")):
+            return True
+
+    return False
+
+
+def append_param_groups(
+    groups: List[dict],
+    named_params: Sequence[Tuple[str, torch.nn.Parameter]],
+    lr: float,
+    wd: float,
+    prefix: str,
+    cfg: dict,
+):
+    decay, no_decay = [], []
+
+    for name, p in named_params:
+        if not p.requires_grad:
+            # Mesmo congelado agora, pode ser descongelado futuramente.
+            # Mantemos no optimizer.
+            pass
+
+        full_name = f"{prefix}.{name}"
+
+        if is_no_decay_param(full_name, p, cfg):
+            no_decay.append(p)
+        else:
+            decay.append(p)
+
+    if decay:
+        groups.append({
+            "params": decay,
+            "lr": float(lr),
+            "weight_decay": float(wd),
+            "group_name": f"{prefix}_decay",
+        })
+
+    if no_decay:
+        groups.append({
+            "params": no_decay,
+            "lr": float(lr),
+            "weight_decay": 0.0,
+            "group_name": f"{prefix}_no_decay",
+        })
+
+
+def build_optimizer(
+    base_model: SegformerForSemanticSegmentation,
+    status_head: CorridorStatusHead,
+    base_lr: float,
+    default_wd: float,
+    cfg: dict,
+) -> torch.optim.Optimizer:
+    wd = cfg.get("weight_decay")
+    wd = float(default_wd if wd is None else wd)
+
+    enc_lr = float(base_lr) * float(cfg.get("encoder_lr_mult", 0.50))
+    dec_lr = float(base_lr) * float(cfg.get("decoder_lr_mult", 1.00))
+    status_lr = float(base_lr) * float(cfg.get("status_lr_mult", 2.00))
+
+    groups: List[dict] = []
+
+    append_param_groups(
+        groups,
+        list(base_model.segformer.named_parameters()),
+        enc_lr,
+        wd,
+        "encoder",
+        cfg,
+    )
+
+    append_param_groups(
+        groups,
+        list(base_model.decode_head.named_parameters()),
+        dec_lr,
+        wd,
+        "decoder",
+        cfg,
+    )
+
+    append_param_groups(
+        groups,
+        list(status_head.named_parameters()),
+        status_lr,
+        wd,
+        "status",
+        cfg,
+    )
+
+    betas = cfg.get("betas", [0.9, 0.999])
+    eps = float(cfg.get("eps", 1e-8))
+
+    opt = torch.optim.AdamW(
+        groups,
+        betas=(float(betas[0]), float(betas[1])),
+        eps=eps,
+    )
+
+    print(
+        f"[OPT] lr encoder={enc_lr:.3e} decoder={dec_lr:.3e} "
+        f"status={status_lr:.3e} wd={wd:.3e}"
+    )
+
+    return opt
+
+
+def build_scheduler(
+    optimizer: torch.optim.Optimizer,
+    total_steps: int,
+    cfg: dict,
+):
+    mode = str(cfg.get("mode", "poly")).lower()
+
+    warmup_ratio = float(cfg.get("warmup_ratio", 0.05))
+    warmup_steps = int(round(total_steps * warmup_ratio))
+    warmup_start = float(cfg.get("warmup_start_factor", 0.10))
+    power = float(cfg.get("poly_power", 1.0))
+    min_lr_ratio = float(cfg.get("min_lr_ratio", 0.02))
+
+    def lr_lambda(step: int):
+        if warmup_steps > 0 and step < warmup_steps:
+            alpha = step / max(1, warmup_steps)
+            return warmup_start + alpha * (1.0 - warmup_start)
+
+        progress = (step - warmup_steps) / max(1, total_steps - warmup_steps)
+        progress = min(max(progress, 0.0), 1.0)
+
+        if mode == "poly":
+            factor = (1.0 - progress) ** power
+        elif mode == "cosine":
+            factor = 0.5 * (1.0 + math.cos(math.pi * progress))
+        elif mode == "constant":
+            factor = 1.0
+        else:
+            raise ValueError(f"scheduler mode desconhecido: {mode}")
+
+        return min_lr_ratio + (1.0 - min_lr_ratio) * factor
+
+    return torch.optim.lr_scheduler.LambdaLR(optimizer, lr_lambda=lr_lambda)
+
+
+def set_encoder_trainable(
+    base_model: SegformerForSemanticSegmentation,
+    trainable: bool,
+):
+    for p in base_model.segformer.parameters():
+        p.requires_grad = bool(trainable)
+
+
+# =============================================================================
+# Training history
+# =============================================================================
+
+def _history_json_safe(value):
+    if isinstance(value, dict):
+        return {str(k): _history_json_safe(v) for k, v in value.items()}
+    if isinstance(value, (list, tuple)):
+        return [_history_json_safe(v) for v in value]
+    if isinstance(value, Path):
+        return str(value)
+    if isinstance(value, np.ndarray):
+        return value.tolist()
+    if isinstance(value, np.generic):
+        return value.item()
+    if torch.is_tensor(value):
+        if value.numel() == 1:
+            return value.detach().cpu().item()
+        return value.detach().cpu().tolist()
+    if isinstance(value, (str, int, float, bool)) or value is None:
+        return value
+    return str(value)
+
+
+def _flatten_history(value, prefix: str = "", out: Optional[Dict[str, Any]] = None) -> Dict[str, Any]:
+    if out is None:
+        out = {}
+    if isinstance(value, dict):
+        for key, item in value.items():
+            child = f"{prefix}.{key}" if prefix else str(key)
+            _flatten_history(item, child, out)
+        return out
+    if isinstance(value, (list, tuple)):
+        for i, item in enumerate(value):
+            child = f"{prefix}.{i}" if prefix else str(i)
+            _flatten_history(item, child, out)
+        if not value and prefix:
+            out[prefix] = ""
+        return out
+    out[prefix] = value
+    return out
+
+
+def _load_history_jsonl(jsonl_path: Path) -> List[dict]:
+    if not jsonl_path.is_file():
+        return []
+    rows = []
+    with jsonl_path.open("r", encoding="utf-8") as f:
+        for lineno, raw in enumerate(f, 1):
+            line = raw.strip()
+            if not line:
+                continue
+            try:
+                item = json.loads(line)
+            except Exception as exc:
+                raise RuntimeError(
+                    f"Historico JSONL corrompido em {jsonl_path}:{lineno}: {exc}"
+                ) from exc
+            if not isinstance(item, dict):
+                raise RuntimeError(f"Historico JSONL invalido em {jsonl_path}:{lineno}")
+            rows.append(item)
+    return rows
+
+
+def _rewrite_history_files(*, records: Sequence[dict], jsonl_path: Path, json_path: Path, csv_path: Path) -> None:
+    records_safe = [_history_json_safe(r) for r in records]
+    records_safe = sorted(records_safe, key=lambda r: int(r.get("epoch", 0)))
+    jsonl_path.parent.mkdir(parents=True, exist_ok=True)
+
+    tmp_jsonl = jsonl_path.with_suffix(jsonl_path.suffix + ".tmp")
+    with tmp_jsonl.open("w", encoding="utf-8") as f:
+        for record in records_safe:
+            f.write(json.dumps(record, ensure_ascii=False, separators=(",", ":")))
+            f.write("\n")
+    os.replace(tmp_jsonl, jsonl_path)
+
+    tmp_json = json_path.with_suffix(json_path.suffix + ".tmp")
+    with tmp_json.open("w", encoding="utf-8") as f:
+        json.dump(records_safe, f, ensure_ascii=False, indent=2)
+    os.replace(tmp_json, json_path)
+
+    flat_rows = [_flatten_history(record) for record in records_safe]
+    fieldnames = []
+    seen = set()
+    for row in flat_rows:
+        for key in row:
+            if key not in seen:
+                seen.add(key)
+                fieldnames.append(key)
+
+    priority = [
+        "epoch", "created_at", "field_score",
+        "operational_eligible", "operational_score_improved",
+        "encoder_trainable",
+        "augmentation_strength", "status_weight",
+        "train.loss", "train.loss_seg", "train.loss_status",
+        "train.seg.nav_iou", "train.seg.nav_f1", "train.status.macro_f1",
+        "val.loss", "val.loss_seg", "val.loss_status",
+        "val.seg.nav_iou", "val.seg.non_nav_iou",
+        "val.seg.nav_precision", "val.seg.nav_recall", "val.seg.nav_f1",
+        "val.seg.unsafe_nav_rate", "val.seg.safety",
+        "val.status.acc", "val.status.macro_f1", "val.status.macro_recall",
+        "epoch_time_s",
+    ]
+    ordered = [key for key in priority if key in seen]
+    ordered += [key for key in fieldnames if key not in ordered]
+
+    tmp_csv = csv_path.with_suffix(csv_path.suffix + ".tmp")
+    with tmp_csv.open("w", newline="", encoding="utf-8-sig") as f:
+        writer = csv.DictWriter(f, fieldnames=ordered, extrasaction="ignore")
+        writer.writeheader()
+        for row in flat_rows:
+            writer.writerow(row)
+    os.replace(tmp_csv, csv_path)
+
+
+def initialize_training_history(*, save_dir: Path, resume: bool, start_epoch: int) -> Tuple[Path, Path, Path]:
+    jsonl_path = save_dir / "metrics_history.jsonl"
+    json_path = save_dir / "metrics_history.json"
+    csv_path = save_dir / "metrics_history.csv"
+
+    if not resume:
+        for p in (jsonl_path, json_path, csv_path):
+            if p.exists():
+                p.unlink()
+        return jsonl_path, json_path, csv_path
+
+    old = _load_history_jsonl(jsonl_path)
+    kept = [row for row in old if int(row.get("epoch", 0)) < int(start_epoch)]
+    if old:
+        _rewrite_history_files(
+            records=kept,
+            jsonl_path=jsonl_path,
+            json_path=json_path,
+            csv_path=csv_path,
+        )
+    return jsonl_path, json_path, csv_path
+
+
+def save_epoch_history(*, record: dict, jsonl_path: Path, json_path: Path, csv_path: Path) -> None:
+    records = _load_history_jsonl(jsonl_path)
+    epoch = int(record["epoch"])
+    records = [row for row in records if int(row.get("epoch", -1)) != epoch]
+    records.append(_history_json_safe(record))
+    _rewrite_history_files(
+        records=records,
+        jsonl_path=jsonl_path,
+        json_path=json_path,
+        csv_path=csv_path,
+    )
+
+
+# =============================================================================
+# Checkpoint
+# =============================================================================
+
+@dataclass
+class EarlyStopState:
+    best: float = -1e9
+    bad_epochs: int = 0
+
+
+def save_checkpoint(
+    path: Path,
+    base_model: SegformerForSemanticSegmentation,
+    status_head: CorridorStatusHead,
+    optimizer: torch.optim.Optimizer,
+    scheduler,
+    scaler: GradScaler,
+    epoch: int,
+    bests: dict,
+    early_stop: EarlyStopState,
+    metadata: dict,
+):
+    payload = {
+        "trainer_version": TRAINER_VERSION,
+        "epoch": int(epoch),
+        "model": base_model.state_dict(),
+        "status_head": status_head.state_dict(),
+        "optimizer": optimizer.state_dict(),
+        "scheduler": scheduler.state_dict() if scheduler is not None else None,
+        "scaler": scaler.state_dict() if scaler is not None else None,
+        "bests": dict(bests),
+        "early_stop": {
+            "best": float(early_stop.best),
+            "bad_epochs": int(early_stop.bad_epochs),
+        },
+        "rng_state": capture_rng_state(),
+        "metadata": metadata,
+    }
+
+    torch.save(payload, path)
+
+
+def load_checkpoint(
+    path: Path,
+    base_model: SegformerForSemanticSegmentation,
+    status_head: CorridorStatusHead,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler=None,
+    scaler: Optional[GradScaler] = None,
+) -> dict:
+    ckpt = torch.load(path, map_location="cpu", weights_only=False)
+
+    base_model.load_state_dict(ckpt["model"], strict=True)
+    status_head.load_state_dict(ckpt["status_head"], strict=True)
+
+    if optimizer is not None and ckpt.get("optimizer") is not None:
+        optimizer.load_state_dict(ckpt["optimizer"])
+
+    if scheduler is not None and ckpt.get("scheduler") is not None:
+        scheduler.load_state_dict(ckpt["scheduler"])
+
+    if scaler is not None and ckpt.get("scaler") is not None:
+        scaler.load_state_dict(ckpt["scaler"])
+
+    restore_rng_state(ckpt.get("rng_state"))
+    return ckpt
+
+
+# =============================================================================
+# Epoch
+# =============================================================================
+
+def current_status_weight(epoch: int, loss_cfg: dict) -> float:
+    target = float(loss_cfg.get("status_weight", 0.40))
+    ramp = int(loss_cfg.get("status_ramp_epochs", 8))
+
+    if ramp <= 0:
+        return target
+
+    alpha = min(1.0, max(0.10, float(epoch) / float(ramp)))
+    return target * alpha
+
+
+def run_epoch(
+    *,
+    base_model: SegformerForSemanticSegmentation,
+    status_head: CorridorStatusHead,
+    loader: DataLoader,
+    device: torch.device,
+    mean: torch.Tensor,
+    std: torch.Tensor,
+    num_seg_classes: int,
+    num_label_classes: int,
+    num_groups: int,
+    group_name_by_id: Dict[int, str],
+    nav_class_id: int,
+    non_nav_class_id: int,
+    seg_class_weights: Optional[torch.Tensor],
+    status_class_weights: Optional[torch.Tensor],
+    loss_cfg: dict,
+    augmenter: Optional[FieldCorridorAugmenter],
+    optimizer: Optional[torch.optim.Optimizer],
+    scheduler,
+    scaler: Optional[GradScaler],
+    amp: bool,
+    train: bool,
+    grad_accum: int,
+    grad_clip_norm: float,
+    epoch: int,
+) -> Dict[str, Any]:
+    base_model.train(train)
+    status_head.train(train)
+
+    if augmenter is not None:
+        augmenter.set_epoch(epoch)
+
+    cm_seg = torch.zeros(
+        (num_seg_classes, num_seg_classes),
+        dtype=torch.int64,
+        device=device,
+    )
+    cm_status = torch.zeros(
+        (num_label_classes, num_label_classes),
+        dtype=torch.int64,
+        device=device,
+    )
+
+    cm_seg_group = torch.zeros(
+        (
+            num_groups,
+            num_seg_classes,
+            num_seg_classes,
+        ),
+        dtype=torch.int64,
+        device=device,
+    )
+
+    cm_status_group = torch.zeros(
+        (
+            num_groups,
+            num_label_classes,
+            num_label_classes,
+        ),
+        dtype=torch.int64,
+        device=device,
+    )
+
+    total_loss = 0.0
+    total_seg = 0.0
+    total_status = 0.0
+    total_ce = 0.0
+    total_dice = 0.0
+    total_unsafe = 0.0
+    n_batches = 0
+
+    status_w = current_status_weight(epoch, loss_cfg)
+    t0 = time.perf_counter()
+
+    if train and optimizer is not None:
+        optimizer.zero_grad(set_to_none=True)
+
+    for step, (imgs, masks, labels, group_ids) in enumerate(loader, start=1):
+        imgs = imgs.to(device, non_blocking=True)
+        masks = masks.to(device, non_blocking=True)
+        labels = labels.to(device, non_blocking=True)
+        group_ids = group_ids.to(device, non_blocking=True)
+
+        if train and augmenter is not None:
+            imgs, masks = augmenter(imgs, masks)
+
+        imgs_norm = (imgs - mean) / std
+
+        with torch.set_grad_enabled(train):
+            with torch.autocast(
+                device_type=device.type,
+                enabled=bool(amp and device.type == "cuda"),
+            ):
+                out = base_model(
+                    pixel_values=imgs_norm,
+                    output_hidden_states=True,
+                    return_dict=True,
+                )
+
+                seg_logits_native = out.logits
+                feat = out.hidden_states[-1]
+
+                status_logits = status_head(feat, seg_logits_native)
+
+                seg_logits = seg_logits_native
+                if seg_logits.shape[-2:] != masks.shape[-2:]:
+                    seg_logits = F.interpolate(
+                        seg_logits,
+                        size=masks.shape[-2:],
+                        mode="bilinear",
+                        align_corners=False,
+                    )
+
+                seg_loss, seg_parts = corridor_seg_loss(
+                    seg_logits,
+                    masks,
+                    class_weights=seg_class_weights,
+                    nav_class_id=nav_class_id,
+                    non_nav_class_id=non_nav_class_id,
+                    num_classes=num_seg_classes,
+                    ignore_index=IGNORE_INDEX,
+                    cfg=loss_cfg,
+                )
+
+                status_loss = corridor_status_loss(
+                    status_logits,
+                    labels,
+                    class_weights=status_class_weights,
+                    label_smoothing=float(loss_cfg.get("status_label_smoothing", 0.03)),
+                )
+
+                loss = seg_loss + status_w * status_loss
+
+                loss_for_backward = loss / max(1, grad_accum)
+
+        if train and optimizer is not None:
+            assert scaler is not None
+
+            scaler.scale(loss_for_backward).backward()
+
+            should_step = (
+                step % max(1, grad_accum) == 0
+                or step == len(loader)
+            )
+
+            if should_step:
+                scaler.unscale_(optimizer)
+
+                if grad_clip_norm > 0:
+                    params = list(base_model.parameters()) + list(status_head.parameters())
+                    torch.nn.utils.clip_grad_norm_(params, max_norm=grad_clip_norm)
+
+                scaler.step(optimizer)
+                scaler.update()
+                optimizer.zero_grad(set_to_none=True)
+
+                if scheduler is not None:
+                    scheduler.step()
+
+        total_loss += float(loss.detach().item())
+        total_seg += float(seg_loss.detach().item())
+        total_status += float(status_loss.detach().item())
+        total_ce += seg_parts["ce"]
+        total_dice += seg_parts["dice"]
+        total_unsafe += seg_parts["unsafe"]
+        n_batches += 1
+
+        with torch.no_grad():
+            pred_seg = torch.argmax(seg_logits, dim=1)
+            update_confusion_matrix(
+                cm_seg,
+                pred_seg,
+                masks,
+                num_seg_classes,
+                IGNORE_INDEX,
+            )
+            update_label_confusion(
+                cm_status,
+                status_logits,
+                labels,
+                num_label_classes,
+            )
+
+            for gid in range(num_groups):
+                sel = group_ids == gid
+
+                if not bool(sel.any()):
+                    continue
+
+                update_confusion_matrix(
+                    cm_seg_group[gid],
+                    pred_seg[sel],
+                    masks[sel],
+                    num_seg_classes,
+                    IGNORE_INDEX,
+                )
+
+                update_label_confusion(
+                    cm_status_group[gid],
+                    status_logits[sel],
+                    labels[sel],
+                    num_label_classes,
+                )
+
+    seg_metrics = segmentation_metrics(
+        cm_seg,
+        nav_class_id=nav_class_id,
+        non_nav_class_id=non_nav_class_id,
+    )
+    stat_metrics = status_metrics(cm_status)
+
+    seg_by_group = {}
+    status_by_group = {}
+
+    for gid in range(num_groups):
+        gname = group_name_by_id.get(
+            gid,
+            str(gid),
+        )
+
+        seg_by_group[gname] = segmentation_metrics(
+            cm_seg_group[gid],
+            nav_class_id=nav_class_id,
+            non_nav_class_id=non_nav_class_id,
+        )
+
+        status_by_group[gname] = status_metrics(
+            cm_status_group[gid]
+        )
+
+    return {
+        "loss": total_loss / max(1, n_batches),
+        "loss_seg": total_seg / max(1, n_batches),
+        "loss_status": total_status / max(1, n_batches),
+        "loss_ce": total_ce / max(1, n_batches),
+        "loss_dice": total_dice / max(1, n_batches),
+        "loss_unsafe": total_unsafe / max(1, n_batches),
+        "status_weight": status_w,
+        "seg": seg_metrics,
+        "status": stat_metrics,
+        "seg_by_group": seg_by_group,
+        "status_by_group": status_by_group,
+        "time_s": time.perf_counter() - t0,
+    }
+
+
+
+# =============================================================================
+# Pipeline contract / preflight
+# =============================================================================
+
+def _stem_set(folder: Path, exts: Tuple[str, ...]) -> set[str]:
+    if not folder.is_dir():
+        return set()
+
+    return {
+        p.stem
+        for p in folder.iterdir()
+        if p.is_file()
+        and p.suffix.lower() in exts
+    }
+
+
+def audit_split_root(
+    split_root: Path,
+    expected_wh: Tuple[int, int],
+    num_seg_classes: int,
+    nav_class_id: int,
+    label_classes: Sequence[str],
+    max_samples: int = 0,
+) -> dict:
+    """
+    Auditoria forte antes de entregar qualquer batch ao modelo.
+    """
+    images_dir = split_root / "images"
+    masks_dir = split_root / "masks"
+    labels_dir = split_root / "labels"
+
+    for folder in (
+        images_dir,
+        masks_dir,
+        labels_dir,
+    ):
+        if not folder.is_dir():
+            raise FileNotFoundError(
+                f"Split incompleto, pasta ausente: {folder}"
+            )
+
+    image_stems = _stem_set(
+        images_dir,
+        (
+            ".png",
+            ".jpg",
+            ".jpeg",
+            ".bmp",
+            ".webp",
+        ),
+    )
+    mask_stems = _stem_set(
+        masks_dir,
+        (".png",),
+    )
+    label_stems = _stem_set(
+        labels_dir,
+        (".json",),
+    )
+
+    if not image_stems:
+        raise RuntimeError(
+            f"Split vazio: {split_root}"
+        )
+
+    if not (
+        image_stems
+        == mask_stems
+        == label_stems
+    ):
+        missing_mask = sorted(
+            image_stems - mask_stems
+        )[:10]
+        missing_label = sorted(
+            image_stems - label_stems
+        )[:10]
+        orphan_mask = sorted(
+            mask_stems - image_stems
+        )[:10]
+        orphan_label = sorted(
+            label_stems - image_stems
+        )[:10]
+
+        raise RuntimeError(
+            f"Contrato quebrado em {split_root}.\n"
+            f"missing_mask={missing_mask}\n"
+            f"missing_label={missing_label}\n"
+            f"orphan_mask={orphan_mask}\n"
+            f"orphan_label={orphan_label}"
+        )
+
+    ordered = sorted(
+        image_stems
+    )
+
+    if max_samples > 0 and len(ordered) > max_samples:
+        idxs = np.linspace(
+            0,
+            len(ordered) - 1,
+            max_samples,
+        ).astype(int)
+        audit_stems = [
+            ordered[int(i)]
+            for i in idxs
+        ]
+    else:
+        audit_stems = ordered
+
+    W, H = int(expected_wh[0]), int(expected_wh[1])
+
+    status_counts_local = Counter()
+    group_counts_local = Counter()
+    joint_counts_local = Counter()
+    nav_pct_values = []
+
+    for stem in audit_stems:
+        # Imagem.
+        image_candidates = [
+            p
+            for p in images_dir.glob(stem + ".*")
+            if p.suffix.lower()
+            in {
+                ".png",
+                ".jpg",
+                ".jpeg",
+                ".bmp",
+                ".webp",
+            }
+        ]
+
+        if len(image_candidates) != 1:
+            raise RuntimeError(
+                f"Esperava 1 imagem para {stem}, encontrei {image_candidates}"
+            )
+
+        img = cv2.imread(
+            str(image_candidates[0]),
+            cv2.IMREAD_COLOR,
+        )
+
+        if img is None:
+            raise RuntimeError(
+                f"Nao consegui ler imagem: {image_candidates[0]}"
+            )
+
+        if img.shape[:2] != (
+            H,
+            W,
+        ):
+            raise RuntimeError(
+                f"Imagem {stem} shape={img.shape[:2]}, esperado={(H,W)}"
+            )
+
+        # Mask ID.
+        mask_path = masks_dir / f"{stem}.png"
+        mask = cv2.imread(
+            str(mask_path),
+            cv2.IMREAD_UNCHANGED,
+        )
+
+        if mask is None:
+            raise RuntimeError(
+                f"Nao consegui ler mask: {mask_path}"
+            )
+
+        if mask.ndim != 2:
+            raise RuntimeError(
+                f"Mask normalizada precisa ser 1 canal IDs: "
+                f"{mask_path} shape={mask.shape}"
+            )
+
+        if mask.shape != (
+            H,
+            W,
+        ):
+            raise RuntimeError(
+                f"Mask {stem} shape={mask.shape}, esperado={(H,W)}"
+            )
+
+        ids = sorted(
+            int(x)
+            for x in np.unique(mask)
+        )
+
+        if not set(ids).issubset(
+            set(
+                range(
+                    num_seg_classes
+                )
+            )
+        ):
+            raise RuntimeError(
+                f"Mask {stem} possui IDs invalidos={ids}"
+            )
+
+        # Label.
+        label_path = labels_dir / f"{stem}.json"
+
+        with label_path.open(
+            "r",
+            encoding="utf-8",
+        ) as f:
+            data = json.load(f)
+
+        label_id = int(
+            data.get(
+                "label_id",
+                -999,
+            )
+        )
+
+        if not (
+            0
+            <= label_id
+            < len(label_classes)
+        ):
+            raise RuntimeError(
+                f"label_id invalido em {label_path}: {label_id}"
+            )
+
+        expected_name = str(
+            label_classes[label_id]
+        )
+        got_name = str(
+            data.get(
+                "estado_corredor",
+                expected_name,
+            )
+        )
+
+        if got_name != expected_name:
+            raise RuntimeError(
+                f"Label inconsistente em {label_path}: "
+                f"id={label_id} => {expected_name}, arquivo={got_name}"
+            )
+
+        group = str(
+            data.get(
+                "split_source_group",
+                data.get(
+                    "group",
+                    "unknown",
+                ),
+            )
+        )
+
+        status_counts_local[label_id] += 1
+        group_counts_local[group] += 1
+        joint_counts_local[
+            (
+                group,
+                label_id,
+            )
+        ] += 1
+
+        # nav_class_id nao e necessario aqui; para auditoria de densidade
+        # basta guardar composicao binaria pelo maior ID se houver 2 classes.
+        if num_seg_classes == 2:
+            nav_pct_values.append(
+                float(
+                    (
+                        mask == int(nav_class_id)
+                    ).mean()
+                    * 100.0
+                )
+            )
+
+    return {
+        "count_total": len(image_stems),
+        "count_audited": len(audit_stems),
+        "status_counts_audited": {
+            str(k): int(v)
+            for k, v in sorted(
+                status_counts_local.items()
+            )
+        },
+        "group_counts_audited": {
+            str(k): int(v)
+            for k, v in sorted(
+                group_counts_local.items()
+            )
+        },
+        "joint_counts_audited": {
+            f"{g}__label_{lid}": int(v)
+            for (
+                g,
+                lid,
+            ), v in sorted(
+                joint_counts_local.items()
+            )
+        },
+        "nav_pct_audited_min": (
+            float(min(nav_pct_values))
+            if nav_pct_values
+            else None
+        ),
+        "nav_pct_audited_max": (
+            float(max(nav_pct_values))
+            if nav_pct_values
+            else None
+        ),
+    }
+
+
+def audit_pipeline_contract(
+    resolution_root: Path,
+    train_root: Path,
+    val_root: Path,
+    expected_wh: Tuple[int, int],
+    num_seg_classes: int,
+    nav_class_id: int,
+    label_classes: Sequence[str],
+    strict_contract: bool,
+    max_samples: int,
+) -> dict:
+    split_report_path = (
+        resolution_root
+        / "split_report.json"
+    )
+    manifest_path = (
+        resolution_root
+        / "split_manifest.csv"
+    )
+
+    if strict_contract:
+        for required in (
+            split_report_path,
+            manifest_path,
+            resolution_root / "norm_stats.json",
+        ):
+            if not required.is_file():
+                raise FileNotFoundError(
+                    f"Artefato obrigatorio da etapa split ausente: {required}"
+                )
+
+    train_audit = audit_split_root(
+        train_root,
+        expected_wh=expected_wh,
+        num_seg_classes=num_seg_classes,
+        nav_class_id=nav_class_id,
+        label_classes=label_classes,
+        max_samples=max_samples,
+    )
+
+    val_audit = audit_split_root(
+        val_root,
+        expected_wh=expected_wh,
+        num_seg_classes=num_seg_classes,
+        nav_class_id=nav_class_id,
+        label_classes=label_classes,
+        max_samples=max_samples,
+    )
+
+    train_stems = _stem_set(
+        train_root / "images",
+        (
+            ".png",
+            ".jpg",
+            ".jpeg",
+            ".bmp",
+            ".webp",
+        ),
+    )
+    val_stems = _stem_set(
+        val_root / "images",
+        (
+            ".png",
+            ".jpg",
+            ".jpeg",
+            ".bmp",
+            ".webp",
+        ),
+    )
+
+    overlap = sorted(
+        train_stems
+        & val_stems
+    )
+
+    if overlap:
+        raise RuntimeError(
+            f"Vazamento por stem entre train e val: {overlap[:20]}"
+        )
+
+    split_report = {}
+
+    if split_report_path.is_file():
+        with split_report_path.open(
+            "r",
+            encoding="utf-8",
+        ) as f:
+            split_report = json.load(f)
+
+        report_wh = split_report.get(
+            "resolution_wh"
+        )
+
+        if (
+            strict_contract
+            and report_wh is not None
+            and [
+                int(report_wh[0]),
+                int(report_wh[1]),
+            ]
+            != [
+                int(expected_wh[0]),
+                int(expected_wh[1]),
+            ]
+        ):
+            raise RuntimeError(
+                f"split_report resolution={report_wh} "
+                f"mas config={list(expected_wh)}"
+            )
+
+        if strict_contract:
+            tr_report = int(
+                split_report.get(
+                    "train_count",
+                    -1,
+                )
+            )
+            va_report = int(
+                split_report.get(
+                    "val_count",
+                    -1,
+                )
+            )
+
+            if (
+                tr_report
+                != train_audit[
+                    "count_total"
+                ]
+                or va_report
+                != val_audit[
+                    "count_total"
+                ]
+            ):
+                raise RuntimeError(
+                    "Contagem do split_report diverge das pastas atuais: "
+                    f"report train/val={tr_report}/{va_report}, "
+                    f"pastas={train_audit['count_total']}/{val_audit['count_total']}"
+                )
+
+    return {
+        "train": train_audit,
+        "val": val_audit,
+        "split_report_path": str(
+            split_report_path
+        ),
+        "split_manifest_path": str(
+            manifest_path
+        ),
+        "split_report": split_report,
+    }
+
+
+# =============================================================================
+# Main
+# =============================================================================
+
+def main():
+    ap = argparse.ArgumentParser(
+        description="Agri Corridor Teacher v2 - SegFormer OAK-D Lite"
+    )
+
+    ap.add_argument("--config", default="config.json")
+
+    ap.add_argument("--epochs", type=int, default=140)
+    ap.add_argument("--batch", type=int, default=4)
+    ap.add_argument("--num_workers", type=int, default=4)
+    ap.add_argument("--seed", type=int, default=42)
+
+    ap.add_argument("--lr", type=float, default=3e-5)
+    ap.add_argument("--wd", type=float, default=1e-2)
+    ap.add_argument("--grad_accum", type=int, default=2)
+
+    ap.add_argument("--amp", action="store_true")
+    ap.add_argument("--amp_val", action="store_true")
+    ap.add_argument(
+        "--grad_ckpt",
+        action="store_true",
+        help="Ativa gradient checkpointing no SegFormer se suportado.",
+    )
+
+    ap.add_argument(
+        "--norm_stats",
+        type=str,
+        default=None,
+        help=(
+            "Override opcional. Default: "
+            "dataset/<W>x<H>/norm_stats.json produzido pelo split."
+        ),
+    )
+    ap.add_argument("--resume", action="store_true")
+    ap.add_argument("--allow_missing_labels", action="store_true")
+
+    args = ap.parse_args()
+
+    seed_everything(args.seed)
+
+    config_path = Path(args.config).resolve()
+    if not config_path.is_file():
+        raise FileNotFoundError(config_path)
+
+    with config_path.open("r", encoding="utf-8") as f:
+        config = json.load(f)
+
+    train_cfg = deep_merge(
+        DEFAULT_TRAINING,
+        config.get("corridor_training", {}),
+    )
+
+    # -------------------------------------------------------------------------
+    # Contrato do problema
+    # -------------------------------------------------------------------------
+
+    camera_root = resolve_project_root(
+        config_path,
+        str(config.get("camera", "oak-d")),
+    )
+    dataset_root = camera_root / "dataset"
+
+    resolution = config.get("resolucao", [1024, 576])
+    W, H = int(resolution[0]), int(resolution[1])
+
+    resolution_root = (
+        dataset_root
+        / f"{W}x{H}"
+    )
+
+    roi_inicio = float(config.get("roi_inicio", 0.0))
+    roi_tamanho = float(config.get("roi_tamanho", 1.0))
+    backbone = str(config.get("backbone", "nvidia/mit-b0"))
+
+    label_classes = config.get("label_classes")
+    if not isinstance(label_classes, list) or len(label_classes) < 2:
+        raise RuntimeError(
+            "config['label_classes'] obrigatorio e deve conter as classes de status."
+        )
+
+    label_name_by_id = {
+        i: str(name)
+        for i, name in enumerate(label_classes)
+    }
+    num_label_classes = len(label_classes)
+
+    labelmap_path = dataset_root / "labelmap.txt"
+    seg_id2label, seg_label2id = read_labelmap(labelmap_path)
+
+    if len(seg_id2label) != 2:
+        raise RuntimeError(
+            f"Este trainer e especializado em navegavel vs nao navegavel. "
+            f"labelmap possui {len(seg_id2label)} classes: {seg_id2label}"
+        )
+
+    if set(seg_id2label.keys()) != {0, 1}:
+        raise RuntimeError(
+            f"Este contrato espera ids de segmentacao 0 e 1; recebido={seg_id2label}. "
+            "Remapeie o labelmap/masks antes do treino."
+        )
+
+    main_class_name = str(config.get("main_class_name", "navegavel")).lower()
+    if main_class_name not in seg_label2id:
+        raise RuntimeError(
+            f"main_class_name='{main_class_name}' nao existe no labelmap {seg_id2label}"
+        )
+
+    nav_class_id = int(seg_label2id[main_class_name])
+    non_nav_ids = [cid for cid in seg_id2label if cid != nav_class_id]
+    if len(non_nav_ids) != 1:
+        raise RuntimeError("Nao foi possivel resolver classe nao navegavel.")
+    non_nav_class_id = int(non_nav_ids[0])
+
+    num_seg_classes = 2
+
+    # -------------------------------------------------------------------------
+    # Saida
+    # -------------------------------------------------------------------------
+
+    model_family = str(config.get("modelo", "segformer_b0"))
+    model_name = str(config.get("model_name", "nav_mit"))
+
+    save_dir = (
+        camera_root
+        / "backup"
+        / model_family
+        / f"{model_name}_corridor_agri_v2"
+    )
+    save_dir.mkdir(parents=True, exist_ok=True)
+
+    last_path = save_dir / "last.pt"
+    best_operational_path = save_dir / "best_operational.pt"
+    best_nav_path = save_dir / "best_nav.pt"
+    best_status_path = save_dir / "best_status.pt"
+
+    # Um treino novo precisa nascer sem checkpoints da rodada anterior.
+    if not args.resume:
+        for stale_path in (
+            best_operational_path,
+            best_nav_path,
+            best_status_path,
+            last_path,
+        ):
+            if stale_path.exists():
+                stale_path.unlink()
+
+        print("[FRESH] checkpoints antigos removidos antes do novo treino.")
+
+    # -------------------------------------------------------------------------
+    # Device / performance
+    # -------------------------------------------------------------------------
+
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+    opt_cfg = train_cfg["optimization"]
+    try:
+        torch.set_float32_matmul_precision(
+            str(opt_cfg.get("matmul_precision", "high"))
+        )
+    except Exception:
+        pass
+
+    if torch.backends.cudnn.is_available():
+        torch.backends.cudnn.benchmark = bool(
+            opt_cfg.get("cudnn_benchmark", True)
+        )
+
+    print("=" * 78)
+    print(f"Agri Corridor Teacher | {TRAINER_VERSION}")
+    print("=" * 78)
+    print(f"Device              : {device}")
+    print(f"Camera root         : {camera_root}")
+    print(f"Dataset root        : {dataset_root}")
+    print(f"Resolution dataset  : {resolution_root}")
+    print(f"Backbone            : {backbone}")
+    print(f"Resolution          : {W}x{H}")
+    print(f"ROI                 : begin={roi_inicio:.3f} size={roi_tamanho:.3f}")
+    print(f"Seg classes         : {seg_id2label}")
+    print(f"Navegavel           : id={nav_class_id} name={seg_id2label[nav_class_id]}")
+    print(f"Nao navegavel       : id={non_nav_class_id} name={seg_id2label[non_nav_class_id]}")
+    print(f"Status classes      : {label_name_by_id}")
+    print(f"Save dir            : {save_dir}")
+    print("=" * 78)
+
+    # -------------------------------------------------------------------------
+    # Dataset
+    # -------------------------------------------------------------------------
+
+    split_train = resolution_root / "split" / "train"
+    split_val = resolution_root / "split" / "val"
+
+    strict_pipeline_contract = bool(
+        train_cfg["data"].get(
+            "strict_pipeline_contract",
+            True,
+        )
+    )
+
+    preflight_max_samples = int(
+        train_cfg["data"].get(
+            "preflight_max_samples",
+            0,
+        )
+    )
+
+    pipeline_audit = audit_pipeline_contract(
+        resolution_root=resolution_root,
+        train_root=split_train,
+        val_root=split_val,
+        expected_wh=(W, H),
+        num_seg_classes=num_seg_classes,
+        nav_class_id=nav_class_id,
+        label_classes=label_classes,
+        strict_contract=strict_pipeline_contract,
+        max_samples=preflight_max_samples,
+    )
+
+    print(
+        f"[PREFLIGHT] train={pipeline_audit['train']['count_total']} "
+        f"val={pipeline_audit['val']['count_total']} "
+        f"audited(train/val)="
+        f"{pipeline_audit['train']['count_audited']}/"
+        f"{pipeline_audit['val']['count_audited']}"
+    )
+    print(
+        f"[PREFLIGHT] train groups="
+        f"{pipeline_audit['train']['group_counts_audited']}"
+    )
+    print(
+        f"[PREFLIGHT] val groups="
+        f"{pipeline_audit['val']['group_counts_audited']}"
+    )
+
+    base_train = ROISegDataset(
+        str(split_train),
+        str(save_dir),
+        roi_inicio,
+        roi_tamanho,
+        W,
+        H,
+        str(labelmap_path),
+    )
+
+    base_val = ROISegDataset(
+        str(split_val),
+        str(save_dir),
+        roi_inicio,
+        roi_tamanho,
+        W,
+        H,
+        str(labelmap_path),
+    )
+
+    strict_labels = bool(train_cfg["data"].get("strict_labels", True))
+    if args.allow_missing_labels:
+        strict_labels = False
+
+    ds_train = CorridorDataset(
+        base_train,
+        split_train,
+        label_classes=label_classes,
+        strict_labels=strict_labels,
+    )
+
+    ds_val = CorridorDataset(
+        base_val,
+        split_val,
+        label_classes=label_classes,
+        strict_labels=strict_labels,
+    )
+
+    print(
+        f"[DATA] train={len(ds_train)} val={len(ds_val)} "
+        f"missing_labels(train/val)="
+        f"{len(ds_train.missing_labels)}/{len(ds_val.missing_labels)}"
+    )
+
+    all_group_names = sorted(
+        set(ds_train.group_names)
+        | set(ds_val.group_names)
+    )
+
+    group_name_by_id = {
+        i: name
+        for i, name in enumerate(
+            all_group_names
+        )
+    }
+    group_id_by_name = {
+        name: i
+        for i, name in group_name_by_id.items()
+    }
+
+    # Reindexa ambos datasets no mesmo espaço de IDs de grupo.
+    ds_train.group_id_by_name = dict(
+        group_id_by_name
+    )
+    ds_train.group_name_by_id = dict(
+        group_name_by_id
+    )
+    ds_train.group_ids = [
+        group_id_by_name[g]
+        for g in ds_train.group_names
+    ]
+
+    ds_val.group_id_by_name = dict(
+        group_id_by_name
+    )
+    ds_val.group_name_by_id = dict(
+        group_name_by_id
+    )
+    ds_val.group_ids = [
+        group_id_by_name[g]
+        for g in ds_val.group_names
+    ]
+
+    num_groups = len(
+        group_name_by_id
+    )
+
+    print(
+        f"[DATA] groups={group_name_by_id}"
+    )
+
+    # -------------------------------------------------------------------------
+    # Normalizacao obrigatoria
+    # -------------------------------------------------------------------------
+
+    if args.norm_stats:
+        norm_stats_path_expected = Path(
+            args.norm_stats
+        ).resolve()
+    else:
+        norm_stats_path_expected = (
+            resolution_root
+            / "norm_stats.json"
+        )
+
+    mean, std, norm_stats_path, norm_stats_raw = load_norm_stats(
+        norm_stats_path_expected,
+        device=device,
+        expected_resolution_wh=(W, H),
+        strict_contract=strict_pipeline_contract,
+    )
+
+    # O split e o norm_stats precisam descrever o dataset atual.
+    if strict_pipeline_contract:
+        train_count_stats = norm_stats_raw.get(
+            "train_count"
+        )
+        val_count_stats = norm_stats_raw.get(
+            "val_count"
+        )
+
+        if (
+            train_count_stats is not None
+            and int(train_count_stats)
+            != int(
+                pipeline_audit[
+                    "train"
+                ][
+                    "count_total"
+                ]
+            )
+        ):
+            raise RuntimeError(
+                "norm_stats train_count diverge do split atual."
+            )
+
+        if (
+            val_count_stats is not None
+            and int(val_count_stats)
+            != int(
+                pipeline_audit[
+                    "val"
+                ][
+                    "count_total"
+                ]
+            )
+        ):
+            raise RuntimeError(
+                "norm_stats val_count diverge do split atual."
+            )
+
+    print(f"[NORM] source={norm_stats_path}")
+    print(
+        f"[NORM] scope={norm_stats_raw.get('split_scope')} "
+        f"images={norm_stats_raw.get('image_count')}"
+    )
+    print(
+        f"[NORM] mean="
+        f"{mean.detach().cpu().view(-1).tolist()}"
+    )
+    print(
+        f"[NORM] std ="
+        f"{std.detach().cpu().view(-1).tolist()}"
+    )
+
+    # -------------------------------------------------------------------------
+    # Sampler / loaders
+    # -------------------------------------------------------------------------
+
+    sampler = build_training_sampler(
+        ds_train,
+        num_label_classes,
+        train_cfg["sampler"],
+    )
+
+    gen = torch.Generator()
+    gen.manual_seed(args.seed)
+
+    loader_kwargs = {
+        "batch_size": args.batch,
+        "num_workers": args.num_workers,
+        "pin_memory": device.type == "cuda",
+        "collate_fn": collate_corridor,
+        "worker_init_fn": seed_worker,
+        "generator": gen,
+    }
+
+    if args.num_workers > 0:
+        loader_kwargs["persistent_workers"] = bool(
+            opt_cfg.get("persistent_workers", True)
+        )
+        loader_kwargs["prefetch_factor"] = int(
+            opt_cfg.get("prefetch_factor", 2)
+        )
+
+    dl_train = DataLoader(
+        ds_train,
+        shuffle=(sampler is None),
+        sampler=sampler,
+        **loader_kwargs,
+    )
+
+    val_kwargs = dict(loader_kwargs)
+    val_kwargs["num_workers"] = max(0, args.num_workers // 2)
+    if val_kwargs["num_workers"] == 0:
+        val_kwargs.pop("persistent_workers", None)
+        val_kwargs.pop("prefetch_factor", None)
+
+    dl_val = DataLoader(
+        ds_val,
+        shuffle=False,
+        sampler=None,
+        **val_kwargs,
+    )
+
+    # -------------------------------------------------------------------------
+    # Modelo
+    # -------------------------------------------------------------------------
+
+    base_model = SegformerForSemanticSegmentation.from_pretrained(
+        backbone,
+        num_labels=num_seg_classes,
+        ignore_mismatched_sizes=True,
+        use_safetensors=True,
+    )
+
+    base_model.config.output_hidden_states = True
+
+    if args.grad_ckpt:
+        try:
+            base_model.gradient_checkpointing_enable()
+            print("[MODEL] gradient checkpointing=ON")
+        except Exception as exc:
+            print(
+                f"[WARN] gradient checkpointing indisponivel: "
+                f"{type(exc).__name__}: {exc}"
+            )
+
+    base_model.to(device)
+
+    with torch.no_grad():
+        dummy = torch.zeros((1, 3, H, W), dtype=torch.float32, device=device)
+        dummy = (dummy - mean) / std
+        out = base_model(
+            pixel_values=dummy,
+            output_hidden_states=True,
+            return_dict=True,
+        )
+        feat = out.hidden_states[-1]
+        feat_ch = int(feat.shape[1])
+        print(
+            f"[MODEL] dummy seg_native={tuple(out.logits.shape)} "
+            f"feat_last={tuple(feat.shape)} feat_ch={feat_ch}"
+        )
+
+    sh_cfg = train_cfg["status_head"]
+    status_head = CorridorStatusHead(
+        feat_ch=feat_ch,
+        num_seg_classes=num_seg_classes,
+        num_label_classes=num_label_classes,
+        pool_hw=(
+            int(sh_cfg.get("pool_h", 3)),
+            int(sh_cfg.get("pool_w", 4)),
+        ),
+        hidden=int(sh_cfg.get("hidden", 256)),
+        dropout=float(sh_cfg.get("dropout", 0.20)),
+        detach_seg_summary=bool(sh_cfg.get("detach_seg_summary", True)),
+        seg_summary=str(sh_cfg.get("seg_summary", "probabilities")),
+    ).to(device)
+
+    print(f"[MODEL] status_head={status_head.export_config()}")
+
+    # -------------------------------------------------------------------------
+    # Loss weights
+    # -------------------------------------------------------------------------
+
+    cw_cfg = train_cfg["class_weighting"]
+
+    seg_class_weights = estimate_seg_class_weights(
+        ds_train,
+        num_classes=num_seg_classes,
+        ignore_index=IGNORE_INDEX,
+        cfg=cw_cfg,
+    ).to(device)
+
+    if bool(cw_cfg.get("status_enabled", True)):
+        status_class_weights = estimate_status_class_weights(
+            ds_train,
+            num_label_classes=num_label_classes,
+            cfg=cw_cfg,
+        ).to(device)
+    else:
+        status_class_weights = None
+
+    # -------------------------------------------------------------------------
+    # Optimizer / scheduler / scaler
+    # -------------------------------------------------------------------------
+
+    optimizer = build_optimizer(
+        base_model,
+        status_head,
+        base_lr=args.lr,
+        default_wd=args.wd,
+        cfg=train_cfg["optimizer"],
+    )
+
+    updates_per_epoch = math.ceil(len(dl_train) / max(1, args.grad_accum))
+    total_steps = max(1, args.epochs * updates_per_epoch)
+
+    scheduler = build_scheduler(
+        optimizer,
+        total_steps=total_steps,
+        cfg=train_cfg["scheduler"],
+    )
+
+    scaler = GradScaler(enabled=bool(args.amp and device.type == "cuda"))
+
+    augmenter = FieldCorridorAugmenter(
+        train_cfg["augmentation"],
+        ignore_index=IGNORE_INDEX,
+    )
+
+    # -------------------------------------------------------------------------
+    # Resume
+    # -------------------------------------------------------------------------
+
+    start_epoch = 1
+    bests = {
+        "operational": -1.0,
+        "nav_iou": -1.0,
+        "status_macro_f1": -1.0,
+    }
+    early = EarlyStopState()
+
+    if args.resume and last_path.is_file():
+        ckpt = load_checkpoint(
+            last_path,
+            base_model,
+            status_head,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            scaler=scaler,
+        )
+
+        start_epoch = int(ckpt.get("epoch", 0)) + 1
+        bests.update(ckpt.get("bests", {}))
+
+        es = ckpt.get("early_stop", {}) or {}
+        early.best = float(es.get("best", bests.get("operational", -1.0)))
+        early.bad_epochs = int(es.get("bad_epochs", 0))
+
+        print(
+            f"[RESUME] epoch={start_epoch} "
+            f"bests={bests} early={early}"
+        )
+
+        resume_meta = ckpt.get("metadata", {}) or {}
+        gate_enabled = bool(
+            train_cfg.get("operational_gate", {}).get("enabled", True)
+        )
+        if (
+            gate_enabled
+            and resume_meta.get("operational_gate_schema")
+            != "agrobot.corridor.operational_gate.v1"
+        ):
+            print(
+                "[RESUME][OP GATE] checkpoint legado sem gate v1; "
+                "best_operational sera reconquistado por uma epoca elegivel."
+            )
+            bests["operational"] = -1.0
+            if best_operational_path.exists():
+                best_operational_path.unlink()
+
+    # -------------------------------------------------------------------------
+    # Historico persistente de aprendizagem
+    # -------------------------------------------------------------------------
+
+    (
+        metrics_jsonl_path,
+        metrics_json_path,
+        metrics_csv_path,
+    ) = initialize_training_history(
+        save_dir=save_dir,
+        resume=bool(args.resume),
+        start_epoch=start_epoch,
+    )
+
+    print(f"[METRICS] JSONL : {metrics_jsonl_path}")
+    print(f"[METRICS] JSON  : {metrics_json_path}")
+    print(f"[METRICS] CSV   : {metrics_csv_path}")
+
+    # -------------------------------------------------------------------------
+    # Metadata permanente para exportacao futura
+    # -------------------------------------------------------------------------
+
+    metadata_base = {
+        "trainer_version": TRAINER_VERSION,
+        "config": config,
+        "corridor_training": train_cfg,
+        "resolution_wh": [W, H],
+        "resolution_root": str(resolution_root),
+        "split_train": str(split_train),
+        "split_val": str(split_val),
+        "split_report_path": pipeline_audit.get("split_report_path"),
+        "split_manifest_path": pipeline_audit.get("split_manifest_path"),
+        "pipeline_audit": {
+            "train": pipeline_audit["train"],
+            "val": pipeline_audit["val"],
+        },
+        "roi_inicio": roi_inicio,
+        "roi_tamanho": roi_tamanho,
+        "backbone": backbone,
+        "seg_id2label": seg_id2label,
+        "nav_class_id": nav_class_id,
+        "non_nav_class_id": non_nav_class_id,
+        "label_name_by_id": label_name_by_id,
+        "group_name_by_id": group_name_by_id,
+        "norm_stats_path": str(norm_stats_path),
+        "norm_channels": norm_stats_raw.get("channels"),
+        "norm_mean": mean.detach().cpu().view(-1).tolist(),
+        "norm_std": std.detach().cpu().view(-1).tolist(),
+        "status_head": status_head.export_config(),
+        "operational_gate_schema": "agrobot.corridor.operational_gate.v1",
+        "operational_gate_policy": copy.deepcopy(
+            train_cfg.get("operational_gate", {})
+        ),
+        "onnx_target_contract": {
+            "input": "rgb_uint8_nhwc",
+            "input_color_order": "RGB",
+            "input_shape": [1, H, W, 3],
+            "outputs": {
+                "seg_ids": [1, H, W],
+                "label_probs": [1, num_label_classes],
+            },
+            "preprocess_in_graph": True,
+            "argmax_in_graph": True,
+        },
+    }
+
+    # -------------------------------------------------------------------------
+    # Train
+    # -------------------------------------------------------------------------
+
+    freeze_epochs = int(
+        train_cfg["optimizer"].get("freeze_encoder_epochs", 2)
+    )
+
+    early_patience = int(
+        train_cfg["checkpoint"].get("early_stop_patience", 18)
+    )
+    early_delta = float(
+        train_cfg["checkpoint"].get("early_stop_min_delta", 3e-4)
+    )
+
+    grad_clip = float(
+        train_cfg["optimization"].get("grad_clip_norm", 1.0)
+    )
+
+    for epoch in range(start_epoch, args.epochs + 1):
+        encoder_trainable = epoch > freeze_epochs
+        set_encoder_trainable(base_model, encoder_trainable)
+
+        if epoch == start_epoch or epoch == freeze_epochs + 1:
+            print(
+                f"[TRAIN] encoder_trainable={encoder_trainable} "
+                f"(freeze_encoder_epochs={freeze_epochs})"
+            )
+
+        print()
+        print("=" * 78)
+        print(
+            f"Epoch {epoch}/{args.epochs} | "
+            f"aug_strength={min(1.0, max(0.20, epoch / max(1, int(train_cfg['augmentation'].get('ramp_epochs', 6))))):.2f} | "
+            f"status_w={current_status_weight(epoch, train_cfg['loss']):.3f}"
+        )
+        print("=" * 78)
+
+        tr = run_epoch(
+            base_model=base_model,
+            status_head=status_head,
+            loader=dl_train,
+            device=device,
+            mean=mean,
+            std=std,
+            num_seg_classes=num_seg_classes,
+            num_label_classes=num_label_classes,
+            num_groups=num_groups,
+            group_name_by_id=group_name_by_id,
+            nav_class_id=nav_class_id,
+            non_nav_class_id=non_nav_class_id,
+            seg_class_weights=seg_class_weights,
+            status_class_weights=status_class_weights,
+            loss_cfg=train_cfg["loss"],
+            augmenter=augmenter,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            scaler=scaler,
+            amp=args.amp,
+            train=True,
+            grad_accum=max(1, args.grad_accum),
+            grad_clip_norm=grad_clip,
+            epoch=epoch,
+        )
+
+        va = run_epoch(
+            base_model=base_model,
+            status_head=status_head,
+            loader=dl_val,
+            device=device,
+            mean=mean,
+            std=std,
+            num_seg_classes=num_seg_classes,
+            num_label_classes=num_label_classes,
+            num_groups=num_groups,
+            group_name_by_id=group_name_by_id,
+            nav_class_id=nav_class_id,
+            non_nav_class_id=non_nav_class_id,
+            seg_class_weights=seg_class_weights,
+            status_class_weights=status_class_weights,
+            loss_cfg=train_cfg["loss"],
+            augmenter=None,
+            optimizer=None,
+            scheduler=None,
+            scaler=None,
+            amp=args.amp_val,
+            train=False,
+            grad_accum=1,
+            grad_clip_norm=0.0,
+            epoch=epoch,
+        )
+
+        op_score = field_score(
+            va["seg"],
+            va["status"],
+            train_cfg["score"],
+        )
+
+        operational_gate = evaluate_operational_gate(
+            epoch=epoch,
+            val_metrics=va,
+            cfg=train_cfg.get("operational_gate", {}),
+            label_name_by_id=label_name_by_id,
+        )
+
+        print(
+            f"TRAIN loss={tr['loss']:.4f} "
+            f"seg={tr['loss_seg']:.4f} status={tr['loss_status']:.4f} "
+            f"navIoU={tr['seg']['nav_iou']:.4f} "
+            f"navF1={tr['seg']['nav_f1']:.4f} "
+            f"statusF1={tr['status']['macro_f1']:.4f} "
+            f"t={tr['time_s']:.1f}s"
+        )
+
+        print(
+            f"VAL   loss={va['loss']:.4f} "
+            f"seg={va['loss_seg']:.4f} status={va['loss_status']:.4f} "
+            f"navIoU={va['seg']['nav_iou']:.4f} "
+            f"nonNavIoU={va['seg']['non_nav_iou']:.4f} "
+            f"navP/R/F1="
+            f"{va['seg']['nav_precision']:.4f}/"
+            f"{va['seg']['nav_recall']:.4f}/"
+            f"{va['seg']['nav_f1']:.4f} "
+            f"unsafeNav={va['seg']['unsafe_nav_rate']:.4f} "
+            f"safety={va['seg']['safety']:.4f} "
+            f"statusAcc={va['status']['acc']:.4f} "
+            f"statusF1={va['status']['macro_f1']:.4f} "
+            f"FIELD_SCORE={op_score:.4f} "
+            f"t={va['time_s']:.1f}s"
+        )
+
+        print(
+            f"[OP GATE] {format_operational_gate(operational_gate)}"
+        )
+
+        print(
+            "STATUS F1    : "
+            + pretty_status(
+                va["status"]["f1_per_class"],
+                label_name_by_id,
+            )
+        )
+        print(
+            "STATUS Recall: "
+            + pretty_status(
+                va["status"]["recall_per_class"],
+                label_name_by_id,
+            )
+        )
+
+        print("VAL POR GRUPO:")
+        for group_name in sorted(
+            va["seg_by_group"]
+        ):
+            gm = va[
+                "seg_by_group"
+            ][group_name]
+            gs = va[
+                "status_by_group"
+            ][group_name]
+
+            print(
+                f"  {group_name:28s} "
+                f"navIoU={gm['nav_iou']:.4f} "
+                f"nonNavIoU={gm['non_nav_iou']:.4f} "
+                f"unsafe={gm['unsafe_nav_rate']:.4f} "
+                f"statusF1={gs['macro_f1']:.4f}"
+            )
+
+        # Atualiza early stop primeiro.
+        if op_score > early.best + early_delta:
+            early.best = op_score
+            early.bad_epochs = 0
+        else:
+            early.bad_epochs += 1
+
+        metadata_epoch = dict(metadata_base)
+        metadata_epoch.update({
+            "epoch": epoch,
+            "val": va,
+            "train": tr,
+            "field_score": op_score,
+            "operational_gate": operational_gate,
+            "operational_eligible": bool(operational_gate["eligible"]),
+        })
+
+        score_improved_operational = op_score > bests["operational"]
+        improved_operational = (
+            bool(operational_gate["eligible"])
+            and score_improved_operational
+        )
+        improved_nav = va["seg"]["nav_iou"] > bests["nav_iou"]
+        improved_status = va["status"]["macro_f1"] > bests["status_macro_f1"]
+
+        # Best operacional.
+        if improved_operational:
+            bests["operational"] = float(op_score)
+            save_checkpoint(
+                best_operational_path,
+                base_model,
+                status_head,
+                optimizer,
+                scheduler,
+                scaler,
+                epoch,
+                bests,
+                early,
+                metadata_epoch,
+            )
+            print(
+                f"[BEST OPERATIONAL] {op_score:.4f} -> "
+                f"{best_operational_path}"
+            )
+
+        elif score_improved_operational and not operational_gate["eligible"]:
+            print(
+                f"[OP CANDIDATE REJECTED] FIELD_SCORE={op_score:.4f} | "
+                f"{format_operational_gate(operational_gate)}"
+            )
+
+        # Best navegacao.
+        if improved_nav:
+            bests["nav_iou"] = float(va["seg"]["nav_iou"])
+            save_checkpoint(
+                best_nav_path,
+                base_model,
+                status_head,
+                optimizer,
+                scheduler,
+                scaler,
+                epoch,
+                bests,
+                early,
+                metadata_epoch,
+            )
+            print(
+                f"[BEST NAV] {bests['nav_iou']:.4f} -> "
+                f"{best_nav_path}"
+            )
+
+        # Best status.
+        if improved_status:
+            bests["status_macro_f1"] = float(
+                va["status"]["macro_f1"]
+            )
+            save_checkpoint(
+                best_status_path,
+                base_model,
+                status_head,
+                optimizer,
+                scheduler,
+                scaler,
+                epoch,
+                bests,
+                early,
+                metadata_epoch,
+            )
+            print(
+                f"[BEST STATUS] {bests['status_macro_f1']:.4f} -> "
+                f"{best_status_path}"
+            )
+
+        # Last sempre por ultimo para carregar os bests atualizados.
+        save_checkpoint(
+            last_path,
+            base_model,
+            status_head,
+            optimizer,
+            scheduler,
+            scaler,
+            epoch,
+            bests,
+            early,
+            metadata_epoch,
+        )
+
+        lrs = {
+            g.get("group_name", str(i)): float(g["lr"])
+            for i, g in enumerate(optimizer.param_groups)
+        }
+        print(f"[LR] {lrs}")
+        print(
+            f"[ES] best={early.best:.5f} "
+            f"bad={early.bad_epochs}/{early_patience}"
+        )
+
+        augmentation_strength = min(
+            1.0,
+            max(
+                0.20,
+                epoch / max(1, int(train_cfg["augmentation"].get("ramp_epochs", 6))),
+            ),
+        )
+
+        gpu_memory = None
+        if device.type == "cuda":
+            gpu_memory = {
+                "allocated_bytes": int(torch.cuda.memory_allocated(device)),
+                "reserved_bytes": int(torch.cuda.memory_reserved(device)),
+                "max_allocated_bytes": int(torch.cuda.max_memory_allocated(device)),
+                "max_reserved_bytes": int(torch.cuda.max_memory_reserved(device)),
+            }
+
+        epoch_history = {
+            "schema": "agrobot.corridor.training.metrics.v2",
+            "trainer_version": TRAINER_VERSION,
+            "created_at": time.strftime("%Y-%m-%dT%H:%M:%S"),
+            "epoch": int(epoch),
+            "epochs_requested": int(args.epochs),
+            "encoder_trainable": bool(encoder_trainable),
+            "augmentation_strength": float(augmentation_strength),
+            "status_weight": float(current_status_weight(epoch, train_cfg["loss"])),
+            "train": tr,
+            "val": va,
+            "field_score": float(op_score),
+            "operational_gate": operational_gate,
+            "operational_eligible": bool(operational_gate["eligible"]),
+            "operational_score_improved": bool(score_improved_operational),
+            "learning_rates": lrs,
+            "best_improved": {
+                "operational": bool(improved_operational),
+                "nav_iou": bool(improved_nav),
+                "status_macro_f1": bool(improved_status),
+            },
+            "bests_after_epoch": dict(bests),
+            "early_stop": {
+                "best": float(early.best),
+                "bad_epochs": int(early.bad_epochs),
+                "patience": int(early_patience),
+                "min_delta": float(early_delta),
+            },
+            "epoch_time_s": float(tr.get("time_s", 0.0) + va.get("time_s", 0.0)),
+            "gpu_memory": gpu_memory,
+        }
+
+        save_epoch_history(
+            record=epoch_history,
+            jsonl_path=metrics_jsonl_path,
+            json_path=metrics_json_path,
+            csv_path=metrics_csv_path,
+        )
+
+        print(
+            f"[METRICS] epoch={epoch} persistida | "
+            f"FIELD_SCORE={op_score:.4f} | "
+            f"OP_ELIGIBLE={operational_gate['eligible']} | "
+            f"{metrics_csv_path.name}"
+        )
+
+        if early.bad_epochs >= early_patience:
+            print(
+                f"[EARLY STOP] FIELD_SCORE sem melhora relevante por "
+                f"{early.bad_epochs} epocas. "
+                f"(gate de elegibilidade e independente)"
+            )
+            break
+
+    print()
+    print("=" * 78)
+    print("Treino finalizado.")
+    print(
+        f"Best operational : {best_operational_path} "
+        f"({'OK' if best_operational_path.is_file() else 'NAO GERADO - nenhum elegivel'})"
+    )
+    print(f"Best nav         : {best_nav_path}")
+    print(f"Best status      : {best_status_path}")
+    print(f"Metrics JSONL    : {metrics_jsonl_path}")
+    print(f"Metrics JSON     : {metrics_json_path}")
+    print(f"Metrics CSV      : {metrics_csv_path}")
+    print("=" * 78)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/Python/OAK/datasets/oak-d/_5_test_corridor.py b/Python/OAK/datasets/oak-d/_5_test_corridor.py
new file mode 100644
index 000000000..8f4e11478
--- /dev/null
+++ b/Python/OAK/datasets/oak-d/_5_test_corridor.py
@@ -0,0 +1,4052 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+"""
+test_corridor_agri_v2.py
+========================
+
+Viewer/teste manual do modelo frontal OAK-D Lite, alinhado ao pipeline 2026-09.
+
+Compatível com checkpoints do:
+    train_corridor_agri_v2.py
+
+Tarefa:
+    1) segmentação binária:
+       - não navegável
+       - navegável
+
+    2) classificação global do estado do corredor:
+       - config["label_classes"]
+
+Fontes suportadas:
+    A) split com Ground Truth:
+       dataset/<W>x<H>/split/train
+       dataset/<W>x<H>/split/val
+       ou qualquer pasta com:
+         images/
+         masks/       opcional
+         labels/      opcional
+
+    B) pasta simples:
+       qualquer pasta contendo PNG/JPG/etc.
+       Sem GT é tratada como preview/imagens externas.
+
+    C) câmera:
+       OAK-D Lite em tempo real.
+       Captura 1080p por padrão para manter paridade com a coleta/produção,
+       e aplica o MESMO preprocess usado para pasta/dataset.
+
+Exemplos:
+    # Val padrão do pipeline atual:
+    python test_corridor_agri_v2.py
+
+    # Train:
+    python test_corridor_agri_v2.py --split train
+
+    # Pasta estruturada:
+    python test_corridor_agri_v2.py --input dataset/1024x576/split/val
+
+    # Pasta só com imagens:
+    python test_corridor_agri_v2.py --input C:/meus_previews
+
+    # Câmera:
+    python test_corridor_agri_v2.py --camera
+
+    # Avaliar todos sem janela:
+    python test_corridor_agri_v2.py --split val --no-view --save-dir outputs_val
+
+Teclas:
+    D / seta direita / SPACE = próxima
+    A / seta esquerda         = anterior
+    S                         = salva painel atual
+    Q / ESC                   = sair
+
+Decisões de projeto:
+    - checkpoint é a fonte de verdade de arquitetura, resolução, status head
+      e normalização;
+    - não usa o LabelHead antigo;
+    - não usa input quadrado 512;
+    - GT é opcional;
+    - camera/pasta/dataset passam pelo mesmo preprocess;
+    - em dataset com GT mostra erro perigoso (falso navegável) separadamente.
+"""
+
+from __future__ import annotations
+
+import argparse
+import csv
+import json
+import math
+import os
+import re
+import time
+from collections import Counter
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Dict, Iterable, List, Optional, Sequence, Tuple
+
+import cv2
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from transformers import SegformerForSemanticSegmentation
+
+
+# =============================================================================
+# Constants
+# =============================================================================
+
+TESTER_VERSION = "agri_corridor_manual_test_v2.0"
+IGNORE_INDEX = 255
+
+IMG_EXTS = {
+    ".png",
+    ".jpg",
+    ".jpeg",
+    ".bmp",
+    ".webp",
+    ".tif",
+    ".tiff",
+}
+
+DEFAULT_OVERLAY_ALPHA = 0.45
+
+
+# =============================================================================
+# Data
+# =============================================================================
+
+@dataclass(frozen=True)
+class SegClass:
+    cid: int
+    name: str
+    rgb: Tuple[int, int, int]
+
+    @property
+    def bgr(self) -> Tuple[int, int, int]:
+        r, g, b = self.rgb
+        return b, g, r
+
+
+@dataclass(frozen=True)
+class Sample:
+    base: str
+    image: Path
+    mask: Optional[Path]
+    label: Optional[Path]
+    group: str
+    source_root: Path
+
+
+@dataclass
+class RuntimeContract:
+    checkpoint_path: Path
+    checkpoint_epoch: int
+    backbone: str
+    resolution_wh: Tuple[int, int]
+    roi_inicio: float
+    roi_tamanho: float
+
+    seg_id2label: Dict[int, str]
+    nav_class_id: int
+    non_nav_class_id: int
+
+    label_name_by_id: Dict[int, str]
+    group_name_by_id: Dict[int, str]
+
+    norm_mean: np.ndarray
+    norm_std: np.ndarray
+    norm_channels: List[str]
+
+    status_head_cfg: dict
+    trainer_version: str
+
+
+# =============================================================================
+# Generic helpers
+# =============================================================================
+
+def natural_key(text: str):
+    parts = re.split(r"(\d+)", str(text))
+    return [
+        int(p) if p.isdigit() else p.lower()
+        for p in parts
+    ]
+
+
+def safe_json_load(path: Path) -> dict:
+    with path.open("r", encoding="utf-8") as f:
+        data = json.load(f)
+
+    if not isinstance(data, dict):
+        raise ValueError(
+            f"Esperava objeto JSON em {path}"
+        )
+
+    return data
+
+
+def find_project_root(config_path: Path) -> Path:
+    root = config_path.resolve().parent
+
+    if not (root / "dataset").is_dir():
+        raise FileNotFoundError(
+            f"dataset/ não encontrado ao lado de {config_path}. "
+            "Execute este script dentro da pasta oak-d ou passe --config correto."
+        )
+
+    return root
+
+
+def map_by_stem(
+    folder: Path,
+    exts: set[str],
+) -> Dict[str, Path]:
+    if not folder.is_dir():
+        return {}
+
+    result: Dict[str, Path] = {}
+
+    for p in sorted(
+        folder.iterdir(),
+        key=lambda x: natural_key(x.name),
+    ):
+        if (
+            not p.is_file()
+            or p.suffix.lower() not in exts
+        ):
+            continue
+
+        if p.stem in result:
+            raise RuntimeError(
+                f"Stem duplicado {p.stem!r} em {folder}: "
+                f"{result[p.stem].name} / {p.name}"
+            )
+
+        result[p.stem] = p
+
+    return result
+
+
+def ensure_bgr(path: Path) -> np.ndarray:
+    img = cv2.imread(
+        str(path),
+        cv2.IMREAD_COLOR,
+    )
+
+    if img is None:
+        raise RuntimeError(
+            f"Falha ao ler imagem: {path}"
+        )
+
+    return img
+
+
+def clamp01(x: float) -> float:
+    return max(
+        0.0,
+        min(
+            1.0,
+            float(x),
+        ),
+    )
+
+
+# =============================================================================
+# Labelmap
+# =============================================================================
+
+def _parse_rgb_triplet(text: str) -> Optional[Tuple[int, int, int]]:
+    nums = re.findall(
+        r"-?\d+",
+        str(text),
+    )
+
+    if len(nums) < 3:
+        return None
+
+    rgb = tuple(
+        int(x)
+        for x in nums[:3]
+    )
+
+    if not all(
+        0 <= x <= 255
+        for x in rgb
+    ):
+        return None
+
+    return rgb  # type: ignore
+
+
+def load_labelmap(
+    path: Path,
+) -> Dict[int, SegClass]:
+    """
+    Formatos aceitos:
+        navegavel: 40,220,80
+        1 navegavel 40 220 80
+        1:navegavel:40,220,80
+    """
+    if not path.is_file():
+        raise FileNotFoundError(
+            f"labelmap não encontrado: {path}"
+        )
+
+    classes: Dict[int, SegClass] = {}
+    next_id = 0
+
+    for raw in path.read_text(
+        encoding="utf-8"
+    ).splitlines():
+        line = raw.strip()
+
+        if (
+            not line
+            or line.startswith("#")
+        ):
+            continue
+
+        lower = line.lower()
+
+        if any(
+            token in lower
+            for token in (
+                "ignore",
+                "void",
+                "background_ignore",
+            )
+        ):
+            continue
+
+        cid: Optional[int] = None
+        name: Optional[str] = None
+        rgb: Optional[
+            Tuple[int, int, int]
+        ] = None
+
+        # "name: R,G,B" ou "id:name:R,G,B"
+        colon = [
+            p.strip()
+            for p in line.split(":")
+        ]
+
+        if len(colon) >= 2:
+            try:
+                possible_id = int(
+                    colon[0]
+                )
+
+                if len(colon) >= 3:
+                    cid = possible_id
+                    name = colon[1]
+                    rgb = _parse_rgb_triplet(
+                        ":".join(colon[2:])
+                    )
+            except ValueError:
+                name = colon[0]
+                rgb = _parse_rgb_triplet(
+                    ":".join(colon[1:])
+                )
+
+        # "id name R G B"
+        if (
+            name is None
+            or rgb is None
+        ):
+            tokens = re.split(
+                r"[\s,;]+",
+                line,
+            )
+            tokens = [
+                x
+                for x in tokens
+                if x
+            ]
+
+            if (
+                len(tokens) >= 5
+                and tokens[0].lstrip("-").isdigit()
+            ):
+                cid = int(tokens[0])
+                name = tokens[1]
+
+                try:
+                    rgb = (
+                        int(tokens[2]),
+                        int(tokens[3]),
+                        int(tokens[4]),
+                    )
+                except Exception:
+                    rgb = None
+
+        if (
+            name is None
+            or rgb is None
+        ):
+            raise ValueError(
+                f"Linha de labelmap não reconhecida: {line!r}"
+            )
+
+        if cid is None:
+            while next_id in classes:
+                next_id += 1
+            cid = next_id
+
+        classes[int(cid)] = SegClass(
+            cid=int(cid),
+            name=str(name),
+            rgb=(
+                int(rgb[0]),
+                int(rgb[1]),
+                int(rgb[2]),
+            ),
+        )
+
+        next_id = max(
+            next_id,
+            int(cid) + 1,
+        )
+
+    if len(classes) != 2:
+        raise RuntimeError(
+            f"Tester de corredor espera exatamente 2 classes; "
+            f"labelmap={classes}"
+        )
+
+    if set(classes) != {0, 1}:
+        raise RuntimeError(
+            f"Tester espera IDs 0/1; recebido={sorted(classes)}"
+        )
+
+    return classes
+
+
+def colorize_ids(
+    mask_ids: np.ndarray,
+    classes: Dict[int, SegClass],
+) -> np.ndarray:
+    out = np.zeros(
+        (
+            mask_ids.shape[0],
+            mask_ids.shape[1],
+            3,
+        ),
+        dtype=np.uint8,
+    )
+
+    for cid, cls in classes.items():
+        out[
+            mask_ids == int(cid)
+        ] = np.asarray(
+            cls.bgr,
+            dtype=np.uint8,
+        )
+
+    out[
+        mask_ids == IGNORE_INDEX
+    ] = np.asarray(
+        (180, 0, 180),
+        dtype=np.uint8,
+    )
+
+    return out
+
+
+def decode_color_mask(
+    raw_bgr: np.ndarray,
+    classes: Dict[int, SegClass],
+) -> np.ndarray:
+    ids = np.full(
+        raw_bgr.shape[:2],
+        IGNORE_INDEX,
+        dtype=np.uint8,
+    )
+
+    for cid, cls in classes.items():
+        bgr = np.asarray(
+            cls.bgr,
+            dtype=np.uint8,
+        )
+
+        match = np.all(
+            raw_bgr[:, :, :3]
+            == bgr[None, None, :],
+            axis=2,
+        )
+
+        ids[match] = int(cid)
+
+    return ids
+
+
+def load_gt_mask(
+    path: Optional[Path],
+    classes: Dict[int, SegClass],
+) -> Optional[np.ndarray]:
+    if path is None:
+        return None
+
+    raw = cv2.imread(
+        str(path),
+        cv2.IMREAD_UNCHANGED,
+    )
+
+    if raw is None:
+        raise RuntimeError(
+            f"Falha ao ler GT mask: {path}"
+        )
+
+    if raw.ndim == 2:
+        return raw.astype(
+            np.uint8,
+            copy=False,
+        )
+
+    if (
+        raw.ndim == 3
+        and raw.shape[2] >= 3
+    ):
+        return decode_color_mask(
+            raw[:, :, :3],
+            classes,
+        )
+
+    raise RuntimeError(
+        f"Formato de GT mask inválido: {path} shape={raw.shape}"
+    )
+
+
+# =============================================================================
+# Discovery
+# =============================================================================
+
+def discover_flat_dataset(
+    root: Path,
+) -> List[Sample]:
+    images = map_by_stem(
+        root / "images",
+        IMG_EXTS,
+    )
+
+    if not images:
+        return []
+
+    masks = map_by_stem(
+        root / "masks",
+        {".png", ".jpg", ".jpeg", ".bmp", ".tif", ".tiff"},
+    )
+
+    labels = map_by_stem(
+        root / "labels",
+        {".json"},
+    )
+
+    result: List[Sample] = []
+
+    for base, image in sorted(
+        images.items(),
+        key=lambda kv: natural_key(kv[0]),
+    ):
+        group = "unknown"
+
+        label = labels.get(base)
+
+        if label is not None:
+            try:
+                data = safe_json_load(
+                    label
+                )
+                group = str(
+                    data.get(
+                        "split_source_group",
+                        data.get(
+                            "group",
+                            "unknown",
+                        ),
+                    )
+                )
+            except Exception:
+                group = "unknown"
+
+        result.append(
+            Sample(
+                base=base,
+                image=image,
+                mask=masks.get(base),
+                label=label,
+                group=group,
+                source_root=root,
+            )
+        )
+
+    return result
+
+
+def discover_grouped_dataset(
+    root: Path,
+) -> List[Sample]:
+    """
+    Aceita:
+        root/group/<grupo>/{images,masks,labels}
+    ou:
+        root/<grupo>/{images,masks,labels}
+    """
+    group_root = (
+        root / "group"
+        if (root / "group").is_dir()
+        else root
+    )
+
+    samples: List[Sample] = []
+
+    image_dirs = sorted(
+        {
+            p
+            for p in group_root.rglob("images")
+            if p.is_dir()
+        },
+        key=lambda p: natural_key(
+            str(p)
+        ),
+    )
+
+    for images_dir in image_dirs:
+        base_dir = images_dir.parent
+
+        try:
+            group = str(
+                base_dir.relative_to(
+                    group_root
+                )
+            ).replace(
+                "\\",
+                "/",
+            )
+        except Exception:
+            group = base_dir.name
+
+        images = map_by_stem(
+            images_dir,
+            IMG_EXTS,
+        )
+        masks = map_by_stem(
+            base_dir / "masks",
+            {".png", ".jpg", ".jpeg", ".bmp", ".tif", ".tiff"},
+        )
+        labels = map_by_stem(
+            base_dir / "labels",
+            {".json"},
+        )
+
+        for stem, image in sorted(
+            images.items(),
+            key=lambda kv: natural_key(kv[0]),
+        ):
+            samples.append(
+                Sample(
+                    base=stem,
+                    image=image,
+                    mask=masks.get(stem),
+                    label=labels.get(stem),
+                    group=group,
+                    source_root=root,
+                )
+            )
+
+    return samples
+
+
+def discover_plain_folder(
+    root: Path,
+) -> List[Sample]:
+    result: List[Sample] = []
+
+    if not root.is_dir():
+        return result
+
+    for p in sorted(
+        root.iterdir(),
+        key=lambda x: natural_key(x.name),
+    ):
+        if (
+            p.is_file()
+            and p.suffix.lower()
+            in IMG_EXTS
+        ):
+            result.append(
+                Sample(
+                    base=p.stem,
+                    image=p,
+                    mask=None,
+                    label=None,
+                    group="external",
+                    source_root=root,
+                )
+            )
+
+    return result
+
+
+def discover_samples(
+    root: Path,
+) -> Tuple[List[Sample], str]:
+    if not root.is_dir():
+        raise FileNotFoundError(
+            f"Entrada não encontrada: {root}"
+        )
+
+    # 1) split novo / pasta estruturada plana.
+    flat = discover_flat_dataset(
+        root
+    )
+
+    if flat:
+        return flat, "dataset_flat"
+
+    # 2) original/group ou normalized/group.
+    grouped = discover_grouped_dataset(
+        root
+    )
+
+    if grouped:
+        return grouped, "dataset_grouped"
+
+    # 3) pasta de imagens.
+    plain = discover_plain_folder(
+        root
+    )
+
+    if plain:
+        return plain, "image_folder"
+
+    raise RuntimeError(
+        f"Nenhuma imagem reconhecida em: {root}"
+    )
+
+
+def read_gt_label(
+    sample: Sample,
+    label_name_by_id: Dict[int, str],
+) -> Tuple[Optional[int], Optional[str]]:
+    if sample.label is None:
+        return None, None
+
+    data = safe_json_load(
+        sample.label
+    )
+
+    lid = data.get(
+        "label_id"
+    )
+
+    name = (
+        data.get("estado_corredor")
+        or data.get("label")
+        or data.get("state")
+    )
+
+    if lid is not None:
+        lid = int(lid)
+
+        if name is None:
+            name = label_name_by_id.get(
+                lid
+            )
+
+    return (
+        int(lid)
+        if lid is not None
+        else None,
+        str(name)
+        if name is not None
+        else None,
+    )
+
+
+# =============================================================================
+# Status head exact contract
+# =============================================================================
+
+class CorridorStatusHead(nn.Module):
+    """
+    Espelho da cabeça usada por train_corridor_agri_v2.py.
+
+    Não é o LabelHead legado:
+      - feature profunda permanece nativa;
+      - AdaptiveAvgPool 3x4 (ou metadata);
+      - resumo da segmentação separado;
+      - concatena vetores compactos.
+    """
+
+    def __init__(
+        self,
+        feat_ch: int,
+        num_seg_classes: int,
+        num_label_classes: int,
+        pool_hw: Tuple[int, int],
+        hidden: int,
+        dropout: float,
+        detach_seg_summary: bool,
+        seg_summary: str,
+    ):
+        super().__init__()
+
+        self.feat_ch = int(
+            feat_ch
+        )
+        self.num_seg_classes = int(
+            num_seg_classes
+        )
+        self.num_label_classes = int(
+            num_label_classes
+        )
+        self.pool_hw = (
+            int(pool_hw[0]),
+            int(pool_hw[1]),
+        )
+        self.detach_seg_summary = bool(
+            detach_seg_summary
+        )
+        self.seg_summary = str(
+            seg_summary
+        ).lower()
+
+        pooled_cells = (
+            self.pool_hw[0]
+            * self.pool_hw[1]
+        )
+
+        in_dim = (
+            self.feat_ch
+            + self.num_seg_classes
+        ) * pooled_cells
+
+        self.feat_pool = nn.AdaptiveAvgPool2d(
+            self.pool_hw
+        )
+        self.seg_pool = nn.AdaptiveAvgPool2d(
+            self.pool_hw
+        )
+
+        self.net = nn.Sequential(
+            nn.Linear(
+                in_dim,
+                int(hidden),
+            ),
+            nn.ReLU(
+                inplace=True
+            ),
+            nn.Dropout(
+                float(dropout)
+            ),
+            nn.Linear(
+                int(hidden),
+                self.num_label_classes,
+            ),
+        )
+
+    def forward(
+        self,
+        feat: torch.Tensor,
+        seg_logits_native: torch.Tensor,
+    ) -> torch.Tensor:
+        feat_grid = self.feat_pool(
+            feat
+        ).flatten(1)
+
+        seg_src = (
+            seg_logits_native.detach()
+            if self.detach_seg_summary
+            else seg_logits_native
+        )
+
+        if (
+            self.seg_summary
+            == "probabilities"
+        ):
+            seg_src = torch.softmax(
+                seg_src,
+                dim=1,
+            )
+        elif (
+            self.seg_summary
+            != "logits"
+        ):
+            raise RuntimeError(
+                f"seg_summary inválido: "
+                f"{self.seg_summary}"
+            )
+
+        seg_grid = self.seg_pool(
+            seg_src
+        ).flatten(1)
+
+        x = torch.cat(
+            [
+                feat_grid,
+                seg_grid,
+            ],
+            dim=1,
+        )
+
+        return self.net(
+            x
+        )
+
+
+# =============================================================================
+# Checkpoint
+# =============================================================================
+
+def default_checkpoint_path(
+    project_root: Path,
+    config: dict,
+    which: str,
+) -> Path:
+    model_family = str(
+        config.get(
+            "modelo",
+            "segformer_b0",
+        )
+    )
+    model_name = str(
+        config.get(
+            "model_name",
+            "nav_mit",
+        )
+    )
+
+    save_dir = (
+        project_root
+        / "backup"
+        / model_family
+        / f"{model_name}_corridor_agri_v2"
+    )
+
+    filename = {
+        "operational": "best_operational.pt",
+        "nav": "best_nav.pt",
+        "status": "best_status.pt",
+        "last": "last.pt",
+    }[which]
+
+    return save_dir / filename
+
+
+def _int_key_dict(
+    value,
+) -> Dict[int, str]:
+    if not isinstance(
+        value,
+        dict,
+    ):
+        return {}
+
+    return {
+        int(k): str(v)
+        for k, v in value.items()
+    }
+
+
+def build_runtime_from_checkpoint(
+    checkpoint_path: Path,
+    config: dict,
+    device: torch.device,
+):
+    if not checkpoint_path.is_file():
+        raise FileNotFoundError(
+            f"Checkpoint não encontrado: {checkpoint_path}"
+        )
+
+    ckpt = torch.load(
+        checkpoint_path,
+        map_location="cpu",
+        weights_only=False,
+    )
+
+    if "model" not in ckpt:
+        raise RuntimeError(
+            "Checkpoint não possui state_dict 'model'."
+        )
+
+    if "status_head" not in ckpt:
+        raise RuntimeError(
+            "Checkpoint não possui 'status_head'. "
+            "Este tester é para train_corridor_agri_v2.py, "
+            "não para o LabelHead antigo."
+        )
+
+    metadata = ckpt.get(
+        "metadata",
+        {},
+    ) or {}
+
+    trainer_version = str(
+        ckpt.get(
+            "trainer_version",
+            metadata.get(
+                "trainer_version",
+                "unknown",
+            ),
+        )
+    )
+
+    resolution = metadata.get(
+        "resolution_wh",
+        config.get(
+            "resolucao",
+            [1024, 576],
+        ),
+    )
+
+    W = int(
+        resolution[0]
+    )
+    H = int(
+        resolution[1]
+    )
+
+    backbone = str(
+        metadata.get(
+            "backbone",
+            config.get(
+                "backbone",
+                "nvidia/mit-b0",
+            ),
+        )
+    )
+
+    seg_id2label = _int_key_dict(
+        metadata.get(
+            "seg_id2label",
+            {},
+        )
+    )
+
+    if not seg_id2label:
+        seg_id2label = {
+            0: "naonavegavel",
+            1: "navegavel",
+        }
+
+    nav_class_id = int(
+        metadata.get(
+            "nav_class_id",
+            1,
+        )
+    )
+    non_nav_class_id = int(
+        metadata.get(
+            "non_nav_class_id",
+            1 - nav_class_id,
+        )
+    )
+
+    label_name_by_id = _int_key_dict(
+        metadata.get(
+            "label_name_by_id",
+            {},
+        )
+    )
+
+    if not label_name_by_id:
+        raw = config.get(
+            "label_classes",
+            [],
+        )
+        label_name_by_id = {
+            i: str(name)
+            for i, name in enumerate(
+                raw
+            )
+        }
+
+    if not label_name_by_id:
+        raise RuntimeError(
+            "Não consegui resolver classes da segunda cabeça."
+        )
+
+    group_name_by_id = _int_key_dict(
+        metadata.get(
+            "group_name_by_id",
+            {},
+        )
+    )
+
+    mean = metadata.get(
+        "norm_mean"
+    )
+    std = metadata.get(
+        "norm_std"
+    )
+    channels = metadata.get(
+        "norm_channels",
+        ["R", "G", "B"],
+    )
+
+    if (
+        mean is None
+        or std is None
+    ):
+        raise RuntimeError(
+            "Checkpoint não possui norm_mean/norm_std em metadata. "
+            "O tester v2 exige o contrato do trainer v2."
+        )
+
+    mean_np = np.asarray(
+        mean,
+        dtype=np.float32,
+    ).reshape(3)
+    std_np = np.asarray(
+        std,
+        dtype=np.float32,
+    ).reshape(3)
+
+    if not np.all(
+        np.isfinite(mean_np)
+    ) or not np.all(
+        np.isfinite(std_np)
+    ):
+        raise RuntimeError(
+            "Checkpoint possui norm stats NaN/Inf."
+        )
+
+    if np.any(
+        std_np <= 0
+    ):
+        raise RuntimeError(
+            f"Checkpoint possui std inválido: {std_np.tolist()}"
+        )
+
+    status_sd = ckpt[
+        "status_head"
+    ]
+
+    status_meta = dict(
+        metadata.get(
+            "status_head",
+            {},
+        ) or {}
+    )
+
+    # Inferências seguras diretamente do state_dict.
+    if "net.0.weight" not in status_sd:
+        raise RuntimeError(
+            "status_head state_dict não possui net.0.weight."
+        )
+
+    hidden = int(
+        status_sd[
+            "net.0.weight"
+        ].shape[0]
+    )
+
+    final_weight_key = None
+
+    for candidate in (
+        "net.3.weight",
+        "net.4.weight",
+    ):
+        if candidate in status_sd:
+            final_weight_key = candidate
+            break
+
+    if final_weight_key is None:
+        raise RuntimeError(
+            "Não consegui inferir saída do status_head."
+        )
+
+    num_label_classes = int(
+        status_sd[
+            final_weight_key
+        ].shape[0]
+    )
+
+    feat_ch = int(
+        status_meta.get(
+            "feat_ch",
+            0,
+        )
+    )
+
+    num_seg_classes = int(
+        status_meta.get(
+            "num_seg_classes",
+            len(seg_id2label),
+        )
+    )
+
+    pool_hw_raw = status_meta.get(
+        "pool_hw",
+        [3, 4],
+    )
+    pool_hw = (
+        int(pool_hw_raw[0]),
+        int(pool_hw_raw[1]),
+    )
+
+    detach_seg_summary = bool(
+        status_meta.get(
+            "detach_seg_summary",
+            True,
+        )
+    )
+    seg_summary = str(
+        status_meta.get(
+            "seg_summary",
+            "probabilities",
+        )
+    )
+
+    base_model = (
+        SegformerForSemanticSegmentation
+        .from_pretrained(
+            backbone,
+            num_labels=num_seg_classes,
+            ignore_mismatched_sizes=True,
+            use_safetensors=True,
+        )
+    )
+
+    base_model.config.output_hidden_states = True
+
+    base_model.load_state_dict(
+        ckpt["model"],
+        strict=True,
+    )
+
+    base_model.to(
+        device
+    ).eval()
+
+    # Metadata antiga pode não ter feat_ch. Descobre com dummy.
+    if feat_ch <= 0:
+        with torch.inference_mode():
+            dummy = torch.zeros(
+                (
+                    1,
+                    3,
+                    H,
+                    W,
+                ),
+                dtype=torch.float32,
+                device=device,
+            )
+
+            mean_t = torch.tensor(
+                mean_np,
+                dtype=torch.float32,
+                device=device,
+            ).view(
+                1,
+                3,
+                1,
+                1,
+            )
+            std_t = torch.tensor(
+                std_np,
+                dtype=torch.float32,
+                device=device,
+            ).view(
+                1,
+                3,
+                1,
+                1,
+            )
+
+            out = base_model(
+                pixel_values=(
+                    dummy
+                    - mean_t
+                ) / std_t,
+                output_hidden_states=True,
+                return_dict=True,
+            )
+
+            feat_ch = int(
+                out.hidden_states[-1].shape[1]
+            )
+
+    status_head = CorridorStatusHead(
+        feat_ch=feat_ch,
+        num_seg_classes=num_seg_classes,
+        num_label_classes=num_label_classes,
+        pool_hw=pool_hw,
+        hidden=hidden,
+        dropout=0.0,  # irrelevante em eval
+        detach_seg_summary=detach_seg_summary,
+        seg_summary=seg_summary,
+    )
+
+    status_head.load_state_dict(
+        status_sd,
+        strict=True,
+    )
+
+    status_head.to(
+        device
+    ).eval()
+
+    contract = RuntimeContract(
+        checkpoint_path=checkpoint_path,
+        checkpoint_epoch=int(
+            ckpt.get(
+                "epoch",
+                -1,
+            )
+        ),
+        backbone=backbone,
+        resolution_wh=(
+            W,
+            H,
+        ),
+        roi_inicio=float(
+            metadata.get(
+                "roi_inicio",
+                config.get(
+                    "roi_inicio",
+                    0.0,
+                ),
+            )
+        ),
+        roi_tamanho=float(
+            metadata.get(
+                "roi_tamanho",
+                config.get(
+                    "roi_tamanho",
+                    1.0,
+                ),
+            )
+        ),
+        seg_id2label=seg_id2label,
+        nav_class_id=nav_class_id,
+        non_nav_class_id=non_nav_class_id,
+        label_name_by_id=label_name_by_id,
+        group_name_by_id=group_name_by_id,
+        norm_mean=mean_np,
+        norm_std=std_np,
+        norm_channels=[
+            str(x)
+            for x in channels
+        ],
+        status_head_cfg={
+            "feat_ch": feat_ch,
+            "num_seg_classes": num_seg_classes,
+            "num_label_classes": num_label_classes,
+            "pool_hw": list(
+                pool_hw
+            ),
+            "hidden": hidden,
+            "detach_seg_summary": detach_seg_summary,
+            "seg_summary": seg_summary,
+        },
+        trainer_version=trainer_version,
+    )
+
+    return (
+        base_model,
+        status_head,
+        contract,
+        ckpt,
+    )
+
+
+# =============================================================================
+# Preprocess / inference
+# =============================================================================
+
+def compute_roi(
+    h: int,
+    roi_inicio: float,
+    roi_tamanho: float,
+) -> Tuple[int, int]:
+    y0 = int(
+        round(
+            h
+            * float(
+                roi_inicio
+            )
+        )
+    )
+    y1 = int(
+        round(
+            h
+            * min(
+                1.0,
+                float(
+                    roi_inicio
+                    + roi_tamanho
+                ),
+            )
+        )
+    )
+
+    y0 = max(
+        0,
+        min(
+            h - 1,
+            y0,
+        ),
+    )
+    y1 = max(
+        y0 + 1,
+        min(
+            h,
+            y1,
+        ),
+    )
+
+    return y0, y1
+
+
+def prepare_input(
+    img_bgr: np.ndarray,
+    contract: RuntimeContract,
+    device: torch.device,
+):
+    h0, w0 = img_bgr.shape[:2]
+
+    y0, y1 = compute_roi(
+        h0,
+        contract.roi_inicio,
+        contract.roi_tamanho,
+    )
+
+    roi_bgr = img_bgr[
+        y0:y1,
+        0:w0,
+    ]
+
+    W, H = contract.resolution_wh
+
+    resized_bgr = cv2.resize(
+        roi_bgr,
+        (
+            W,
+            H,
+        ),
+        interpolation=cv2.INTER_AREA,
+    )
+
+    rgb = cv2.cvtColor(
+        resized_bgr,
+        cv2.COLOR_BGR2RGB,
+    )
+
+    arr = (
+        rgb.astype(
+            np.float32
+        )
+        / 255.0
+    )
+
+    arr = np.transpose(
+        arr,
+        (
+            2,
+            0,
+            1,
+        ),
+    )
+
+    mean = contract.norm_mean.reshape(
+        3,
+        1,
+        1,
+    )
+    std = contract.norm_std.reshape(
+        3,
+        1,
+        1,
+    )
+
+    arr = (
+        arr
+        - mean
+    ) / np.maximum(
+        std,
+        1e-6,
+    )
+
+    x = torch.from_numpy(
+        np.ascontiguousarray(
+            arr
+        )
+    ).unsqueeze(0).to(
+        device,
+        non_blocking=True,
+    )
+
+    return (
+        x,
+        {
+            "y0": y0,
+            "y1": y1,
+            "source_wh": (
+                w0,
+                h0,
+            ),
+            "roi_wh": (
+                w0,
+                y1 - y0,
+            ),
+        },
+    )
+
+
+@torch.inference_mode()
+def infer(
+    base_model,
+    status_head,
+    img_bgr: np.ndarray,
+    contract: RuntimeContract,
+    device: torch.device,
+) -> dict:
+    x, geometry = prepare_input(
+        img_bgr,
+        contract,
+        device,
+    )
+
+    if device.type == "cuda":
+        torch.cuda.synchronize()
+
+    t0 = time.perf_counter()
+
+    out = base_model(
+        pixel_values=x,
+        output_hidden_states=True,
+        return_dict=True,
+    )
+
+    seg_logits_native = out.logits
+    feat = out.hidden_states[-1]
+
+    status_logits = status_head(
+        feat,
+        seg_logits_native,
+    )
+
+    W, H = contract.resolution_wh
+
+    seg_logits_model = seg_logits_native
+
+    if (
+        seg_logits_model.shape[-2:]
+        != (
+            H,
+            W,
+        )
+    ):
+        seg_logits_model = F.interpolate(
+            seg_logits_model,
+            size=(
+                H,
+                W,
+            ),
+            mode="bilinear",
+            align_corners=False,
+        )
+
+    pred_model = torch.argmax(
+        seg_logits_model,
+        dim=1,
+    )[0]
+
+    status_probs = torch.softmax(
+        status_logits,
+        dim=1,
+    )[0]
+
+    if device.type == "cuda":
+        torch.cuda.synchronize()
+
+    infer_ms = (
+        time.perf_counter()
+        - t0
+    ) * 1000.0
+
+    pred_model_np = (
+        pred_model
+        .detach()
+        .cpu()
+        .numpy()
+        .astype(
+            np.uint8
+        )
+    )
+
+    probs_np = (
+        status_probs
+        .detach()
+        .cpu()
+        .numpy()
+        .astype(
+            np.float32
+        )
+    )
+
+    label_id = int(
+        np.argmax(
+            probs_np
+        )
+    )
+
+    # Volta a segmentação para a geometria original para visualização.
+    w0, h0 = geometry[
+        "source_wh"
+    ]
+    y0 = geometry["y0"]
+    y1 = geometry["y1"]
+
+    pred_roi_source = cv2.resize(
+        pred_model_np,
+        (
+            w0,
+            y1 - y0,
+        ),
+        interpolation=cv2.INTER_NEAREST,
+    )
+
+    pred_full = np.full(
+        (
+            h0,
+            w0,
+        ),
+        IGNORE_INDEX,
+        dtype=np.uint8,
+    )
+
+    pred_full[
+        y0:y1,
+        0:w0,
+    ] = pred_roi_source
+
+    return {
+        "pred_model": pred_model_np,
+        "pred_full": pred_full,
+        "status_probs": probs_np,
+        "label_id": label_id,
+        "label_conf": float(
+            probs_np[
+                label_id
+            ]
+        ),
+        "infer_ms": float(
+            infer_ms
+        ),
+        "geometry": geometry,
+    }
+
+
+def gt_to_model_space(
+    gt_full: Optional[np.ndarray],
+    geometry: dict,
+    model_wh: Tuple[int, int],
+) -> Optional[np.ndarray]:
+    if gt_full is None:
+        return None
+
+    y0 = int(
+        geometry["y0"]
+    )
+    y1 = int(
+        geometry["y1"]
+    )
+
+    roi = gt_full[
+        y0:y1,
+        :,
+    ]
+
+    W, H = model_wh
+
+    return cv2.resize(
+        roi,
+        (
+            W,
+            H,
+        ),
+        interpolation=cv2.INTER_NEAREST,
+    )
+
+
+# =============================================================================
+# Metrics
+# =============================================================================
+
+def confusion_from_arrays(
+    pred: np.ndarray,
+    gt: np.ndarray,
+    num_classes: int,
+) -> np.ndarray:
+    valid = (
+        gt != IGNORE_INDEX
+    )
+
+    p = pred[
+        valid
+    ].astype(
+        np.int64
+    )
+    t = gt[
+        valid
+    ].astype(
+        np.int64
+    )
+
+    valid2 = (
+        (t >= 0)
+        & (t < num_classes)
+        & (p >= 0)
+        & (p < num_classes)
+    )
+
+    p = p[
+        valid2
+    ]
+    t = t[
+        valid2
+    ]
+
+    cm = np.zeros(
+        (
+            num_classes,
+            num_classes,
+        ),
+        dtype=np.int64,
+    )
+
+    if t.size:
+        idx = (
+            t
+            * num_classes
+            + p
+        )
+
+        bins = np.bincount(
+            idx,
+            minlength=(
+                num_classes
+                * num_classes
+            ),
+        )
+
+        cm += bins.reshape(
+            num_classes,
+            num_classes,
+        )
+
+    return cm
+
+
+def metrics_from_cm(
+    cm: np.ndarray,
+    nav_class_id: int,
+    non_nav_class_id: int,
+) -> dict:
+    cm = cm.astype(
+        np.float64
+    )
+
+    tp = np.diag(
+        cm
+    )
+    fp = cm.sum(
+        axis=0
+    ) - tp
+    fn = cm.sum(
+        axis=1
+    ) - tp
+
+    iou = tp / np.maximum(
+        tp + fp + fn,
+        1e-12,
+    )
+
+    precision = tp / np.maximum(
+        tp + fp,
+        1e-12,
+    )
+    recall = tp / np.maximum(
+        tp + fn,
+        1e-12,
+    )
+
+    f1 = (
+        2
+        * precision
+        * recall
+        / np.maximum(
+            precision + recall,
+            1e-12,
+        )
+    )
+
+    acc = (
+        tp.sum()
+        / max(
+            1.0,
+            cm.sum(),
+        )
+    )
+
+    gt_non_nav = cm[
+        non_nav_class_id,
+        :,
+    ].sum()
+
+    unsafe = (
+        cm[
+            non_nav_class_id,
+            nav_class_id,
+        ]
+        / max(
+            1.0,
+            gt_non_nav,
+        )
+    )
+
+    return {
+        "acc": float(
+            acc
+        ),
+        "miou": float(
+            iou.mean()
+        ),
+        "iou_per_class": iou.tolist(),
+        "f1_per_class": f1.tolist(),
+        "nav_iou": float(
+            iou[
+                nav_class_id
+            ]
+        ),
+        "nav_f1": float(
+            f1[
+                nav_class_id
+            ]
+        ),
+        "non_nav_iou": float(
+            iou[
+                non_nav_class_id
+            ]
+        ),
+        "unsafe_nav_rate": float(
+            unsafe
+        ),
+        "safety": float(
+            1.0 - unsafe
+        ),
+    }
+
+
+def status_metrics_from_cm(
+    cm: np.ndarray,
+) -> dict:
+    cm = cm.astype(
+        np.float64
+    )
+
+    tp = np.diag(
+        cm
+    )
+    fp = cm.sum(
+        axis=0
+    ) - tp
+    fn = cm.sum(
+        axis=1
+    ) - tp
+    support = cm.sum(
+        axis=1
+    )
+
+    precision = tp / np.maximum(
+        tp + fp,
+        1e-12,
+    )
+    recall = tp / np.maximum(
+        tp + fn,
+        1e-12,
+    )
+    f1 = (
+        2
+        * precision
+        * recall
+        / np.maximum(
+            precision + recall,
+            1e-12,
+        )
+    )
+
+    present = (
+        support > 0
+    )
+
+    macro_f1 = (
+        float(
+            f1[
+                present
+            ].mean()
+        )
+        if np.any(
+            present
+        )
+        else 0.0
+    )
+
+    acc = (
+        float(
+            tp.sum()
+            / max(
+                1.0,
+                cm.sum(),
+            )
+        )
+    )
+
+    return {
+        "acc": acc,
+        "macro_f1": macro_f1,
+        "f1_per_class": f1.tolist(),
+        "recall_per_class": recall.tolist(),
+        "support_per_class": support.tolist(),
+    }
+
+
+# =============================================================================
+# Visualization
+# =============================================================================
+
+def overlay_ids(
+    img_bgr: np.ndarray,
+    mask_ids: np.ndarray,
+    classes: Dict[int, SegClass],
+    alpha: float,
+) -> np.ndarray:
+    colored = colorize_ids(
+        mask_ids,
+        classes,
+    )
+
+    valid = (
+        mask_ids
+        != IGNORE_INDEX
+    )
+
+    out = img_bgr.copy()
+
+    if np.any(
+        valid
+    ):
+        blended = cv2.addWeighted(
+            img_bgr,
+            1.0 - alpha,
+            colored,
+            alpha,
+            0.0,
+        )
+
+        out[
+            valid
+        ] = blended[
+            valid
+        ]
+
+    return out
+
+
+def error_visual(
+    img_bgr: np.ndarray,
+    gt: Optional[np.ndarray],
+    pred: np.ndarray,
+    nav_class_id: int,
+    non_nav_class_id: int,
+) -> np.ndarray:
+    """
+    Vermelho = falso navegável (perigoso)
+    Amarelo   = navegável bloqueado
+    Verde     = acerto de classe navegável
+    Cinza     = demais acertos
+    Magenta   = GT ignore/desconhecido
+    """
+    out = (
+        img_bgr.astype(
+            np.float32
+        )
+        * 0.45
+    ).astype(
+        np.uint8
+    )
+
+    if gt is None:
+        return out
+
+    if pred.shape != gt.shape:
+        pred = cv2.resize(
+            pred,
+            (
+                gt.shape[1],
+                gt.shape[0],
+            ),
+            interpolation=cv2.INTER_NEAREST,
+        )
+
+    valid = (
+        gt != IGNORE_INDEX
+    )
+
+    unsafe = (
+        valid
+        & (
+            gt
+            == non_nav_class_id
+        )
+        & (
+            pred
+            == nav_class_id
+        )
+    )
+
+    blocked = (
+        valid
+        & (
+            gt
+            == nav_class_id
+        )
+        & (
+            pred
+            == non_nav_class_id
+        )
+    )
+
+    correct_nav = (
+        valid
+        & (
+            gt
+            == nav_class_id
+        )
+        & (
+            pred
+            == nav_class_id
+        )
+    )
+
+    correct_other = (
+        valid
+        & (
+            gt
+            == pred
+        )
+        & ~correct_nav
+    )
+
+    out[
+        correct_other
+    ] = (
+        90,
+        90,
+        90,
+    )
+    out[
+        correct_nav
+    ] = (
+        60,
+        155,
+        60,
+    )
+    out[
+        blocked
+    ] = (
+        0,
+        210,
+        255,
+    )
+    out[
+        unsafe
+    ] = (
+        0,
+        0,
+        255,
+    )
+    out[
+        ~valid
+    ] = (
+        180,
+        0,
+        180,
+    )
+
+    return out
+
+
+def put_title(
+    image: np.ndarray,
+    title: str,
+) -> np.ndarray:
+    out = image.copy()
+
+    overlay = out.copy()
+    cv2.rectangle(
+        overlay,
+        (
+            0,
+            0,
+        ),
+        (
+            out.shape[1],
+            42,
+        ),
+        (
+            15,
+            15,
+            15,
+        ),
+        -1,
+    )
+
+    out = cv2.addWeighted(
+        overlay,
+        0.78,
+        out,
+        0.22,
+        0.0,
+    )
+
+    cv2.putText(
+        out,
+        str(title),
+        (
+            12,
+            29,
+        ),
+        cv2.FONT_HERSHEY_SIMPLEX,
+        0.72,
+        (
+            245,
+            245,
+            245,
+        ),
+        2,
+        cv2.LINE_AA,
+    )
+
+    return out
+
+
+def resize_letterbox(
+    image: np.ndarray,
+    width: int,
+    height: int,
+) -> np.ndarray:
+    h, w = image.shape[:2]
+
+    scale = min(
+        width / max(
+            1,
+            w,
+        ),
+        height / max(
+            1,
+            h,
+        ),
+    )
+
+    nw = max(
+        1,
+        int(
+            round(
+                w
+                * scale
+            )
+        ),
+    )
+    nh = max(
+        1,
+        int(
+            round(
+                h
+                * scale
+            )
+        ),
+    )
+
+    resized = cv2.resize(
+        image,
+        (
+            nw,
+            nh,
+        ),
+        interpolation=(
+            cv2.INTER_AREA
+            if scale < 1.0
+            else cv2.INTER_LINEAR
+        ),
+    )
+
+    canvas = np.zeros(
+        (
+            height,
+            width,
+            3,
+        ),
+        dtype=np.uint8,
+    )
+    canvas[:] = (
+        24,
+        24,
+        24,
+    )
+
+    x0 = (
+        width
+        - nw
+    ) // 2
+    y0 = (
+        height
+        - nh
+    ) // 2
+
+    canvas[
+        y0:y0 + nh,
+        x0:x0 + nw,
+    ] = resized
+
+    return canvas
+
+
+def make_status_panel(
+    probs: np.ndarray,
+    names: Dict[int, str],
+    width: int,
+    height: int,
+    pred_id: int,
+    gt_id: Optional[int],
+    gt_name: Optional[str],
+    label_conf: float,
+    infer_ms: float,
+    seg_metrics: Optional[dict],
+    group: str,
+    filename: str,
+) -> np.ndarray:
+    panel = np.zeros(
+        (
+            height,
+            width,
+            3,
+        ),
+        dtype=np.uint8,
+    )
+    panel[:] = (
+        29,
+        29,
+        29,
+    )
+
+    pred_name = names.get(
+        int(pred_id),
+        f"label_{pred_id}",
+    )
+
+    gt_display = (
+        gt_name
+        or (
+            names.get(
+                int(gt_id),
+                f"label_{gt_id}",
+            )
+            if gt_id is not None
+            else "-"
+        )
+    )
+
+    if gt_id is None:
+        verdict = "SEM GT"
+    else:
+        verdict = (
+            "OK"
+            if int(gt_id)
+            == int(pred_id)
+            else "MISS"
+        )
+
+    header_lines = [
+        f"{group}/{filename}",
+        (
+            f"STATUS: {pred_name} ({pred_id}) "
+            f"conf={label_conf:.3f} | "
+            f"GT={gt_display} | {verdict}"
+        ),
+        f"infer={infer_ms:.2f} ms",
+    ]
+
+    if seg_metrics is not None:
+        header_lines.append(
+            (
+                f"SEG navIoU={seg_metrics['nav_iou']:.3f} "
+                f"nonNavIoU={seg_metrics['non_nav_iou']:.3f} "
+                f"unsafeNav={seg_metrics['unsafe_nav_rate']:.4f} "
+                f"acc={seg_metrics['acc']:.3f}"
+            )
+        )
+
+    y = 28
+
+    for line in header_lines:
+        cv2.putText(
+            panel,
+            line,
+            (
+                14,
+                y,
+            ),
+            cv2.FONT_HERSHEY_SIMPLEX,
+            0.58,
+            (
+                240,
+                240,
+                240,
+            ),
+            1,
+            cv2.LINE_AA,
+        )
+        y += 25
+
+    y += 8
+
+    # Ordena visualmente pelo ID para manter mapa mental estável.
+    n = len(
+        probs
+    )
+    usable_h = max(
+        1,
+        height
+        - y
+        - 10,
+    )
+    row_h = max(
+        24,
+        min(
+            38,
+            usable_h
+            // max(
+                1,
+                n,
+            ),
+        ),
+    )
+
+    label_w = min(
+        260,
+        max(
+            150,
+            width
+            // 5,
+        ),
+    )
+
+    bar_x0 = (
+        label_w
+        + 24
+    )
+    bar_x1 = (
+        width
+        - 90
+    )
+    bar_total = max(
+        20,
+        bar_x1
+        - bar_x0
+    )
+
+    for i in range(n):
+        p = float(
+            probs[i]
+        )
+
+        name = names.get(
+            i,
+            f"label_{i}",
+        )
+
+        is_pred = (
+            i
+            == int(
+                pred_id
+            )
+        )
+        is_gt = (
+            gt_id is not None
+            and i
+            == int(
+                gt_id
+            )
+        )
+
+        txt = f"{i} {name}"
+
+        if is_pred:
+            txt += "  < PRED"
+        if is_gt:
+            txt += "  < GT"
+
+        yy = (
+            y
+            + i
+            * row_h
+        )
+
+        cv2.putText(
+            panel,
+            txt,
+            (
+                14,
+                yy + 18,
+            ),
+            cv2.FONT_HERSHEY_SIMPLEX,
+            0.48,
+            (
+                245,
+                245,
+                245,
+            ),
+            1,
+            cv2.LINE_AA,
+        )
+
+        cv2.rectangle(
+            panel,
+            (
+                bar_x0,
+                yy + 5,
+            ),
+            (
+                bar_x0
+                + bar_total,
+                yy + row_h - 7,
+            ),
+            (
+                58,
+                58,
+                58,
+            ),
+            -1,
+        )
+
+        bw = int(
+            round(
+                bar_total
+                * clamp01(
+                    p
+                )
+            )
+        )
+
+        # Sem codificar semanticamente por cor. Destaque vem do contorno/texto.
+        cv2.rectangle(
+            panel,
+            (
+                bar_x0,
+                yy + 5,
+            ),
+            (
+                bar_x0 + bw,
+                yy + row_h - 7,
+            ),
+            (
+                135,
+                170,
+                105,
+            ),
+            -1,
+        )
+
+        if is_pred:
+            cv2.rectangle(
+                panel,
+                (
+                    bar_x0 - 2,
+                    yy + 3,
+                ),
+                (
+                    bar_x0
+                    + bar_total
+                    + 2,
+                    yy + row_h - 5,
+                ),
+                (
+                    245,
+                    245,
+                    245,
+                ),
+                1,
+            )
+
+        cv2.putText(
+            panel,
+            f"{p:.3f}",
+            (
+                bar_x1 + 8,
+                yy + 18,
+            ),
+            cv2.FONT_HERSHEY_SIMPLEX,
+            0.48,
+            (
+                235,
+                235,
+                235,
+            ),
+            1,
+            cv2.LINE_AA,
+        )
+
+    return panel
+
+
+def make_panel(
+    *,
+    sample: Sample,
+    img_bgr: np.ndarray,
+    gt_full: Optional[np.ndarray],
+    pred_full: np.ndarray,
+    classes: Dict[int, SegClass],
+    contract: RuntimeContract,
+    gt_label_id: Optional[int],
+    gt_label_name: Optional[str],
+    result: dict,
+    sample_seg_metrics: Optional[dict],
+    alpha: float,
+    max_width: int,
+) -> np.ndarray:
+    pred_overlay = overlay_ids(
+        img_bgr,
+        pred_full,
+        classes,
+        alpha,
+    )
+
+    # Para pastas externas/legadas, tolera GT em resolução diferente da imagem.
+    # A métrica continua usando model-space; isto aqui é somente visualização.
+    gt_display = gt_full
+
+    if (
+        gt_display is not None
+        and gt_display.shape[:2] != img_bgr.shape[:2]
+    ):
+        gt_display = cv2.resize(
+            gt_display,
+            (
+                img_bgr.shape[1],
+                img_bgr.shape[0],
+            ),
+            interpolation=cv2.INTER_NEAREST,
+        )
+
+    has_gt = (
+        gt_display is not None
+    )
+
+    tiles: List[
+        Tuple[str, np.ndarray]
+    ] = [
+        (
+            "ORIGINAL",
+            img_bgr,
+        ),
+    ]
+
+    if has_gt:
+        gt_overlay = overlay_ids(
+            img_bgr,
+            gt_display,
+            classes,
+            alpha,
+        )
+
+        err = error_visual(
+            img_bgr,
+            gt_display,
+            pred_full,
+            contract.nav_class_id,
+            contract.non_nav_class_id,
+        )
+
+        tiles.extend([
+            (
+                "GT",
+                gt_overlay,
+            ),
+            (
+                "PRED",
+                pred_overlay,
+            ),
+            (
+                "ERRO | vermelho=falso navegável",
+                err,
+            ),
+        ])
+    else:
+        pred_color = colorize_ids(
+            pred_full,
+            classes,
+        )
+
+        tiles.extend([
+            (
+                "PRED",
+                pred_overlay,
+            ),
+            (
+                "MÁSCARA PRED",
+                pred_color,
+            ),
+        ])
+
+    cols = len(
+        tiles
+    )
+
+    tile_w = max(
+        300,
+        int(
+            max_width
+            / cols
+        ),
+    )
+
+    aspect = (
+        img_bgr.shape[0]
+        / max(
+            1,
+            img_bgr.shape[1],
+        )
+    )
+
+    tile_h = max(
+        240,
+        int(
+            tile_w
+            * aspect
+        ),
+    )
+
+    top = np.concatenate(
+        [
+            put_title(
+                resize_letterbox(
+                    tile,
+                    tile_w,
+                    tile_h,
+                ),
+                title,
+            )
+            for title, tile in tiles
+        ],
+        axis=1,
+    )
+
+    status_h = max(
+        300,
+        110
+        + 34
+        * len(
+            contract.label_name_by_id
+        ),
+    )
+
+    status = make_status_panel(
+        probs=result[
+            "status_probs"
+        ],
+        names=contract.label_name_by_id,
+        width=top.shape[1],
+        height=status_h,
+        pred_id=result[
+            "label_id"
+        ],
+        gt_id=gt_label_id,
+        gt_name=gt_label_name,
+        label_conf=result[
+            "label_conf"
+        ],
+        infer_ms=result[
+            "infer_ms"
+        ],
+        seg_metrics=sample_seg_metrics,
+        group=sample.group,
+        filename=sample.image.name,
+    )
+
+    return np.concatenate(
+        [
+            top,
+            status,
+        ],
+        axis=0,
+    )
+
+
+# =============================================================================
+# Session metrics
+# =============================================================================
+
+class SessionMetrics:
+    def __init__(
+        self,
+        num_seg_classes: int,
+        num_label_classes: int,
+        nav_class_id: int,
+        non_nav_class_id: int,
+    ):
+        self.num_seg_classes = int(
+            num_seg_classes
+        )
+        self.num_label_classes = int(
+            num_label_classes
+        )
+        self.nav_class_id = int(
+            nav_class_id
+        )
+        self.non_nav_class_id = int(
+            non_nav_class_id
+        )
+
+        self.records: Dict[
+            str,
+            dict,
+        ] = {}
+
+    def put(
+        self,
+        key: str,
+        record: dict,
+    ):
+        # Interactive pode voltar/avançar. Uma amostra conta uma única vez.
+        self.records[
+            str(key)
+        ] = record
+
+    def summary(self) -> dict:
+        cm_seg = np.zeros(
+            (
+                self.num_seg_classes,
+                self.num_seg_classes,
+            ),
+            dtype=np.int64,
+        )
+        cm_status = np.zeros(
+            (
+                self.num_label_classes,
+                self.num_label_classes,
+            ),
+            dtype=np.int64,
+        )
+
+        infer_ms = []
+        groups = Counter()
+
+        for record in self.records.values():
+            if (
+                record.get(
+                    "cm_seg"
+                )
+                is not None
+            ):
+                cm_seg += record[
+                    "cm_seg"
+                ]
+
+            gt_label_id = record.get(
+                "gt_label_id"
+            )
+            pred_label_id = record.get(
+                "pred_label_id"
+            )
+
+            if (
+                gt_label_id is not None
+                and pred_label_id is not None
+                and 0
+                <= int(
+                    gt_label_id
+                )
+                < self.num_label_classes
+                and 0
+                <= int(
+                    pred_label_id
+                )
+                < self.num_label_classes
+            ):
+                cm_status[
+                    int(
+                        gt_label_id
+                    ),
+                    int(
+                        pred_label_id
+                    ),
+                ] += 1
+
+            infer_ms.append(
+                float(
+                    record.get(
+                        "infer_ms",
+                        0.0,
+                    )
+                )
+            )
+            groups[
+                str(
+                    record.get(
+                        "group",
+                        "unknown",
+                    )
+                )
+            ] += 1
+
+        seg = (
+            metrics_from_cm(
+                cm_seg,
+                self.nav_class_id,
+                self.non_nav_class_id,
+            )
+            if cm_seg.sum()
+            > 0
+            else None
+        )
+
+        status = (
+            status_metrics_from_cm(
+                cm_status
+            )
+            if cm_status.sum()
+            > 0
+            else None
+        )
+
+        return {
+            "samples_unique": len(
+                self.records
+            ),
+            "groups": dict(
+                sorted(
+                    groups.items()
+                )
+            ),
+            "infer_ms_mean": (
+                float(
+                    np.mean(
+                        infer_ms
+                    )
+                )
+                if infer_ms
+                else None
+            ),
+            "infer_ms_p95": (
+                float(
+                    np.percentile(
+                        infer_ms,
+                        95,
+                    )
+                )
+                if infer_ms
+                else None
+            ),
+            "seg": seg,
+            "status": status,
+            "cm_seg": cm_seg.tolist(),
+            "cm_status": cm_status.tolist(),
+        }
+
+
+def save_records_csv(
+    path: Path,
+    records: Dict[str, dict],
+):
+    path.parent.mkdir(
+        parents=True,
+        exist_ok=True,
+    )
+
+    fields = [
+        "base",
+        "group",
+        "image",
+        "infer_ms",
+        "gt_label_id",
+        "pred_label_id",
+        "label_conf",
+        "label_ok",
+        "seg_acc",
+        "nav_iou",
+        "non_nav_iou",
+        "unsafe_nav_rate",
+    ]
+
+    with path.open(
+        "w",
+        newline="",
+        encoding="utf-8-sig",
+    ) as f:
+        writer = csv.DictWriter(
+            f,
+            fieldnames=fields,
+        )
+        writer.writeheader()
+
+        for record in records.values():
+            writer.writerow({
+                key: record.get(
+                    key
+                )
+                for key in fields
+            })
+
+
+# =============================================================================
+# Camera
+# =============================================================================
+
+class OakCamera:
+    """
+    Compatibilidade simples com DepthAI moderno e legado.
+
+    O objetivo é fornecer frame BGR 1080p ao mesmo preprocess usado pelos arquivos.
+    """
+
+    def __init__(
+        self,
+        fps: float,
+        resolution: str,
+    ):
+        self.fps = float(
+            fps
+        )
+        self.resolution = str(
+            resolution
+        ).lower()
+
+        self.mode = None
+        self.pipeline = None
+        self.device = None
+        self.queue = None
+
+    def start(self):
+        try:
+            import depthai as dai
+        except ImportError as exc:
+            raise ImportError(
+                "Modo câmera exige depthai instalado."
+            ) from exc
+
+        self.dai = dai
+
+        target_size = (
+            (1920, 1080)
+            if self.resolution == "1080p"
+            else (
+                1280,
+                720,
+            )
+        )
+
+        # API nova, igual ao estilo do script anterior.
+        try:
+            pipeline = dai.Pipeline()
+
+            cam = pipeline.create(
+                dai.node.Camera
+            ).build()
+
+            out = cam.requestOutput(
+                size=target_size,
+                type=dai.ImgFrame.Type.BGR888p,
+                fps=self.fps,
+            )
+
+            queue = out.createOutputQueue(
+                maxSize=4,
+                blocking=False,
+            )
+
+            pipeline.start()
+
+            self.mode = "modern"
+            self.pipeline = pipeline
+            self.queue = queue
+
+            print(
+                f"[CAM] DepthAI modern API | "
+                f"{target_size[0]}x{target_size[1]} @ {self.fps:g}"
+            )
+            return self
+
+        except Exception as modern_exc:
+            print(
+                f"[CAM] modern API indisponível: "
+                f"{type(modern_exc).__name__}: {modern_exc}"
+            )
+
+        # API clássica.
+        pipeline = dai.Pipeline()
+
+        cam = pipeline.create(
+            dai.node.ColorCamera
+        )
+        xout = pipeline.create(
+            dai.node.XLinkOut
+        )
+        xout.setStreamName(
+            "rgb"
+        )
+
+        try:
+            cam.setBoardSocket(
+                dai.CameraBoardSocket.CAM_A
+            )
+        except Exception:
+            pass
+
+        if self.resolution == "1080p":
+            cam.setResolution(
+                dai.ColorCameraProperties.SensorResolution.THE_1080_P
+            )
+        else:
+            cam.setResolution(
+                dai.ColorCameraProperties.SensorResolution.THE_720_P
+            )
+
+        cam.setFps(
+            self.fps
+        )
+        cam.setInterleaved(
+            False
+        )
+
+        # getCvFrame entrega BGR ao host.
+        cam.video.link(
+            xout.input
+        )
+
+        device = dai.Device(
+            pipeline
+        )
+        queue = device.getOutputQueue(
+            name="rgb",
+            maxSize=4,
+            blocking=False,
+        )
+
+        self.mode = "classic"
+        self.pipeline = pipeline
+        self.device = device
+        self.queue = queue
+
+        print(
+            f"[CAM] DepthAI classic API | "
+            f"{target_size[0]}x{target_size[1]} @ {self.fps:g}"
+        )
+
+        return self
+
+    def read(self) -> np.ndarray:
+        if self.queue is None:
+            raise RuntimeError(
+                "Câmera não inicializada."
+            )
+
+        if self.mode == "modern":
+            frame = self.queue.get()
+        else:
+            frame = self.queue.get()
+
+        img = frame.getCvFrame()
+
+        if img is None:
+            raise RuntimeError(
+                "Frame OAK vazio."
+            )
+
+        return img
+
+    def running(self) -> bool:
+        if self.mode == "modern":
+            try:
+                return bool(
+                    self.pipeline.isRunning()
+                )
+            except Exception:
+                return True
+
+        return True
+
+    def close(self):
+        try:
+            if (
+                self.mode == "modern"
+                and self.pipeline is not None
+            ):
+                self.pipeline.stop()
+        except Exception:
+            pass
+
+        try:
+            if self.device is not None:
+                self.device.close()
+        except Exception:
+            pass
+
+
+# =============================================================================
+# Console
+# =============================================================================
+
+def print_contract(
+    contract: RuntimeContract,
+    device: torch.device,
+):
+    print("=" * 92)
+    print(
+        f"Agri Corridor Manual Test | {TESTER_VERSION}"
+    )
+    print("=" * 92)
+    print(
+        f"Device             : {device}"
+    )
+    print(
+        f"Checkpoint         : {contract.checkpoint_path}"
+    )
+    print(
+        f"Epoch              : {contract.checkpoint_epoch}"
+    )
+    print(
+        f"Trainer            : {contract.trainer_version}"
+    )
+    print(
+        f"Backbone           : {contract.backbone}"
+    )
+    print(
+        f"Resolution         : "
+        f"{contract.resolution_wh[0]}x{contract.resolution_wh[1]}"
+    )
+    print(
+        f"ROI                : "
+        f"begin={contract.roi_inicio:.3f} "
+        f"size={contract.roi_tamanho:.3f}"
+    )
+    print(
+        f"Seg classes        : {contract.seg_id2label}"
+    )
+    print(
+        f"Nav / NonNav       : "
+        f"{contract.nav_class_id} / {contract.non_nav_class_id}"
+    )
+    print(
+        f"Status classes     : {contract.label_name_by_id}"
+    )
+    print(
+        f"Status head        : {contract.status_head_cfg}"
+    )
+    print(
+        f"Norm channels      : {contract.norm_channels}"
+    )
+    print(
+        f"Norm mean          : {contract.norm_mean.tolist()}"
+    )
+    print(
+        f"Norm std           : {contract.norm_std.tolist()}"
+    )
+    print("=" * 92)
+
+
+def print_summary(
+    summary: dict,
+    contract: RuntimeContract,
+):
+    print()
+    print("=" * 92)
+    print("RESUMO DA SESSÃO")
+    print("=" * 92)
+    print(
+        f"Amostras únicas    : {summary['samples_unique']}"
+    )
+    print(
+        f"Grupos             : {summary['groups']}"
+    )
+
+    if (
+        summary.get(
+            "infer_ms_mean"
+        )
+        is not None
+    ):
+        print(
+            f"Inferência          : "
+            f"mean={summary['infer_ms_mean']:.2f} ms "
+            f"p95={summary['infer_ms_p95']:.2f} ms"
+        )
+
+    seg = summary.get(
+        "seg"
+    )
+
+    if seg:
+        print(
+            f"SEG                : "
+            f"acc={seg['acc']:.4f} "
+            f"mIoU={seg['miou']:.4f} "
+            f"navIoU={seg['nav_iou']:.4f} "
+            f"nonNavIoU={seg['non_nav_iou']:.4f} "
+            f"unsafeNav={seg['unsafe_nav_rate']:.5f} "
+            f"safety={seg['safety']:.5f}"
+        )
+
+    status = summary.get(
+        "status"
+    )
+
+    if status:
+        print(
+            f"STATUS             : "
+            f"acc={status['acc']:.4f} "
+            f"macroF1={status['macro_f1']:.4f}"
+        )
+
+        for i, f1 in enumerate(
+            status[
+                "f1_per_class"
+            ]
+        ):
+            print(
+                f"  {i:2d} "
+                f"{contract.label_name_by_id.get(i, str(i)):28s} "
+                f"F1={float(f1):.4f} "
+                f"support={int(status['support_per_class'][i])}"
+            )
+
+    print("=" * 92)
+
+
+# =============================================================================
+# Main
+# =============================================================================
+
+def main():
+    ap = argparse.ArgumentParser(
+        description=(
+            "Teste manual OAK-D Lite: dataset, pasta de imagens ou câmera."
+        )
+    )
+
+    ap.add_argument(
+        "--config",
+        default="config.json",
+    )
+
+    ap.add_argument(
+        "--ckpt",
+        default=None,
+        help="Checkpoint .pt explícito.",
+    )
+
+    ap.add_argument(
+        "--checkpoint",
+        default="operational",
+        choices=[
+            "operational",
+            "nav",
+            "status",
+            "last",
+        ],
+        help=(
+            "Checkpoint padrão quando --ckpt não for informado."
+        ),
+    )
+
+    ap.add_argument(
+        "--split",
+        default="val",
+        choices=[
+            "train",
+            "val",
+            "test",
+        ],
+        help=(
+            "Split padrão dentro de dataset/<W>x<H>/split."
+        ),
+    )
+
+    ap.add_argument(
+        "--input",
+        default=None,
+        help=(
+            "Pasta estruturada ou pasta simples de imagens. "
+            "Se omitido usa --split."
+        ),
+    )
+
+    ap.add_argument(
+        "--camera",
+        action="store_true",
+        help="Usa OAK-D Lite em tempo real.",
+    )
+
+    ap.add_argument(
+        "--camera-fps",
+        type=float,
+        default=20.0,
+    )
+
+    ap.add_argument(
+        "--camera-res",
+        default="1080p",
+        choices=[
+            "1080p",
+            "720p",
+        ],
+    )
+
+    ap.add_argument(
+        "--alpha",
+        type=float,
+        default=DEFAULT_OVERLAY_ALPHA,
+    )
+
+    ap.add_argument(
+        "--max-width",
+        type=int,
+        default=1800,
+        help="Largura aproximada máxima do painel.",
+    )
+
+    ap.add_argument(
+        "--start",
+        type=int,
+        default=0,
+    )
+
+    ap.add_argument(
+        "--limit",
+        type=int,
+        default=0,
+        help="0 = todas.",
+    )
+
+    ap.add_argument(
+        "--no-view",
+        action="store_true",
+        help="Processa lote sem abrir janela.",
+    )
+
+    ap.add_argument(
+        "--save-dir",
+        default=None,
+        help="Salva painéis + resultados.",
+    )
+
+    args = ap.parse_args()
+
+    if args.camera and args.input:
+        raise ValueError(
+            "Use --camera OU --input, não ambos."
+        )
+
+    config_path = Path(
+        args.config
+    ).resolve()
+
+    if not config_path.is_file():
+        raise FileNotFoundError(
+            config_path
+        )
+
+    config = safe_json_load(
+        config_path
+    )
+
+    project_root = find_project_root(
+        config_path
+    )
+
+    device = torch.device(
+        "cuda"
+        if torch.cuda.is_available()
+        else "cpu"
+    )
+
+    if args.ckpt:
+        checkpoint_path = Path(
+            args.ckpt
+        ).resolve()
+    else:
+        checkpoint_path = default_checkpoint_path(
+            project_root,
+            config,
+            args.checkpoint,
+        )
+
+    (
+        base_model,
+        status_head,
+        contract,
+        checkpoint_raw,
+    ) = build_runtime_from_checkpoint(
+        checkpoint_path,
+        config,
+        device,
+    )
+
+    print_contract(
+        contract,
+        device,
+    )
+
+    labelmap_path = (
+        project_root
+        / "dataset"
+        / "labelmap.txt"
+    )
+
+    classes = load_labelmap(
+        labelmap_path
+    )
+
+    # Sanity entre checkpoint e labelmap atual.
+    for cid, expected_name in contract.seg_id2label.items():
+        got = classes.get(
+            int(cid)
+        )
+
+        if (
+            got is not None
+            and got.name.lower()
+            != str(
+                expected_name
+            ).lower()
+        ):
+            print(
+                f"[WARN] labelmap atual id={cid} name={got.name!r} "
+                f"difere do checkpoint {expected_name!r}"
+            )
+
+    save_dir = (
+        Path(
+            args.save_dir
+        ).resolve()
+        if args.save_dir
+        else None
+    )
+
+    if save_dir is not None:
+        save_dir.mkdir(
+            parents=True,
+            exist_ok=True,
+        )
+
+    session = SessionMetrics(
+        num_seg_classes=len(
+            contract.seg_id2label
+        ),
+        num_label_classes=len(
+            contract.label_name_by_id
+        ),
+        nav_class_id=contract.nav_class_id,
+        non_nav_class_id=contract.non_nav_class_id,
+    )
+
+    # -------------------------------------------------------------------------
+    # Camera
+    # -------------------------------------------------------------------------
+
+    if args.camera:
+        if args.no_view:
+            raise ValueError(
+                "--camera com --no-view não faz sentido neste tester manual."
+            )
+
+        camera = OakCamera(
+            fps=args.camera_fps,
+            resolution=args.camera_res,
+        ).start()
+
+        win = "Agrobot | Corridor Test | CAMERA"
+        cv2.namedWindow(
+            win,
+            cv2.WINDOW_NORMAL,
+        )
+
+        frame_index = 0
+        fps_ema = None
+        last_t = time.perf_counter()
+
+        try:
+            while camera.running():
+                frame = camera.read()
+
+                result = infer(
+                    base_model,
+                    status_head,
+                    frame,
+                    contract,
+                    device,
+                )
+
+                now = time.perf_counter()
+                inst_fps = (
+                    1.0
+                    / max(
+                        1e-6,
+                        now - last_t,
+                    )
+                )
+                last_t = now
+
+                fps_ema = (
+                    inst_fps
+                    if fps_ema is None
+                    else (
+                        0.85
+                        * fps_ema
+                        + 0.15
+                        * inst_fps
+                    )
+                )
+
+                fake_sample = Sample(
+                    base=f"camera_{frame_index:08d}",
+                    image=Path(
+                        f"camera_{frame_index:08d}.png"
+                    ),
+                    mask=None,
+                    label=None,
+                    group="camera",
+                    source_root=project_root,
+                )
+
+                panel = make_panel(
+                    sample=fake_sample,
+                    img_bgr=frame,
+                    gt_full=None,
+                    pred_full=result[
+                        "pred_full"
+                    ],
+                    classes=classes,
+                    contract=contract,
+                    gt_label_id=None,
+                    gt_label_name=None,
+                    result=result,
+                    sample_seg_metrics=None,
+                    alpha=args.alpha,
+                    max_width=args.max_width,
+                )
+
+                fps_text = (
+                    f"CAM FPS={fps_ema:.1f} | "
+                    f"infer={result['infer_ms']:.2f} ms | "
+                    f"Q/ESC sair | S salvar"
+                )
+
+                cv2.putText(
+                    panel,
+                    fps_text,
+                    (
+                        12,
+                        panel.shape[0] - 12,
+                    ),
+                    cv2.FONT_HERSHEY_SIMPLEX,
+                    0.55,
+                    (
+                        245,
+                        245,
+                        245,
+                    ),
+                    1,
+                    cv2.LINE_AA,
+                )
+
+                cv2.imshow(
+                    win,
+                    panel,
+                )
+
+                key = (
+                    cv2.waitKey(
+                        1
+                    )
+                    & 0xFF
+                )
+
+                if key in (
+                    ord("q"),
+                    ord("Q"),
+                    27,
+                ):
+                    break
+
+                if (
+                    key
+                    in (
+                        ord("s"),
+                        ord("S"),
+                    )
+                    and save_dir
+                    is not None
+                ):
+                    out = (
+                        save_dir
+                        / f"camera_{frame_index:08d}.jpg"
+                    )
+                    cv2.imwrite(
+                        str(out),
+                        panel,
+                        [
+                            cv2.IMWRITE_JPEG_QUALITY,
+                            92,
+                        ],
+                    )
+                    print(
+                        f"[SAVE] {out}"
+                    )
+
+                frame_index += 1
+
+        finally:
+            camera.close()
+            cv2.destroyAllWindows()
+
+        return
+
+    # -------------------------------------------------------------------------
+    # Finite source
+    # -------------------------------------------------------------------------
+
+    W, H = contract.resolution_wh
+    resolution_root = (
+        project_root
+        / "dataset"
+        / f"{W}x{H}"
+    )
+
+    if args.input:
+        input_root = Path(
+            args.input
+        ).resolve()
+    else:
+        input_root = (
+            resolution_root
+            / "split"
+            / args.split
+        )
+
+        if (
+            not input_root.is_dir()
+            and args.split
+            == "test"
+        ):
+            print(
+                "[WARN] split/test não existe; usando split/val."
+            )
+            input_root = (
+                resolution_root
+                / "split"
+                / "val"
+            )
+
+    samples, source_kind = discover_samples(
+        input_root
+    )
+
+    start = max(
+        0,
+        min(
+            len(
+                samples
+            ) - 1,
+            int(
+                args.start
+            ),
+        ),
+    )
+
+    if args.limit > 0:
+        samples = samples[
+            start:
+            start
+            + int(
+                args.limit
+            )
+        ]
+        idx = 0
+    else:
+        idx = start
+
+    print(
+        f"[SOURCE] kind={source_kind} root={input_root}"
+    )
+    print(
+        f"[SOURCE] samples={len(samples)}"
+    )
+
+    has_any_gt_mask = any(
+        s.mask is not None
+        for s in samples
+    )
+    has_any_gt_label = any(
+        s.label is not None
+        for s in samples
+    )
+
+    print(
+        f"[SOURCE] GT mask={has_any_gt_mask} "
+        f"GT label={has_any_gt_label}"
+    )
+
+    win = "Agrobot | Corridor Test | FILES"
+
+    if not args.no_view:
+        cv2.namedWindow(
+            win,
+            cv2.WINDOW_NORMAL,
+        )
+
+    while True:
+        if not samples:
+            break
+
+        sample = samples[
+            idx
+        ]
+
+        img = ensure_bgr(
+            sample.image
+        )
+
+        gt_full = load_gt_mask(
+            sample.mask,
+            classes,
+        )
+
+        (
+            gt_label_id,
+            gt_label_name,
+        ) = read_gt_label(
+            sample,
+            contract.label_name_by_id,
+        )
+
+        result = infer(
+            base_model,
+            status_head,
+            img,
+            contract,
+            device,
+        )
+
+        gt_model = gt_to_model_space(
+            gt_full,
+            result[
+                "geometry"
+            ],
+            contract.resolution_wh,
+        )
+
+        cm_seg = None
+        sample_seg_metrics = None
+
+        if gt_model is not None:
+            cm_seg = confusion_from_arrays(
+                result[
+                    "pred_model"
+                ],
+                gt_model,
+                num_classes=len(
+                    contract.seg_id2label
+                ),
+            )
+
+            sample_seg_metrics = metrics_from_cm(
+                cm_seg,
+                contract.nav_class_id,
+                contract.non_nav_class_id,
+            )
+
+        pred_label_id = int(
+            result[
+                "label_id"
+            ]
+        )
+
+        label_ok = (
+            None
+            if gt_label_id is None
+            else (
+                int(
+                    gt_label_id
+                )
+                == pred_label_id
+            )
+        )
+
+        record = {
+            "base": sample.base,
+            "group": sample.group,
+            "image": str(
+                sample.image
+            ),
+            "infer_ms": float(
+                result[
+                    "infer_ms"
+                ]
+            ),
+            "gt_label_id": gt_label_id,
+            "pred_label_id": pred_label_id,
+            "label_conf": float(
+                result[
+                    "label_conf"
+                ]
+            ),
+            "label_ok": label_ok,
+            "cm_seg": cm_seg,
+            "seg_acc": (
+                sample_seg_metrics[
+                    "acc"
+                ]
+                if sample_seg_metrics
+                else None
+            ),
+            "nav_iou": (
+                sample_seg_metrics[
+                    "nav_iou"
+                ]
+                if sample_seg_metrics
+                else None
+            ),
+            "non_nav_iou": (
+                sample_seg_metrics[
+                    "non_nav_iou"
+                ]
+                if sample_seg_metrics
+                else None
+            ),
+            "unsafe_nav_rate": (
+                sample_seg_metrics[
+                    "unsafe_nav_rate"
+                ]
+                if sample_seg_metrics
+                else None
+            ),
+        }
+
+        key_id = (
+            f"{sample.group}/{sample.base}"
+        )
+
+        session.put(
+            key_id,
+            record,
+        )
+
+        panel = make_panel(
+            sample=sample,
+            img_bgr=img,
+            gt_full=gt_full,
+            pred_full=result[
+                "pred_full"
+            ],
+            classes=classes,
+            contract=contract,
+            gt_label_id=gt_label_id,
+            gt_label_name=gt_label_name,
+            result=result,
+            sample_seg_metrics=sample_seg_metrics,
+            alpha=args.alpha,
+            max_width=args.max_width,
+        )
+
+        if save_dir is not None:
+            auto_save = bool(
+                args.no_view
+            )
+
+            if auto_save:
+                out = (
+                    save_dir
+                    / (
+                        f"{idx:06d}_"
+                        f"{sample.group.replace('/', '_')}_"
+                        f"{sample.base}.jpg"
+                    )
+                )
+                cv2.imwrite(
+                    str(out),
+                    panel,
+                    [
+                        cv2.IMWRITE_JPEG_QUALITY,
+                        92,
+                    ],
+                )
+
+        if args.no_view:
+            print(
+                f"[{idx + 1:5d}/{len(samples):5d}] "
+                f"{sample.group}/{sample.base} | "
+                f"status="
+                f"{contract.label_name_by_id.get(pred_label_id, pred_label_id)} "
+                f"conf={result['label_conf']:.3f} "
+                f"infer={result['infer_ms']:.2f}ms"
+                + (
+                    (
+                        f" | navIoU={sample_seg_metrics['nav_iou']:.3f} "
+                        f"unsafe={sample_seg_metrics['unsafe_nav_rate']:.4f}"
+                    )
+                    if sample_seg_metrics
+                    else ""
+                )
+                + (
+                    (
+                        f" | label={'OK' if label_ok else 'MISS'}"
+                    )
+                    if label_ok
+                    is not None
+                    else ""
+                )
+            )
+
+            idx += 1
+
+            if idx >= len(
+                samples
+            ):
+                break
+
+            continue
+
+        cv2.imshow(
+            win,
+            panel,
+        )
+
+        key = (
+            cv2.waitKey(
+                0
+            )
+            & 0xFF
+        )
+
+        if key in (
+            ord("q"),
+            ord("Q"),
+            27,
+        ):
+            break
+
+        if key in (
+            ord("a"),
+            ord("A"),
+            81,
+        ):
+            idx = (
+                idx - 1
+            ) % len(
+                samples
+            )
+            continue
+
+        if key in (
+            ord("d"),
+            ord("D"),
+            83,
+            ord(" "),
+        ):
+            idx = (
+                idx + 1
+            ) % len(
+                samples
+            )
+            continue
+
+        if key in (
+            ord("s"),
+            ord("S"),
+        ):
+            target_dir = (
+                save_dir
+                if save_dir
+                is not None
+                else (
+                    project_root
+                    / "test_outputs"
+                )
+            )
+            target_dir.mkdir(
+                parents=True,
+                exist_ok=True,
+            )
+
+            out = (
+                target_dir
+                / (
+                    f"{sample.group.replace('/', '_')}_"
+                    f"{sample.base}.jpg"
+                )
+            )
+            cv2.imwrite(
+                str(out),
+                panel,
+                [
+                    cv2.IMWRITE_JPEG_QUALITY,
+                    92,
+                ],
+            )
+
+            print(
+                f"[SAVE] {out}"
+            )
+            continue
+
+    if not args.no_view:
+        cv2.destroyAllWindows()
+
+    summary = session.summary()
+
+    print_summary(
+        summary,
+        contract,
+    )
+
+    if save_dir is not None:
+        json_path = (
+            save_dir
+            / "session_summary.json"
+        )
+
+        with json_path.open(
+            "w",
+            encoding="utf-8",
+        ) as f:
+            json.dump(
+                summary,
+                f,
+                ensure_ascii=False,
+                indent=2,
+            )
+
+        csv_path = (
+            save_dir
+            / "results.csv"
+        )
+        save_records_csv(
+            csv_path,
+            session.records,
+        )
+
+        print(
+            f"[SAVE] {json_path}"
+        )
+        print(
+            f"[SAVE] {csv_path}"
+        )
+
+
+if __name__ == "__main__":
+    main()
diff --git a/Python/OAK/datasets/oak-d/_6_export_corridor_onnx.py b/Python/OAK/datasets/oak-d/_6_export_corridor_onnx.py
new file mode 100644
index 000000000..af82ffd03
--- /dev/null
+++ b/Python/OAK/datasets/oak-d/_6_export_corridor_onnx.py
@@ -0,0 +1,3285 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+"""
+export_corridor_agri_onnx_v2.py
+===============================
+
+Exportador ONNX de CAMPO para o modelo frontal OAK-D Lite.
+
+Compatível com checkpoints gerados por:
+    train_corridor_agri_v2.py
+
+Contrato de campo v2
+--------------------
+
+INPUT:
+    name   : rgb_u8_nhwc
+    dtype  : uint8
+    layout : NHWC
+    shape  : [1, H, W, 3]
+    color  : RGB
+    range  : [0, 255]
+
+O GRAFO FAZ:
+    uint8 -> float32
+    NHWC -> NCHW
+    / 255
+    normalize(mean/std do checkpoint)
+    SegFormer
+    CorridorStatusHead 3x4
+    resize bilinear dos logits semânticos para HxW
+    argmax semântico
+    softmax do status
+
+OUTPUTS:
+    seg_ids
+        dtype : int32
+        shape : [1, H, W]
+
+    label_probs
+        dtype : float32
+        shape : [1, K]
+
+Motivação:
+    O Visual Worker não precisa mais fazer em Python:
+      - astype(float32)
+      - /255
+      - transpose HWC->CHW
+      - normalize
+      - resize dos logits
+      - argmax da segmentação
+      - softmax da cabeça global
+
+    Ele só precisa:
+      1) fornecer RGB uint8 na resolução do contrato;
+      2) ler seg_ids;
+      3) fazer argmax de apenas K probabilidades para obter o estado global.
+
+IMPORTANTE:
+    - batch fixo = 1;
+    - resolução fixa;
+    - checkpoint é a fonte da verdade;
+    - norm_stats externo NÃO é necessário no runtime;
+    - este exportador NÃO suporta o LabelHead antigo;
+    - este exportador NÃO exporta semantic_logits por padrão;
+    - AdaptiveAvgPool2d da StatusHead é convertido, SOMENTE no wrapper de
+      export fixo, para AvgPool2d estático matematicamente equivalente;
+    - a verificação PyTorch x ONNX usa a cabeça ORIGINAL do treinamento;
+    - o objetivo é o menor contrato útil e estável para campo.
+
+Exemplos:
+
+    python export_corridor_agri_onnx_v2.py
+
+    python export_corridor_agri_onnx_v2.py ^
+        --checkpoint nav
+
+    python export_corridor_agri_onnx_v2.py ^
+        --ckpt backup/segformer_b0/nav_mit_corridor_agri_v2/best_operational.pt ^
+        --out backup/segformer_b0/nav_mit_corridor_agri_v2/best_operational.field_v2.onnx
+
+Validação:
+    Após exportar, o script tenta:
+      - onnx.checker;
+      - shape inference;
+      - ONNX Runtime;
+      - comparação PyTorch x ONNX em input sintético, com tolerância OOD própria;
+      - comparação em algumas imagens reais de split/val, com tolerância mais rígida.
+
+    Para testar especificamente com TensorRT EP:
+        --verify-provider tensorrt
+
+    Para pular ONNX Runtime:
+        --no-ort-verify
+"""
+
+from __future__ import annotations
+
+import argparse
+import copy
+import hashlib
+import json
+import os
+import shutil
+import time
+from dataclasses import dataclass
+from datetime import datetime
+from pathlib import Path
+from typing import Dict, List, Optional, Sequence, Tuple
+
+import cv2
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from transformers import SegformerForSemanticSegmentation
+
+
+# =============================================================================
+# Constants
+# =============================================================================
+
+EXPORTER_VERSION = "agrobot_corridor_onnx_exporter_v2.1_fixed_pool"
+CONTRACT_SCHEMA = "agrobot.visual.corridor.onnx.v2"
+DEFAULT_OPSET = 17
+
+INPUT_NAME = "rgb_u8_nhwc"
+SEG_OUTPUT_NAME = "seg_ids"
+STATUS_OUTPUT_NAME = "label_probs"
+
+
+# =============================================================================
+# Generic helpers
+# =============================================================================
+
+def now_iso() -> str:
+    return datetime.now().isoformat(
+        timespec="seconds"
+    )
+
+
+def load_json(
+    path: Path,
+) -> dict:
+    with path.open(
+        "r",
+        encoding="utf-8",
+    ) as f:
+        data = json.load(
+            f
+        )
+
+    if not isinstance(
+        data,
+        dict,
+    ):
+        raise ValueError(
+            f"Esperava objeto JSON em {path}"
+        )
+
+    return data
+
+
+def save_json(
+    path: Path,
+    data: dict,
+) -> None:
+    path.parent.mkdir(
+        parents=True,
+        exist_ok=True,
+    )
+
+    tmp = path.with_suffix(
+        path.suffix
+        + ".tmp"
+    )
+
+    with tmp.open(
+        "w",
+        encoding="utf-8",
+    ) as f:
+        json.dump(
+            data,
+            f,
+            ensure_ascii=False,
+            indent=2,
+        )
+
+    os.replace(
+        tmp,
+        path,
+    )
+
+
+def sha256_file(
+    path: Path,
+    chunk_size: int = 4 * 1024 * 1024,
+) -> str:
+    h = hashlib.sha256()
+
+    with path.open(
+        "rb"
+    ) as f:
+        while True:
+            block = f.read(
+                chunk_size
+            )
+
+            if not block:
+                break
+
+            h.update(
+                block
+            )
+
+    return h.hexdigest()
+
+
+def int_key_dict(
+    value,
+) -> Dict[int, str]:
+    if not isinstance(
+        value,
+        dict,
+    ):
+        return {}
+
+    return {
+        int(k): str(v)
+        for k, v in value.items()
+    }
+
+
+def resolve_project_root(
+    config_path: Path,
+) -> Path:
+    root = config_path.resolve().parent
+
+    if not (
+        root
+        / "dataset"
+    ).is_dir():
+        raise FileNotFoundError(
+            f"dataset/ não encontrado ao lado de {config_path}. "
+            "Execute dentro da pasta oak-d ou passe --config correto."
+        )
+
+    return root
+
+
+# =============================================================================
+# Status head, ESPELHO do trainer v2
+# =============================================================================
+
+class CorridorStatusHead(
+    nn.Module
+):
+    """
+    Deve permanecer compatível com train_corridor_agri_v2.py.
+
+    A cabeça:
+      - mantém feature profunda em resolução nativa;
+      - faz AdaptiveAvgPool espacial;
+      - resume separadamente feature e informação semântica;
+      - concatena vetores compactos;
+      - MLP final para estado global.
+    """
+
+    def __init__(
+        self,
+        feat_ch: int,
+        num_seg_classes: int,
+        num_label_classes: int,
+        pool_hw: Tuple[int, int],
+        hidden: int,
+        dropout: float,
+        detach_seg_summary: bool,
+        seg_summary: str,
+    ):
+        super().__init__()
+
+        self.feat_ch = int(
+            feat_ch
+        )
+        self.num_seg_classes = int(
+            num_seg_classes
+        )
+        self.num_label_classes = int(
+            num_label_classes
+        )
+        self.pool_hw = (
+            int(
+                pool_hw[
+                    0
+                ]
+            ),
+            int(
+                pool_hw[
+                    1
+                ]
+            ),
+        )
+        self.detach_seg_summary = bool(
+            detach_seg_summary
+        )
+        self.seg_summary = str(
+            seg_summary
+        ).lower()
+
+        pooled_cells = (
+            self.pool_hw[
+                0
+            ]
+            * self.pool_hw[
+                1
+            ]
+        )
+
+        in_dim = (
+            self.feat_ch
+            + self.num_seg_classes
+        ) * pooled_cells
+
+        self.feat_pool = nn.AdaptiveAvgPool2d(
+            self.pool_hw
+        )
+        self.seg_pool = nn.AdaptiveAvgPool2d(
+            self.pool_hw
+        )
+
+        self.net = nn.Sequential(
+            nn.Linear(
+                in_dim,
+                int(
+                    hidden
+                ),
+            ),
+            nn.ReLU(
+                inplace=True
+            ),
+            nn.Dropout(
+                float(
+                    dropout
+                )
+            ),
+            nn.Linear(
+                int(
+                    hidden
+                ),
+                self.num_label_classes,
+            ),
+        )
+
+    def forward(
+        self,
+        feat: torch.Tensor,
+        seg_logits_native: torch.Tensor,
+    ) -> torch.Tensor:
+        feat_grid = self.feat_pool(
+            feat
+        ).flatten(
+            1
+        )
+
+        seg_src = (
+            seg_logits_native.detach()
+            if self.detach_seg_summary
+            else seg_logits_native
+        )
+
+        if (
+            self.seg_summary
+            == "probabilities"
+        ):
+            seg_src = torch.softmax(
+                seg_src,
+                dim=1,
+            )
+        elif (
+            self.seg_summary
+            != "logits"
+        ):
+            raise RuntimeError(
+                f"seg_summary inválido: "
+                f"{self.seg_summary}"
+            )
+
+        seg_grid = self.seg_pool(
+            seg_src
+        ).flatten(
+            1
+        )
+
+        x = torch.cat(
+            [
+                feat_grid,
+                seg_grid,
+            ],
+            dim=1,
+        )
+
+        return self.net(
+            x
+        )
+
+
+# =============================================================================
+# ONNX export adaptation for fixed-shape status pooling
+# =============================================================================
+
+@dataclass(frozen=True)
+class StatusPoolExportPlan:
+    """
+    Descreve a substituição export-only do AdaptiveAvgPool2d.
+
+    O trainer continua usando AdaptiveAvgPool2d(pool_hw).
+    Para o ONNX de campo, cuja resolução é FIXA, quando a dimensão nativa é
+    exatamente divisível por pool_hw, o AdaptiveAvgPool2d é matematicamente
+    equivalente a AvgPool2d(kernel=stride=input/output).
+
+    Isso evita a limitação do exporter TorchScript clássico, que pode perder
+    H/W estáticos depois de Softmax e falhar ao simbolizar adaptive_avg_pool2d.
+    """
+
+    feat_hw: Tuple[int, int]
+    seg_hw: Tuple[int, int]
+    pool_hw: Tuple[int, int]
+
+    feat_kernel_hw: Tuple[int, int]
+    seg_kernel_hw: Tuple[int, int]
+
+    method: str = "fixed_avgpool_exact"
+
+
+def _exact_pool_kernel(
+    *,
+    input_hw: Tuple[int, int],
+    output_hw: Tuple[int, int],
+    name: str,
+) -> Tuple[int, int]:
+    ih, iw = (
+        int(input_hw[0]),
+        int(input_hw[1]),
+    )
+
+    oh, ow = (
+        int(output_hw[0]),
+        int(output_hw[1]),
+    )
+
+    if (
+        ih <= 0
+        or iw <= 0
+        or oh <= 0
+        or ow <= 0
+    ):
+        raise RuntimeError(
+            f"Dimensões inválidas para {name}: "
+            f"input={input_hw} output={output_hw}"
+        )
+
+    if (
+        ih % oh != 0
+        or iw % ow != 0
+    ):
+        raise RuntimeError(
+            f"{name}: não é possível substituir AdaptiveAvgPool2d"
+            f"{output_hw} por AvgPool2d EXATO porque input={input_hw} "
+            "não é divisível célula-a-célula. "
+            "Não vou exportar uma aproximação silenciosa."
+        )
+
+    return (
+        ih // oh,
+        iw // ow,
+    )
+
+
+def discover_status_native_shapes(
+    *,
+    base_model: nn.Module,
+    H: int,
+    W: int,
+    device: torch.device,
+    mean: Sequence[float],
+    std: Sequence[float],
+) -> Tuple[
+    Tuple[int, int],
+    Tuple[int, int],
+]:
+    """
+    Descobre as formas espaciais reais produzidas PELO checkpoint reconstruído.
+
+    Não hardcodamos B0 nem 1024x576 aqui. Se backbone/resolução mudarem, o
+    exporter recalcula o plano e só aceita a adaptação quando ela for exata.
+    """
+    dummy = torch.zeros(
+        (
+            1,
+            3,
+            int(H),
+            int(W),
+        ),
+        dtype=torch.float32,
+        device=device,
+    )
+
+    mean_t = torch.tensor(
+        list(mean),
+        dtype=torch.float32,
+        device=device,
+    ).view(
+        1,
+        3,
+        1,
+        1,
+    )
+
+    std_t = torch.tensor(
+        list(std),
+        dtype=torch.float32,
+        device=device,
+    ).view(
+        1,
+        3,
+        1,
+        1,
+    )
+
+    with torch.inference_mode():
+        out = base_model(
+            pixel_values=(
+                dummy
+                - mean_t
+            ) / std_t,
+            output_hidden_states=True,
+            return_dict=True,
+        )
+
+    seg_logits_native = out.logits
+    feat = out.hidden_states[
+        -1
+    ]
+
+    feat_hw = (
+        int(feat.shape[-2]),
+        int(feat.shape[-1]),
+    )
+
+    seg_hw = (
+        int(seg_logits_native.shape[-2]),
+        int(seg_logits_native.shape[-1]),
+    )
+
+    return (
+        feat_hw,
+        seg_hw,
+    )
+
+
+def build_export_status_head(
+    *,
+    reference_status_head: CorridorStatusHead,
+    feat_hw: Tuple[int, int],
+    seg_hw: Tuple[int, int],
+    device: torch.device,
+) -> Tuple[
+    CorridorStatusHead,
+    StatusPoolExportPlan,
+]:
+    """
+    Faz cópia SOMENTE da cabeça pequena.
+
+    A cabeça original permanece intacta para a validação PyTorch de referência.
+    A cópia recebe AvgPool2d estático exato, amigável a ONNX/TensorRT.
+    """
+    pool_hw = (
+        int(reference_status_head.pool_hw[0]),
+        int(reference_status_head.pool_hw[1]),
+    )
+
+    feat_kernel = _exact_pool_kernel(
+        input_hw=feat_hw,
+        output_hw=pool_hw,
+        name="feat_pool",
+    )
+
+    seg_kernel = _exact_pool_kernel(
+        input_hw=seg_hw,
+        output_hw=pool_hw,
+        name="seg_pool",
+    )
+
+    export_head = copy.deepcopy(
+        reference_status_head
+    )
+
+    export_head.feat_pool = nn.AvgPool2d(
+        kernel_size=feat_kernel,
+        stride=feat_kernel,
+        padding=0,
+        ceil_mode=False,
+        count_include_pad=True,
+    )
+
+    export_head.seg_pool = nn.AvgPool2d(
+        kernel_size=seg_kernel,
+        stride=seg_kernel,
+        padding=0,
+        ceil_mode=False,
+        count_include_pad=True,
+    )
+
+    export_head.to(
+        device
+    ).eval()
+
+    plan = StatusPoolExportPlan(
+        feat_hw=feat_hw,
+        seg_hw=seg_hw,
+        pool_hw=pool_hw,
+        feat_kernel_hw=feat_kernel,
+        seg_kernel_hw=seg_kernel,
+    )
+
+    return (
+        export_head,
+        plan,
+    )
+
+
+def verify_export_wrapper_equivalence(
+    *,
+    reference_wrapper: nn.Module,
+    export_wrapper: nn.Module,
+    H: int,
+    W: int,
+    device: torch.device,
+    max_prob_abs: float = 1e-6,
+) -> dict:
+    """
+    Confirma ANTES do ONNX que a adaptação export-only preserva a função.
+
+    A segmentação precisa ser bit-a-bit idêntica.
+    As probabilidades de status podem ter erro FP minúsculo por ordem de soma.
+    """
+    gen = torch.Generator(
+        device=device
+    )
+
+    gen.manual_seed(
+        20260915
+    )
+
+    dummy = torch.randint(
+        low=0,
+        high=256,
+        size=(
+            1,
+            int(H),
+            int(W),
+            3,
+        ),
+        dtype=torch.uint8,
+        device=device,
+        generator=gen,
+    )
+
+    reference_wrapper.eval()
+    export_wrapper.eval()
+
+    with torch.inference_mode():
+        ref_seg, ref_probs = reference_wrapper(
+            dummy
+        )
+
+        exp_seg, exp_probs = export_wrapper(
+            dummy
+        )
+
+    seg_equal = bool(
+        torch.equal(
+            ref_seg,
+            exp_seg,
+        )
+    )
+
+    prob_abs = torch.abs(
+        ref_probs.float()
+        - exp_probs.float()
+    )
+
+    prob_max_abs = float(
+        prob_abs.max().item()
+    )
+
+    prob_mean_abs = float(
+        prob_abs.mean().item()
+    )
+
+    label_ref = int(
+        torch.argmax(
+            ref_probs[
+                0
+            ]
+        ).item()
+    )
+
+    label_exp = int(
+        torch.argmax(
+            exp_probs[
+                0
+            ]
+        ).item()
+    )
+
+    passed = (
+        seg_equal
+        and label_ref
+        == label_exp
+        and prob_max_abs
+        <= float(
+            max_prob_abs
+        )
+    )
+
+    result = {
+        "passed": bool(
+            passed
+        ),
+        "seg_equal": bool(
+            seg_equal
+        ),
+        "label_same": bool(
+            label_ref
+            == label_exp
+        ),
+        "label_reference": label_ref,
+        "label_export": label_exp,
+        "status_prob_max_abs": prob_max_abs,
+        "status_prob_mean_abs": prob_mean_abs,
+        "max_prob_abs_allowed": float(
+            max_prob_abs
+        ),
+    }
+
+    print(
+        "[EXPORT-ADAPT] PT original x PT export-friendly | "
+        f"seg_equal={seg_equal} "
+        f"label={label_ref}/{label_exp} "
+        f"pMaxAbs={prob_max_abs:.9g} "
+        f"{'OK' if passed else 'FAIL'}"
+    )
+
+    if not passed:
+        raise RuntimeError(
+            "A adaptação export-only do status head não preservou "
+            f"a saída dentro da tolerância: {result}"
+        )
+
+    return result
+
+
+# =============================================================================
+# Checkpoint contract
+# =============================================================================
+
+@dataclass
+class CheckpointContract:
+    checkpoint_path: Path
+    trainer_version: str
+    epoch: int
+    bests: dict
+
+    backbone: str
+    resolution_wh: Tuple[int, int]
+
+    roi_inicio: float
+    roi_tamanho: float
+
+    seg_id2label: Dict[int, str]
+    nav_class_id: int
+    non_nav_class_id: int
+
+    label_name_by_id: Dict[int, str]
+    group_name_by_id: Dict[int, str]
+
+    norm_channels: List[str]
+    norm_mean: List[float]
+    norm_std: List[float]
+
+    status_head: dict
+    checkpoint_metadata: dict
+
+
+def default_checkpoint_path(
+    project_root: Path,
+    config: dict,
+    which: str,
+) -> Path:
+    model_family = str(
+        config.get(
+            "modelo",
+            "segformer_b0",
+        )
+    )
+
+    model_name = str(
+        config.get(
+            "model_name",
+            "nav_mit",
+        )
+    )
+
+    save_dir = (
+        project_root
+        / "backup"
+        / model_family
+        / f"{model_name}_corridor_agri_v2"
+    )
+
+    filename = {
+        "operational": "best_operational.pt",
+        "nav": "best_nav.pt",
+        "status": "best_status.pt",
+        "last": "last.pt",
+    }[
+        which
+    ]
+
+    return (
+        save_dir
+        / filename
+    )
+
+
+def load_checkpoint_contract(
+    checkpoint_path: Path,
+    config: dict,
+) -> Tuple[dict, CheckpointContract]:
+    if not checkpoint_path.is_file():
+        raise FileNotFoundError(
+            f"Checkpoint não encontrado: {checkpoint_path}"
+        )
+
+    ckpt = torch.load(
+        checkpoint_path,
+        map_location="cpu",
+        weights_only=False,
+    )
+
+    if "model" not in ckpt:
+        raise RuntimeError(
+            "Checkpoint sem chave 'model'."
+        )
+
+    if "status_head" not in ckpt:
+        raise RuntimeError(
+            "Checkpoint sem chave 'status_head'. "
+            "O export v2 não aceita o antigo aux_head/LabelHead."
+        )
+
+    metadata = ckpt.get(
+        "metadata",
+        {},
+    ) or {}
+
+    trainer_version = str(
+        ckpt.get(
+            "trainer_version",
+            metadata.get(
+                "trainer_version",
+                "unknown",
+            ),
+        )
+    )
+
+    resolution = metadata.get(
+        "resolution_wh",
+        config.get(
+            "resolucao",
+            [1024, 576],
+        ),
+    )
+
+    W = int(
+        resolution[
+            0
+        ]
+    )
+    H = int(
+        resolution[
+            1
+        ]
+    )
+
+    backbone = str(
+        metadata.get(
+            "backbone",
+            config.get(
+                "backbone",
+                "nvidia/mit-b0",
+            ),
+        )
+    )
+
+    seg_id2label = int_key_dict(
+        metadata.get(
+            "seg_id2label",
+            {},
+        )
+    )
+
+    if not seg_id2label:
+        raise RuntimeError(
+            "Checkpoint v2 sem metadata.seg_id2label."
+        )
+
+    nav_class_id = int(
+        metadata.get(
+            "nav_class_id"
+        )
+    )
+    non_nav_class_id = int(
+        metadata.get(
+            "non_nav_class_id"
+        )
+    )
+
+    label_name_by_id = int_key_dict(
+        metadata.get(
+            "label_name_by_id",
+            {},
+        )
+    )
+
+    if not label_name_by_id:
+        raise RuntimeError(
+            "Checkpoint v2 sem metadata.label_name_by_id."
+        )
+
+    group_name_by_id = int_key_dict(
+        metadata.get(
+            "group_name_by_id",
+            {},
+        )
+    )
+
+    norm_channels = [
+        str(
+            x
+        ).upper()
+        for x in metadata.get(
+            "norm_channels",
+            [],
+        )
+    ]
+
+    norm_mean = [
+        float(
+            x
+        )
+        for x in metadata.get(
+            "norm_mean",
+            [],
+        )
+    ]
+
+    norm_std = [
+        float(
+            x
+        )
+        for x in metadata.get(
+            "norm_std",
+            [],
+        )
+    ]
+
+    if (
+        norm_channels
+        != [
+            "R",
+            "G",
+            "B",
+        ]
+    ):
+        raise RuntimeError(
+            f"Checkpoint precisa norm_channels RGB; "
+            f"recebido={norm_channels}"
+        )
+
+    if (
+        len(
+            norm_mean
+        )
+        != 3
+        or len(
+            norm_std
+        )
+        != 3
+    ):
+        raise RuntimeError(
+            f"Checkpoint possui norm stats inválidos: "
+            f"mean={norm_mean} std={norm_std}"
+        )
+
+    if (
+        not np.all(
+            np.isfinite(
+                np.asarray(
+                    norm_mean
+                    + norm_std,
+                    dtype=np.float64,
+                )
+            )
+        )
+        or any(
+            s <= 0
+            for s in norm_std
+        )
+    ):
+        raise RuntimeError(
+            "Checkpoint possui mean/std inválidos."
+        )
+
+    status_meta = dict(
+        metadata.get(
+            "status_head",
+            {},
+        )
+        or {}
+    )
+
+    if not status_meta:
+        raise RuntimeError(
+            "Checkpoint v2 sem metadata.status_head."
+        )
+
+    target_contract = (
+        metadata.get(
+            "onnx_target_contract",
+            {},
+        )
+        or {}
+    )
+
+    if target_contract:
+        expected_input = str(
+            target_contract.get(
+                "input",
+                "rgb_uint8_nhwc",
+            )
+        )
+
+        if (
+            expected_input
+            != "rgb_uint8_nhwc"
+        ):
+            raise RuntimeError(
+                f"Checkpoint pede ONNX input "
+                f"{expected_input!r}, mas exporter v2 "
+                "implementa rgb_uint8_nhwc."
+            )
+
+    contract = CheckpointContract(
+        checkpoint_path=checkpoint_path,
+        trainer_version=trainer_version,
+        epoch=int(
+            ckpt.get(
+                "epoch",
+                -1,
+            )
+        ),
+        bests=dict(
+            ckpt.get(
+                "bests",
+                {},
+            )
+            or {}
+        ),
+        backbone=backbone,
+        resolution_wh=(
+            W,
+            H,
+        ),
+        roi_inicio=float(
+            metadata.get(
+                "roi_inicio",
+                config.get(
+                    "roi_inicio",
+                    0.0,
+                ),
+            )
+        ),
+        roi_tamanho=float(
+            metadata.get(
+                "roi_tamanho",
+                config.get(
+                    "roi_tamanho",
+                    1.0,
+                ),
+            )
+        ),
+        seg_id2label=seg_id2label,
+        nav_class_id=nav_class_id,
+        non_nav_class_id=non_nav_class_id,
+        label_name_by_id=label_name_by_id,
+        group_name_by_id=group_name_by_id,
+        norm_channels=norm_channels,
+        norm_mean=norm_mean,
+        norm_std=norm_std,
+        status_head=status_meta,
+        checkpoint_metadata=metadata,
+    )
+
+    return (
+        ckpt,
+        contract,
+    )
+
+
+# =============================================================================
+# Build PyTorch runtime
+# =============================================================================
+
+def infer_hidden_from_status_state(
+    status_state: dict,
+) -> int:
+    key = "net.0.weight"
+
+    if key not in status_state:
+        raise RuntimeError(
+            f"status_head state_dict sem {key}"
+        )
+
+    return int(
+        status_state[
+            key
+        ].shape[
+            0
+        ]
+    )
+
+
+def infer_num_status_classes(
+    status_state: dict,
+) -> int:
+    for key in (
+        "net.3.weight",
+        "net.4.weight",
+    ):
+        if key in status_state:
+            return int(
+                status_state[
+                    key
+                ].shape[
+                    0
+                ]
+            )
+
+    raise RuntimeError(
+        "Não consegui inferir saída do status_head."
+    )
+
+
+def discover_feat_ch(
+    base_model: nn.Module,
+    H: int,
+    W: int,
+    device: torch.device,
+    mean: Sequence[float],
+    std: Sequence[float],
+) -> int:
+    dummy = torch.zeros(
+        (
+            1,
+            3,
+            H,
+            W,
+        ),
+        dtype=torch.float32,
+        device=device,
+    )
+
+    mean_t = torch.tensor(
+        mean,
+        dtype=torch.float32,
+        device=device,
+    ).view(
+        1,
+        3,
+        1,
+        1,
+    )
+
+    std_t = torch.tensor(
+        std,
+        dtype=torch.float32,
+        device=device,
+    ).view(
+        1,
+        3,
+        1,
+        1,
+    )
+
+    with torch.inference_mode():
+        out = base_model(
+            pixel_values=(
+                dummy
+                - mean_t
+            ) / std_t,
+            output_hidden_states=True,
+            return_dict=True,
+        )
+
+    return int(
+        out.hidden_states[
+            -1
+        ].shape[
+            1
+        ]
+    )
+
+
+def build_models(
+    ckpt: dict,
+    contract: CheckpointContract,
+    device: torch.device,
+) -> Tuple[
+    SegformerForSemanticSegmentation,
+    CorridorStatusHead,
+]:
+    num_seg_classes = len(
+        contract.seg_id2label
+    )
+
+    num_status_classes = len(
+        contract.label_name_by_id
+    )
+
+    status_num_from_state = infer_num_status_classes(
+        ckpt[
+            "status_head"
+        ]
+    )
+
+    if (
+        status_num_from_state
+        != num_status_classes
+    ):
+        raise RuntimeError(
+            f"status classes metadata={num_status_classes}, "
+            f"state_dict={status_num_from_state}"
+        )
+
+    base_model = (
+        SegformerForSemanticSegmentation
+        .from_pretrained(
+            contract.backbone,
+            num_labels=num_seg_classes,
+            ignore_mismatched_sizes=True,
+            use_safetensors=True,
+        )
+    )
+
+    base_model.config.output_hidden_states = True
+
+    base_model.load_state_dict(
+        ckpt[
+            "model"
+        ],
+        strict=True,
+    )
+
+    base_model.to(
+        device
+    ).eval()
+
+    status_meta = contract.status_head
+
+    feat_ch = int(
+        status_meta.get(
+            "feat_ch",
+            0,
+        )
+    )
+
+    W, H = contract.resolution_wh
+
+    if feat_ch <= 0:
+        feat_ch = discover_feat_ch(
+            base_model,
+            H,
+            W,
+            device,
+            contract.norm_mean,
+            contract.norm_std,
+        )
+
+    pool_hw_raw = status_meta.get(
+        "pool_hw"
+    )
+
+    if pool_hw_raw is None:
+        # Trainer salva pool_h/pool_w no config e pool_hw no export_config.
+        pool_h = int(
+            status_meta.get(
+                "pool_h",
+                3,
+            )
+        )
+        pool_w = int(
+            status_meta.get(
+                "pool_w",
+                4,
+            )
+        )
+        pool_hw = (
+            pool_h,
+            pool_w,
+        )
+    else:
+        pool_hw = (
+            int(
+                pool_hw_raw[
+                    0
+                ]
+            ),
+            int(
+                pool_hw_raw[
+                    1
+                ]
+            ),
+        )
+
+    hidden = infer_hidden_from_status_state(
+        ckpt[
+            "status_head"
+        ]
+    )
+
+    status_head = CorridorStatusHead(
+        feat_ch=feat_ch,
+        num_seg_classes=num_seg_classes,
+        num_label_classes=num_status_classes,
+        pool_hw=pool_hw,
+        hidden=hidden,
+        dropout=0.0,
+        detach_seg_summary=bool(
+            status_meta.get(
+                "detach_seg_summary",
+                True,
+            )
+        ),
+        seg_summary=str(
+            status_meta.get(
+                "seg_summary",
+                "probabilities",
+            )
+        ),
+    )
+
+    status_head.load_state_dict(
+        ckpt[
+            "status_head"
+        ],
+        strict=True,
+    )
+
+    status_head.to(
+        device
+    ).eval()
+
+    return (
+        base_model,
+        status_head,
+    )
+
+
+# =============================================================================
+# Field ONNX wrapper
+# =============================================================================
+
+class CorridorFieldOnnx(
+    nn.Module
+):
+    """
+    CONTRATO FIXO DE CAMPO.
+
+    Entrada:
+        RGB uint8 NHWC na resolução do modelo.
+
+    Saída:
+        seg_ids int32 HxW
+        label_probs float32 K
+    """
+
+    def __init__(
+        self,
+        base_model: nn.Module,
+        status_head: CorridorStatusHead,
+        resolution_wh: Tuple[int, int],
+        norm_mean: Sequence[float],
+        norm_std: Sequence[float],
+    ):
+        super().__init__()
+
+        self.base_model = base_model
+        self.status_head = status_head
+
+        self.W = int(
+            resolution_wh[
+                0
+            ]
+        )
+        self.H = int(
+            resolution_wh[
+                1
+            ]
+        )
+
+        mean = torch.tensor(
+            list(
+                norm_mean
+            ),
+            dtype=torch.float32,
+        ).view(
+            1,
+            3,
+            1,
+            1,
+        )
+
+        std = torch.tensor(
+            list(
+                norm_std
+            ),
+            dtype=torch.float32,
+        ).view(
+            1,
+            3,
+            1,
+            1,
+        )
+
+        self.register_buffer(
+            "norm_mean",
+            mean,
+        )
+        self.register_buffer(
+            "norm_std",
+            std,
+        )
+
+    def forward(
+        self,
+        rgb_u8_nhwc: torch.Tensor,
+    ):
+        # uint8 network I/O -> float.
+        x = rgb_u8_nhwc.to(
+            dtype=torch.float32
+        )
+
+        # [0,255] -> [0,1]
+        x = x / 255.0
+
+        # NHWC -> NCHW
+        x = x.permute(
+            0,
+            3,
+            1,
+            2,
+        )
+
+        # Normalização EXATA do treinamento.
+        x = (
+            x
+            - self.norm_mean
+        ) / self.norm_std
+
+        out = self.base_model(
+            pixel_values=x,
+            output_hidden_states=True,
+            return_dict=True,
+        )
+
+        seg_logits_native = out.logits
+        feat = out.hidden_states[
+            -1
+        ]
+
+        status_logits = self.status_head(
+            feat,
+            seg_logits_native,
+        )
+
+        # Treinamento/validação fazem resize dos LOGITS e só depois argmax.
+        seg_logits_full = F.interpolate(
+            seg_logits_native,
+            size=(
+                self.H,
+                self.W,
+            ),
+            mode="bilinear",
+            align_corners=False,
+        )
+
+        # ArgMax ONNX nasce como int64. Cast explícito para int32 deixa
+        # o output menor e adequado ao runtime TensorRT.
+        seg_ids = torch.argmax(
+            seg_logits_full,
+            dim=1,
+        ).to(
+            torch.int32
+        )
+
+        label_probs = torch.softmax(
+            status_logits,
+            dim=1,
+        )
+
+        return (
+            seg_ids,
+            label_probs,
+        )
+
+
+# =============================================================================
+# Input preparation for verification ONLY
+# =============================================================================
+
+def image_to_contract_input(
+    image_path: Path,
+    contract: CheckpointContract,
+) -> np.ndarray:
+    """
+    Converte um arquivo qualquer para o tensor EXTERNO esperado pelo ONNX.
+
+    Isto é só para verificação do export.
+
+    O ONNX NÃO contém crop/resize geométrico de câmera:
+      - aplica ROI da source;
+      - resize para W,H;
+      - converte BGR OpenCV -> RGB;
+      - mantém uint8 NHWC.
+    """
+    bgr = cv2.imread(
+        str(
+            image_path
+        ),
+        cv2.IMREAD_COLOR,
+    )
+
+    if bgr is None:
+        raise RuntimeError(
+            f"Falha ao ler imagem: {image_path}"
+        )
+
+    h, w = bgr.shape[
+        :2
+    ]
+
+    y0 = int(
+        round(
+            h
+            * contract.roi_inicio
+        )
+    )
+
+    y1 = int(
+        round(
+            h
+            * min(
+                1.0,
+                contract.roi_inicio
+                + contract.roi_tamanho,
+            )
+        )
+    )
+
+    y0 = max(
+        0,
+        min(
+            h - 1,
+            y0,
+        ),
+    )
+
+    y1 = max(
+        y0 + 1,
+        min(
+            h,
+            y1,
+        ),
+    )
+
+    roi = bgr[
+        y0:y1,
+        :,
+    ]
+
+    W, H = contract.resolution_wh
+
+    resized = cv2.resize(
+        roi,
+        (
+            W,
+            H,
+        ),
+        interpolation=cv2.INTER_AREA,
+    )
+
+    rgb = cv2.cvtColor(
+        resized,
+        cv2.COLOR_BGR2RGB,
+    )
+
+    return np.ascontiguousarray(
+        rgb[
+            None,
+            ...,
+        ],
+        dtype=np.uint8,
+    )
+
+
+def find_verification_images(
+    project_root: Path,
+    contract: CheckpointContract,
+    explicit_dir: Optional[Path],
+    count: int,
+) -> List[Path]:
+    if count <= 0:
+        return []
+
+    if explicit_dir is not None:
+        roots = [
+            explicit_dir
+        ]
+    else:
+        W, H = contract.resolution_wh
+
+        roots = [
+            (
+                project_root
+                / "dataset"
+                / f"{W}x{H}"
+                / "split"
+                / "val"
+                / "images"
+            ),
+            (
+                project_root
+                / "dataset"
+                / f"{W}x{H}"
+                / "group"
+            ),
+        ]
+
+    candidates: List[
+        Path
+    ] = []
+
+    exts = {
+        ".png",
+        ".jpg",
+        ".jpeg",
+        ".bmp",
+        ".webp",
+    }
+
+    for root in roots:
+        if not root.is_dir():
+            continue
+
+        if (
+            root.name
+            == "images"
+        ):
+            found = [
+                p
+                for p in root.iterdir()
+                if (
+                    p.is_file()
+                    and p.suffix.lower()
+                    in exts
+                )
+            ]
+        else:
+            found = [
+                p
+                for p in root.rglob(
+                    "*"
+                )
+                if (
+                    p.is_file()
+                    and p.parent.name
+                    == "images"
+                    and p.suffix.lower()
+                    in exts
+                )
+            ]
+
+        candidates.extend(
+            found
+        )
+
+        if candidates:
+            break
+
+    candidates = sorted(
+        set(
+            candidates
+        ),
+        key=lambda p: str(
+            p
+        ).lower(),
+    )
+
+    if not candidates:
+        return []
+
+    if len(
+        candidates
+    ) <= count:
+        return candidates
+
+    idxs = np.linspace(
+        0,
+        len(
+            candidates
+        ) - 1,
+        count,
+    ).astype(
+        int
+    )
+
+    return [
+        candidates[
+            int(
+                i
+            )
+        ]
+        for i in idxs
+    ]
+
+
+# =============================================================================
+# ONNX metadata
+# =============================================================================
+
+def contract_core_dict(
+    *,
+    checkpoint_contract: CheckpointContract,
+    checkpoint_sha256: str,
+    opset: int,
+    export_pool_plan: Optional[
+        StatusPoolExportPlan
+    ] = None,
+) -> dict:
+    W, H = checkpoint_contract.resolution_wh
+
+    return {
+        "schema": CONTRACT_SCHEMA,
+        "contract_version": 2,
+        "exporter_version": EXPORTER_VERSION,
+        "created_at": now_iso(),
+
+        "model": {
+            "kind": "agrobot_visual_corridor",
+            "task": "semantic_navigation_plus_corridor_status",
+            "backbone": checkpoint_contract.backbone,
+            "trainer_version": checkpoint_contract.trainer_version,
+            "checkpoint_epoch": checkpoint_contract.epoch,
+            "checkpoint_bests": checkpoint_contract.bests,
+            "checkpoint_sha256": checkpoint_sha256,
+        },
+
+        "input": {
+            "name": INPUT_NAME,
+            "dtype": "uint8",
+            "shape": [
+                1,
+                H,
+                W,
+                3,
+            ],
+            "layout": "NHWC",
+            "color_order": "RGB",
+            "value_range": [
+                0,
+                255,
+            ],
+            "batch": 1,
+            "geometry": {
+                "input_is_already_model_resolution": True,
+                "source_roi_inicio": checkpoint_contract.roi_inicio,
+                "source_roi_tamanho": checkpoint_contract.roi_tamanho,
+                "source_to_model_resize": "INTER_AREA_recommended",
+                "model_resolution_wh": [
+                    W,
+                    H,
+                ],
+            },
+        },
+
+        "preprocess_in_graph": {
+            "enabled": True,
+            "cast": "uint8_to_float32",
+            "scale": "x/255.0",
+            "layout": "NHWC_to_NCHW",
+            "normalize": "(x-mean)/std",
+            "channels": checkpoint_contract.norm_channels,
+            "mean": checkpoint_contract.norm_mean,
+            "std": checkpoint_contract.norm_std,
+        },
+
+        "outputs": {
+            SEG_OUTPUT_NAME: {
+                "name": SEG_OUTPUT_NAME,
+                "dtype": "int32",
+                "shape": [
+                    1,
+                    H,
+                    W,
+                ],
+                "meaning": "semantic_class_id",
+                "postprocess_in_graph": (
+                    "native_logits -> bilinear_resize_to_input -> argmax -> int32"
+                ),
+                "id2label": {
+                    str(
+                        k
+                    ): v
+                    for k, v in checkpoint_contract.seg_id2label.items()
+                },
+                "nav_class_id": checkpoint_contract.nav_class_id,
+                "non_nav_class_id": checkpoint_contract.non_nav_class_id,
+            },
+
+            STATUS_OUTPUT_NAME: {
+                "name": STATUS_OUTPUT_NAME,
+                "dtype": "float32",
+                "shape": [
+                    1,
+                    len(
+                        checkpoint_contract.label_name_by_id
+                    ),
+                ],
+                "meaning": "softmax_probability_per_corridor_state",
+                "decision_rule": "argmax(label_probs[0])",
+                "id2label": {
+                    str(
+                        k
+                    ): v
+                    for k, v in checkpoint_contract.label_name_by_id.items()
+                },
+            },
+        },
+
+        "status_head": checkpoint_contract.status_head,
+
+        "export_adaptation": (
+            {
+                "status_pool": {
+                    "method": export_pool_plan.method,
+                    "training_operator": "AdaptiveAvgPool2d",
+                    "onnx_operator": "AveragePool",
+                    "mathematically_exact_for_fixed_shape": True,
+                    "feat_hw": list(
+                        export_pool_plan.feat_hw
+                    ),
+                    "seg_hw": list(
+                        export_pool_plan.seg_hw
+                    ),
+                    "pool_hw": list(
+                        export_pool_plan.pool_hw
+                    ),
+                    "feat_kernel_hw": list(
+                        export_pool_plan.feat_kernel_hw
+                    ),
+                    "seg_kernel_hw": list(
+                        export_pool_plan.seg_kernel_hw
+                    ),
+                }
+            }
+            if export_pool_plan
+            is not None
+            else {}
+        ),
+
+        "runtime_contract": {
+            "fixed_batch": True,
+            "fixed_spatial_shape": True,
+            "dynamic_axes": False,
+            "external_normalization_required": False,
+            "external_semantic_argmax_required": False,
+            "external_semantic_resize_required": False,
+            "external_status_softmax_required": False,
+            "only_external_status_argmax_required": True,
+        },
+
+        "onnx": {
+            "opset": int(
+                opset
+            ),
+            "input_names": [
+                INPUT_NAME
+            ],
+            "output_names": [
+                SEG_OUTPUT_NAME,
+                STATUS_OUTPUT_NAME,
+            ],
+        },
+    }
+
+
+def embed_onnx_metadata(
+    onnx_path: Path,
+    core_contract: dict,
+) -> bool:
+    try:
+        import onnx
+    except ImportError:
+        return False
+
+    model = onnx.load(
+        str(
+            onnx_path
+        )
+    )
+
+    del model.metadata_props[
+        :
+    ]
+
+    metadata = {
+        "agrobot.schema": CONTRACT_SCHEMA,
+        "agrobot.contract_version": "2",
+        "agrobot.exporter_version": EXPORTER_VERSION,
+        "agrobot.input_name": INPUT_NAME,
+        "agrobot.input_dtype": "uint8",
+        "agrobot.input_layout": "NHWC",
+        "agrobot.input_color_order": "RGB",
+        "agrobot.seg_output_name": SEG_OUTPUT_NAME,
+        "agrobot.status_output_name": STATUS_OUTPUT_NAME,
+        "agrobot.contract_json": json.dumps(
+            core_contract,
+            ensure_ascii=False,
+            separators=(
+                ",",
+                ":",
+            ),
+        ),
+    }
+
+    for key, value in metadata.items():
+        prop = model.metadata_props.add()
+        prop.key = str(
+            key
+        )
+        prop.value = str(
+            value
+        )
+
+    onnx.save(
+        model,
+        str(
+            onnx_path
+        ),
+    )
+
+    return True
+
+
+# =============================================================================
+# Export
+# =============================================================================
+
+def export_onnx(
+    wrapper: CorridorFieldOnnx,
+    out_path: Path,
+    H: int,
+    W: int,
+    device: torch.device,
+    opset: int,
+) -> None:
+    wrapper.to(
+        device
+    ).eval()
+
+    # Input real do campo é uint8.
+    dummy = torch.randint(
+        low=0,
+        high=256,
+        size=(
+            1,
+            H,
+            W,
+            3,
+        ),
+        dtype=torch.uint8,
+        device=device,
+    )
+
+    with torch.inference_mode():
+        seg_ids, label_probs = wrapper(
+            dummy
+        )
+
+    print(
+        f"[CHECK] {INPUT_NAME}: "
+        f"shape={tuple(dummy.shape)} dtype={dummy.dtype}"
+    )
+    print(
+        f"[CHECK] {SEG_OUTPUT_NAME}: "
+        f"shape={tuple(seg_ids.shape)} dtype={seg_ids.dtype}"
+    )
+    print(
+        f"[CHECK] {STATUS_OUTPUT_NAME}: "
+        f"shape={tuple(label_probs.shape)} dtype={label_probs.dtype}"
+    )
+
+    out_path.parent.mkdir(
+        parents=True,
+        exist_ok=True,
+    )
+
+    kwargs = dict(
+        model=wrapper,
+        args=(
+            dummy,
+        ),
+        f=str(
+            out_path
+        ),
+        export_params=True,
+        opset_version=int(
+            opset
+        ),
+        do_constant_folding=True,
+        input_names=[
+            INPUT_NAME
+        ],
+        output_names=[
+            SEG_OUTPUT_NAME,
+            STATUS_OUTPUT_NAME,
+        ],
+        dynamic_axes=None,
+    )
+
+    print(
+        "[EXPORT] torch.onnx.export "
+        f"opset={opset} fixed_batch=1 fixed_shape={H}x{W}"
+    )
+
+    # Para TensorRT 10.x, o exporter clássico continua uma ponte previsível.
+    # Em PyTorch novo, dynamo pode ser default; forçamos legacy quando aceito.
+    try:
+        torch.onnx.export(
+            **kwargs,
+            dynamo=False,
+        )
+    except TypeError:
+        torch.onnx.export(
+            **kwargs
+        )
+
+
+# =============================================================================
+# ONNX checks
+# =============================================================================
+
+def check_onnx(
+    path: Path,
+) -> dict:
+    result = {
+        "onnx_installed": False,
+        "checker_ok": None,
+        "shape_inference_ok": None,
+        "inputs": [],
+        "outputs": [],
+        "error": None,
+    }
+
+    try:
+        import onnx
+    except ImportError:
+        print(
+            "[WARN] pacote onnx não instalado; "
+            "checker/shape inference pulados."
+        )
+        return result
+
+    result[
+        "onnx_installed"
+    ] = True
+
+    try:
+        model = onnx.load(
+            str(
+                path
+            )
+        )
+
+        onnx.checker.check_model(
+            model
+        )
+
+        result[
+            "checker_ok"
+        ] = True
+
+        try:
+            inferred = (
+                onnx.shape_inference
+                .infer_shapes(
+                    model
+                )
+            )
+
+            result[
+                "shape_inference_ok"
+            ] = True
+
+            graph = inferred.graph
+
+        except Exception as exc:
+            result[
+                "shape_inference_ok"
+            ] = False
+            graph = model.graph
+
+            print(
+                f"[WARN] ONNX shape inference: {exc}"
+            )
+
+        result[
+            "inputs"
+        ] = [
+            x.name
+            for x in graph.input
+        ]
+
+        result[
+            "outputs"
+        ] = [
+            x.name
+            for x in graph.output
+        ]
+
+        print(
+            "[OK] onnx.checker passou."
+        )
+        print(
+            f"[ONNX] inputs={result['inputs']}"
+        )
+        print(
+            f"[ONNX] outputs={result['outputs']}"
+        )
+
+    except Exception as exc:
+        result[
+            "checker_ok"
+        ] = False
+        result[
+            "error"
+        ] = (
+            f"{type(exc).__name__}: {exc}"
+        )
+
+        raise
+
+    return result
+
+
+def create_ort_session(
+    onnx_path: Path,
+    provider_mode: str,
+):
+    try:
+        import onnxruntime as ort
+    except ImportError:
+        return None, {
+            "installed": False,
+            "available": [],
+            "requested": provider_mode,
+            "active": [],
+        }
+
+    available = ort.get_available_providers()
+
+    provider_mode = str(
+        provider_mode
+    ).lower()
+
+    if provider_mode == "auto":
+        if (
+            "CUDAExecutionProvider"
+            in available
+        ):
+            providers = [
+                "CUDAExecutionProvider",
+                "CPUExecutionProvider",
+            ]
+        else:
+            providers = [
+                "CPUExecutionProvider"
+            ]
+
+    elif provider_mode == "cuda":
+        providers = [
+            "CUDAExecutionProvider",
+            "CPUExecutionProvider",
+        ]
+
+    elif provider_mode == "cpu":
+        providers = [
+            "CPUExecutionProvider"
+        ]
+
+    elif provider_mode == "tensorrt":
+        cache_dir = (
+            onnx_path.parent
+            / "trt_cache_field_v2"
+        )
+
+        cache_dir.mkdir(
+            parents=True,
+            exist_ok=True,
+        )
+
+        trt_options = {
+            "trt_engine_cache_enable": True,
+            "trt_engine_cache_path": str(
+                cache_dir
+            ),
+            "trt_timing_cache_enable": True,
+            "trt_timing_cache_path": str(
+                cache_dir
+            ),
+            "trt_fp16_enable": True,
+        }
+
+        providers = [
+            (
+                "TensorrtExecutionProvider",
+                trt_options,
+            ),
+            "CUDAExecutionProvider",
+            "CPUExecutionProvider",
+        ]
+
+    else:
+        raise ValueError(
+            provider_mode
+        )
+
+    providers_ok = []
+
+    for p in providers:
+        name = (
+            p[
+                0
+            ]
+            if isinstance(
+                p,
+                tuple,
+            )
+            else p
+        )
+
+        if name in available:
+            providers_ok.append(
+                p
+            )
+
+    if not providers_ok:
+        raise RuntimeError(
+            f"Nenhum provider ORT utilizável. "
+            f"requested={provider_mode} available={available}"
+        )
+
+    options = ort.SessionOptions()
+    options.graph_optimization_level = (
+        ort.GraphOptimizationLevel
+        .ORT_ENABLE_ALL
+    )
+
+    session = ort.InferenceSession(
+        str(
+            onnx_path
+        ),
+        sess_options=options,
+        providers=providers_ok,
+    )
+
+    active = session.get_providers()
+
+    if (
+        provider_mode
+        == "cuda"
+        and "CUDAExecutionProvider"
+        not in active
+    ):
+        raise RuntimeError(
+            f"CUDA solicitado, mas providers ativos={active}"
+        )
+
+    if (
+        provider_mode
+        == "tensorrt"
+        and "TensorrtExecutionProvider"
+        not in active
+    ):
+        raise RuntimeError(
+            f"TensorRT solicitado, mas providers ativos={active}"
+        )
+
+    info = {
+        "installed": True,
+        "available": list(
+            available
+        ),
+        "requested": provider_mode,
+        "active": list(
+            active
+        ),
+    }
+
+    return (
+        session,
+        info,
+    )
+
+
+def run_pytorch_wrapper(
+    wrapper: CorridorFieldOnnx,
+    input_np: np.ndarray,
+    device: torch.device,
+):
+    x = torch.from_numpy(
+        np.ascontiguousarray(
+            input_np
+        )
+    ).to(
+        device
+    )
+
+    with torch.inference_mode():
+        seg, probs = wrapper(
+            x
+        )
+
+    return (
+        seg.detach()
+        .cpu()
+        .numpy(),
+        probs.detach()
+        .cpu()
+        .numpy(),
+    )
+
+
+def compare_outputs(
+    pt_seg: np.ndarray,
+    pt_probs: np.ndarray,
+    ort_seg: np.ndarray,
+    ort_probs: np.ndarray,
+) -> dict:
+    if (
+        pt_seg.shape
+        != ort_seg.shape
+    ):
+        raise RuntimeError(
+            f"shape seg divergiu: PT={pt_seg.shape} ORT={ort_seg.shape}"
+        )
+
+    if (
+        pt_probs.shape
+        != ort_probs.shape
+    ):
+        raise RuntimeError(
+            f"shape probs divergiu: PT={pt_probs.shape} ORT={ort_probs.shape}"
+        )
+
+    seg_agreement = float(
+        np.mean(
+            pt_seg
+            == ort_seg
+        )
+    )
+
+    probs_abs = np.abs(
+        pt_probs.astype(
+            np.float64
+        )
+        - ort_probs.astype(
+            np.float64
+        )
+    )
+
+    max_abs = float(
+        probs_abs.max()
+    )
+
+    mean_abs = float(
+        probs_abs.mean()
+    )
+
+    pt_label = int(
+        np.argmax(
+            pt_probs[
+                0
+            ]
+        )
+    )
+    ort_label = int(
+        np.argmax(
+            ort_probs[
+                0
+            ]
+        )
+    )
+
+    return {
+        "seg_agreement": seg_agreement,
+        "label_same": (
+            pt_label
+            == ort_label
+        ),
+        "pt_label": pt_label,
+        "ort_label": ort_label,
+        "label_probs_max_abs": max_abs,
+        "label_probs_mean_abs": mean_abs,
+    }
+
+
+def verify_with_ort(
+    *,
+    onnx_path: Path,
+    wrapper: CorridorFieldOnnx,
+    contract: CheckpointContract,
+    project_root: Path,
+    device: torch.device,
+    provider_mode: str,
+    verify_samples: int,
+    verify_dir: Optional[Path],
+    min_seg_agreement: float,
+    max_prob_abs: float,
+    max_prob_abs_synthetic: float,
+) -> dict:
+    session, provider_info = create_ort_session(
+        onnx_path,
+        provider_mode,
+    )
+
+    if session is None:
+        print(
+            "[WARN] onnxruntime não instalado; "
+            "verificação PT x ORT pulada."
+        )
+
+        return {
+            "provider": provider_info,
+            "executed": False,
+            "cases": [],
+            "passed": None,
+        }
+
+    print(
+        f"[ORT] providers ativos: "
+        f"{provider_info['active']}"
+    )
+
+    ort_input = session.get_inputs()[
+        0
+    ]
+
+    ort_outputs = [
+        x.name
+        for x in session.get_outputs()
+    ]
+
+    if (
+        ort_input.name
+        != INPUT_NAME
+    ):
+        raise RuntimeError(
+            f"ORT input inesperado: {ort_input.name}"
+        )
+
+    if ort_outputs != [
+        SEG_OUTPUT_NAME,
+        STATUS_OUTPUT_NAME,
+    ]:
+        raise RuntimeError(
+            f"ORT outputs inesperados: {ort_outputs}"
+        )
+
+    W, H = contract.resolution_wh
+
+    cases = []
+
+    # Caso sintético.
+    rng = np.random.default_rng(
+        20260914
+    )
+
+    synthetic = rng.integers(
+        0,
+        256,
+        size=(
+            1,
+            H,
+            W,
+            3,
+        ),
+        dtype=np.uint8,
+    )
+
+    source_cases = [
+        (
+            "synthetic_random",
+            synthetic,
+        )
+    ]
+
+    images = find_verification_images(
+        project_root=project_root,
+        contract=contract,
+        explicit_dir=verify_dir,
+        count=verify_samples,
+    )
+
+    for image_path in images:
+        source_cases.append(
+            (
+                str(
+                    image_path
+                ),
+                image_to_contract_input(
+                    image_path,
+                    contract,
+                ),
+            )
+        )
+
+    # Warmup ORT/CUDA fora das métricas.
+    #
+    # A primeira session.run() pode pagar criação de kernels, alocações e
+    # otimizações do CUDA EP. Isso não representa steady-state do modelo.
+    # Usamos o caso sintético apenas para aquecer e descartamos esse tempo.
+    if source_cases:
+        warmup_input = source_cases[0][1]
+
+        for _ in range(2):
+            session.run(
+                [
+                    SEG_OUTPUT_NAME,
+                    STATUS_OUTPUT_NAME,
+                ],
+                {
+                    INPUT_NAME: warmup_input
+                },
+            )
+
+        print(
+            "[ORT] warmup concluído (2 inferências não contabilizadas)."
+        )
+
+    all_passed = True
+
+    for name, input_np in source_cases:
+        pt_seg, pt_probs = run_pytorch_wrapper(
+            wrapper,
+            input_np,
+            device,
+        )
+
+        t0 = time.perf_counter()
+
+        ort_seg, ort_probs = session.run(
+            [
+                SEG_OUTPUT_NAME,
+                STATUS_OUTPUT_NAME,
+            ],
+            {
+                INPUT_NAME: input_np
+            },
+        )
+
+        ort_ms = (
+            time.perf_counter()
+            - t0
+        ) * 1000.0
+
+        cmp = compare_outputs(
+            pt_seg,
+            pt_probs,
+            ort_seg,
+            ort_probs,
+        )
+
+        is_synthetic = (
+            name
+            == "synthetic_random"
+        )
+
+        prob_threshold = float(
+            max_prob_abs_synthetic
+            if is_synthetic
+            else max_prob_abs
+        )
+
+        passed = (
+            cmp[
+                "seg_agreement"
+            ]
+            >= float(
+                min_seg_agreement
+            )
+            and cmp[
+                "label_same"
+            ]
+            and cmp[
+                "label_probs_max_abs"
+            ]
+            <= prob_threshold
+        )
+
+        all_passed = (
+            all_passed
+            and passed
+        )
+
+        row = {
+            "source": name,
+            "ort_ms": float(
+                ort_ms
+            ),
+            **cmp,
+            "case_kind": (
+                "synthetic"
+                if is_synthetic
+                else "real"
+            ),
+            "prob_abs_threshold": float(
+                prob_threshold
+            ),
+            "passed": bool(
+                passed
+            ),
+        }
+
+        cases.append(
+            row
+        )
+
+        print(
+            f"[VERIFY] {name} | "
+            f"seg={cmp['seg_agreement'] * 100.0:.5f}% "
+            f"label={cmp['pt_label']}/{cmp['ort_label']} "
+            f"pMaxAbs={cmp['label_probs_max_abs']:.7g} "
+            f"(lim={prob_threshold:.7g}) "
+            f"ORT={ort_ms:.2f}ms | "
+            f"{'OK' if passed else 'FAIL'}"
+        )
+
+    if not all_passed:
+        raise RuntimeError(
+            "Verificação PyTorch x ONNX falhou. "
+            "Não considere este ONNX pronto para campo."
+        )
+
+    return {
+        "provider": provider_info,
+        "executed": True,
+        "cases": cases,
+        "passed": True,
+        "thresholds": {
+            "min_seg_agreement": float(
+                min_seg_agreement
+            ),
+            "max_label_probs_abs_real": float(
+                max_prob_abs
+            ),
+            "max_label_probs_abs_synthetic": float(
+                max_prob_abs_synthetic
+            ),
+            "require_same_status_argmax": True,
+        },
+    }
+
+
+# =============================================================================
+# Optional trtexec parse/build validation
+# =============================================================================
+
+def find_trtexec(
+    explicit: Optional[
+        str
+    ],
+) -> Optional[
+    Path
+]:
+    if explicit:
+        p = Path(
+            explicit
+        )
+
+        if p.is_file():
+            return p.resolve()
+
+    found = shutil.which(
+        "trtexec"
+    )
+
+    if found:
+        return Path(
+            found
+        ).resolve()
+
+    return None
+
+
+# =============================================================================
+# Main
+# =============================================================================
+
+def main():
+    ap = argparse.ArgumentParser(
+        description=(
+            "Export ONNX campo v2: RGB uint8 NHWC -> seg_ids + label_probs."
+        )
+    )
+
+    ap.add_argument(
+        "--config",
+        default="config.json",
+    )
+
+    ap.add_argument(
+        "--ckpt",
+        default=None,
+        help="Checkpoint .pt explícito.",
+    )
+
+    ap.add_argument(
+        "--checkpoint",
+        default="operational",
+        choices=[
+            "operational",
+            "nav",
+            "status",
+            "last",
+        ],
+        help="Checkpoint automático quando --ckpt não é informado.",
+    )
+
+    ap.add_argument(
+        "--out",
+        default=None,
+        help="Destino ONNX. Default fica ao lado do checkpoint.",
+    )
+
+    ap.add_argument(
+        "--device",
+        default="cuda",
+        choices=[
+            "cuda",
+            "cpu",
+        ],
+    )
+
+    ap.add_argument(
+        "--opset",
+        type=int,
+        default=DEFAULT_OPSET,
+    )
+
+    ap.add_argument(
+        "--verify-provider",
+        default="auto",
+        choices=[
+            "auto",
+            "cuda",
+            "cpu",
+            "tensorrt",
+        ],
+        help=(
+            "Provider ORT usado no sanity check. "
+            "auto prefere CUDA, não TensorRT."
+        ),
+    )
+
+    ap.add_argument(
+        "--verify-samples",
+        type=int,
+        default=3,
+        help="Quantidade de imagens reais do val além do caso sintético.",
+    )
+
+    ap.add_argument(
+        "--verify-dir",
+        default=None,
+        help="Pasta de imagens para verificação opcional.",
+    )
+
+    ap.add_argument(
+        "--min-seg-agreement",
+        type=float,
+        default=0.999,
+        help="Acordo mínimo PT x ONNX na máscara final.",
+    )
+
+    ap.add_argument(
+        "--max-prob-abs",
+        type=float,
+        default=2e-4,
+        help=(
+            "Erro absoluto máximo aceitável nas probabilidades de status "
+            "para imagens reais."
+        ),
+    )
+
+    ap.add_argument(
+        "--max-prob-abs-synthetic",
+        type=float,
+        default=1e-3,
+        help=(
+            "Erro absoluto máximo no caso synthetic_random. "
+            "Ruído uniforme 0..255 é OOD e usa tolerância numérica separada."
+        ),
+    )
+
+    ap.add_argument(
+        "--no-ort-verify",
+        action="store_true",
+    )
+
+    ap.add_argument(
+        "--trtexec",
+        default=None,
+        help="Caminho opcional do trtexec; apenas registrado no contrato.",
+    )
+
+    args = ap.parse_args()
+
+    config_path = Path(
+        args.config
+    ).resolve()
+
+    if not config_path.is_file():
+        raise FileNotFoundError(
+            f"Config não encontrado: {config_path}"
+        )
+
+    config = load_json(
+        config_path
+    )
+
+    project_root = resolve_project_root(
+        config_path
+    )
+
+    if args.ckpt:
+        checkpoint_path = Path(
+            args.ckpt
+        ).resolve()
+    else:
+        checkpoint_path = default_checkpoint_path(
+            project_root,
+            config,
+            args.checkpoint,
+        )
+
+    ckpt, contract = load_checkpoint_contract(
+        checkpoint_path,
+        config,
+    )
+
+    use_cuda = (
+        args.device
+        == "cuda"
+        and torch.cuda.is_available()
+    )
+
+    if (
+        args.device
+        == "cuda"
+        and not use_cuda
+    ):
+        print(
+            "[WARN] CUDA solicitado, mas indisponível. "
+            "Exportando em CPU."
+        )
+
+    device = torch.device(
+        "cuda"
+        if use_cuda
+        else "cpu"
+    )
+
+    if args.out:
+        out_path = Path(
+            args.out
+        ).resolve()
+    else:
+        stem = checkpoint_path.stem
+
+        out_path = (
+            checkpoint_path.parent
+            / f"{stem}.field_v2.onnx"
+        )
+
+    contract_path = out_path.with_suffix(
+        ".contract.json"
+    )
+
+    checkpoint_sha = sha256_file(
+        checkpoint_path
+    )
+
+    W, H = contract.resolution_wh
+
+    print("=" * 96)
+    print(
+        "AGROBOT | EXPORT ONNX CAMPO V2"
+    )
+    print("=" * 96)
+    print(
+        f"Config              : {config_path}"
+    )
+    print(
+        f"Checkpoint          : {checkpoint_path}"
+    )
+    print(
+        f"Checkpoint SHA256   : {checkpoint_sha}"
+    )
+    print(
+        f"Trainer             : {contract.trainer_version}"
+    )
+    print(
+        f"Epoch               : {contract.epoch}"
+    )
+    print(
+        f"Bests               : {contract.bests}"
+    )
+    print(
+        f"Backbone            : {contract.backbone}"
+    )
+    print(
+        f"Resolution          : {W}x{H}"
+    )
+    print(
+        f"Seg classes         : {contract.seg_id2label}"
+    )
+    print(
+        f"Nav / NonNav        : "
+        f"{contract.nav_class_id} / {contract.non_nav_class_id}"
+    )
+    print(
+        f"Status classes      : {contract.label_name_by_id}"
+    )
+    print(
+        f"Status head         : {contract.status_head}"
+    )
+    print(
+        f"Norm mean           : {contract.norm_mean}"
+    )
+    print(
+        f"Norm std            : {contract.norm_std}"
+    )
+    print(
+        f"Device export       : {device}"
+    )
+    print(
+        f"Opset               : {args.opset}"
+    )
+    print(
+        f"ONNX                : {out_path}"
+    )
+    print("=" * 96)
+
+    print(
+        "[MODEL] reconstruindo SegFormer + CorridorStatusHead..."
+    )
+
+    base_model, status_head = build_models(
+        ckpt,
+        contract,
+        device,
+    )
+
+    # Wrapper de REFERÊNCIA: arquitetura exatamente igual ao treino.
+    reference_wrapper = CorridorFieldOnnx(
+        base_model=base_model,
+        status_head=status_head,
+        resolution_wh=contract.resolution_wh,
+        norm_mean=contract.norm_mean,
+        norm_std=contract.norm_std,
+    ).to(
+        device
+    ).eval()
+
+    # -------------------------------------------------------------------------
+    # Adaptação export-only do AdaptiveAvgPool2d.
+    #
+    # O ONNX de campo possui forma espacial fixa. Descobrimos as formas nativas
+    # do checkpoint e substituímos SOMENTE NA CÓPIA da cabeça:
+    #
+    #     AdaptiveAvgPool2d(pool_hw)
+    #
+    # por:
+    #
+    #     AvgPool2d(kernel=input_hw/pool_hw, stride=kernel)
+    #
+    # quando a divisão é exata. O wrapper original permanece intacto e será
+    # usado depois para verificar PyTorch-original x ONNX.
+    # -------------------------------------------------------------------------
+
+    feat_hw, seg_hw = discover_status_native_shapes(
+        base_model=base_model,
+        H=H,
+        W=W,
+        device=device,
+        mean=contract.norm_mean,
+        std=contract.norm_std,
+    )
+
+    export_status_head, export_pool_plan = build_export_status_head(
+        reference_status_head=status_head,
+        feat_hw=feat_hw,
+        seg_hw=seg_hw,
+        device=device,
+    )
+
+    print(
+        "[EXPORT-ADAPT] status pooling | "
+        f"feat {export_pool_plan.feat_hw} -> {export_pool_plan.pool_hw} "
+        f"kernel={export_pool_plan.feat_kernel_hw} | "
+        f"seg {export_pool_plan.seg_hw} -> {export_pool_plan.pool_hw} "
+        f"kernel={export_pool_plan.seg_kernel_hw}"
+    )
+
+    export_wrapper = CorridorFieldOnnx(
+        base_model=base_model,
+        status_head=export_status_head,
+        resolution_wh=contract.resolution_wh,
+        norm_mean=contract.norm_mean,
+        norm_std=contract.norm_std,
+    ).to(
+        device
+    ).eval()
+
+    export_adaptation_check = verify_export_wrapper_equivalence(
+        reference_wrapper=reference_wrapper,
+        export_wrapper=export_wrapper,
+        H=H,
+        W=W,
+        device=device,
+        max_prob_abs=1e-6,
+    )
+
+    # -------------------------------------------------------------------------
+    # Export
+    # -------------------------------------------------------------------------
+
+    export_onnx(
+        wrapper=export_wrapper,
+        out_path=out_path,
+        H=H,
+        W=W,
+        device=device,
+        opset=int(
+            args.opset
+        ),
+    )
+
+    print(
+        f"[OK] ONNX exportado: {out_path}"
+    )
+
+    # -------------------------------------------------------------------------
+    # Checker
+    # -------------------------------------------------------------------------
+
+    onnx_check = check_onnx(
+        out_path
+    )
+
+    # -------------------------------------------------------------------------
+    # Contract + metadata embed
+    # -------------------------------------------------------------------------
+
+    core_contract = contract_core_dict(
+        checkpoint_contract=contract,
+        checkpoint_sha256=checkpoint_sha,
+        opset=int(
+            args.opset
+        ),
+        export_pool_plan=export_pool_plan,
+    )
+
+    metadata_embedded = embed_onnx_metadata(
+        out_path,
+        core_contract,
+    )
+
+    if metadata_embedded:
+        print(
+            "[OK] contrato básico embutido em ONNX metadata_props."
+        )
+    else:
+        print(
+            "[WARN] metadata não embutida porque pacote onnx não está instalado."
+        )
+
+    # Embedding modifies the file, so checker once more when possible.
+    if metadata_embedded:
+        onnx_check = check_onnx(
+            out_path
+        )
+
+    # -------------------------------------------------------------------------
+    # ORT verification
+    # -------------------------------------------------------------------------
+
+    verify_result = {
+        "executed": False,
+        "passed": None,
+    }
+
+    if not args.no_ort_verify:
+        verify_dir = (
+            Path(
+                args.verify_dir
+            ).resolve()
+            if args.verify_dir
+            else None
+        )
+
+        verify_result = verify_with_ort(
+            onnx_path=out_path,
+            wrapper=reference_wrapper,
+            contract=contract,
+            project_root=project_root,
+            device=device,
+            provider_mode=args.verify_provider,
+            verify_samples=max(
+                0,
+                int(
+                    args.verify_samples
+                ),
+            ),
+            verify_dir=verify_dir,
+            min_seg_agreement=float(
+                args.min_seg_agreement
+            ),
+            max_prob_abs=float(
+                args.max_prob_abs
+            ),
+            max_prob_abs_synthetic=float(
+                args.max_prob_abs_synthetic
+            ),
+        )
+
+    # -------------------------------------------------------------------------
+    # Final hashes / full sidecar
+    # -------------------------------------------------------------------------
+
+    onnx_sha = sha256_file(
+        out_path
+    )
+
+    trtexec_path = find_trtexec(
+        args.trtexec
+    )
+
+    full_contract = {
+        **core_contract,
+
+        "artifact": {
+            "onnx_path": str(
+                out_path
+            ),
+            "onnx_sha256": onnx_sha,
+            "contract_path": str(
+                contract_path
+            ),
+            "checkpoint_path": str(
+                checkpoint_path
+            ),
+            "checkpoint_sha256": checkpoint_sha,
+            "metadata_embedded_in_onnx": bool(
+                metadata_embedded
+            ),
+        },
+
+        "validation": {
+            "export_adaptation_pytorch_equivalence": export_adaptation_check,
+            "onnx_check": onnx_check,
+            "ort": verify_result,
+        },
+
+        "deployment": {
+            "target": "VisualWorker / ONNX Runtime / TensorRT EP",
+            "recommended_precision": "FP16 TensorRT engine",
+            "trtexec_detected": (
+                str(
+                    trtexec_path
+                )
+                if trtexec_path
+                is not None
+                else None
+            ),
+            "runtime_steps": [
+                "obtain RGB uint8 frame",
+                (
+                    f"crop source ROI [{contract.roi_inicio}, "
+                    f"{contract.roi_tamanho}] if source frame is larger"
+                ),
+                (
+                    f"resize source image to {W}x{H} "
+                    "before ONNX input"
+                ),
+                (
+                    f"feed {INPUT_NAME} as contiguous uint8 NHWC "
+                    f"shape [1,{H},{W},3]"
+                ),
+                f"read {SEG_OUTPUT_NAME} int32 [1,{H},{W}]",
+                (
+                    f"read {STATUS_OUTPUT_NAME} float32 "
+                    f"[1,{len(contract.label_name_by_id)}]"
+                ),
+                "status_id = argmax(label_probs[0])",
+                "status_conf = label_probs[0,status_id]",
+            ],
+        },
+    }
+
+    save_json(
+        contract_path,
+        full_contract,
+    )
+
+    print()
+    print("=" * 96)
+    print(
+        "EXPORT FINALIZADO"
+    )
+    print("=" * 96)
+    print(
+        f"ONNX                : {out_path}"
+    )
+    print(
+        f"ONNX SHA256         : {onnx_sha}"
+    )
+    print(
+        f"Contrato            : {contract_path}"
+    )
+    print(
+        f"Input               : "
+        f"{INPUT_NAME} uint8 [1,{H},{W},3] RGB NHWC"
+    )
+    print(
+        f"Output SEG          : "
+        f"{SEG_OUTPUT_NAME} int32 [1,{H},{W}]"
+    )
+    print(
+        f"Output STATUS       : "
+        f"{STATUS_OUTPUT_NAME} float32 "
+        f"[1,{len(contract.label_name_by_id)}]"
+    )
+    print(
+        f"ORT verify          : "
+        f"{verify_result.get('passed')}"
+    )
+    print("=" * 96)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/Python/OAK/datasets/oak-d/_7_benchmark_corridor.py b/Python/OAK/datasets/oak-d/_7_benchmark_corridor.py
new file mode 100644
index 000000000..e60733d3b
--- /dev/null
+++ b/Python/OAK/datasets/oak-d/_7_benchmark_corridor.py
@@ -0,0 +1,3589 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+"""
+benchmark_corridor_field_v2.py
+==============================
+
+Benchmark de ponta a ponta do modelo frontal de corredor.
+
+Fecha o pipeline:
+
+    CAPTURE
+      ↓
+    ANNOTATE
+      ↓
+    NORMALIZE
+      ↓
+    SPLIT
+      ↓
+    TRAIN
+      ↓
+    TEST / REVIEW
+      ↓
+    EXPORT ONNX
+      ↓
+    BENCHMARK  ← este script
+      ↓
+    VISUAL WORKER
+
+O benchmark foi dividido em camadas para não esconder gargalos.
+
+FASES OFFLINE
+-------------
+1) host_prepare
+    Simula o preparo do frame bruto:
+        BGR source
+        ROI vertical
+        resize INTER_AREA
+        BGR -> RGB
+        contiguous uint8 NHWC
+        batch [1,H,W,3]
+
+2) pytorch_core
+    Mede SegFormer + CorridorStatusHead + resize logits + argmax + softmax.
+    Entrada já está normalizada float32 NCHW.
+    É a melhor aproximação do "modelo puro".
+
+3) pytorch_field
+    Mede o wrapper de campo PyTorch:
+        uint8 NHWC
+        cast
+        /255
+        transpose
+        mean/std
+        modelo
+        resize
+        argmax
+        status softmax
+
+4) onnx_cuda
+    Mede session.run() do ONNX de campo com CUDA EP.
+    Inclui o contrato real:
+        uint8 host -> GPU -> outputs CPU
+
+5) onnx_tensorrt
+    Mede o MESMO ONNX com TensorRT EP.
+    Registra:
+        criação da sessão
+        primeira inferência
+        warmup
+        steady-state
+
+PIPELINE REAL COM OAK-D
+-----------------------
+Opcional com --camera.
+
+Threads independentes:
+
+    CaptureThread
+        OAK-D Lite BGR 1080p
+             ↓ latest-slot
+    PrepareThread
+        ROI + resize + BGR→RGB + contiguous
+             ↓ latest-slot
+    InferThread
+        ONNX Runtime provider escolhido
+             ↓
+        seg_ids + label_probs
+
+A política latest-frame é intencional:
+    se o consumidor atrasar, frames antigos são descartados.
+    Isso evita que a inferência processe uma fila envelhecida.
+
+Métricas de câmera:
+    capture_fps
+    prepare_fps
+    output_fps
+
+    capture_frames
+    prepared_frames
+    inferred_frames
+
+    dropped_before_prepare
+    dropped_before_infer
+
+    prepare_ms:
+        mean / p50 / p95 / p99 / max
+
+    inference_ms:
+        mean / p50 / p95 / p99 / max
+
+    age_before_infer_ms
+    wait_after_prepare_ms
+    capture_to_result_ms
+
+ARTEFATOS
+---------
+benchmark_corridor/<run_name>/
+    benchmark_summary.json
+    benchmark_results.csv
+    camera_pipeline_samples.csv       # se --camera
+    environment.json
+
+USO RÁPIDO
+----------
+Benchmark offline completo:
+
+    python benchmark_corridor_field_v2.py
+
+Adicionar pipeline real da OAK-D com TensorRT:
+
+    python benchmark_corridor_field_v2.py --camera
+
+Pipeline real usando CUDA EP:
+
+    python benchmark_corridor_field_v2.py ^
+        --camera ^
+        --pipeline_provider cuda
+
+Medir cold-start do TensorRT:
+
+    python benchmark_corridor_field_v2.py ^
+        --clear_trt_cache
+
+Rodar mais tempo:
+
+    python benchmark_corridor_field_v2.py ^
+        --iterations 500 ^
+        --warmup 50 ^
+        --camera ^
+        --camera_duration 60
+
+Somente fases específicas:
+
+    python benchmark_corridor_field_v2.py ^
+        --modes pytorch_core,onnx_tensorrt
+
+OBSERVAÇÕES
+-----------
+- O script usa test_corridor_agri_v2.py para reconstruir o checkpoint.
+- Usa export_corridor_agri_onnx_v2.py para o wrapper de campo PyTorch.
+- O ONNX default é:
+      <checkpoint>.field_v2.onnx
+- O benchmark não compila o ONNX. Rode o export primeiro.
+"""
+
+from __future__ import annotations
+
+import argparse
+import csv
+import gc
+import importlib.util
+import json
+import os
+import platform
+import queue
+import shutil
+import socket
+import statistics
+import sys
+import threading
+import time
+from dataclasses import dataclass, asdict
+from datetime import datetime
+from pathlib import Path
+from typing import Any, Dict, Iterable, List, Optional, Sequence, Tuple
+
+import cv2
+import numpy as np
+import torch
+import torch.nn.functional as F
+
+
+# =============================================================================
+# Constants
+# =============================================================================
+
+BENCHMARK_VERSION = "agrobot_corridor_field_benchmark_v2.0"
+
+DEFAULT_MODES = (
+    "host_prepare",
+    "pytorch_core",
+    "pytorch_field",
+    "onnx_cuda",
+    "onnx_tensorrt",
+)
+
+VALID_MODES = {
+    "host_prepare",
+    "pytorch_core",
+    "pytorch_field",
+    "onnx_cpu",
+    "onnx_cuda",
+    "onnx_tensorrt",
+}
+
+INPUT_NAME = "rgb_u8_nhwc"
+SEG_OUTPUT_NAME = "seg_ids"
+STATUS_OUTPUT_NAME = "label_probs"
+
+
+# =============================================================================
+# Generic helpers
+# =============================================================================
+
+def now_iso() -> str:
+    return datetime.now().isoformat(
+        timespec="seconds"
+    )
+
+
+def timestamp_run_name() -> str:
+    return datetime.now().strftime(
+        "%Y%m%d_%H%M%S"
+    )
+
+
+def safe_json_load(
+    path: Path,
+) -> dict:
+    with path.open(
+        "r",
+        encoding="utf-8",
+    ) as f:
+        data = json.load(
+            f
+        )
+
+    if not isinstance(
+        data,
+        dict,
+    ):
+        raise ValueError(
+            f"Esperava objeto JSON: {path}"
+        )
+
+    return data
+
+
+def save_json(
+    path: Path,
+    data: dict,
+) -> None:
+    path.parent.mkdir(
+        parents=True,
+        exist_ok=True,
+    )
+
+    tmp = path.with_suffix(
+        path.suffix
+        + ".tmp"
+    )
+
+    with tmp.open(
+        "w",
+        encoding="utf-8",
+    ) as f:
+        json.dump(
+            data,
+            f,
+            ensure_ascii=False,
+            indent=2,
+        )
+
+    os.replace(
+        tmp,
+        path,
+    )
+
+
+def write_csv(
+    path: Path,
+    rows: Sequence[dict],
+) -> None:
+    path.parent.mkdir(
+        parents=True,
+        exist_ok=True,
+    )
+
+    if not rows:
+        path.write_text(
+            "",
+            encoding="utf-8",
+        )
+        return
+
+    fields = []
+    seen = set()
+
+    for row in rows:
+        for key in row:
+            if key not in seen:
+                seen.add(
+                    key
+                )
+                fields.append(
+                    key
+                )
+
+    with path.open(
+        "w",
+        newline="",
+        encoding="utf-8-sig",
+    ) as f:
+        writer = csv.DictWriter(
+            f,
+            fieldnames=fields,
+            extrasaction="ignore",
+        )
+        writer.writeheader()
+        writer.writerows(
+            rows
+        )
+
+
+def load_module_from_file(
+    script_path: Path,
+    module_name: str,
+):
+    if not script_path.is_file():
+        raise FileNotFoundError(
+            f"Script dependência não encontrado: {script_path}"
+        )
+
+    spec = importlib.util.spec_from_file_location(
+        module_name,
+        str(
+            script_path
+        ),
+    )
+
+    if (
+        spec is None
+        or spec.loader is None
+    ):
+        raise RuntimeError(
+            f"Não consegui importar {script_path}"
+        )
+
+    module = importlib.util.module_from_spec(
+        spec
+    )
+
+    sys.modules[
+        module_name
+    ] = module
+
+    spec.loader.exec_module(
+        module
+    )
+
+    return module
+
+
+def resolve_project_root(
+    config_path: Path,
+) -> Path:
+    root = config_path.resolve().parent
+
+    if not (
+        root
+        / "dataset"
+    ).is_dir():
+        raise FileNotFoundError(
+            f"dataset/ não encontrado ao lado de {config_path}"
+        )
+
+    return root
+
+
+def parse_modes(
+    text: str,
+) -> List[str]:
+    result = []
+
+    for raw in str(
+        text
+    ).split(
+        ","
+    ):
+        mode = raw.strip().lower()
+
+        if not mode:
+            continue
+
+        if mode == "all":
+            for item in DEFAULT_MODES:
+                if item not in result:
+                    result.append(
+                        item
+                    )
+            continue
+
+        if mode not in VALID_MODES:
+            raise ValueError(
+                f"Modo desconhecido: {mode}. "
+                f"Válidos={sorted(VALID_MODES)}"
+            )
+
+        if mode not in result:
+            result.append(
+                mode
+            )
+
+    if not result:
+        raise ValueError(
+            "Nenhum modo selecionado."
+        )
+
+    return result
+
+
+def percentile(
+    values: Sequence[float],
+    q: float,
+) -> Optional[float]:
+    if not values:
+        return None
+
+    return float(
+        np.percentile(
+            np.asarray(
+                values,
+                dtype=np.float64,
+            ),
+            q,
+        )
+    )
+
+
+def stats_ms(
+    values: Sequence[float],
+) -> dict:
+    if not values:
+        return {
+            "n": 0,
+            "min": None,
+            "mean": None,
+            "p50": None,
+            "p95": None,
+            "p99": None,
+            "max": None,
+            "std": None,
+        }
+
+    arr = np.asarray(
+        values,
+        dtype=np.float64,
+    )
+
+    return {
+        "n": int(
+            arr.size
+        ),
+        "min": float(
+            arr.min()
+        ),
+        "mean": float(
+            arr.mean()
+        ),
+        "p50": float(
+            np.percentile(
+                arr,
+                50,
+            )
+        ),
+        "p95": float(
+            np.percentile(
+                arr,
+                95,
+            )
+        ),
+        "p99": float(
+            np.percentile(
+                arr,
+                99,
+            )
+        ),
+        "max": float(
+            arr.max()
+        ),
+        "std": float(
+            arr.std()
+        ),
+    }
+
+
+def fps_from_ms(
+    mean_ms: Optional[float],
+) -> Optional[float]:
+    if (
+        mean_ms is None
+        or mean_ms <= 0
+    ):
+        return None
+
+    return float(
+        1000.0
+        / mean_ms
+    )
+
+
+def flatten_stats_row(
+    name: str,
+    kind: str,
+    stat: dict,
+    extra: Optional[dict] = None,
+) -> dict:
+    row = {
+        "name": name,
+        "kind": kind,
+        "n": stat.get(
+            "n"
+        ),
+        "mean_ms": stat.get(
+            "mean"
+        ),
+        "p50_ms": stat.get(
+            "p50"
+        ),
+        "p95_ms": stat.get(
+            "p95"
+        ),
+        "p99_ms": stat.get(
+            "p99"
+        ),
+        "min_ms": stat.get(
+            "min"
+        ),
+        "max_ms": stat.get(
+            "max"
+        ),
+        "std_ms": stat.get(
+            "std"
+        ),
+        "fps_from_mean": fps_from_ms(
+            stat.get(
+                "mean"
+            )
+        ),
+    }
+
+    if extra:
+        row.update(
+            extra
+        )
+
+    return row
+
+
+def synchronize_cuda(
+    device: torch.device,
+) -> None:
+    if device.type == "cuda":
+        torch.cuda.synchronize()
+
+
+# =============================================================================
+# Environment
+# =============================================================================
+
+def collect_environment() -> dict:
+    env = {
+        "created_at": now_iso(),
+        "benchmark_version": BENCHMARK_VERSION,
+        "hostname": socket.gethostname(),
+        "platform": platform.platform(),
+        "python": sys.version,
+        "opencv": cv2.__version__,
+        "numpy": np.__version__,
+        "torch": torch.__version__,
+        "torch_cuda_available": bool(
+            torch.cuda.is_available()
+        ),
+        "torch_cuda_version": torch.version.cuda,
+        "cudnn_version": (
+            torch.backends.cudnn.version()
+            if torch.backends.cudnn.is_available()
+            else None
+        ),
+    }
+
+    if torch.cuda.is_available():
+        env.update({
+            "gpu_name": torch.cuda.get_device_name(
+                0
+            ),
+            "gpu_capability": list(
+                torch.cuda.get_device_capability(
+                    0
+                )
+            ),
+            "gpu_total_memory_bytes": int(
+                torch.cuda.get_device_properties(
+                    0
+                ).total_memory
+            ),
+        })
+
+    try:
+        import onnxruntime as ort
+
+        env[
+            "onnxruntime"
+        ] = ort.__version__
+
+        env[
+            "onnxruntime_available_providers"
+        ] = ort.get_available_providers()
+
+    except Exception as exc:
+        env[
+            "onnxruntime"
+        ] = None
+
+        env[
+            "onnxruntime_error"
+        ] = (
+            f"{type(exc).__name__}: {exc}"
+        )
+
+    try:
+        import depthai as dai
+
+        env[
+            "depthai"
+        ] = getattr(
+            dai,
+            "__version__",
+            "unknown",
+        )
+    except Exception:
+        env[
+            "depthai"
+        ] = None
+
+    return env
+
+
+# =============================================================================
+# Geometry / preprocess
+# =============================================================================
+
+def compute_roi(
+    h: int,
+    roi_inicio: float,
+    roi_tamanho: float,
+) -> Tuple[int, int]:
+    y0 = int(
+        round(
+            h
+            * float(
+                roi_inicio
+            )
+        )
+    )
+
+    y1 = int(
+        round(
+            h
+            * min(
+                1.0,
+                float(
+                    roi_inicio
+                    + roi_tamanho
+                ),
+            )
+        )
+    )
+
+    y0 = max(
+        0,
+        min(
+            h - 1,
+            y0,
+        ),
+    )
+
+    y1 = max(
+        y0 + 1,
+        min(
+            h,
+            y1,
+        ),
+    )
+
+    return (
+        y0,
+        y1,
+    )
+
+
+def prepare_rgb_u8_nhwc(
+    frame_bgr: np.ndarray,
+    *,
+    resolution_wh: Tuple[int, int],
+    roi_inicio: float,
+    roi_tamanho: float,
+) -> np.ndarray:
+    """
+    Contrato EXTERNO do ONNX v2.
+
+    frame BGR source
+        -> ROI vertical
+        -> resize INTER_AREA
+        -> RGB
+        -> uint8 contiguous
+        -> [1,H,W,3]
+    """
+    if (
+        frame_bgr.ndim
+        != 3
+        or frame_bgr.shape[
+            2
+        ]
+        < 3
+    ):
+        raise ValueError(
+            f"Frame BGR inválido: shape={frame_bgr.shape}"
+        )
+
+    h, w = frame_bgr.shape[
+        :2
+    ]
+
+    y0, y1 = compute_roi(
+        h,
+        roi_inicio,
+        roi_tamanho,
+    )
+
+    roi = frame_bgr[
+        y0:y1,
+        :,
+        :3,
+    ]
+
+    W, H = (
+        int(
+            resolution_wh[
+                0
+            ]
+        ),
+        int(
+            resolution_wh[
+                1
+            ]
+        ),
+    )
+
+    if roi.shape[
+        1
+    ] == W and roi.shape[
+        0
+    ] == H:
+        resized = roi
+    else:
+        resized = cv2.resize(
+            roi,
+            (
+                W,
+                H,
+            ),
+            interpolation=cv2.INTER_AREA,
+        )
+
+    rgb = cv2.cvtColor(
+        resized,
+        cv2.COLOR_BGR2RGB,
+    )
+
+    return np.ascontiguousarray(
+        rgb[
+            None,
+            ...,
+        ],
+        dtype=np.uint8,
+    )
+
+
+def make_synthetic_source_frame(
+    source_wh: Tuple[int, int],
+    seed: int = 20260914,
+) -> np.ndarray:
+    W, H = (
+        int(
+            source_wh[
+                0
+            ]
+        ),
+        int(
+            source_wh[
+                1
+            ]
+        ),
+    )
+
+    rng = np.random.default_rng(
+        seed
+    )
+
+    return rng.integers(
+        0,
+        256,
+        size=(
+            H,
+            W,
+            3,
+        ),
+        dtype=np.uint8,
+    )
+
+
+def find_reference_image(
+    project_root: Path,
+    resolution_wh: Tuple[int, int],
+) -> Optional[Path]:
+    W, H = resolution_wh
+
+    candidates = [
+        (
+            project_root
+            / "dataset"
+            / f"{W}x{H}"
+            / "split"
+            / "val"
+            / "images"
+        ),
+        (
+            project_root
+            / "dataset"
+            / f"{W}x{H}"
+            / "split"
+            / "train"
+            / "images"
+        ),
+    ]
+
+    exts = {
+        ".png",
+        ".jpg",
+        ".jpeg",
+        ".bmp",
+        ".webp",
+    }
+
+    for folder in candidates:
+        if not folder.is_dir():
+            continue
+
+        images = sorted(
+            [
+                p
+                for p in folder.iterdir()
+                if (
+                    p.is_file()
+                    and p.suffix.lower()
+                    in exts
+                )
+            ],
+            key=lambda p: p.name.lower(),
+        )
+
+        if images:
+            return images[
+                0
+            ]
+
+    return None
+
+
+# =============================================================================
+# PyTorch inputs and calls
+# =============================================================================
+
+def uint8_contract_to_normalized_nchw(
+    input_np: np.ndarray,
+    *,
+    mean: Sequence[float],
+    std: Sequence[float],
+    device: torch.device,
+) -> torch.Tensor:
+    x = torch.from_numpy(
+        np.ascontiguousarray(
+            input_np
+        )
+    ).to(
+        device,
+        non_blocking=True,
+    )
+
+    x = x.to(
+        torch.float32
+    )
+
+    x = (
+        x
+        / 255.0
+    )
+
+    x = x.permute(
+        0,
+        3,
+        1,
+        2,
+    )
+
+    mean_t = torch.tensor(
+        list(
+            mean
+        ),
+        dtype=torch.float32,
+        device=device,
+    ).view(
+        1,
+        3,
+        1,
+        1,
+    )
+
+    std_t = torch.tensor(
+        list(
+            std
+        ),
+        dtype=torch.float32,
+        device=device,
+    ).view(
+        1,
+        3,
+        1,
+        1,
+    )
+
+    return (
+        x
+        - mean_t
+    ) / std_t
+
+
+@torch.inference_mode()
+def pytorch_core_call(
+    base_model,
+    status_head,
+    x_norm: torch.Tensor,
+    resolution_wh: Tuple[int, int],
+):
+    out = base_model(
+        pixel_values=x_norm,
+        output_hidden_states=True,
+        return_dict=True,
+    )
+
+    seg_logits_native = out.logits
+    feat = out.hidden_states[
+        -1
+    ]
+
+    status_logits = status_head(
+        feat,
+        seg_logits_native,
+    )
+
+    W, H = resolution_wh
+
+    seg_logits_full = F.interpolate(
+        seg_logits_native,
+        size=(
+            H,
+            W,
+        ),
+        mode="bilinear",
+        align_corners=False,
+    )
+
+    seg_ids = torch.argmax(
+        seg_logits_full,
+        dim=1,
+    )
+
+    label_probs = torch.softmax(
+        status_logits,
+        dim=1,
+    )
+
+    return (
+        seg_ids,
+        label_probs,
+    )
+
+
+# =============================================================================
+# Microbench harness
+# =============================================================================
+
+def benchmark_callable(
+    *,
+    name: str,
+    call,
+    warmup: int,
+    iterations: int,
+    sync_before_after: Optional[
+        callable
+    ] = None,
+) -> dict:
+    """
+    Mede wall time por chamada.
+
+    Para PyTorch CUDA:
+        sync_before_after = torch.cuda.synchronize
+
+    Para ORT session.run:
+        chamada é síncrona do ponto de vista host, portanto não precisa.
+    """
+    print(
+        f"[BENCH] {name}: warmup={warmup} iterations={iterations}"
+    )
+
+    first_ms = None
+
+    if sync_before_after:
+        sync_before_after()
+
+    t0 = time.perf_counter()
+
+    call()
+
+    if sync_before_after:
+        sync_before_after()
+
+    first_ms = (
+        time.perf_counter()
+        - t0
+    ) * 1000.0
+
+    for _ in range(
+        max(
+            0,
+            int(
+                warmup
+            ),
+        )
+    ):
+        call()
+
+    if sync_before_after:
+        sync_before_after()
+
+    samples = []
+
+    for _ in range(
+        max(
+            1,
+            int(
+                iterations
+            ),
+        )
+    ):
+        if sync_before_after:
+            sync_before_after()
+
+        t0 = time.perf_counter()
+
+        call()
+
+        if sync_before_after:
+            sync_before_after()
+
+        samples.append(
+            (
+                time.perf_counter()
+                - t0
+            )
+            * 1000.0
+        )
+
+    stat = stats_ms(
+        samples
+    )
+
+    print(
+        f"  first={first_ms:.3f}ms | "
+        f"mean={stat['mean']:.3f}ms | "
+        f"p50={stat['p50']:.3f} | "
+        f"p95={stat['p95']:.3f} | "
+        f"p99={stat['p99']:.3f} | "
+        f"fps≈{fps_from_ms(stat['mean']):.2f}"
+    )
+
+    return {
+        "name": name,
+        "first_ms": float(
+            first_ms
+        ),
+        "latency_ms": stat,
+        "fps_from_mean": fps_from_ms(
+            stat[
+                "mean"
+            ]
+        ),
+    }
+
+
+# =============================================================================
+# ONNX Runtime
+# =============================================================================
+
+def ort_provider_name(
+    provider_mode: str,
+) -> str:
+    return {
+        "cpu": "CPUExecutionProvider",
+        "cuda": "CUDAExecutionProvider",
+        "tensorrt": "TensorrtExecutionProvider",
+    }[
+        provider_mode
+    ]
+
+
+def create_ort_session(
+    *,
+    onnx_path: Path,
+    provider_mode: str,
+    trt_cache_dir: Path,
+    trt_fp16: bool,
+) -> Tuple[
+    Any,
+    dict,
+]:
+    import onnxruntime as ort
+
+    available = ort.get_available_providers()
+
+    provider_mode = str(
+        provider_mode
+    ).lower()
+
+    if provider_mode == "cpu":
+        requested = [
+            "CPUExecutionProvider"
+        ]
+
+    elif provider_mode == "cuda":
+        requested = [
+            "CUDAExecutionProvider",
+            "CPUExecutionProvider",
+        ]
+
+    elif provider_mode == "tensorrt":
+        trt_cache_dir.mkdir(
+            parents=True,
+            exist_ok=True,
+        )
+
+        trt_options = {
+            "trt_engine_cache_enable": True,
+            "trt_engine_cache_path": str(
+                trt_cache_dir
+            ),
+            "trt_timing_cache_enable": True,
+            "trt_timing_cache_path": str(
+                trt_cache_dir
+            ),
+            "trt_fp16_enable": bool(
+                trt_fp16
+            ),
+        }
+
+        requested = [
+            (
+                "TensorrtExecutionProvider",
+                trt_options,
+            ),
+            "CUDAExecutionProvider",
+            "CPUExecutionProvider",
+        ]
+
+    else:
+        raise ValueError(
+            provider_mode
+        )
+
+    filtered = []
+
+    for provider in requested:
+        name = (
+            provider[
+                0
+            ]
+            if isinstance(
+                provider,
+                tuple,
+            )
+            else provider
+        )
+
+        if name in available:
+            filtered.append(
+                provider
+            )
+
+    target_name = ort_provider_name(
+        provider_mode
+    )
+
+    filtered_names = [
+        p[
+            0
+        ]
+        if isinstance(
+            p,
+            tuple,
+        )
+        else p
+        for p in filtered
+    ]
+
+    if target_name not in filtered_names:
+        raise RuntimeError(
+            f"Provider {target_name} indisponível. "
+            f"ORT disponíveis={available}"
+        )
+
+    options = ort.SessionOptions()
+
+    options.graph_optimization_level = (
+        ort.GraphOptimizationLevel
+        .ORT_ENABLE_ALL
+    )
+
+    t0 = time.perf_counter()
+
+    session = ort.InferenceSession(
+        str(
+            onnx_path
+        ),
+        sess_options=options,
+        providers=filtered,
+    )
+
+    build_s = (
+        time.perf_counter()
+        - t0
+    )
+
+    active = session.get_providers()
+
+    if target_name not in active:
+        raise RuntimeError(
+            f"Provider solicitado não ficou ativo. "
+            f"requested={provider_mode} active={active}"
+        )
+
+    info = {
+        "provider_mode": provider_mode,
+        "target_provider": target_name,
+        "available_providers": list(
+            available
+        ),
+        "active_providers": list(
+            active
+        ),
+        "session_create_seconds": float(
+            build_s
+        ),
+        "trt_cache_dir": (
+            str(
+                trt_cache_dir
+            )
+            if provider_mode
+            == "tensorrt"
+            else None
+        ),
+        "trt_fp16": (
+            bool(
+                trt_fp16
+            )
+            if provider_mode
+            == "tensorrt"
+            else None
+        ),
+    }
+
+    return (
+        session,
+        info,
+    )
+
+
+def benchmark_ort(
+    *,
+    onnx_path: Path,
+    provider_mode: str,
+    input_np: np.ndarray,
+    warmup: int,
+    iterations: int,
+    trt_cache_dir: Path,
+    trt_fp16: bool,
+) -> dict:
+    print()
+    print(
+        f"[ORT] criando sessão provider={provider_mode}"
+    )
+
+    session, session_info = create_ort_session(
+        onnx_path=onnx_path,
+        provider_mode=provider_mode,
+        trt_cache_dir=trt_cache_dir,
+        trt_fp16=trt_fp16,
+    )
+
+    print(
+        f"  sessão={session_info['session_create_seconds']:.3f}s "
+        f"active={session_info['active_providers']}"
+    )
+
+    inputs = session.get_inputs()
+
+    outputs = session.get_outputs()
+
+    if len(
+        inputs
+    ) != 1:
+        raise RuntimeError(
+            f"ONNX esperado 1 input, recebeu {[x.name for x in inputs]}"
+        )
+
+    if inputs[
+        0
+    ].name != INPUT_NAME:
+        raise RuntimeError(
+            f"Input ONNX inesperado: {inputs[0].name}"
+        )
+
+    output_names = [
+        x.name
+        for x in outputs
+    ]
+
+    if output_names != [
+        SEG_OUTPUT_NAME,
+        STATUS_OUTPUT_NAME,
+    ]:
+        raise RuntimeError(
+            f"Outputs ONNX inesperados: {output_names}"
+        )
+
+    def call():
+        return session.run(
+            [
+                SEG_OUTPUT_NAME,
+                STATUS_OUTPUT_NAME,
+            ],
+            {
+                INPUT_NAME: input_np
+            },
+        )
+
+    result = benchmark_callable(
+        name=f"onnx_{provider_mode}",
+        call=call,
+        warmup=warmup,
+        iterations=iterations,
+        sync_before_after=None,
+    )
+
+    # Um sanity simples dos outputs.
+    seg, probs = call()
+
+    result.update({
+        "provider": session_info,
+        "seg_shape": list(
+            seg.shape
+        ),
+        "seg_dtype": str(
+            seg.dtype
+        ),
+        "status_shape": list(
+            probs.shape
+        ),
+        "status_dtype": str(
+            probs.dtype
+        ),
+        "status_argmax": int(
+            np.argmax(
+                probs[
+                    0
+                ]
+            )
+        ),
+        "status_confidence": float(
+            np.max(
+                probs[
+                    0
+                ]
+            )
+        ),
+    })
+
+    return (
+        result,
+        session,
+    )
+
+
+# =============================================================================
+# Camera pipeline data
+# =============================================================================
+
+@dataclass
+class CapturedFrame:
+    frame_id: int
+    t_capture: float
+    frame_bgr: np.ndarray
+
+
+@dataclass
+class PreparedFrame:
+    prep_id: int
+    source_frame_id: int
+
+    t_capture: float
+    t_prepare_start: float
+    t_prepare_end: float
+
+    input_np: np.ndarray
+
+
+class LatestSlot:
+    """
+    Slot single-item com versionamento.
+
+    put() substitui sempre o item anterior.
+    O consumidor pede get_after(version), então nunca espera por frames antigos.
+    """
+
+    def __init__(
+        self,
+    ):
+        self._lock = threading.Lock()
+        self._cond = threading.Condition(
+            self._lock
+        )
+
+        self._version = 0
+        self._item = None
+        self._closed = False
+
+    def put(
+        self,
+        item,
+    ) -> int:
+        with self._cond:
+            if self._closed:
+                return self._version
+
+            self._version += 1
+            self._item = item
+
+            self._cond.notify_all()
+
+            return self._version
+
+    def get_after(
+        self,
+        last_version: int,
+        timeout: float = 0.5,
+    ):
+        deadline = (
+            time.perf_counter()
+            + float(
+                timeout
+            )
+        )
+
+        with self._cond:
+            while (
+                not self._closed
+                and self._version
+                <= last_version
+            ):
+                remaining = (
+                    deadline
+                    - time.perf_counter()
+                )
+
+                if remaining <= 0:
+                    return (
+                        last_version,
+                        None,
+                    )
+
+                self._cond.wait(
+                    timeout=remaining
+                )
+
+            if (
+                self._version
+                <= last_version
+            ):
+                return (
+                    last_version,
+                    None,
+                )
+
+            return (
+                self._version,
+                self._item,
+            )
+
+    def close(
+        self,
+    ):
+        with self._cond:
+            self._closed = True
+            self._cond.notify_all()
+
+
+# =============================================================================
+# OAK camera
+# =============================================================================
+
+class OakCamera:
+    """
+    BGR host frames.
+
+    Tenta API moderna e depois clássica.
+    """
+
+    def __init__(
+        self,
+        fps: float,
+        resolution: str,
+    ):
+        self.fps = float(
+            fps
+        )
+
+        self.resolution = str(
+            resolution
+        ).lower()
+
+        self.mode = None
+        self.pipeline = None
+        self.device = None
+        self.queue = None
+        self.dai = None
+
+    def start(
+        self,
+    ):
+        try:
+            import depthai as dai
+        except ImportError as exc:
+            raise ImportError(
+                "Modo câmera exige depthai."
+            ) from exc
+
+        self.dai = dai
+
+        target_size = (
+            (
+                1920,
+                1080,
+            )
+            if self.resolution
+            == "1080p"
+            else (
+                1280,
+                720,
+            )
+        )
+
+        # API moderna.
+        try:
+            pipeline = dai.Pipeline()
+
+            cam = pipeline.create(
+                dai.node.Camera
+            ).build()
+
+            out = cam.requestOutput(
+                size=target_size,
+                type=dai.ImgFrame.Type.BGR888p,
+                fps=self.fps,
+            )
+
+            q = out.createOutputQueue(
+                maxSize=4,
+                blocking=False,
+            )
+
+            pipeline.start()
+
+            self.mode = "modern"
+            self.pipeline = pipeline
+            self.queue = q
+
+            print(
+                f"[CAM] modern API "
+                f"{target_size[0]}x{target_size[1]} "
+                f"@{self.fps:g}"
+            )
+
+            return self
+
+        except Exception as exc:
+            print(
+                f"[CAM] modern API indisponível: "
+                f"{type(exc).__name__}: {exc}"
+            )
+
+        # API clássica.
+        pipeline = dai.Pipeline()
+
+        cam = pipeline.create(
+            dai.node.ColorCamera
+        )
+
+        xout = pipeline.create(
+            dai.node.XLinkOut
+        )
+
+        xout.setStreamName(
+            "rgb"
+        )
+
+        try:
+            cam.setBoardSocket(
+                dai.CameraBoardSocket.CAM_A
+            )
+        except Exception:
+            pass
+
+        if (
+            self.resolution
+            == "1080p"
+        ):
+            cam.setResolution(
+                dai.ColorCameraProperties
+                .SensorResolution
+                .THE_1080_P
+            )
+        else:
+            cam.setResolution(
+                dai.ColorCameraProperties
+                .SensorResolution
+                .THE_720_P
+            )
+
+        cam.setFps(
+            self.fps
+        )
+
+        cam.setInterleaved(
+            False
+        )
+
+        cam.video.link(
+            xout.input
+        )
+
+        device = dai.Device(
+            pipeline
+        )
+
+        q = device.getOutputQueue(
+            name="rgb",
+            maxSize=4,
+            blocking=False,
+        )
+
+        self.mode = "classic"
+        self.pipeline = pipeline
+        self.device = device
+        self.queue = q
+
+        print(
+            f"[CAM] classic API "
+            f"{target_size[0]}x{target_size[1]} "
+            f"@{self.fps:g}"
+        )
+
+        return self
+
+    def read(
+        self,
+    ) -> np.ndarray:
+        if self.queue is None:
+            raise RuntimeError(
+                "Câmera não inicializada."
+            )
+
+        frame = self.queue.get()
+
+        img = frame.getCvFrame()
+
+        if img is None:
+            raise RuntimeError(
+                "Frame OAK vazio."
+            )
+
+        return img
+
+    def close(
+        self,
+    ):
+        try:
+            if (
+                self.mode
+                == "modern"
+                and self.pipeline
+                is not None
+            ):
+                self.pipeline.stop()
+        except Exception:
+            pass
+
+        try:
+            if self.device is not None:
+                self.device.close()
+        except Exception:
+            pass
+
+
+# =============================================================================
+# Threaded camera benchmark
+# =============================================================================
+
+class CameraPipelineBenchmark:
+    def __init__(
+        self,
+        *,
+        session,
+        contract,
+        camera_fps: float,
+        camera_resolution: str,
+        warmup_s: float,
+        duration_s: float,
+    ):
+        self.session = session
+        self.contract = contract
+
+        self.camera_fps = float(
+            camera_fps
+        )
+
+        self.camera_resolution = str(
+            camera_resolution
+        )
+
+        self.warmup_s = float(
+            warmup_s
+        )
+
+        self.duration_s = float(
+            duration_s
+        )
+
+        self.raw_slot = LatestSlot()
+        self.prepared_slot = LatestSlot()
+
+        self.stop_event = threading.Event()
+
+        self.capture_count_total = 0
+        self.prepare_count_total = 0
+        self.infer_count_total = 0
+
+        self.capture_times_measure = []
+        self.prepare_times_measure = []
+        self.infer_times_measure = []
+
+        self.samples = []
+
+        self.dropped_before_prepare_measure = 0
+        self.dropped_before_infer_measure = 0
+
+        self.errors = []
+
+        self.measure_start = None
+        self.measure_end = None
+
+        self.camera = None
+
+    def in_measurement(
+        self,
+        t: Optional[
+            float
+        ] = None,
+    ) -> bool:
+        if (
+            self.measure_start
+            is None
+            or self.measure_end
+            is None
+        ):
+            return False
+
+        if t is None:
+            t = time.perf_counter()
+
+        return (
+            self.measure_start
+            <= t
+            <= self.measure_end
+        )
+
+    def capture_loop(
+        self,
+    ):
+        frame_id = 0
+
+        try:
+            while not self.stop_event.is_set():
+                frame = self.camera.read()
+
+                t_capture = time.perf_counter()
+
+                item = CapturedFrame(
+                    frame_id=frame_id,
+                    t_capture=t_capture,
+                    frame_bgr=frame,
+                )
+
+                self.raw_slot.put(
+                    item
+                )
+
+                self.capture_count_total += 1
+
+                if self.in_measurement(
+                    t_capture
+                ):
+                    self.capture_times_measure.append(
+                        t_capture
+                    )
+
+                frame_id += 1
+
+        except Exception as exc:
+            self.errors.append(
+                "CaptureThread: "
+                f"{type(exc).__name__}: {exc}"
+            )
+            self.stop_event.set()
+
+        finally:
+            self.raw_slot.close()
+
+    def prepare_loop(
+        self,
+    ):
+        last_version = 0
+        prep_id = 0
+        last_source_measure_id = None
+
+        try:
+            while not self.stop_event.is_set():
+                version, item = self.raw_slot.get_after(
+                    last_version,
+                    timeout=0.5,
+                )
+
+                if item is None:
+                    continue
+
+                last_version = version
+
+                t0 = time.perf_counter()
+
+                input_np = prepare_rgb_u8_nhwc(
+                    item.frame_bgr,
+                    resolution_wh=self.contract.resolution_wh,
+                    roi_inicio=self.contract.roi_inicio,
+                    roi_tamanho=self.contract.roi_tamanho,
+                )
+
+                t1 = time.perf_counter()
+
+                prepared = PreparedFrame(
+                    prep_id=prep_id,
+                    source_frame_id=item.frame_id,
+                    t_capture=item.t_capture,
+                    t_prepare_start=t0,
+                    t_prepare_end=t1,
+                    input_np=input_np,
+                )
+
+                self.prepared_slot.put(
+                    prepared
+                )
+
+                self.prepare_count_total += 1
+
+                if self.in_measurement(
+                    t1
+                ):
+                    if last_source_measure_id is not None:
+                        gap = (
+                            item.frame_id
+                            - last_source_measure_id
+                            - 1
+                        )
+
+                        if gap > 0:
+                            self.dropped_before_prepare_measure += int(
+                                gap
+                            )
+
+                    last_source_measure_id = int(
+                        item.frame_id
+                    )
+
+                    self.prepare_times_measure.append(
+                        t1
+                    )
+
+                prep_id += 1
+
+        except Exception as exc:
+            self.errors.append(
+                "PrepareThread: "
+                f"{type(exc).__name__}: {exc}"
+            )
+            self.stop_event.set()
+
+        finally:
+            self.prepared_slot.close()
+
+    def infer_loop(
+        self,
+    ):
+        last_version = 0
+
+        prev_prep_measure_id = None
+
+        try:
+            while not self.stop_event.is_set():
+                version, item = self.prepared_slot.get_after(
+                    last_version,
+                    timeout=0.5,
+                )
+
+                if item is None:
+                    continue
+
+                last_version = version
+
+                t_infer_start = time.perf_counter()
+
+                seg_ids, label_probs = self.session.run(
+                    [
+                        SEG_OUTPUT_NAME,
+                        STATUS_OUTPUT_NAME,
+                    ],
+                    {
+                        INPUT_NAME: item.input_np
+                    },
+                )
+
+                # Visual Worker ainda precisa apenas desta decisão minúscula.
+                status_id = int(
+                    np.argmax(
+                        label_probs[
+                            0
+                        ]
+                    )
+                )
+
+                status_conf = float(
+                    label_probs[
+                        0,
+                        status_id,
+                    ]
+                )
+
+                t_done = time.perf_counter()
+
+                self.infer_count_total += 1
+
+                if self.in_measurement(
+                    t_done
+                ):
+                    if prev_prep_measure_id is not None:
+                        gap = (
+                            item.prep_id
+                            - prev_prep_measure_id
+                            - 1
+                        )
+
+                        if gap > 0:
+                            self.dropped_before_infer_measure += int(
+                                gap
+                            )
+
+                    prev_prep_measure_id = int(
+                        item.prep_id
+                    )
+
+                    self.infer_times_measure.append(
+                        t_done
+                    )
+
+                    self.samples.append({
+                        "source_frame_id": int(
+                            item.source_frame_id
+                        ),
+                        "prep_id": int(
+                            item.prep_id
+                        ),
+
+                        "prepare_ms": float(
+                            (
+                                item.t_prepare_end
+                                - item.t_prepare_start
+                            )
+                            * 1000.0
+                        ),
+
+                        "age_at_prepare_start_ms": float(
+                            (
+                                item.t_prepare_start
+                                - item.t_capture
+                            )
+                            * 1000.0
+                        ),
+
+                        "wait_after_prepare_ms": float(
+                            (
+                                t_infer_start
+                                - item.t_prepare_end
+                            )
+                            * 1000.0
+                        ),
+
+                        "age_before_infer_ms": float(
+                            (
+                                t_infer_start
+                                - item.t_capture
+                            )
+                            * 1000.0
+                        ),
+
+                        "inference_ms": float(
+                            (
+                                t_done
+                                - t_infer_start
+                            )
+                            * 1000.0
+                        ),
+
+                        "capture_to_result_ms": float(
+                            (
+                                t_done
+                                - item.t_capture
+                            )
+                            * 1000.0
+                        ),
+
+                        "status_id": status_id,
+                        "status_confidence": status_conf,
+
+                        # Snapshots acumulados.
+                        "dropped_before_prepare_accum": int(
+                            self.dropped_before_prepare_measure
+                        ),
+                        "dropped_before_infer_accum": int(
+                            self.dropped_before_infer_measure
+                        ),
+                    })
+
+        except Exception as exc:
+            self.errors.append(
+                "InferThread: "
+                f"{type(exc).__name__}: {exc}"
+            )
+            self.stop_event.set()
+
+    def run(
+        self,
+    ) -> dict:
+        self.camera = OakCamera(
+            fps=self.camera_fps,
+            resolution=self.camera_resolution,
+        ).start()
+
+        capture_thread = threading.Thread(
+            target=self.capture_loop,
+            name="CaptureThread",
+            daemon=True,
+        )
+
+        prepare_thread = threading.Thread(
+            target=self.prepare_loop,
+            name="PrepareThread",
+            daemon=True,
+        )
+
+        infer_thread = threading.Thread(
+            target=self.infer_loop,
+            name="InferThread",
+            daemon=True,
+        )
+
+        t_pipeline_start = time.perf_counter()
+
+        capture_thread.start()
+        prepare_thread.start()
+        infer_thread.start()
+
+        print(
+            f"[CAM BENCH] warmup={self.warmup_s:.1f}s"
+        )
+
+        while (
+            time.perf_counter()
+            - t_pipeline_start
+            < self.warmup_s
+        ):
+            if self.stop_event.is_set():
+                break
+
+            time.sleep(
+                0.05
+            )
+
+        self.measure_start = time.perf_counter()
+        self.measure_end = (
+            self.measure_start
+            + self.duration_s
+        )
+
+        print(
+            f"[CAM BENCH] medição={self.duration_s:.1f}s"
+        )
+
+        while (
+            time.perf_counter()
+            < self.measure_end
+        ):
+            if self.stop_event.is_set():
+                break
+
+            time.sleep(
+                0.05
+            )
+
+        self.stop_event.set()
+
+        self.raw_slot.close()
+        self.prepared_slot.close()
+
+        capture_thread.join(
+            timeout=3.0
+        )
+        prepare_thread.join(
+            timeout=3.0
+        )
+        infer_thread.join(
+            timeout=3.0
+        )
+
+        self.camera.close()
+
+        actual_measure_duration = max(
+            1e-6,
+            min(
+                time.perf_counter(),
+                self.measure_end,
+            )
+            - self.measure_start
+        )
+
+        prepare_ms = [
+            float(
+                r[
+                    "prepare_ms"
+                ]
+            )
+            for r in self.samples
+        ]
+
+        infer_ms = [
+            float(
+                r[
+                    "inference_ms"
+                ]
+            )
+            for r in self.samples
+        ]
+
+        age_prepare = [
+            float(
+                r[
+                    "age_at_prepare_start_ms"
+                ]
+            )
+            for r in self.samples
+        ]
+
+        wait_after_prepare = [
+            float(
+                r[
+                    "wait_after_prepare_ms"
+                ]
+            )
+            for r in self.samples
+        ]
+
+        age_before_infer = [
+            float(
+                r[
+                    "age_before_infer_ms"
+                ]
+            )
+            for r in self.samples
+        ]
+
+        total_latency = [
+            float(
+                r[
+                    "capture_to_result_ms"
+                ]
+            )
+            for r in self.samples
+        ]
+
+        dropped_before_prepare = int(
+            self.dropped_before_prepare_measure
+        )
+
+        dropped_before_infer = int(
+            self.dropped_before_infer_measure
+        )
+
+        summary = {
+            "warmup_seconds": float(
+                self.warmup_s
+            ),
+            "measurement_seconds": float(
+                actual_measure_duration
+            ),
+
+            "capture_frames_measure": int(
+                len(
+                    self.capture_times_measure
+                )
+            ),
+            "prepare_frames_measure": int(
+                len(
+                    self.prepare_times_measure
+                )
+            ),
+            "infer_frames_measure": int(
+                len(
+                    self.samples
+                )
+            ),
+
+            "capture_fps": float(
+                len(
+                    self.capture_times_measure
+                )
+                / actual_measure_duration
+            ),
+            "prepare_fps": float(
+                len(
+                    self.prepare_times_measure
+                )
+                / actual_measure_duration
+            ),
+            "output_fps": float(
+                len(
+                    self.samples
+                )
+                / actual_measure_duration
+            ),
+
+            "dropped_before_prepare": dropped_before_prepare,
+            "dropped_before_infer": dropped_before_infer,
+
+            "prepare_ms": stats_ms(
+                prepare_ms
+            ),
+            "age_at_prepare_start_ms": stats_ms(
+                age_prepare
+            ),
+            "wait_after_prepare_ms": stats_ms(
+                wait_after_prepare
+            ),
+            "age_before_infer_ms": stats_ms(
+                age_before_infer
+            ),
+            "inference_ms": stats_ms(
+                infer_ms
+            ),
+            "capture_to_result_ms": stats_ms(
+                total_latency
+            ),
+
+            "errors": list(
+                self.errors
+            ),
+        }
+
+        return summary
+
+
+# =============================================================================
+# Main
+# =============================================================================
+
+def main():
+    ap = argparse.ArgumentParser(
+        description=(
+            "Benchmark de campo do modelo frontal: "
+            "PyTorch, ONNX CUDA, TensorRT e OAK-D threaded."
+        )
+    )
+
+    ap.add_argument(
+        "--config",
+        default="config.json",
+    )
+
+    ap.add_argument(
+        "--test_script",
+        default=None,
+        help="Default: test_corridor_agri_v2.py ao lado.",
+    )
+
+    ap.add_argument(
+        "--export_script",
+        default=None,
+        help="Default: export_corridor_agri_onnx_v2.py ao lado.",
+    )
+
+    ap.add_argument(
+        "--ckpt",
+        default=None,
+    )
+
+    ap.add_argument(
+        "--checkpoint",
+        default="operational",
+        choices=[
+            "operational",
+            "nav",
+            "status",
+            "last",
+        ],
+    )
+
+    ap.add_argument(
+        "--onnx",
+        default=None,
+        help=(
+            "Default: <checkpoint>.field_v2.onnx"
+        ),
+    )
+
+    ap.add_argument(
+        "--modes",
+        default="all",
+        help=(
+            "CSV: host_prepare,pytorch_core,pytorch_field,"
+            "onnx_cpu,onnx_cuda,onnx_tensorrt ou all."
+        ),
+    )
+
+    ap.add_argument(
+        "--iterations",
+        type=int,
+        default=200,
+    )
+
+    ap.add_argument(
+        "--warmup",
+        type=int,
+        default=30,
+    )
+
+    ap.add_argument(
+        "--source_width",
+        type=int,
+        default=1920,
+    )
+
+    ap.add_argument(
+        "--source_height",
+        type=int,
+        default=1080,
+    )
+
+    ap.add_argument(
+        "--device",
+        default="cuda",
+        choices=[
+            "cuda",
+            "cpu",
+        ],
+    )
+
+    ap.add_argument(
+        "--trt_fp16",
+        action="store_true",
+        default=True,
+    )
+
+    ap.add_argument(
+        "--trt_fp32",
+        action="store_true",
+        help="Desliga FP16 do TensorRT.",
+    )
+
+    ap.add_argument(
+        "--trt_cache",
+        default=None,
+    )
+
+    ap.add_argument(
+        "--clear_trt_cache",
+        action="store_true",
+        help="Mede cold-start removendo o cache TRT antes de criar sessão.",
+    )
+
+    ap.add_argument(
+        "--camera",
+        action="store_true",
+        help="Adiciona benchmark threaded real com OAK-D Lite.",
+    )
+
+    ap.add_argument(
+        "--pipeline_provider",
+        default="tensorrt",
+        choices=[
+            "cpu",
+            "cuda",
+            "tensorrt",
+        ],
+    )
+
+    ap.add_argument(
+        "--camera_fps",
+        type=float,
+        default=30.0,
+    )
+
+    ap.add_argument(
+        "--camera_res",
+        default="1080p",
+        choices=[
+            "1080p",
+            "720p",
+        ],
+    )
+
+    ap.add_argument(
+        "--camera_warmup",
+        type=float,
+        default=5.0,
+    )
+
+    ap.add_argument(
+        "--camera_duration",
+        type=float,
+        default=30.0,
+    )
+
+    ap.add_argument(
+        "--out_root",
+        default="benchmark_corridor",
+    )
+
+    ap.add_argument(
+        "--run_name",
+        default=None,
+    )
+
+    args = ap.parse_args()
+
+    if args.iterations <= 0:
+        ap.error(
+            "--iterations deve ser > 0."
+        )
+
+    if args.warmup < 0:
+        ap.error(
+            "--warmup deve ser >= 0."
+        )
+
+    modes = parse_modes(
+        args.modes
+    )
+
+    here = Path(
+        __file__
+    ).resolve()
+
+    config_path = Path(
+        args.config
+    ).resolve()
+
+    if not config_path.is_file():
+        raise FileNotFoundError(
+            config_path
+        )
+
+    config = safe_json_load(
+        config_path
+    )
+
+    project_root = resolve_project_root(
+        config_path
+    )
+
+    test_script = (
+        Path(
+            args.test_script
+        ).resolve()
+        if args.test_script
+        else here.with_name(
+            "test_corridor_agri_v2.py"
+        )
+    )
+
+    export_script = (
+        Path(
+            args.export_script
+        ).resolve()
+        if args.export_script
+        else here.with_name(
+            "export_corridor_agri_onnx_v2.py"
+        )
+    )
+
+    test_mod = load_module_from_file(
+        test_script,
+        "agrobot_test_corridor_v2_bench",
+    )
+
+    export_mod = load_module_from_file(
+        export_script,
+        "agrobot_export_corridor_v2_bench",
+    )
+
+    if args.ckpt:
+        checkpoint_path = Path(
+            args.ckpt
+        ).resolve()
+    else:
+        checkpoint_path = test_mod.default_checkpoint_path(
+            project_root,
+            config,
+            args.checkpoint,
+        )
+
+    if args.onnx:
+        onnx_path = Path(
+            args.onnx
+        ).resolve()
+    else:
+        onnx_path = (
+            checkpoint_path.parent
+            / f"{checkpoint_path.stem}.field_v2.onnx"
+        )
+
+    run_name = (
+        args.run_name
+        or (
+            f"{checkpoint_path.stem}_"
+            f"{timestamp_run_name()}"
+        )
+    )
+
+    out_root = Path(
+        args.out_root
+    )
+
+    if not out_root.is_absolute():
+        out_root = (
+            project_root
+            / out_root
+        )
+
+    run_dir = (
+        out_root
+        / run_name
+    ).resolve()
+
+    run_dir.mkdir(
+        parents=True,
+        exist_ok=True,
+    )
+
+    environment = collect_environment()
+
+    save_json(
+        run_dir
+        / "environment.json",
+        environment,
+    )
+
+    use_cuda = (
+        args.device
+        == "cuda"
+        and torch.cuda.is_available()
+    )
+
+    if (
+        args.device
+        == "cuda"
+        and not use_cuda
+    ):
+        print(
+            "[WARN] CUDA solicitado mas indisponível. "
+            "PyTorch benchmark usará CPU."
+        )
+
+    device = torch.device(
+        "cuda"
+        if use_cuda
+        else "cpu"
+    )
+
+    # -------------------------------------------------------------------------
+    # Load checkpoint only if required.
+    # -------------------------------------------------------------------------
+
+    needs_pytorch = any(
+        mode.startswith(
+            "pytorch_"
+        )
+        for mode in modes
+    )
+
+    base_model = None
+    status_head = None
+    contract = None
+    checkpoint_raw = None
+    field_wrapper = None
+
+    # Mesmo quando só ONNX/camera, contract vem do checkpoint para geometria.
+    (
+        base_model_loaded,
+        status_head_loaded,
+        contract,
+        checkpoint_raw,
+    ) = test_mod.build_runtime_from_checkpoint(
+        checkpoint_path,
+        config,
+        device,
+    )
+
+    base_model = base_model_loaded
+    status_head = status_head_loaded
+
+    W, H = contract.resolution_wh
+
+    # Wrapper exatamente igual ao export.
+    field_wrapper = export_mod.CorridorFieldOnnx(
+        base_model=base_model,
+        status_head=status_head,
+        resolution_wh=contract.resolution_wh,
+        norm_mean=contract.norm_mean.tolist()
+        if hasattr(
+            contract.norm_mean,
+            "tolist",
+        )
+        else contract.norm_mean,
+        norm_std=contract.norm_std.tolist()
+        if hasattr(
+            contract.norm_std,
+            "tolist",
+        )
+        else contract.norm_std,
+    ).to(
+        device
+    ).eval()
+
+    print("=" * 100)
+    print(
+        "AGROBOT | CORRIDOR FIELD BENCHMARK V2"
+    )
+    print("=" * 100)
+    print(
+        f"Run                : {run_name}"
+    )
+    print(
+        f"Checkpoint         : {checkpoint_path}"
+    )
+    print(
+        f"ONNX               : {onnx_path}"
+    )
+    print(
+        f"Resolution         : {W}x{H}"
+    )
+    print(
+        f"Source synth       : "
+        f"{args.source_width}x{args.source_height}"
+    )
+    print(
+        f"Device PyTorch     : {device}"
+    )
+    print(
+        f"Modes              : {modes}"
+    )
+    print(
+        f"Iterations/warmup  : "
+        f"{args.iterations}/{args.warmup}"
+    )
+    print(
+        f"Camera             : {args.camera}"
+    )
+    print("=" * 100)
+
+    if any(
+        mode.startswith(
+            "onnx_"
+        )
+        for mode in modes
+    ) or args.camera:
+        if not onnx_path.is_file():
+            raise FileNotFoundError(
+                f"ONNX não encontrado: {onnx_path}\n"
+                "Rode export_corridor_agri_onnx_v2.py primeiro."
+            )
+
+    # -------------------------------------------------------------------------
+    # Build reference input.
+    # -------------------------------------------------------------------------
+
+    ref_image_path = find_reference_image(
+        project_root,
+        contract.resolution_wh,
+    )
+
+    if ref_image_path is not None:
+        ref_bgr = cv2.imread(
+            str(
+                ref_image_path
+            ),
+            cv2.IMREAD_COLOR,
+        )
+
+        if ref_bgr is None:
+            ref_bgr = make_synthetic_source_frame(
+                (
+                    args.source_width,
+                    args.source_height,
+                )
+            )
+
+            ref_source = "synthetic_raw"
+        else:
+            ref_source = str(
+                ref_image_path
+            )
+    else:
+        ref_bgr = make_synthetic_source_frame(
+            (
+                args.source_width,
+                args.source_height,
+            )
+        )
+
+        ref_source = "synthetic_raw"
+
+    # Para host_prepare precisamos sempre uma source grande.
+    synthetic_raw = make_synthetic_source_frame(
+        (
+            args.source_width,
+            args.source_height,
+        )
+    )
+
+    input_np = prepare_rgb_u8_nhwc(
+        ref_bgr,
+        resolution_wh=contract.resolution_wh,
+        roi_inicio=contract.roi_inicio,
+        roi_tamanho=contract.roi_tamanho,
+    )
+
+    print(
+        f"[INPUT] referência={ref_source}"
+    )
+    print(
+        f"[INPUT] contract={input_np.shape} {input_np.dtype}"
+    )
+
+    results: Dict[
+        str,
+        dict
+    ] = {}
+
+    csv_rows = []
+
+    # -------------------------------------------------------------------------
+    # Host prepare
+    # -------------------------------------------------------------------------
+
+    if "host_prepare" in modes:
+        def host_prepare_call():
+            return prepare_rgb_u8_nhwc(
+                synthetic_raw,
+                resolution_wh=contract.resolution_wh,
+                roi_inicio=contract.roi_inicio,
+                roi_tamanho=contract.roi_tamanho,
+            )
+
+        result = benchmark_callable(
+            name="host_prepare",
+            call=host_prepare_call,
+            warmup=args.warmup,
+            iterations=args.iterations,
+        )
+
+        result[
+            "source_shape"
+        ] = list(
+            synthetic_raw.shape
+        )
+
+        result[
+            "target_shape"
+        ] = list(
+            input_np.shape
+        )
+
+        results[
+            "host_prepare"
+        ] = result
+
+        csv_rows.append(
+            flatten_stats_row(
+                "host_prepare",
+                "host",
+                result[
+                    "latency_ms"
+                ],
+                {
+                    "first_ms": result[
+                        "first_ms"
+                    ],
+                },
+            )
+        )
+
+    # -------------------------------------------------------------------------
+    # PyTorch core
+    # -------------------------------------------------------------------------
+
+    x_norm = uint8_contract_to_normalized_nchw(
+        input_np,
+        mean=(
+            contract.norm_mean.tolist()
+            if hasattr(
+                contract.norm_mean,
+                "tolist",
+            )
+            else contract.norm_mean
+        ),
+        std=(
+            contract.norm_std.tolist()
+            if hasattr(
+                contract.norm_std,
+                "tolist",
+            )
+            else contract.norm_std
+        ),
+        device=device,
+    )
+
+    if "pytorch_core" in modes:
+        def pt_core_call():
+            return pytorch_core_call(
+                base_model,
+                status_head,
+                x_norm,
+                contract.resolution_wh,
+            )
+
+        result = benchmark_callable(
+            name="pytorch_core",
+            call=pt_core_call,
+            warmup=args.warmup,
+            iterations=args.iterations,
+            sync_before_after=(
+                lambda: synchronize_cuda(
+                    device
+                )
+            ),
+        )
+
+        result[
+            "input_shape"
+        ] = list(
+            x_norm.shape
+        )
+
+        result[
+            "input_dtype"
+        ] = str(
+            x_norm.dtype
+        )
+
+        results[
+            "pytorch_core"
+        ] = result
+
+        csv_rows.append(
+            flatten_stats_row(
+                "pytorch_core",
+                "model",
+                result[
+                    "latency_ms"
+                ],
+                {
+                    "first_ms": result[
+                        "first_ms"
+                    ],
+                },
+            )
+        )
+
+    # -------------------------------------------------------------------------
+    # PyTorch field wrapper
+    # -------------------------------------------------------------------------
+
+    if "pytorch_field" in modes:
+        x_u8 = torch.from_numpy(
+            np.ascontiguousarray(
+                input_np
+            )
+        ).to(
+            device,
+            non_blocking=True,
+        )
+
+        def pt_field_call():
+            with torch.inference_mode():
+                return field_wrapper(
+                    x_u8
+                )
+
+        result = benchmark_callable(
+            name="pytorch_field",
+            call=pt_field_call,
+            warmup=args.warmup,
+            iterations=args.iterations,
+            sync_before_after=(
+                lambda: synchronize_cuda(
+                    device
+                )
+            ),
+        )
+
+        result[
+            "input_shape"
+        ] = list(
+            x_u8.shape
+        )
+
+        result[
+            "input_dtype"
+        ] = str(
+            x_u8.dtype
+        )
+
+        results[
+            "pytorch_field"
+        ] = result
+
+        csv_rows.append(
+            flatten_stats_row(
+                "pytorch_field",
+                "field_reference",
+                result[
+                    "latency_ms"
+                ],
+                {
+                    "first_ms": result[
+                        "first_ms"
+                    ],
+                },
+            )
+        )
+
+    # -------------------------------------------------------------------------
+    # Release PyTorch before ORT/TensorRT.
+    #
+    # Não queremos medir TensorRT competindo por VRAM com uma segunda cópia
+    # inteira do mesmo modelo carregada no PyTorch.
+    # -------------------------------------------------------------------------
+
+    if any(
+        mode.startswith(
+            "onnx_"
+        )
+        for mode in modes
+    ) or args.camera:
+        try:
+            del x_norm
+        except Exception:
+            pass
+
+        try:
+            del x_u8
+        except Exception:
+            pass
+
+        try:
+            del field_wrapper
+        except Exception:
+            pass
+
+        try:
+            del base_model
+        except Exception:
+            pass
+
+        try:
+            del status_head
+        except Exception:
+            pass
+
+        gc.collect()
+
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
+            torch.cuda.synchronize()
+
+        print(
+            "[MEM] PyTorch runtime liberado antes dos benchmarks ORT/TensorRT."
+        )
+
+    # -------------------------------------------------------------------------
+    # TRT cache
+    # -------------------------------------------------------------------------
+
+    trt_cache_dir = (
+        Path(
+            args.trt_cache
+        ).resolve()
+        if args.trt_cache
+        else (
+            onnx_path.parent
+            / "trt_cache_field_v2"
+        )
+    )
+
+    if args.clear_trt_cache:
+        print(
+            f"[TRT] removendo cache para cold-start: {trt_cache_dir}"
+        )
+
+        if trt_cache_dir.exists():
+            shutil.rmtree(
+                trt_cache_dir
+            )
+
+    trt_fp16 = (
+        not bool(
+            args.trt_fp32
+        )
+    )
+
+    ort_sessions = {}
+
+    # -------------------------------------------------------------------------
+    # ONNX providers
+    # -------------------------------------------------------------------------
+
+    provider_modes = []
+
+    if "onnx_cpu" in modes:
+        provider_modes.append(
+            "cpu"
+        )
+
+    if "onnx_cuda" in modes:
+        provider_modes.append(
+            "cuda"
+        )
+
+    if "onnx_tensorrt" in modes:
+        provider_modes.append(
+            "tensorrt"
+        )
+
+    for provider_mode in provider_modes:
+        try:
+            result, session = benchmark_ort(
+                onnx_path=onnx_path,
+                provider_mode=provider_mode,
+                input_np=input_np,
+                warmup=args.warmup,
+                iterations=args.iterations,
+                trt_cache_dir=trt_cache_dir,
+                trt_fp16=trt_fp16,
+            )
+
+            key = (
+                f"onnx_{provider_mode}"
+            )
+
+            results[
+                key
+            ] = result
+
+            ort_sessions[
+                provider_mode
+            ] = session
+
+            csv_rows.append(
+                flatten_stats_row(
+                    key,
+                    "onnxruntime",
+                    result[
+                        "latency_ms"
+                    ],
+                    {
+                        "first_ms": result[
+                            "first_ms"
+                        ],
+                        "session_create_seconds": result[
+                            "provider"
+                        ][
+                            "session_create_seconds"
+                        ],
+                        "active_providers": "|".join(
+                            result[
+                                "provider"
+                            ][
+                                "active_providers"
+                            ]
+                        ),
+                    },
+                )
+            )
+
+        except Exception as exc:
+            key = (
+                f"onnx_{provider_mode}"
+            )
+
+            print(
+                f"[WARN] {key} indisponível/falhou: "
+                f"{type(exc).__name__}: {exc}"
+            )
+
+            results[
+                key
+            ] = {
+                "error": (
+                    f"{type(exc).__name__}: {exc}"
+                )
+            }
+
+            csv_rows.append({
+                "name": key,
+                "kind": "onnxruntime",
+                "error": results[
+                    key
+                ][
+                    "error"
+                ],
+            })
+
+    # -------------------------------------------------------------------------
+    # Camera threaded pipeline
+    # -------------------------------------------------------------------------
+
+    camera_summary = None
+    camera_samples = []
+
+    if args.camera:
+        provider_mode = str(
+            args.pipeline_provider
+        )
+
+        print()
+        print("=" * 100)
+        print(
+            f"CAMERA PIPELINE | provider={provider_mode}"
+        )
+        print("=" * 100)
+
+        session = ort_sessions.get(
+            provider_mode
+        )
+
+        session_info = None
+
+        if session is None:
+            session, session_info = create_ort_session(
+                onnx_path=onnx_path,
+                provider_mode=provider_mode,
+                trt_cache_dir=trt_cache_dir,
+                trt_fp16=trt_fp16,
+            )
+        else:
+            session_info = results.get(
+                f"onnx_{provider_mode}",
+                {},
+            ).get(
+                "provider"
+            )
+
+        cam_bench = CameraPipelineBenchmark(
+            session=session,
+            contract=contract,
+            camera_fps=args.camera_fps,
+            camera_resolution=args.camera_res,
+            warmup_s=args.camera_warmup,
+            duration_s=args.camera_duration,
+        )
+
+        camera_summary = cam_bench.run()
+
+        camera_summary[
+            "provider"
+        ] = session_info
+
+        camera_summary[
+            "camera_resolution"
+        ] = args.camera_res
+
+        camera_summary[
+            "camera_requested_fps"
+        ] = float(
+            args.camera_fps
+        )
+
+        camera_samples = cam_bench.samples
+
+        results[
+            "camera_pipeline"
+        ] = camera_summary
+
+        print()
+        print(
+            "[CAM RESULT] "
+            f"capture={camera_summary['capture_fps']:.2f} FPS | "
+            f"prepare={camera_summary['prepare_fps']:.2f} FPS | "
+            f"output={camera_summary['output_fps']:.2f} FPS"
+        )
+
+        if camera_summary[
+            "inference_ms"
+        ][
+            "mean"
+        ] is not None:
+            print(
+                "[CAM RESULT] "
+                f"infer mean/p95="
+                f"{camera_summary['inference_ms']['mean']:.2f}/"
+                f"{camera_summary['inference_ms']['p95']:.2f} ms"
+            )
+
+        if camera_summary[
+            "capture_to_result_ms"
+        ][
+            "mean"
+        ] is not None:
+            print(
+                "[CAM RESULT] "
+                f"E2E capture→result mean/p95="
+                f"{camera_summary['capture_to_result_ms']['mean']:.2f}/"
+                f"{camera_summary['capture_to_result_ms']['p95']:.2f} ms"
+            )
+
+        print(
+            "[CAM RESULT] "
+            f"drops raw/prepared="
+            f"{camera_summary['dropped_before_prepare']}/"
+            f"{camera_summary['dropped_before_infer']}"
+        )
+
+        if camera_summary[
+            "errors"
+        ]:
+            print(
+                f"[CAM WARN] errors={camera_summary['errors']}"
+            )
+
+        csv_rows.append({
+            "name": "camera_pipeline",
+            "kind": "threaded_camera",
+            "capture_fps": camera_summary[
+                "capture_fps"
+            ],
+            "prepare_fps": camera_summary[
+                "prepare_fps"
+            ],
+            "output_fps": camera_summary[
+                "output_fps"
+            ],
+            "prepare_mean_ms": camera_summary[
+                "prepare_ms"
+            ][
+                "mean"
+            ],
+            "prepare_p95_ms": camera_summary[
+                "prepare_ms"
+            ][
+                "p95"
+            ],
+            "infer_mean_ms": camera_summary[
+                "inference_ms"
+            ][
+                "mean"
+            ],
+            "infer_p95_ms": camera_summary[
+                "inference_ms"
+            ][
+                "p95"
+            ],
+            "e2e_mean_ms": camera_summary[
+                "capture_to_result_ms"
+            ][
+                "mean"
+            ],
+            "e2e_p95_ms": camera_summary[
+                "capture_to_result_ms"
+            ][
+                "p95"
+            ],
+            "dropped_before_prepare": camera_summary[
+                "dropped_before_prepare"
+            ],
+            "dropped_before_infer": camera_summary[
+                "dropped_before_infer"
+            ],
+        })
+
+    # -------------------------------------------------------------------------
+    # GPU memory snapshot
+    # -------------------------------------------------------------------------
+
+    gpu_memory = None
+
+    if torch.cuda.is_available():
+        synchronize_cuda(
+            torch.device(
+                "cuda"
+            )
+        )
+
+        gpu_memory = {
+            "allocated_bytes": int(
+                torch.cuda.memory_allocated(
+                    0
+                )
+            ),
+            "reserved_bytes": int(
+                torch.cuda.memory_reserved(
+                    0
+                )
+            ),
+            "max_allocated_bytes": int(
+                torch.cuda.max_memory_allocated(
+                    0
+                )
+            ),
+            "max_reserved_bytes": int(
+                torch.cuda.max_memory_reserved(
+                    0
+                )
+            ),
+        }
+
+    # -------------------------------------------------------------------------
+    # Summary
+    # -------------------------------------------------------------------------
+
+    summary = {
+        "schema": "agrobot.visual.corridor.benchmark.v2",
+        "benchmark_version": BENCHMARK_VERSION,
+        "created_at": now_iso(),
+
+        "run_name": run_name,
+
+        "project_root": str(
+            project_root
+        ),
+        "config": str(
+            config_path
+        ),
+        "checkpoint": str(
+            checkpoint_path
+        ),
+        "onnx": str(
+            onnx_path
+        ),
+
+        "contract": {
+            "resolution_wh": [
+                int(
+                    W
+                ),
+                int(
+                    H
+                ),
+            ],
+            "roi_inicio": float(
+                contract.roi_inicio
+            ),
+            "roi_tamanho": float(
+                contract.roi_tamanho
+            ),
+            "nav_class_id": int(
+                contract.nav_class_id
+            ),
+            "non_nav_class_id": int(
+                contract.non_nav_class_id
+            ),
+            "status_classes": {
+                str(
+                    k
+                ): v
+                for k, v in contract.label_name_by_id.items()
+            },
+            "input_name": INPUT_NAME,
+            "outputs": [
+                SEG_OUTPUT_NAME,
+                STATUS_OUTPUT_NAME,
+            ],
+        },
+
+        "settings": {
+            "modes": modes,
+            "iterations": int(
+                args.iterations
+            ),
+            "warmup": int(
+                args.warmup
+            ),
+            "trt_fp16": bool(
+                trt_fp16
+            ),
+            "trt_cache_dir": str(
+                trt_cache_dir
+            ),
+            "clear_trt_cache": bool(
+                args.clear_trt_cache
+            ),
+            "camera": bool(
+                args.camera
+            ),
+            "pipeline_provider": str(
+                args.pipeline_provider
+            ),
+        },
+
+        "reference_input": {
+            "source": ref_source,
+            "shape": list(
+                input_np.shape
+            ),
+            "dtype": str(
+                input_np.dtype
+            ),
+        },
+
+        "environment": environment,
+        "gpu_memory": gpu_memory,
+
+        "results": results,
+    }
+
+    save_json(
+        run_dir
+        / "benchmark_summary.json",
+        summary,
+    )
+
+    write_csv(
+        run_dir
+        / "benchmark_results.csv",
+        csv_rows,
+    )
+
+    if camera_samples:
+        write_csv(
+            run_dir
+            / "camera_pipeline_samples.csv",
+            camera_samples,
+        )
+
+    # -------------------------------------------------------------------------
+    # Compact ranking
+    # -------------------------------------------------------------------------
+
+    print()
+    print("=" * 100)
+    print(
+        "BENCHMARK FINALIZADO"
+    )
+    print("=" * 100)
+
+    ranking = []
+
+    for name, result in results.items():
+        if (
+            isinstance(
+                result,
+                dict
+            )
+            and "latency_ms"
+            in result
+            and result[
+                "latency_ms"
+            ].get(
+                "mean"
+            )
+            is not None
+        ):
+            ranking.append(
+                (
+                    name,
+                    float(
+                        result[
+                            "latency_ms"
+                        ][
+                            "mean"
+                        ]
+                    ),
+                )
+            )
+
+    for name, mean_ms in sorted(
+        ranking,
+        key=lambda x: x[
+            1
+        ],
+    ):
+        print(
+            f"{name:22s} "
+            f"{mean_ms:9.3f} ms  "
+            f"{1000.0 / mean_ms:8.2f} FPS"
+        )
+
+    if camera_summary is not None:
+        print("-" * 100)
+        print(
+            f"{'camera_pipeline':22s} "
+            f"output={camera_summary['output_fps']:.2f} FPS | "
+            f"E2E p50={camera_summary['capture_to_result_ms']['p50']:.2f} ms | "
+            f"p95={camera_summary['capture_to_result_ms']['p95']:.2f} ms"
+        )
+
+    print("-" * 100)
+    print(
+        f"Summary : {run_dir / 'benchmark_summary.json'}"
+    )
+    print(
+        f"CSV     : {run_dir / 'benchmark_results.csv'}"
+    )
+
+    if camera_samples:
+        print(
+            f"Camera  : {run_dir / 'camera_pipeline_samples.csv'}"
+        )
+
+    print("=" * 100)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/Python/OAK/datasets/oak-d/audit/review_corridor_dataset_v2.py b/Python/OAK/datasets/oak-d/audit/review_corridor_dataset_v2.py
new file mode 100644
index 000000000..38e876314
--- /dev/null
+++ b/Python/OAK/datasets/oak-d/audit/review_corridor_dataset_v2.py
@@ -0,0 +1,5125 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+"""
+review_corridor_dataset_v2.py
+=============================
+
+Auditoria / model-mining do dataset frontal de corredor.
+
+Objetivo
+--------
+Rodar o checkpoint treinado sobre TODO o dataset normalizado, comparar:
+
+    1) segmentação GT vs predição;
+    2) estado global GT vs segunda cabeça;
+
+e ranquear as amostras que mais merecem revisão humana.
+
+IMPORTANTE
+----------
+Discordância modelo x GT NÃO prova que a anotação está errada.
+
+O modelo é usado como um "segundo revisor":
+    - conflito de alta confiança no interior de uma região;
+    - falso navegável de alta confiança;
+    - estado global diferente com alta confiança;
+    - composição semântica incompatível com o grupo anotado;
+
+são sinais úteis para encontrar:
+    - máscaras antigas ruins;
+    - labels globais incorretos;
+    - inconsistências de anotação;
+    - ou, igualmente importante, falhas sistemáticas do próprio modelo.
+
+O review_score é SOMENTE um ranking heurístico.
+Não é probabilidade estatística de erro humano.
+
+Backend
+-------
+Este auditor usa o checkpoint PyTorch v2 de propósito.
+
+O ONNX de CAMPO v2 foi desenhado para ser mínimo:
+    rgb_u8_nhwc -> seg_ids + label_probs
+
+Ele não expõe semantic logits/probabilities por pixel, portanto não é adequado
+para "high-confidence model mining".
+
+Não vamos poluir o contrato de campo só para auditoria. Se futuramente quisermos
+auditoria ONNX acelerada, criaremos um export AUDIT separado.
+
+Dependência
+-----------
+Por padrão carrega dinamicamente:
+    test_corridor_agri_v2.py
+
+Assim reutilizamos EXATAMENTE:
+    - reconstrução SegFormer;
+    - CorridorStatusHead 3x4;
+    - contrato do checkpoint;
+    - preprocess;
+    - labelmap;
+    - descoberta de dataset.
+
+Entrada padrão
+--------------
+    dataset/<W>x<H>/group/
+
+Estrutura esperada:
+    group/
+      navegavel/
+        images/
+        masks/
+        labels/
+      naonavegavel/
+        ...
+      naonavegave_navegavel/
+        ...
+
+Também aceita split/train ou split/val.
+
+Saídas
+------
+    dataset/revisao_corridor/
+      reports/<run_name>/
+        review_report.csv
+        review_candidates.csv
+        review_summary.json
+        confusion_semantic.csv
+        confusion_status.csv
+        README_REVIEW.txt
+
+      group/<grupo>/
+        original_images/
+        original_masks/
+        original_labels/
+        predictions/
+        panels/                 # opcional
+        final_masks/            # PRESERVADO pelo --clear_review
+        review_order.csv
+
+      review_source_manifest.csv
+
+Exemplos
+--------
+1) Auditoria completa + export dos casos tecnicamente suspeitos:
+
+    python review_corridor_dataset_v2.py ^
+        --clear_review ^
+        --save_panels
+
+2) Só gerar relatórios:
+
+    python review_corridor_dataset_v2.py ^
+        --report_only ^
+        --run_name best_operational_audit
+
+3) Pegar os 25 casos mais suspeitos de cada grupo x estado:
+
+    python review_corridor_dataset_v2.py ^
+        --top_k_per_stratum 25 ^
+        --clear_review ^
+        --save_panels
+
+4) Exportar somente score >= 70/100:
+
+    python review_corridor_dataset_v2.py ^
+        --export_mode score ^
+        --min_suspicion_pct 70 ^
+        --clear_review
+
+Critérios principais
+--------------------
+SEGMENTAÇÃO:
+    semantic_disagree_frac
+    highconf_disagree_frac
+    interior_disagree_frac
+    highconf_interior_disagree_frac
+    unsafe_nav_frac
+    highconf_unsafe_nav_frac
+    blocked_nav_frac
+    nav_iou
+    non_nav_iou
+
+SEGUNDA CABEÇA:
+    gt_status
+    pred_status
+    status_confidence
+    gt_status_probability
+    status_margin
+    status_mismatch
+
+COMPOSIÇÃO:
+    gt_group
+    pred_group
+    group_mismatch
+
+Fronteira vs interior
+---------------------
+Discordâncias muito perto da borda da máscara recebem menos peso no ranking.
+Conflitos de alta confiança LONGE da borda recebem mais peso.
+
+Isso é importante para separar:
+    "1 ou 2 px de contorno"
+de:
+    "o modelo vê uma região inteira diferente da anotação".
+"""
+
+from __future__ import annotations
+
+import argparse
+import csv
+import gc
+import importlib.util
+import json
+import math
+import os
+import re
+import shutil
+import sys
+import time
+from collections import Counter, defaultdict
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Dict, Iterable, List, Optional, Sequence, Tuple
+
+import cv2
+import numpy as np
+import torch
+import torch.nn.functional as F
+
+
+# =============================================================================
+# Constants
+# =============================================================================
+
+REVIEWER_VERSION = "agrobot_corridor_dataset_review_v2.0"
+REVIEW_SCHEMA = "agrobot.visual.corridor.dataset_review.v2"
+IGNORE_INDEX = 255
+
+IMAGE_EXTS = {
+    ".png",
+    ".jpg",
+    ".jpeg",
+    ".bmp",
+    ".webp",
+    ".tif",
+    ".tiff",
+}
+
+
+# =============================================================================
+# Generic helpers
+# =============================================================================
+
+def now_iso() -> str:
+    from datetime import datetime
+
+    return datetime.now().isoformat(
+        timespec="seconds"
+    )
+
+
+def load_json(
+    path: Path,
+) -> dict:
+    with path.open(
+        "r",
+        encoding="utf-8",
+    ) as f:
+        data = json.load(
+            f
+        )
+
+    if not isinstance(
+        data,
+        dict,
+    ):
+        raise ValueError(
+            f"Esperava objeto JSON: {path}"
+        )
+
+    return data
+
+
+def save_json(
+    path: Path,
+    data: dict,
+) -> None:
+    path.parent.mkdir(
+        parents=True,
+        exist_ok=True,
+    )
+
+    tmp = path.with_suffix(
+        path.suffix
+        + ".tmp"
+    )
+
+    with tmp.open(
+        "w",
+        encoding="utf-8",
+    ) as f:
+        json.dump(
+            data,
+            f,
+            ensure_ascii=False,
+            indent=2,
+        )
+
+    os.replace(
+        tmp,
+        path,
+    )
+
+
+def write_csv(
+    path: Path,
+    rows: Sequence[dict],
+    fieldnames: Optional[
+        Sequence[str]
+    ] = None,
+) -> None:
+    path.parent.mkdir(
+        parents=True,
+        exist_ok=True,
+    )
+
+    if fieldnames is None:
+        keys = []
+        seen = set()
+
+        for row in rows:
+            for key in row:
+                if key not in seen:
+                    seen.add(
+                        key
+                    )
+                    keys.append(
+                        key
+                    )
+
+        fieldnames = keys
+
+    with path.open(
+        "w",
+        newline="",
+        encoding="utf-8-sig",
+    ) as f:
+        writer = csv.DictWriter(
+            f,
+            fieldnames=list(
+                fieldnames
+            ),
+            extrasaction="ignore",
+        )
+
+        writer.writeheader()
+
+        for row in rows:
+            writer.writerow(
+                row
+            )
+
+
+def safe_float(
+    value,
+    default: float = 0.0,
+) -> float:
+    try:
+        x = float(
+            value
+        )
+
+        return (
+            x
+            if math.isfinite(
+                x
+            )
+            else default
+        )
+    except Exception:
+        return default
+
+
+def clear_dir(
+    path: Path,
+) -> None:
+    if path.exists():
+        shutil.rmtree(
+            path
+        )
+
+    path.mkdir(
+        parents=True,
+        exist_ok=True,
+    )
+
+
+def natural_key(
+    text: str,
+):
+    parts = re.split(
+        r"(\d+)",
+        str(
+            text
+        ),
+    )
+
+    return [
+        int(
+            p
+        )
+        if p.isdigit()
+        else p.lower()
+        for p in parts
+    ]
+
+
+# =============================================================================
+# Load shared tester
+# =============================================================================
+
+def load_test_module(
+    script_path: Path,
+):
+    if not script_path.is_file():
+        raise FileNotFoundError(
+            f"Tester base não encontrado: {script_path}\n"
+            "Coloque review_corridor_dataset_v2.py ao lado de "
+            "test_corridor_agri_v2.py ou informe --test_script."
+        )
+
+    spec = importlib.util.spec_from_file_location(
+        "agrobot_corridor_test_v2",
+        str(
+            script_path
+        ),
+    )
+
+    if (
+        spec is None
+        or spec.loader is None
+    ):
+        raise RuntimeError(
+            f"Não consegui importar tester: {script_path}"
+        )
+
+    module = importlib.util.module_from_spec(
+        spec
+    )
+
+    sys.modules[
+        spec.name
+    ] = module
+
+    spec.loader.exec_module(
+        module
+    )
+
+    required = (
+        "safe_json_load",
+        "find_project_root",
+        "load_labelmap",
+        "discover_samples",
+        "default_checkpoint_path",
+        "build_runtime_from_checkpoint",
+        "prepare_input",
+        "load_gt_mask",
+        "read_gt_label",
+        "colorize_ids",
+    )
+
+    missing = [
+        name
+        for name in required
+        if not hasattr(
+            module,
+            name,
+        )
+    ]
+
+    if missing:
+        raise RuntimeError(
+            f"Tester base incompatível; faltando={missing}"
+        )
+
+    return module
+
+
+# =============================================================================
+# Metrics
+# =============================================================================
+
+def confusion_matrix_np(
+    pred: np.ndarray,
+    gt: np.ndarray,
+    num_classes: int,
+    ignore_id: int,
+) -> np.ndarray:
+    if (
+        pred.shape
+        != gt.shape
+    ):
+        pred = cv2.resize(
+            pred.astype(
+                np.uint8
+            ),
+            (
+                gt.shape[
+                    1
+                ],
+                gt.shape[
+                    0
+                ],
+            ),
+            interpolation=cv2.INTER_NEAREST,
+        )
+
+    valid = (
+        gt
+        != int(
+            ignore_id
+        )
+    )
+
+    pred_v = pred[
+        valid
+    ].astype(
+        np.int64
+    )
+    gt_v = gt[
+        valid
+    ].astype(
+        np.int64
+    )
+
+    valid2 = (
+        (gt_v >= 0)
+        & (
+            gt_v
+            < num_classes
+        )
+        & (
+            pred_v
+            >= 0
+        )
+        & (
+            pred_v
+            < num_classes
+        )
+    )
+
+    pred_v = pred_v[
+        valid2
+    ]
+    gt_v = gt_v[
+        valid2
+    ]
+
+    cm = np.zeros(
+        (
+            num_classes,
+            num_classes,
+        ),
+        dtype=np.int64,
+    )
+
+    if gt_v.size:
+        idx = (
+            gt_v
+            * num_classes
+            + pred_v
+        )
+
+        bins = np.bincount(
+            idx,
+            minlength=(
+                num_classes
+                * num_classes
+            ),
+        )
+
+        cm += bins.reshape(
+            num_classes,
+            num_classes,
+        )
+
+    return cm
+
+
+def binary_metrics_from_cm(
+    cm: np.ndarray,
+    nav_id: int,
+    non_nav_id: int,
+) -> dict:
+    cmf = cm.astype(
+        np.float64
+    )
+
+    tp = np.diag(
+        cmf
+    )
+    fp = cmf.sum(
+        axis=0
+    ) - tp
+    fn = cmf.sum(
+        axis=1
+    ) - tp
+
+    iou = tp / np.maximum(
+        tp
+        + fp
+        + fn,
+        1e-12,
+    )
+
+    precision = tp / np.maximum(
+        tp
+        + fp,
+        1e-12,
+    )
+
+    recall = tp / np.maximum(
+        tp
+        + fn,
+        1e-12,
+    )
+
+    f1 = (
+        2.0
+        * precision
+        * recall
+        / np.maximum(
+            precision
+            + recall,
+            1e-12,
+        )
+    )
+
+    acc = float(
+        tp.sum()
+        / max(
+            1.0,
+            cmf.sum(),
+        )
+    )
+
+    gt_non_nav = float(
+        cmf[
+            non_nav_id,
+            :,
+        ].sum()
+    )
+
+    unsafe = float(
+        cmf[
+            non_nav_id,
+            nav_id,
+        ]
+        / max(
+            1.0,
+            gt_non_nav,
+        )
+    )
+
+    gt_nav = float(
+        cmf[
+            nav_id,
+            :,
+        ].sum()
+    )
+
+    blocked = float(
+        cmf[
+            nav_id,
+            non_nav_id,
+        ]
+        / max(
+            1.0,
+            gt_nav,
+        )
+    )
+
+    return {
+        "acc": acc,
+        "miou": float(
+            np.mean(
+                iou
+            )
+        ),
+        "nav_iou": float(
+            iou[
+                nav_id
+            ]
+        ),
+        "non_nav_iou": float(
+            iou[
+                non_nav_id
+            ]
+        ),
+        "nav_f1": float(
+            f1[
+                nav_id
+            ]
+        ),
+        "non_nav_f1": float(
+            f1[
+                non_nav_id
+            ]
+        ),
+        "unsafe_nav_rate": unsafe,
+        "blocked_nav_rate": blocked,
+    }
+
+
+def status_metrics_from_cm(
+    cm: np.ndarray,
+) -> dict:
+    cmf = cm.astype(
+        np.float64
+    )
+
+    tp = np.diag(
+        cmf
+    )
+    fp = cmf.sum(
+        axis=0
+    ) - tp
+    fn = cmf.sum(
+        axis=1
+    ) - tp
+    support = cmf.sum(
+        axis=1
+    )
+
+    precision = tp / np.maximum(
+        tp
+        + fp,
+        1e-12,
+    )
+    recall = tp / np.maximum(
+        tp
+        + fn,
+        1e-12,
+    )
+    f1 = (
+        2.0
+        * precision
+        * recall
+        / np.maximum(
+            precision
+            + recall,
+            1e-12,
+        )
+    )
+
+    present = support > 0
+
+    return {
+        "acc": float(
+            tp.sum()
+            / max(
+                1.0,
+                cmf.sum(),
+            )
+        ),
+        "macro_f1": (
+            float(
+                f1[
+                    present
+                ].mean()
+            )
+            if np.any(
+                present
+            )
+            else 0.0
+        ),
+        "precision": precision.tolist(),
+        "recall": recall.tolist(),
+        "f1": f1.tolist(),
+        "support": support.tolist(),
+    }
+
+
+def write_cm_csv(
+    path: Path,
+    cm: np.ndarray,
+    labels: Sequence[str],
+) -> None:
+    rows = []
+
+    for gi, gt_name in enumerate(
+        labels
+    ):
+        row = {
+            "gt\\pred": gt_name
+        }
+
+        for pi, pred_name in enumerate(
+            labels
+        ):
+            row[
+                pred_name
+            ] = int(
+                cm[
+                    gi,
+                    pi,
+                ]
+            )
+
+        rows.append(
+            row
+        )
+
+    write_csv(
+        path,
+        rows,
+        [
+            "gt\\pred",
+            *labels,
+        ],
+    )
+
+
+# =============================================================================
+# Boundary analysis
+# =============================================================================
+
+def make_boundary_band(
+    gt: np.ndarray,
+    ignore_id: int,
+    radius_px: int,
+) -> np.ndarray:
+    """
+    Banda ao redor das transições de classe do GT.
+
+    A ideia é não tratar erro de 1..N px na borda com o mesmo peso de uma
+    discordância no interior de uma região.
+    """
+    valid = (
+        gt
+        != int(
+            ignore_id
+        )
+    )
+
+    if not np.any(
+        valid
+    ):
+        return np.zeros_like(
+            gt,
+            dtype=bool,
+        )
+
+    # Gradiente binário de classe.
+    # Para duas classes, qualquer vizinho diferente marca fronteira.
+    gt_u8 = gt.astype(
+        np.uint8,
+        copy=False,
+    )
+
+    boundary = np.zeros_like(
+        valid,
+        dtype=np.uint8,
+    )
+
+    neighbors = (
+        (-1, 0),
+        (1, 0),
+        (0, -1),
+        (0, 1),
+        (-1, -1),
+        (-1, 1),
+        (1, -1),
+        (1, 1),
+    )
+
+    for dy, dx in neighbors:
+        shifted = np.roll(
+            gt_u8,
+            shift=(
+                dy,
+                dx,
+            ),
+            axis=(
+                0,
+                1,
+            ),
+        )
+
+        shifted_valid = np.roll(
+            valid,
+            shift=(
+                dy,
+                dx,
+            ),
+            axis=(
+                0,
+                1,
+            ),
+        )
+
+        diff = (
+            valid
+            & shifted_valid
+            & (
+                shifted
+                != gt_u8
+            )
+        )
+
+        boundary[
+            diff
+        ] = 1
+
+    if radius_px > 0:
+        k = (
+            2
+            * int(
+                radius_px
+            )
+            + 1
+        )
+
+        kernel = np.ones(
+            (
+                k,
+                k,
+            ),
+            dtype=np.uint8,
+        )
+
+        boundary = cv2.dilate(
+            boundary,
+            kernel,
+            iterations=1,
+        )
+
+    return (
+        boundary.astype(
+            bool
+        )
+        & valid
+    )
+
+
+# =============================================================================
+# Semantic signals
+# =============================================================================
+
+def derive_composition_group(
+    pred: np.ndarray,
+    nav_id: int,
+    non_nav_id: int,
+    min_fraction: float,
+) -> Tuple[str, float]:
+    valid = (
+        (pred == nav_id)
+        | (
+            pred
+            == non_nav_id
+        )
+    )
+
+    n = int(
+        valid.sum()
+    )
+
+    if n <= 0:
+        return (
+            "unknown",
+            0.0,
+        )
+
+    nav_frac = float(
+        (
+            pred[
+                valid
+            ]
+            == nav_id
+        ).mean()
+    )
+
+    if nav_frac <= float(
+        min_fraction
+    ):
+        group = "naonavegavel"
+
+    elif (
+        1.0
+        - nav_frac
+    ) <= float(
+        min_fraction
+    ):
+        group = "navegavel"
+
+    else:
+        group = "naonavegave_navegavel"
+
+    return (
+        group,
+        nav_frac,
+    )
+
+
+def semantic_sample_signals(
+    *,
+    pred: np.ndarray,
+    probs: np.ndarray,
+    gt: np.ndarray,
+    nav_id: int,
+    non_nav_id: int,
+    ignore_id: int,
+    high_confidence: float,
+    uncertainty_margin: float,
+    boundary_radius: int,
+) -> dict:
+    if pred.shape != gt.shape:
+        pred = cv2.resize(
+            pred.astype(
+                np.uint8
+            ),
+            (
+                gt.shape[
+                    1
+                ],
+                gt.shape[
+                    0
+                ],
+            ),
+            interpolation=cv2.INTER_NEAREST,
+        )
+
+    if probs.shape[
+        1:
+    ] != gt.shape:
+        resized = []
+
+        for c in range(
+            probs.shape[
+                0
+            ]
+        ):
+            resized.append(
+                cv2.resize(
+                    probs[
+                        c
+                    ].astype(
+                        np.float32
+                    ),
+                    (
+                        gt.shape[
+                            1
+                        ],
+                        gt.shape[
+                            0
+                        ],
+                    ),
+                    interpolation=cv2.INTER_LINEAR,
+                )
+            )
+
+        probs = np.stack(
+            resized,
+            axis=0,
+        )
+
+    valid = (
+        gt
+        != int(
+            ignore_id
+        )
+    )
+
+    valid_n = int(
+        valid.sum()
+    )
+
+    if valid_n <= 0:
+        raise RuntimeError(
+            "Amostra sem pixels GT válidos."
+        )
+
+    top1 = np.max(
+        probs,
+        axis=0,
+    )
+
+    if probs.shape[
+        0
+    ] >= 2:
+        part = np.partition(
+            probs,
+            kth=(
+                probs.shape[
+                    0
+                ]
+                - 2
+            ),
+            axis=0,
+        )
+        top2 = part[
+            -2
+        ]
+    else:
+        top2 = np.zeros_like(
+            top1
+        )
+
+    margin = (
+        top1
+        - top2
+    )
+
+    wrong = (
+        valid
+        & (
+            pred
+            != gt
+        )
+    )
+
+    correct = (
+        valid
+        & (
+            pred
+            == gt
+        )
+    )
+
+    highconf = (
+        top1
+        >= float(
+            high_confidence
+        )
+    )
+
+    highconf_wrong = (
+        wrong
+        & highconf
+    )
+
+    uncertain = (
+        valid
+        & (
+            margin
+            <= float(
+                uncertainty_margin
+            )
+        )
+    )
+
+    boundary = make_boundary_band(
+        gt,
+        ignore_id,
+        boundary_radius,
+    )
+
+    interior = (
+        valid
+        & ~boundary
+    )
+
+    interior_wrong = (
+        wrong
+        & interior
+    )
+
+    hc_interior_wrong = (
+        highconf_wrong
+        & interior
+    )
+
+    unsafe = (
+        valid
+        & (
+            gt
+            == non_nav_id
+        )
+        & (
+            pred
+            == nav_id
+        )
+    )
+
+    blocked = (
+        valid
+        & (
+            gt
+            == nav_id
+        )
+        & (
+            pred
+            == non_nav_id
+        )
+    )
+
+    hc_unsafe = (
+        unsafe
+        & highconf
+    )
+
+    hc_unsafe_interior = (
+        hc_unsafe
+        & interior
+    )
+
+    hc_blocked = (
+        blocked
+        & highconf
+    )
+
+    hc_blocked_interior = (
+        hc_blocked
+        & interior
+    )
+
+    def frac(
+        mask: np.ndarray,
+        denom: Optional[
+            int
+        ] = None,
+    ) -> float:
+        d = (
+            valid_n
+            if denom is None
+            else int(
+                denom
+            )
+        )
+
+        return float(
+            int(
+                mask.sum()
+            )
+            / max(
+                1,
+                d,
+            )
+        )
+
+    gt_non_nav_n = int(
+        (
+            valid
+            & (
+                gt
+                == non_nav_id
+            )
+        ).sum()
+    )
+
+    gt_nav_n = int(
+        (
+            valid
+            & (
+                gt
+                == nav_id
+            )
+        ).sum()
+    )
+
+    correct_n = int(
+        correct.sum()
+    )
+
+    wrong_n = int(
+        wrong.sum()
+    )
+
+    mean_conf_correct = (
+        float(
+            top1[
+                correct
+            ].mean()
+        )
+        if correct_n
+        else 0.0
+    )
+
+    mean_conf_wrong = (
+        float(
+            top1[
+                wrong
+            ].mean()
+        )
+        if wrong_n
+        else 0.0
+    )
+
+    return {
+        "valid_pixels": valid_n,
+
+        "wrong_pixels": wrong_n,
+        "semantic_disagree_frac": frac(
+            wrong
+        ),
+
+        "boundary_pixels": int(
+            boundary.sum()
+        ),
+        "boundary_frac": frac(
+            boundary
+        ),
+
+        "interior_pixels": int(
+            interior.sum()
+        ),
+        "interior_wrong_pixels": int(
+            interior_wrong.sum()
+        ),
+        "interior_disagree_frac": frac(
+            interior_wrong
+        ),
+
+        "highconf_wrong_pixels": int(
+            highconf_wrong.sum()
+        ),
+        "highconf_disagree_frac": frac(
+            highconf_wrong
+        ),
+
+        "highconf_interior_wrong_pixels": int(
+            hc_interior_wrong.sum()
+        ),
+        "highconf_interior_disagree_frac": frac(
+            hc_interior_wrong
+        ),
+
+        "uncertain_pixels": int(
+            uncertain.sum()
+        ),
+        "uncertain_frac": frac(
+            uncertain
+        ),
+
+        "mean_conf_correct": mean_conf_correct,
+        "mean_conf_wrong": mean_conf_wrong,
+
+        "unsafe_nav_pixels": int(
+            unsafe.sum()
+        ),
+        "unsafe_nav_frac_all": frac(
+            unsafe
+        ),
+        "unsafe_nav_frac_gt_nonnav": frac(
+            unsafe,
+            gt_non_nav_n,
+        ),
+
+        "highconf_unsafe_nav_pixels": int(
+            hc_unsafe.sum()
+        ),
+        "highconf_unsafe_nav_frac_all": frac(
+            hc_unsafe
+        ),
+        "highconf_unsafe_nav_frac_gt_nonnav": frac(
+            hc_unsafe,
+            gt_non_nav_n,
+        ),
+
+        "highconf_unsafe_nav_interior_pixels": int(
+            hc_unsafe_interior.sum()
+        ),
+        "highconf_unsafe_nav_interior_frac_all": frac(
+            hc_unsafe_interior
+        ),
+
+        "blocked_nav_pixels": int(
+            blocked.sum()
+        ),
+        "blocked_nav_frac_all": frac(
+            blocked
+        ),
+        "blocked_nav_frac_gt_nav": frac(
+            blocked,
+            gt_nav_n,
+        ),
+
+        "highconf_blocked_nav_pixels": int(
+            hc_blocked.sum()
+        ),
+        "highconf_blocked_nav_frac_all": frac(
+            hc_blocked
+        ),
+
+        "highconf_blocked_nav_interior_pixels": int(
+            hc_blocked_interior.sum()
+        ),
+        "highconf_blocked_nav_interior_frac_all": frac(
+            hc_blocked_interior
+        ),
+    }
+
+
+# =============================================================================
+# Status signals
+# =============================================================================
+
+def status_sample_signals(
+    probs: np.ndarray,
+    gt_label_id: Optional[int],
+) -> dict:
+    probs = np.asarray(
+        probs,
+        dtype=np.float64,
+    ).reshape(
+        -1
+    )
+
+    pred_id = int(
+        np.argmax(
+            probs
+        )
+    )
+
+    pred_conf = float(
+        probs[
+            pred_id
+        ]
+    )
+
+    if probs.size >= 2:
+        sorted_probs = np.sort(
+            probs
+        )
+        top2 = float(
+            sorted_probs[
+                -2
+            ]
+        )
+    else:
+        top2 = 0.0
+
+    margin = float(
+        pred_conf
+        - top2
+    )
+
+    if (
+        gt_label_id is None
+        or not (
+            0
+            <= int(
+                gt_label_id
+            )
+            < probs.size
+        )
+    ):
+        gt_prob = None
+        mismatch = None
+    else:
+        gt_prob = float(
+            probs[
+                int(
+                    gt_label_id
+                )
+            ]
+        )
+        mismatch = (
+            pred_id
+            != int(
+                gt_label_id
+            )
+        )
+
+    return {
+        "pred_status_id": pred_id,
+        "status_confidence": pred_conf,
+        "status_margin": margin,
+        "gt_status_probability": gt_prob,
+        "status_mismatch": (
+            int(
+                mismatch
+            )
+            if mismatch is not None
+            else None
+        ),
+    }
+
+
+# =============================================================================
+# Review score
+# =============================================================================
+
+def _sat(
+    value: float,
+    reference: float,
+) -> float:
+    return float(
+        min(
+            max(
+                float(
+                    value
+                )
+                / max(
+                    float(
+                        reference
+                    ),
+                    1e-12,
+                ),
+                0.0,
+            ),
+            1.0,
+        )
+    )
+
+
+def compute_review_score(
+    row: dict,
+) -> Tuple[
+    float,
+    dict,
+]:
+    """
+    Score heurístico para RANKING.
+
+    Filosofia:
+      - alta confiança pesa muito;
+      - conflito no interior pesa mais que borda;
+      - falso navegável tem peso operacional maior;
+      - status global é um eixo independente;
+      - composição grupo ajuda pouco, apenas como evidência adicional.
+
+    Não é probabilidade.
+    """
+    seg_disagree = _sat(
+        row.get(
+            "semantic_disagree_frac",
+            0.0,
+        ),
+        0.18,
+    )
+
+    hc_disagree = _sat(
+        row.get(
+            "highconf_disagree_frac",
+            0.0,
+        ),
+        0.035,
+    )
+
+    hc_interior = _sat(
+        row.get(
+            "highconf_interior_disagree_frac",
+            0.0,
+        ),
+        0.018,
+    )
+
+    hc_unsafe = _sat(
+        row.get(
+            "highconf_unsafe_nav_interior_frac_all",
+            0.0,
+        ),
+        0.006,
+    )
+
+    seg_score = (
+        0.15
+        * seg_disagree
+        + 0.28
+        * hc_disagree
+        + 0.32
+        * hc_interior
+        + 0.25
+        * hc_unsafe
+    )
+
+    status_mismatch = bool(
+        row.get(
+            "status_mismatch",
+            0,
+        )
+        == 1
+    )
+
+    status_conf = safe_float(
+        row.get(
+            "status_confidence"
+        ),
+        0.0,
+    )
+
+    gt_status_prob = row.get(
+        "gt_status_probability"
+    )
+
+    if status_mismatch:
+        status_score = min(
+            1.0,
+            (
+                0.40
+                + 0.60
+                * _sat(
+                    status_conf
+                    - 0.50,
+                    0.45,
+                )
+            ),
+        )
+
+        if (
+            gt_status_prob
+            is not None
+        ):
+            status_score = max(
+                status_score,
+                _sat(
+                    0.35
+                    - float(
+                        gt_status_prob
+                    ),
+                    0.35,
+                ),
+            )
+    else:
+        status_score = 0.0
+
+    group_score = (
+        1.0
+        if int(
+            row.get(
+                "group_mismatch",
+                0,
+            )
+        )
+        == 1
+        else 0.0
+    )
+
+    # União suave de evidências.
+    combined = (
+        1.0
+        - (
+            1.0
+            - 0.82
+            * seg_score
+        )
+        * (
+            1.0
+            - 0.68
+            * status_score
+        )
+        * (
+            1.0
+            - 0.12
+            * group_score
+        )
+    )
+
+    combined = float(
+        min(
+            max(
+                combined,
+                0.0,
+            ),
+            1.0,
+        )
+    )
+
+    parts = {
+        "seg_review_score": float(
+            seg_score
+        ),
+        "status_review_score": float(
+            status_score
+        ),
+        "group_review_score": float(
+            group_score
+        ),
+    }
+
+    return (
+        combined,
+        parts,
+    )
+
+
+def build_reasons(
+    row: dict,
+    *,
+    min_disagree_frac: float,
+    min_highconf_disagree_frac: float,
+    min_highconf_interior_frac: float,
+    min_highconf_unsafe_frac: float,
+    min_conflict_pixels: int,
+    status_confidence: float,
+    min_status_confidence: float,
+    min_group_disagree_frac: float,
+) -> List[
+    str
+]:
+    reasons: List[
+        str
+    ] = []
+
+    if safe_float(
+        row.get(
+            "semantic_disagree_frac"
+        )
+    ) >= float(
+        min_disagree_frac
+    ):
+        reasons.append(
+            "SEMANTIC_DISAGREE"
+        )
+
+    hc_pixels = int(
+        row.get(
+            "highconf_wrong_pixels",
+            0,
+        )
+    )
+
+    if (
+        safe_float(
+            row.get(
+                "highconf_disagree_frac"
+            )
+        )
+        >= float(
+            min_highconf_disagree_frac
+        )
+        and hc_pixels
+        >= int(
+            min_conflict_pixels
+        )
+    ):
+        reasons.append(
+            "HIGH_CONF_GT_CONFLICT"
+        )
+
+    hc_int_pixels = int(
+        row.get(
+            "highconf_interior_wrong_pixels",
+            0,
+        )
+    )
+
+    if (
+        safe_float(
+            row.get(
+                "highconf_interior_disagree_frac"
+            )
+        )
+        >= float(
+            min_highconf_interior_frac
+        )
+        and hc_int_pixels
+        >= int(
+            min_conflict_pixels
+        )
+    ):
+        reasons.append(
+            "HIGH_CONF_INTERIOR_CONFLICT"
+        )
+
+    unsafe_pixels = int(
+        row.get(
+            "highconf_unsafe_nav_interior_pixels",
+            0,
+        )
+    )
+
+    if (
+        safe_float(
+            row.get(
+                "highconf_unsafe_nav_interior_frac_all"
+            )
+        )
+        >= float(
+            min_highconf_unsafe_frac
+        )
+        and unsafe_pixels
+        >= max(
+            16,
+            int(
+                min_conflict_pixels
+                // 4
+            ),
+        )
+    ):
+        reasons.append(
+            "HIGH_CONF_UNSAFE_NAV"
+        )
+
+    if (
+        int(
+            row.get(
+                "status_mismatch",
+                0,
+            )
+            or 0
+        )
+        == 1
+    ):
+        if (
+            float(
+                status_confidence
+            )
+            >= float(
+                min_status_confidence
+            )
+        ):
+            reasons.append(
+                "HIGH_CONF_STATUS_MISMATCH"
+            )
+        else:
+            reasons.append(
+                "STATUS_MISMATCH"
+            )
+
+    if (
+        int(
+            row.get(
+                "group_mismatch",
+                0,
+            )
+            or 0
+        )
+        == 1
+        and safe_float(
+            row.get(
+                "semantic_disagree_frac"
+            )
+        )
+        >= float(
+            min_group_disagree_frac
+        )
+    ):
+        reasons.append(
+            "GROUP_COMPOSITION_MISMATCH"
+        )
+
+    return list(
+        dict.fromkeys(
+            reasons
+        )
+    )
+
+
+# =============================================================================
+# Split manifest
+# =============================================================================
+
+def load_split_manifest(
+    resolution_root: Path,
+) -> Dict[
+    Tuple[
+        str,
+        str,
+    ],
+    str,
+]:
+    path = (
+        resolution_root
+        / "split_manifest.csv"
+    )
+
+    if not path.is_file():
+        return {}
+
+    result = {}
+
+    with path.open(
+        "r",
+        newline="",
+        encoding="utf-8-sig",
+    ) as f:
+        reader = csv.DictReader(
+            f
+        )
+
+        for row in reader:
+            group = str(
+                row.get(
+                    "group",
+                    "",
+                )
+            )
+            base = str(
+                row.get(
+                    "base",
+                    "",
+                )
+            )
+            split = str(
+                row.get(
+                    "split",
+                    "",
+                )
+            )
+
+            if (
+                group
+                and base
+                and split
+            ):
+                result[
+                    (
+                        group,
+                        base,
+                    )
+                ] = split
+
+    return result
+
+
+# =============================================================================
+# Original-source resolution
+# =============================================================================
+
+LEGACY_BASE_SUFFIXES = (
+    "_rgb",
+    "_image",
+    "_img",
+    "_frame",
+    "_mask",
+    "_masks",
+    "_seg",
+    "_segment",
+    "_segmentacao",
+    "_label",
+    "_labels",
+)
+
+
+def canonical_base(
+    stem: str,
+) -> str:
+    out = str(
+        stem
+    )
+
+    changed = True
+
+    while changed:
+        changed = False
+        low = out.lower()
+
+        for suffix in LEGACY_BASE_SUFFIXES:
+            if (
+                low.endswith(
+                    suffix
+                )
+                and len(
+                    out
+                )
+                > len(
+                    suffix
+                )
+            ):
+                out = out[
+                    :-len(
+                        suffix
+                    )
+                ]
+                changed = True
+                break
+
+    return out
+
+
+def index_folder_canonical(
+    folder: Path,
+    exts: set[
+        str
+    ],
+) -> Dict[
+    str,
+    List[
+        Path
+    ],
+]:
+    out: Dict[
+        str,
+        List[
+            Path
+        ],
+    ] = defaultdict(
+        list
+    )
+
+    if not folder.is_dir():
+        return out
+
+    for p in folder.iterdir():
+        if (
+            p.is_file()
+            and p.suffix.lower()
+            in exts
+        ):
+            out[
+                canonical_base(
+                    p.stem
+                )
+            ].append(
+                p
+            )
+
+    return out
+
+
+def resolve_path_value(
+    raw_value,
+    *,
+    dataset_root: Path,
+    label_path: Path,
+) -> Optional[
+    Path
+]:
+    if raw_value is None:
+        return None
+
+    txt = str(
+        raw_value
+    ).strip()
+
+    if not txt:
+        return None
+
+    raw = Path(
+        txt
+    )
+
+    candidates = []
+
+    if raw.is_absolute():
+        candidates.append(
+            raw
+        )
+    else:
+        candidates.extend([
+            dataset_root
+            / raw,
+            label_path.parent
+            / raw,
+            label_path.parent.parent.parent.parent
+            / raw,
+        ])
+
+    for p in candidates:
+        try:
+            rp = p.resolve()
+        except Exception:
+            rp = p
+
+        if rp.is_file():
+            return rp
+
+    return None
+
+
+def resolve_original_source(
+    *,
+    sample,
+    dataset_root: Path,
+) -> dict:
+    """
+    Prioridade:
+      1) provenance source_image/source_mask/source_label do JSON normalizado;
+      2) busca em dataset/original/group/<grupo> usando base canônica.
+
+    Não escolhe colisão silenciosamente.
+    """
+    normalized_label = (
+        Path(
+            sample.label
+        )
+        if sample.label is not None
+        else None
+    )
+
+    if (
+        normalized_label
+        is not None
+        and normalized_label.is_file()
+    ):
+        data = load_json(
+            normalized_label
+        )
+
+        source_image = resolve_path_value(
+            data.get(
+                "source_image"
+            ),
+            dataset_root=dataset_root,
+            label_path=normalized_label,
+        )
+
+        source_mask = resolve_path_value(
+            data.get(
+                "source_mask"
+            ),
+            dataset_root=dataset_root,
+            label_path=normalized_label,
+        )
+
+        source_label = resolve_path_value(
+            data.get(
+                "source_label"
+            ),
+            dataset_root=dataset_root,
+            label_path=normalized_label,
+        )
+
+        if (
+            source_image is not None
+            and source_mask is not None
+        ):
+            return {
+                "mode": "normalized_provenance",
+                "image": source_image,
+                "mask": source_mask,
+                "label": (
+                    source_label
+                    if source_label is not None
+                    else normalized_label
+                ),
+            }
+
+    # Fallback deterministic to original/group.
+    group = str(
+        sample.group
+    )
+
+    group_root = (
+        dataset_root
+        / "original"
+        / "group"
+        / group
+    )
+
+    images = index_folder_canonical(
+        group_root
+        / "images",
+        IMAGE_EXTS,
+    )
+
+    masks = index_folder_canonical(
+        group_root
+        / "masks",
+        IMAGE_EXTS,
+    )
+
+    labels = index_folder_canonical(
+        group_root
+        / "labels",
+        {
+            ".json"
+        },
+    )
+
+    base = canonical_base(
+        sample.base
+    )
+
+    im = images.get(
+        base,
+        [],
+    )
+    mk = masks.get(
+        base,
+        [],
+    )
+    lb = labels.get(
+        base,
+        [],
+    )
+
+    if (
+        len(
+            im
+        )
+        != 1
+        or len(
+            mk
+        )
+        != 1
+    ):
+        raise RuntimeError(
+            f"Fonte original ambígua/ausente para "
+            f"{sample.group}/{sample.base}: "
+            f"images={len(im)} masks={len(mk)} labels={len(lb)}"
+        )
+
+    label = (
+        lb[
+            0
+        ]
+        if len(
+            lb
+        )
+        == 1
+        else normalized_label
+    )
+
+    return {
+        "mode": "original_group_canonical",
+        "image": im[
+            0
+        ],
+        "mask": mk[
+            0
+        ],
+        "label": label,
+    }
+
+
+# =============================================================================
+# Physical review tree
+# =============================================================================
+
+def clear_generated_review(
+    review_group_root: Path,
+) -> None:
+    """
+    Limpa artefatos regeneráveis e PRESERVA final_masks/.
+    """
+    if not review_group_root.exists():
+        review_group_root.mkdir(
+            parents=True,
+            exist_ok=True,
+        )
+        return
+
+    for group_dir in review_group_root.iterdir():
+        if not group_dir.is_dir():
+            continue
+
+        for name in (
+            "original_images",
+            "original_masks",
+            "original_labels",
+            "predictions",
+            "panels",
+        ):
+            p = (
+                group_dir
+                / name
+            )
+
+            if p.exists():
+                shutil.rmtree(
+                    p
+                )
+
+        order = (
+            group_dir
+            / "review_order.csv"
+        )
+
+        if order.exists():
+            order.unlink()
+
+
+def write_png(
+    path: Path,
+    image: np.ndarray,
+) -> None:
+    path.parent.mkdir(
+        parents=True,
+        exist_ok=True,
+    )
+
+    ok = cv2.imwrite(
+        str(
+            path
+        ),
+        image,
+        [
+            cv2.IMWRITE_PNG_COMPRESSION,
+            2,
+        ],
+    )
+
+    if not ok:
+        raise RuntimeError(
+            f"Falha ao salvar PNG: {path}"
+        )
+
+
+def ensure_bgr3(
+    image: np.ndarray,
+) -> np.ndarray:
+    if image.ndim == 2:
+        return cv2.cvtColor(
+            image,
+            cv2.COLOR_GRAY2BGR,
+        )
+
+    if (
+        image.ndim
+        == 3
+        and image.shape[
+            2
+        ]
+        == 4
+    ):
+        return cv2.cvtColor(
+            image,
+            cv2.COLOR_BGRA2BGR,
+        )
+
+    if (
+        image.ndim
+        == 3
+        and image.shape[
+            2
+        ]
+        >= 3
+    ):
+        return image[
+            :,
+            :,
+            :3,
+        ]
+
+    raise RuntimeError(
+        f"Imagem inválida shape={image.shape}"
+    )
+
+
+def error_map_bgr(
+    *,
+    gt: np.ndarray,
+    pred: np.ndarray,
+    nav_id: int,
+    non_nav_id: int,
+    ignore_id: int,
+) -> np.ndarray:
+    if pred.shape != gt.shape:
+        pred = cv2.resize(
+            pred.astype(
+                np.uint8
+            ),
+            (
+                gt.shape[
+                    1
+                ],
+                gt.shape[
+                    0
+                ],
+            ),
+            interpolation=cv2.INTER_NEAREST,
+        )
+
+    out = np.zeros(
+        (
+            gt.shape[
+                0
+            ],
+            gt.shape[
+                1
+            ],
+            3,
+        ),
+        dtype=np.uint8,
+    )
+
+    out[
+        :
+    ] = (
+        60,
+        60,
+        60,
+    )
+
+    valid = (
+        gt
+        != int(
+            ignore_id
+        )
+    )
+
+    correct = (
+        valid
+        & (
+            gt
+            == pred
+        )
+    )
+
+    correct_nav = (
+        correct
+        & (
+            gt
+            == nav_id
+        )
+    )
+
+    unsafe = (
+        valid
+        & (
+            gt
+            == non_nav_id
+        )
+        & (
+            pred
+            == nav_id
+        )
+    )
+
+    blocked = (
+        valid
+        & (
+            gt
+            == nav_id
+        )
+        & (
+            pred
+            == non_nav_id
+        )
+    )
+
+    out[
+        correct
+    ] = (
+        90,
+        90,
+        90,
+    )
+    out[
+        correct_nav
+    ] = (
+        60,
+        155,
+        60,
+    )
+    out[
+        blocked
+    ] = (
+        0,
+        210,
+        255,
+    )
+    out[
+        unsafe
+    ] = (
+        0,
+        0,
+        255,
+    )
+    out[
+        ~valid
+    ] = (
+        180,
+        0,
+        180,
+    )
+
+    return out
+
+
+def add_title(
+    image: np.ndarray,
+    title: str,
+) -> np.ndarray:
+    out = image.copy()
+
+    cv2.rectangle(
+        out,
+        (
+            0,
+            0,
+        ),
+        (
+            out.shape[
+                1
+            ],
+            38,
+        ),
+        (
+            18,
+            18,
+            18,
+        ),
+        -1,
+    )
+
+    cv2.putText(
+        out,
+        title,
+        (
+            10,
+            26,
+        ),
+        cv2.FONT_HERSHEY_SIMPLEX,
+        0.62,
+        (
+            245,
+            245,
+            245,
+        ),
+        1,
+        cv2.LINE_AA,
+    )
+
+    return out
+
+
+def letterbox(
+    image: np.ndarray,
+    target_w: int,
+    target_h: int,
+) -> np.ndarray:
+    h, w = image.shape[
+        :2
+    ]
+
+    scale = min(
+        target_w
+        / max(
+            1,
+            w,
+        ),
+        target_h
+        / max(
+            1,
+            h,
+        ),
+    )
+
+    nw = max(
+        1,
+        int(
+            round(
+                w
+                * scale
+            )
+        ),
+    )
+    nh = max(
+        1,
+        int(
+            round(
+                h
+                * scale
+            )
+        ),
+    )
+
+    resized = cv2.resize(
+        image,
+        (
+            nw,
+            nh,
+        ),
+        interpolation=(
+            cv2.INTER_AREA
+            if scale
+            < 1
+            else cv2.INTER_LINEAR
+        ),
+    )
+
+    canvas = np.zeros(
+        (
+            target_h,
+            target_w,
+            3,
+        ),
+        dtype=np.uint8,
+    )
+
+    canvas[
+        :
+    ] = (
+        24,
+        24,
+        24,
+    )
+
+    x0 = (
+        target_w
+        - nw
+    ) // 2
+    y0 = (
+        target_h
+        - nh
+    ) // 2
+
+    canvas[
+        y0:
+        y0
+        + nh,
+        x0:
+        x0
+        + nw,
+    ] = resized
+
+    return canvas
+
+
+def make_review_panel(
+    *,
+    original_bgr: np.ndarray,
+    gt_color_bgr: np.ndarray,
+    pred_color_bgr: np.ndarray,
+    error_bgr: np.ndarray,
+    row: dict,
+    label_name_by_id: Dict[
+        int,
+        str,
+    ],
+    width: int = 1800,
+) -> np.ndarray:
+    h, w = original_bgr.shape[
+        :2
+    ]
+
+    gt_overlay = cv2.addWeighted(
+        original_bgr,
+        0.58,
+        gt_color_bgr,
+        0.42,
+        0.0,
+    )
+
+    pred_overlay = cv2.addWeighted(
+        original_bgr,
+        0.58,
+        pred_color_bgr,
+        0.42,
+        0.0,
+    )
+
+    cols = 4
+    tile_w = max(
+        280,
+        int(
+            width
+            / cols
+        ),
+    )
+    tile_h = max(
+        230,
+        int(
+            tile_w
+            * h
+            / max(
+                1,
+                w,
+            )
+        ),
+    )
+
+    tiles = [
+        (
+            "ORIGINAL",
+            original_bgr,
+        ),
+        (
+            "GT",
+            gt_overlay,
+        ),
+        (
+            "PRED",
+            pred_overlay,
+        ),
+        (
+            "ERRO | vermelho = falso navegável",
+            error_bgr,
+        ),
+    ]
+
+    top = np.concatenate(
+        [
+            add_title(
+                letterbox(
+                    img,
+                    tile_w,
+                    tile_h,
+                ),
+                title,
+            )
+            for title, img in tiles
+        ],
+        axis=1,
+    )
+
+    panel_h = 220
+
+    info = np.zeros(
+        (
+            panel_h,
+            top.shape[
+                1
+            ],
+            3,
+        ),
+        dtype=np.uint8,
+    )
+
+    info[
+        :
+    ] = (
+        28,
+        28,
+        28,
+    )
+
+    gt_status_id = row.get(
+        "gt_status_id"
+    )
+
+    pred_status_id = row.get(
+        "pred_status_id"
+    )
+
+    gt_status = (
+        label_name_by_id.get(
+            int(
+                gt_status_id
+            ),
+            str(
+                gt_status_id
+            ),
+        )
+        if gt_status_id
+        is not None
+        else "-"
+    )
+
+    pred_status = (
+        label_name_by_id.get(
+            int(
+                pred_status_id
+            ),
+            str(
+                pred_status_id
+            ),
+        )
+        if pred_status_id
+        is not None
+        else "-"
+    )
+
+    lines = [
+        (
+            f"{row['group']}/{row['base']} | "
+            f"split={row.get('split_origin','?')} | "
+            f"score={row['suspicion_pct']:.1f}/100"
+        ),
+        (
+            f"SEG: navIoU={row['nav_iou']:.4f} "
+            f"nonNavIoU={row['non_nav_iou']:.4f} "
+            f"disagree={100.0*row['semantic_disagree_frac']:.2f}% "
+            f"HC-interior={100.0*row['highconf_interior_disagree_frac']:.3f}%"
+        ),
+        (
+            f"RISK: unsafe={100.0*row['unsafe_nav_frac_gt_nonnav']:.3f}% "
+            f"HC-unsafe-interior={100.0*row['highconf_unsafe_nav_interior_frac_all']:.3f}%"
+        ),
+        (
+            f"STATUS: GT={gt_status} | PRED={pred_status} "
+            f"conf={row['status_confidence']:.3f} "
+            f"GTprob={safe_float(row.get('gt_status_probability'),0.0):.3f}"
+        ),
+        (
+            f"GROUP: GT={row['group']} | PRED={row['pred_group']} "
+            f"| reasons={row.get('review_reasons','')}"
+        ),
+    ]
+
+    y = 34
+
+    for line in lines:
+        cv2.putText(
+            info,
+            line,
+            (
+                14,
+                y,
+            ),
+            cv2.FONT_HERSHEY_SIMPLEX,
+            0.62,
+            (
+                240,
+                240,
+                240,
+            ),
+            1,
+            cv2.LINE_AA,
+        )
+
+        y += 37
+
+    return np.concatenate(
+        [
+            top,
+            info,
+        ],
+        axis=0,
+    )
+
+
+# =============================================================================
+# Inference
+# =============================================================================
+
+@torch.inference_mode()
+def infer_audit(
+    *,
+    base_module,
+    base_model,
+    status_head,
+    contract,
+    image_bgr: np.ndarray,
+    device: torch.device,
+    use_amp: bool,
+) -> dict:
+    x, geometry = base_module.prepare_input(
+        image_bgr,
+        contract,
+        device,
+    )
+
+    if (
+        device.type
+        == "cuda"
+    ):
+        torch.cuda.synchronize()
+
+    t0 = time.perf_counter()
+
+    amp_enabled = (
+        bool(
+            use_amp
+        )
+        and device.type
+        == "cuda"
+    )
+
+    with torch.autocast(
+        device_type=(
+            "cuda"
+            if device.type
+            == "cuda"
+            else "cpu"
+        ),
+        dtype=torch.float16,
+        enabled=amp_enabled,
+    ):
+        out = base_model(
+            pixel_values=x,
+            output_hidden_states=True,
+            return_dict=True,
+        )
+
+        seg_logits_native = out.logits
+        feat = out.hidden_states[
+            -1
+        ]
+
+        status_logits = status_head(
+            feat,
+            seg_logits_native,
+        )
+
+        W, H = contract.resolution_wh
+
+        seg_logits_full = F.interpolate(
+            seg_logits_native,
+            size=(
+                H,
+                W,
+            ),
+            mode="bilinear",
+            align_corners=False,
+        )
+
+        seg_probs = torch.softmax(
+            seg_logits_full.float(),
+            dim=1,
+        )
+
+        pred = torch.argmax(
+            seg_probs,
+            dim=1,
+        )
+
+        status_probs = torch.softmax(
+            status_logits.float(),
+            dim=1,
+        )
+
+    if (
+        device.type
+        == "cuda"
+    ):
+        torch.cuda.synchronize()
+
+    infer_ms = (
+        time.perf_counter()
+        - t0
+    ) * 1000.0
+
+    return {
+        "pred": (
+            pred[
+                0
+            ]
+            .detach()
+            .cpu()
+            .numpy()
+            .astype(
+                np.uint8
+            )
+        ),
+        "seg_probs": (
+            seg_probs[
+                0
+            ]
+            .detach()
+            .cpu()
+            .numpy()
+            .astype(
+                np.float32
+            )
+        ),
+        "status_probs": (
+            status_probs[
+                0
+            ]
+            .detach()
+            .cpu()
+            .numpy()
+            .astype(
+                np.float32
+            )
+        ),
+        "geometry": geometry,
+        "infer_ms": float(
+            infer_ms
+        ),
+    }
+
+
+# =============================================================================
+# Candidate export
+# =============================================================================
+
+def save_review_item(
+    *,
+    row: dict,
+    sample,
+    pred_ids: np.ndarray,
+    gt_ids: np.ndarray,
+    source: dict,
+    review_group_root: Path,
+    base_module,
+    classes,
+    contract,
+    save_panels: bool,
+) -> dict:
+    group = str(
+        sample.group
+    )
+
+    group_root = (
+        review_group_root
+        / group
+    )
+
+    dirs = {
+        "images": group_root
+        / "original_images",
+        "masks": group_root
+        / "original_masks",
+        "labels": group_root
+        / "original_labels",
+        "predictions": group_root
+        / "predictions",
+        "panels": group_root
+        / "panels",
+        "final_masks": group_root
+        / "final_masks",
+    }
+
+    for name, folder in dirs.items():
+        if (
+            name
+            == "panels"
+            and not save_panels
+        ):
+            continue
+
+        folder.mkdir(
+            parents=True,
+            exist_ok=True,
+        )
+
+    src_image = Path(
+        source[
+            "image"
+        ]
+    )
+    src_mask = Path(
+        source[
+            "mask"
+        ]
+    )
+    src_label = (
+        Path(
+            source[
+                "label"
+            ]
+        )
+        if source.get(
+            "label"
+        )
+        is not None
+        else None
+    )
+
+    original_bgr = cv2.imread(
+        str(
+            src_image
+        ),
+        cv2.IMREAD_COLOR,
+    )
+
+    if original_bgr is None:
+        raise RuntimeError(
+            f"Falha imagem original: {src_image}"
+        )
+
+    original_mask_native = cv2.imread(
+        str(
+            src_mask
+        ),
+        cv2.IMREAD_UNCHANGED,
+    )
+
+    if original_mask_native is None:
+        raise RuntimeError(
+            f"Falha mask original: {src_mask}"
+        )
+
+    original_mask_bgr = ensure_bgr3(
+        original_mask_native
+    )
+
+    oh, ow = original_bgr.shape[
+        :2
+    ]
+
+    mh, mw = original_mask_bgr.shape[
+        :2
+    ]
+
+    if (
+        oh,
+        ow,
+    ) != (
+        mh,
+        mw,
+    ):
+        raise RuntimeError(
+            f"Imagem/mask original geometria diferente: "
+            f"{src_image}={ow}x{oh}, {src_mask}={mw}x{mh}"
+        )
+
+    # O normalize do corredor só faz resize mantendo aspect ratio.
+    pred_original_ids = cv2.resize(
+        pred_ids.astype(
+            np.uint8
+        ),
+        (
+            ow,
+            oh,
+        ),
+        interpolation=cv2.INTER_NEAREST,
+    )
+
+    gt_original_ids = cv2.resize(
+        gt_ids.astype(
+            np.uint8
+        ),
+        (
+            ow,
+            oh,
+        ),
+        interpolation=cv2.INTER_NEAREST,
+    )
+
+    pred_color_bgr = base_module.colorize_ids(
+        pred_original_ids,
+        classes,
+    )
+
+    gt_color_bgr = base_module.colorize_ids(
+        gt_original_ids,
+        classes,
+    )
+
+    error_bgr = error_map_bgr(
+        gt=gt_original_ids,
+        pred=pred_original_ids,
+        nav_id=contract.nav_class_id,
+        non_nav_id=contract.non_nav_class_id,
+        ignore_id=IGNORE_INDEX,
+    )
+
+    image_out = (
+        dirs[
+            "images"
+        ]
+        / f"{sample.base}{src_image.suffix.lower()}"
+    )
+
+    mask_out = (
+        dirs[
+            "masks"
+        ]
+        / f"{sample.base}{src_mask.suffix.lower()}"
+    )
+
+    pred_out = (
+        dirs[
+            "predictions"
+        ]
+        / f"{sample.base}.png"
+    )
+
+    shutil.copy2(
+        src_image,
+        image_out,
+    )
+
+    shutil.copy2(
+        src_mask,
+        mask_out,
+    )
+
+    label_out = None
+
+    if (
+        src_label is not None
+        and src_label.is_file()
+    ):
+        label_out = (
+            dirs[
+                "labels"
+            ]
+            / f"{sample.base}{src_label.suffix.lower()}"
+        )
+
+        shutil.copy2(
+            src_label,
+            label_out,
+        )
+
+    write_png(
+        pred_out,
+        pred_color_bgr,
+    )
+
+    panel_out = None
+
+    if save_panels:
+        panel = make_review_panel(
+            original_bgr=original_bgr,
+            gt_color_bgr=gt_color_bgr,
+            pred_color_bgr=pred_color_bgr,
+            error_bgr=error_bgr,
+            row=row,
+            label_name_by_id=contract.label_name_by_id,
+        )
+
+        panel_out = (
+            dirs[
+                "panels"
+            ]
+            / f"{sample.base}.jpg"
+        )
+
+        ok = cv2.imwrite(
+            str(
+                panel_out
+            ),
+            panel,
+            [
+                cv2.IMWRITE_JPEG_QUALITY,
+                92,
+            ],
+        )
+
+        if not ok:
+            raise RuntimeError(
+                f"Falha panel: {panel_out}"
+            )
+
+    return {
+        "group": group,
+        "base": str(
+            sample.base
+        ),
+        "source_mode": str(
+            source[
+                "mode"
+            ]
+        ),
+        "source_image": str(
+            src_image
+        ),
+        "source_mask": str(
+            src_mask
+        ),
+        "source_label": (
+            str(
+                src_label
+            )
+            if src_label
+            is not None
+            else None
+        ),
+        "review_image": str(
+            image_out
+        ),
+        "review_mask": str(
+            mask_out
+        ),
+        "review_label": (
+            str(
+                label_out
+            )
+            if label_out
+            is not None
+            else None
+        ),
+        "review_prediction": str(
+            pred_out
+        ),
+        "review_panel": (
+            str(
+                panel_out
+            )
+            if panel_out
+            is not None
+            else None
+        ),
+        "width": int(
+            ow
+        ),
+        "height": int(
+            oh
+        ),
+        "prediction_geometry": (
+            "normalized_model_mask resized with INTER_NEAREST "
+            "to original annotation geometry"
+        ),
+    }
+
+
+# =============================================================================
+# Summary helpers
+# =============================================================================
+
+def summarize_rows(
+    rows: Sequence[
+        dict
+    ],
+    selected_keys: set[
+        Tuple[
+            str,
+            str,
+        ]
+    ],
+) -> dict:
+    if not rows:
+        return {}
+
+    scores = np.asarray(
+        [
+            safe_float(
+                r.get(
+                    "review_score"
+                )
+            )
+            for r in rows
+        ],
+        dtype=np.float64,
+    )
+
+    disag = np.asarray(
+        [
+            safe_float(
+                r.get(
+                    "semantic_disagree_frac"
+                )
+            )
+            for r in rows
+        ],
+        dtype=np.float64,
+    )
+
+    hc = np.asarray(
+        [
+            safe_float(
+                r.get(
+                    "highconf_disagree_frac"
+                )
+            )
+            for r in rows
+        ],
+        dtype=np.float64,
+    )
+
+    hc_int = np.asarray(
+        [
+            safe_float(
+                r.get(
+                    "highconf_interior_disagree_frac"
+                )
+            )
+            for r in rows
+        ],
+        dtype=np.float64,
+    )
+
+    unsafe = np.asarray(
+        [
+            safe_float(
+                r.get(
+                    "highconf_unsafe_nav_interior_frac_all"
+                )
+            )
+            for r in rows
+        ],
+        dtype=np.float64,
+    )
+
+    selected = sum(
+        (
+            str(
+                r[
+                    "group"
+                ]
+            ),
+            str(
+                r[
+                    "base"
+                ]
+            ),
+        )
+        in selected_keys
+        for r in rows
+    )
+
+    return {
+        "samples": len(
+            rows
+        ),
+        "selected": int(
+            selected
+        ),
+        "selected_pct": float(
+            100.0
+            * selected
+            / max(
+                1,
+                len(
+                    rows
+                ),
+            )
+        ),
+        "review_score_mean": float(
+            scores.mean()
+        ),
+        "review_score_p95": float(
+            np.percentile(
+                scores,
+                95,
+            )
+        ),
+        "review_score_max": float(
+            scores.max()
+        ),
+        "semantic_disagree_mean": float(
+            disag.mean()
+        ),
+        "semantic_disagree_p95": float(
+            np.percentile(
+                disag,
+                95,
+            )
+        ),
+        "highconf_disagree_mean": float(
+            hc.mean()
+        ),
+        "highconf_interior_mean": float(
+            hc_int.mean()
+        ),
+        "highconf_unsafe_interior_mean": float(
+            unsafe.mean()
+        ),
+        "status_mismatch_count": int(
+            sum(
+                int(
+                    r.get(
+                        "status_mismatch",
+                        0,
+                    )
+                    or 0
+                )
+                for r in rows
+            )
+        ),
+    }
+
+
+# =============================================================================
+# Main
+# =============================================================================
+
+def main():
+    ap = argparse.ArgumentParser(
+        description=(
+            "Auditoria modelo x GT do dataset frontal de corredor."
+        )
+    )
+
+    ap.add_argument(
+        "--config",
+        default="config.json",
+    )
+
+    ap.add_argument(
+        "--test_script",
+        default=None,
+        help=(
+            "Default: test_corridor_agri_v2.py ao lado deste script."
+        ),
+    )
+
+    ap.add_argument(
+        "--ckpt",
+        default=None,
+        help="Checkpoint .pt explícito.",
+    )
+
+    ap.add_argument(
+        "--checkpoint",
+        default="operational",
+        choices=[
+            "operational",
+            "nav",
+            "status",
+            "last",
+        ],
+    )
+
+    ap.add_argument(
+        "--root",
+        default=None,
+        help=(
+            "Default: dataset/<W>x<H>/group."
+        ),
+    )
+
+    ap.add_argument(
+        "--device",
+        default="cuda",
+        choices=[
+            "cuda",
+            "cpu",
+        ],
+    )
+
+    ap.add_argument(
+        "--no_amp",
+        action="store_true",
+    )
+
+    ap.add_argument(
+        "--out_root",
+        default="dataset/revisao_corridor",
+    )
+
+    ap.add_argument(
+        "--run_name",
+        default=None,
+    )
+
+    ap.add_argument(
+        "--report_only",
+        action="store_true",
+    )
+
+    ap.add_argument(
+        "--save_panels",
+        action="store_true",
+    )
+
+    ap.add_argument(
+        "--clear_review",
+        action="store_true",
+        help=(
+            "Limpa artefatos regeneráveis, preservando final_masks."
+        ),
+    )
+
+    ap.add_argument(
+        "--clear_review_all",
+        action="store_true",
+        help=(
+            "PERIGOSO: apaga toda a árvore física de revisão."
+        ),
+    )
+
+    ap.add_argument(
+        "--export_mode",
+        default="flagged",
+        choices=[
+            "flagged",
+            "score",
+            "all",
+        ],
+    )
+
+    ap.add_argument(
+        "--min_suspicion_pct",
+        type=float,
+        default=65.0,
+        help=(
+            "Usado em --export_mode score. Ranking heurístico 0..100."
+        ),
+    )
+
+    ap.add_argument(
+        "--top_k_per_stratum",
+        type=int,
+        default=0,
+        help=(
+            "Se >0, seleciona os K scores maiores de cada grupo x GT status."
+        ),
+    )
+
+    ap.add_argument(
+        "--max_review_per_stratum",
+        type=int,
+        default=0,
+        help="Cap final por grupo x status. 0=sem limite.",
+    )
+
+    ap.add_argument(
+        "--max_samples",
+        type=int,
+        default=0,
+        help="Debug. 0=todas.",
+    )
+
+    ap.add_argument(
+        "--gc_every",
+        type=int,
+        default=50,
+    )
+
+    # Semantic mining.
+    ap.add_argument(
+        "--high_confidence",
+        type=float,
+        default=0.90,
+    )
+
+    ap.add_argument(
+        "--uncertainty_margin",
+        type=float,
+        default=0.10,
+    )
+
+    ap.add_argument(
+        "--boundary_radius",
+        type=int,
+        default=4,
+        help="Raio px da banda de fronteira do GT.",
+    )
+
+    ap.add_argument(
+        "--composition_min_fraction",
+        type=float,
+        default=0.005,
+        help=(
+            "0.005=0.5%%. Menor que isso é tratado como composição pura."
+        ),
+    )
+
+    # Technical flags.
+    ap.add_argument(
+        "--min_disagree_frac",
+        type=float,
+        default=0.05,
+    )
+
+    ap.add_argument(
+        "--min_highconf_disagree_frac",
+        type=float,
+        default=0.002,
+    )
+
+    ap.add_argument(
+        "--min_highconf_interior_frac",
+        type=float,
+        default=0.001,
+    )
+
+    ap.add_argument(
+        "--min_highconf_unsafe_frac",
+        type=float,
+        default=0.0005,
+    )
+
+    ap.add_argument(
+        "--min_conflict_pixels",
+        type=int,
+        default=150,
+    )
+
+    ap.add_argument(
+        "--min_status_confidence",
+        type=float,
+        default=0.80,
+    )
+
+    ap.add_argument(
+        "--min_group_disagree_frac",
+        type=float,
+        default=0.05,
+    )
+
+    args = ap.parse_args()
+
+    if (
+        args.clear_review
+        and args.clear_review_all
+    ):
+        ap.error(
+            "Use apenas --clear_review OU --clear_review_all."
+        )
+
+    if not (
+        0.0
+        <= float(
+            args.min_suspicion_pct
+        )
+        <= 100.0
+    ):
+        ap.error(
+            "--min_suspicion_pct deve ficar entre 0 e 100."
+        )
+
+    if not (
+        0.0
+        < float(
+            args.high_confidence
+        )
+        <= 1.0
+    ):
+        ap.error(
+            "--high_confidence inválido."
+        )
+
+    # -------------------------------------------------------------------------
+    # Shared tester
+    # -------------------------------------------------------------------------
+
+    here = Path(
+        __file__
+    ).resolve()
+
+    test_script = (
+        Path(
+            args.test_script
+        ).resolve()
+        if args.test_script
+        else here.with_name(
+            "test_corridor_agri_v2.py"
+        )
+    )
+
+    base = load_test_module(
+        test_script
+    )
+
+    config_path = Path(
+        args.config
+    ).resolve()
+
+    if not config_path.is_file():
+        raise FileNotFoundError(
+            config_path
+        )
+
+    config = base.safe_json_load(
+        config_path
+    )
+
+    project_root = base.find_project_root(
+        config_path
+    )
+
+    dataset_root = (
+        project_root
+        / "dataset"
+    )
+
+    # -------------------------------------------------------------------------
+    # Checkpoint
+    # -------------------------------------------------------------------------
+
+    if args.ckpt:
+        checkpoint_path = Path(
+            args.ckpt
+        ).resolve()
+    else:
+        checkpoint_path = base.default_checkpoint_path(
+            project_root,
+            config,
+            args.checkpoint,
+        )
+
+    use_cuda = (
+        args.device
+        == "cuda"
+        and torch.cuda.is_available()
+    )
+
+    if (
+        args.device
+        == "cuda"
+        and not use_cuda
+    ):
+        print(
+            "[WARN] CUDA indisponível; usando CPU."
+        )
+
+    device = torch.device(
+        "cuda"
+        if use_cuda
+        else "cpu"
+    )
+
+    (
+        base_model,
+        status_head,
+        contract,
+        checkpoint_raw,
+    ) = base.build_runtime_from_checkpoint(
+        checkpoint_path,
+        config,
+        device,
+    )
+
+    W, H = contract.resolution_wh
+
+    resolution_root = (
+        dataset_root
+        / f"{W}x{H}"
+    )
+
+    root = (
+        Path(
+            args.root
+        ).resolve()
+        if args.root
+        else (
+            resolution_root
+            / "group"
+        )
+    )
+
+    if not root.is_dir():
+        raise FileNotFoundError(
+            f"Dataset auditável não encontrado: {root}"
+        )
+
+    classes = base.load_labelmap(
+        dataset_root
+        / "labelmap.txt"
+    )
+
+    samples, source_kind = base.discover_samples(
+        root
+    )
+
+    samples = [
+        s
+        for s in samples
+        if (
+            s.mask
+            is not None
+            and s.label
+            is not None
+        )
+    ]
+
+    if args.max_samples > 0:
+        samples = samples[
+            :int(
+                args.max_samples
+            )
+        ]
+
+    if not samples:
+        raise RuntimeError(
+            "Nenhuma tripleta image+mask+label encontrada."
+        )
+
+    split_origin_map = load_split_manifest(
+        resolution_root
+    )
+
+    # -------------------------------------------------------------------------
+    # Output
+    # -------------------------------------------------------------------------
+
+    out_root = Path(
+        args.out_root
+    )
+
+    if not out_root.is_absolute():
+        out_root = (
+            project_root
+            / out_root
+        )
+
+    out_root = out_root.resolve()
+
+    review_group_root = (
+        out_root
+        / "group"
+    )
+
+    run_name = (
+        args.run_name
+        or checkpoint_path.stem
+    )
+
+    report_dir = (
+        out_root
+        / "reports"
+        / run_name
+    )
+
+    report_dir.mkdir(
+        parents=True,
+        exist_ok=True,
+    )
+
+    if not args.report_only:
+        if args.clear_review_all:
+            print(
+                "[WARN] Limpando revisão COMPLETA, inclusive final_masks."
+            )
+            clear_dir(
+                review_group_root
+            )
+        elif args.clear_review:
+            print(
+                "[INFO] Limpando artefatos regeneráveis; final_masks preservada."
+            )
+            clear_generated_review(
+                review_group_root
+            )
+
+    # -------------------------------------------------------------------------
+    # Print contract
+    # -------------------------------------------------------------------------
+
+    print("=" * 94)
+    print(
+        "AGROBOT | CORRIDOR DATASET REVIEW / MODEL MINING"
+    )
+    print("=" * 94)
+    print(
+        f"Reviewer           : {REVIEWER_VERSION}"
+    )
+    print(
+        f"Tester base        : {test_script}"
+    )
+    print(
+        f"Dataset root       : {root}"
+    )
+    print(
+        f"Source kind        : {source_kind}"
+    )
+    print(
+        f"Samples            : {len(samples)}"
+    )
+    print(
+        f"Checkpoint         : {checkpoint_path}"
+    )
+    print(
+        f"Epoch              : {contract.checkpoint_epoch}"
+    )
+    print(
+        f"Resolution         : {W}x{H}"
+    )
+    print(
+        f"Device             : {device}"
+    )
+    print(
+        f"AMP                : {not args.no_amp and device.type == 'cuda'}"
+    )
+    print(
+        f"High confidence    : {args.high_confidence:.3f}"
+    )
+    print(
+        f"Boundary radius    : {args.boundary_radius}px"
+    )
+    print(
+        f"Export mode        : {args.export_mode}"
+    )
+    print(
+        f"Report only        : {args.report_only}"
+    )
+    print("=" * 94)
+
+    # -------------------------------------------------------------------------
+    # Audit pass
+    # -------------------------------------------------------------------------
+
+    num_seg_classes = len(
+        contract.seg_id2label
+    )
+
+    num_status_classes = len(
+        contract.label_name_by_id
+    )
+
+    cm_seg_total = np.zeros(
+        (
+            num_seg_classes,
+            num_seg_classes,
+        ),
+        dtype=np.int64,
+    )
+
+    cm_status_total = np.zeros(
+        (
+            num_status_classes,
+            num_status_classes,
+        ),
+        dtype=np.int64,
+    )
+
+    rows: List[
+        dict
+    ] = []
+
+    infer_ms_sum = 0.0
+
+    t0_all = time.perf_counter()
+
+    for i, sample in enumerate(
+        samples,
+        1,
+    ):
+        image_bgr = cv2.imread(
+            str(
+                sample.image
+            ),
+            cv2.IMREAD_COLOR,
+        )
+
+        if image_bgr is None:
+            raise RuntimeError(
+                f"Imagem inválida: {sample.image}"
+            )
+
+        gt = base.load_gt_mask(
+            sample.mask,
+            classes,
+        )
+
+        if gt is None:
+            raise RuntimeError(
+                f"GT ausente: {sample.mask}"
+            )
+
+        (
+            gt_status_id,
+            gt_status_name,
+        ) = base.read_gt_label(
+            sample,
+            contract.label_name_by_id,
+        )
+
+        infer = infer_audit(
+            base_module=base,
+            base_model=base_model,
+            status_head=status_head,
+            contract=contract,
+            image_bgr=image_bgr,
+            device=device,
+            use_amp=not args.no_amp,
+        )
+
+        infer_ms_sum += float(
+            infer[
+                "infer_ms"
+            ]
+        )
+
+        # Dataset normalizado deveria já estar HxW.
+        # Mesmo assim deixa robusto.
+        pred = infer[
+            "pred"
+        ]
+
+        seg_probs = infer[
+            "seg_probs"
+        ]
+
+        if gt.shape != pred.shape:
+            gt_eval = cv2.resize(
+                gt.astype(
+                    np.uint8
+                ),
+                (
+                    pred.shape[
+                        1
+                    ],
+                    pred.shape[
+                        0
+                    ],
+                ),
+                interpolation=cv2.INTER_NEAREST,
+            )
+        else:
+            gt_eval = gt
+
+        cm_sample = confusion_matrix_np(
+            pred,
+            gt_eval,
+            num_seg_classes,
+            IGNORE_INDEX,
+        )
+
+        cm_seg_total += cm_sample
+
+        metric_sample = binary_metrics_from_cm(
+            cm_sample,
+            contract.nav_class_id,
+            contract.non_nav_class_id,
+        )
+
+        sem = semantic_sample_signals(
+            pred=pred,
+            probs=seg_probs,
+            gt=gt_eval,
+            nav_id=contract.nav_class_id,
+            non_nav_id=contract.non_nav_class_id,
+            ignore_id=IGNORE_INDEX,
+            high_confidence=float(
+                args.high_confidence
+            ),
+            uncertainty_margin=float(
+                args.uncertainty_margin
+            ),
+            boundary_radius=int(
+                args.boundary_radius
+            ),
+        )
+
+        status = status_sample_signals(
+            infer[
+                "status_probs"
+            ],
+            gt_status_id,
+        )
+
+        pred_status_id = int(
+            status[
+                "pred_status_id"
+            ]
+        )
+
+        if (
+            gt_status_id
+            is not None
+            and 0
+            <= int(
+                gt_status_id
+            )
+            < num_status_classes
+        ):
+            cm_status_total[
+                int(
+                    gt_status_id
+                ),
+                pred_status_id,
+            ] += 1
+
+        pred_group, pred_nav_frac = derive_composition_group(
+            pred,
+            contract.nav_class_id,
+            contract.non_nav_class_id,
+            float(
+                args.composition_min_fraction
+            ),
+        )
+
+        gt_group = str(
+            sample.group
+        )
+
+        group_mismatch = (
+            pred_group
+            != gt_group
+        )
+
+        split_origin = split_origin_map.get(
+            (
+                gt_group,
+                str(
+                    sample.base
+                ),
+            ),
+            (
+                "unknown"
+                if (
+                    "split"
+                    not in str(
+                        root
+                    ).lower()
+                )
+                else root.name
+            ),
+        )
+
+        row = {
+            "index": i - 1,
+            "group": gt_group,
+            "base": str(
+                sample.base
+            ),
+            "split_origin": split_origin,
+
+            "image_path": str(
+                sample.image
+            ),
+            "mask_path": str(
+                sample.mask
+            ),
+            "label_path": str(
+                sample.label
+            ),
+
+            "inference_ms": float(
+                infer[
+                    "infer_ms"
+                ]
+            ),
+
+            **metric_sample,
+            **sem,
+
+            "gt_status_id": (
+                int(
+                    gt_status_id
+                )
+                if gt_status_id
+                is not None
+                else None
+            ),
+            "gt_status_name": gt_status_name,
+
+            **status,
+
+            "pred_status_name": contract.label_name_by_id.get(
+                pred_status_id,
+                str(
+                    pred_status_id
+                ),
+            ),
+
+            "pred_group": pred_group,
+            "pred_nav_frac": float(
+                pred_nav_frac
+            ),
+            "group_mismatch": int(
+                group_mismatch
+            ),
+        }
+
+        score, score_parts = compute_review_score(
+            row
+        )
+
+        row[
+            "review_score"
+        ] = float(
+            score
+        )
+        row[
+            "suspicion_pct"
+        ] = float(
+            score
+            * 100.0
+        )
+
+        row.update(
+            score_parts
+        )
+
+        reasons = build_reasons(
+            row,
+            min_disagree_frac=float(
+                args.min_disagree_frac
+            ),
+            min_highconf_disagree_frac=float(
+                args.min_highconf_disagree_frac
+            ),
+            min_highconf_interior_frac=float(
+                args.min_highconf_interior_frac
+            ),
+            min_highconf_unsafe_frac=float(
+                args.min_highconf_unsafe_frac
+            ),
+            min_conflict_pixels=int(
+                args.min_conflict_pixels
+            ),
+            status_confidence=float(
+                row[
+                    "status_confidence"
+                ]
+            ),
+            min_status_confidence=float(
+                args.min_status_confidence
+            ),
+            min_group_disagree_frac=float(
+                args.min_group_disagree_frac
+            ),
+        )
+
+        row[
+            "review_reasons"
+        ] = ";".join(
+            reasons
+        )
+
+        row[
+            "flagged"
+        ] = int(
+            bool(
+                reasons
+            )
+        )
+
+        row[
+            "stratum"
+        ] = (
+            f"{gt_group}"
+            f"__"
+            f"{gt_status_name or 'unknown'}"
+        )
+
+        rows.append(
+            row
+        )
+
+        del (
+            image_bgr,
+            gt,
+            gt_eval,
+            pred,
+            seg_probs,
+            infer,
+        )
+
+        if (
+            args.gc_every
+            > 0
+            and i
+            % int(
+                args.gc_every
+            )
+            == 0
+        ):
+            gc.collect()
+
+            if (
+                device.type
+                == "cuda"
+            ):
+                torch.cuda.empty_cache()
+
+        if (
+            i
+            % 50
+            == 0
+            or i
+            == len(
+                samples
+            )
+        ):
+            flagged_now = sum(
+                int(
+                    r[
+                        "flagged"
+                    ]
+                )
+                for r in rows
+            )
+
+            print(
+                f"[{i:5d}/{len(samples):5d}] "
+                f"flagged={flagged_now:4d} "
+                f"avg_inf={infer_ms_sum / i:.1f}ms "
+                f"last_score={score*100.0:5.1f} "
+                f"{sample.group}/{sample.base}"
+            )
+
+    elapsed = (
+        time.perf_counter()
+        - t0_all
+    )
+
+    # -------------------------------------------------------------------------
+    # Candidate selection
+    # -------------------------------------------------------------------------
+
+    by_stratum: Dict[
+        str,
+        List[
+            dict
+        ],
+    ] = defaultdict(
+        list
+    )
+
+    for row in rows:
+        by_stratum[
+            str(
+                row[
+                    "stratum"
+                ]
+            )
+        ].append(
+            row
+        )
+
+    selected_rows: List[
+        dict
+    ] = []
+
+    for stratum, srows in sorted(
+        by_stratum.items()
+    ):
+        ordered = sorted(
+            srows,
+            key=lambda r: float(
+                r[
+                    "review_score"
+                ]
+            ),
+            reverse=True,
+        )
+
+        if args.top_k_per_stratum > 0:
+            chosen = ordered[
+                :int(
+                    args.top_k_per_stratum
+                )
+            ]
+
+            for row in chosen:
+                if not row[
+                    "review_reasons"
+                ]:
+                    row[
+                        "review_reasons"
+                    ] = "TOP_K_STRATUM"
+                    row[
+                        "flagged"
+                    ] = 1
+
+        elif (
+            args.export_mode
+            == "all"
+        ):
+            chosen = ordered
+
+            for row in chosen:
+                if not row[
+                    "review_reasons"
+                ]:
+                    row[
+                        "review_reasons"
+                    ] = "EXPORT_ALL"
+
+        elif (
+            args.export_mode
+            == "score"
+        ):
+            thr = (
+                float(
+                    args.min_suspicion_pct
+                )
+                / 100.0
+            )
+
+            chosen = [
+                row
+                for row in ordered
+                if float(
+                    row[
+                        "review_score"
+                    ]
+                )
+                >= thr
+            ]
+
+            for row in chosen:
+                if not row[
+                    "review_reasons"
+                ]:
+                    row[
+                        "review_reasons"
+                    ] = "SUSPICION_SCORE"
+
+        else:
+            chosen = [
+                row
+                for row in ordered
+                if int(
+                    row[
+                        "flagged"
+                    ]
+                )
+                == 1
+            ]
+
+        if (
+            args.max_review_per_stratum
+            > 0
+        ):
+            chosen = chosen[
+                :int(
+                    args.max_review_per_stratum
+                )
+            ]
+
+        for rank, row in enumerate(
+            chosen,
+            1,
+        ):
+            row[
+                "review_rank_stratum"
+            ] = rank
+
+        selected_rows.extend(
+            chosen
+        )
+
+    selected_rows = sorted(
+        selected_rows,
+        key=lambda r: (
+            -float(
+                r[
+                    "review_score"
+                ]
+            ),
+            str(
+                r[
+                    "stratum"
+                ]
+            ),
+            natural_key(
+                r[
+                    "base"
+                ]
+            ),
+        ),
+    )
+
+    for rank, row in enumerate(
+        selected_rows,
+        1,
+    ):
+        row[
+            "review_rank_global"
+        ] = rank
+
+    selected_keys = {
+        (
+            str(
+                r[
+                    "group"
+                ]
+            ),
+            str(
+                r[
+                    "base"
+                ]
+            ),
+        )
+        for r in selected_rows
+    }
+
+    # -------------------------------------------------------------------------
+    # Physical export, memory-safe second pass
+    # -------------------------------------------------------------------------
+
+    source_manifest_rows = []
+
+    if not args.report_only:
+        print()
+        print(
+            f"[EXPORT] candidatos físicos: {len(selected_rows)}"
+        )
+
+        sample_by_key = {
+            (
+                str(
+                    s.group
+                ),
+                str(
+                    s.base
+                ),
+            ): s
+            for s in samples
+        }
+
+        selected_by_group = defaultdict(
+            list
+        )
+
+        for row in selected_rows:
+            selected_by_group[
+                str(
+                    row[
+                        "group"
+                    ]
+                )
+            ].append(
+                row
+            )
+
+        for j, row in enumerate(
+            selected_rows,
+            1,
+        ):
+            key = (
+                str(
+                    row[
+                        "group"
+                    ]
+                ),
+                str(
+                    row[
+                        "base"
+                    ]
+                ),
+            )
+
+            sample = sample_by_key[
+                key
+            ]
+
+            try:
+                source = resolve_original_source(
+                    sample=sample,
+                    dataset_root=dataset_root,
+                )
+
+                image_bgr = cv2.imread(
+                    str(
+                        sample.image
+                    ),
+                    cv2.IMREAD_COLOR,
+                )
+
+                if image_bgr is None:
+                    raise RuntimeError(
+                        sample.image
+                    )
+
+                gt_ids = base.load_gt_mask(
+                    sample.mask,
+                    classes,
+                )
+
+                if gt_ids is None:
+                    raise RuntimeError(
+                        sample.mask
+                    )
+
+                infer = infer_audit(
+                    base_module=base,
+                    base_model=base_model,
+                    status_head=status_head,
+                    contract=contract,
+                    image_bgr=image_bgr,
+                    device=device,
+                    use_amp=not args.no_amp,
+                )
+
+                pred_ids = infer[
+                    "pred"
+                ]
+
+                exported = save_review_item(
+                    row=row,
+                    sample=sample,
+                    pred_ids=pred_ids,
+                    gt_ids=gt_ids,
+                    source=source,
+                    review_group_root=review_group_root,
+                    base_module=base,
+                    classes=classes,
+                    contract=contract,
+                    save_panels=bool(
+                        args.save_panels
+                    ),
+                )
+
+                row[
+                    "physical_export_ok"
+                ] = 1
+                row[
+                    "physical_export_error"
+                ] = ""
+
+                source_manifest_rows.append({
+                    **exported,
+                    "review_score": row[
+                        "review_score"
+                    ],
+                    "suspicion_pct": row[
+                        "suspicion_pct"
+                    ],
+                    "review_reasons": row[
+                        "review_reasons"
+                    ],
+                })
+
+            except Exception as exc:
+                row[
+                    "physical_export_ok"
+                ] = 0
+                row[
+                    "physical_export_error"
+                ] = (
+                    f"{type(exc).__name__}: {exc}"
+                )
+
+                print(
+                    f"[EXPORT][WARN] "
+                    f"{row['group']}/{row['base']}: "
+                    f"{row['physical_export_error']}"
+                )
+
+            if (
+                args.gc_every
+                > 0
+                and j
+                % int(
+                    args.gc_every
+                )
+                == 0
+            ):
+                gc.collect()
+
+                if (
+                    device.type
+                    == "cuda"
+                ):
+                    torch.cuda.empty_cache()
+
+            if (
+                j
+                % 50
+                == 0
+                or j
+                == len(
+                    selected_rows
+                )
+            ):
+                print(
+                    f"  exportados {j}/{len(selected_rows)}"
+                )
+
+        # review_order.csv por grupo.
+        for group, grows in selected_by_group.items():
+            order_path = (
+                review_group_root
+                / group
+                / "review_order.csv"
+            )
+
+            write_csv(
+                order_path,
+                sorted(
+                    grows,
+                    key=lambda r: int(
+                        r.get(
+                            "review_rank_global",
+                            999999,
+                        )
+                    ),
+                ),
+            )
+
+        write_csv(
+            out_root
+            / "review_source_manifest.csv",
+            source_manifest_rows,
+        )
+
+    # -------------------------------------------------------------------------
+    # Reports
+    # -------------------------------------------------------------------------
+
+    report_path = (
+        report_dir
+        / "review_report.csv"
+    )
+
+    candidates_path = (
+        report_dir
+        / "review_candidates.csv"
+    )
+
+    write_csv(
+        report_path,
+        rows,
+    )
+
+    write_csv(
+        candidates_path,
+        selected_rows,
+    )
+
+    semantic_labels = [
+        contract.seg_id2label[
+            i
+        ]
+        for i in sorted(
+            contract.seg_id2label
+        )
+    ]
+
+    status_labels = [
+        contract.label_name_by_id[
+            i
+        ]
+        for i in sorted(
+            contract.label_name_by_id
+        )
+    ]
+
+    write_cm_csv(
+        report_dir
+        / "confusion_semantic.csv",
+        cm_seg_total,
+        semantic_labels,
+    )
+
+    write_cm_csv(
+        report_dir
+        / "confusion_status.csv",
+        cm_status_total,
+        status_labels,
+    )
+
+    global_seg = binary_metrics_from_cm(
+        cm_seg_total,
+        contract.nav_class_id,
+        contract.non_nav_class_id,
+    )
+
+    global_status = status_metrics_from_cm(
+        cm_status_total
+    )
+
+    by_group = defaultdict(
+        list
+    )
+    by_status = defaultdict(
+        list
+    )
+    by_split = defaultdict(
+        list
+    )
+
+    for row in rows:
+        by_group[
+            str(
+                row[
+                    "group"
+                ]
+            )
+        ].append(
+            row
+        )
+
+        by_status[
+            str(
+                row.get(
+                    "gt_status_name",
+                    "unknown",
+                )
+            )
+        ].append(
+            row
+        )
+
+        by_split[
+            str(
+                row.get(
+                    "split_origin",
+                    "unknown",
+                )
+            )
+        ].append(
+            row
+        )
+
+    group_summary = {
+        key: summarize_rows(
+            value,
+            selected_keys,
+        )
+        for key, value in sorted(
+            by_group.items()
+        )
+    }
+
+    status_summary = {
+        key: summarize_rows(
+            value,
+            selected_keys,
+        )
+        for key, value in sorted(
+            by_status.items()
+        )
+    }
+
+    split_summary = {
+        key: summarize_rows(
+            value,
+            selected_keys,
+        )
+        for key, value in sorted(
+            by_split.items()
+        )
+    }
+
+    stratum_summary = {
+        key: summarize_rows(
+            value,
+            selected_keys,
+        )
+        for key, value in sorted(
+            by_stratum.items()
+        )
+    }
+
+    all_summary = summarize_rows(
+        rows,
+        selected_keys,
+    )
+
+    source_export_errors = [
+        {
+            "group": row[
+                "group"
+            ],
+            "base": row[
+                "base"
+            ],
+            "error": row.get(
+                "physical_export_error",
+                "",
+            ),
+        }
+        for row in selected_rows
+        if int(
+            row.get(
+                "physical_export_ok",
+                1,
+            )
+            or 0
+        )
+        == 0
+    ]
+
+    summary = {
+        "schema": REVIEW_SCHEMA,
+        "reviewer_version": REVIEWER_VERSION,
+        "created_at": now_iso(),
+
+        "note": (
+            "Model-vs-GT disagreement is a review signal, not proof of annotation error. "
+            "review_score/suspicion_pct are heuristic ranking values, not probabilities."
+        ),
+
+        "config": str(
+            config_path
+        ),
+        "test_script": str(
+            test_script
+        ),
+        "checkpoint": str(
+            checkpoint_path
+        ),
+        "checkpoint_epoch": contract.checkpoint_epoch,
+        "trainer_version": contract.trainer_version,
+
+        "dataset_root": str(
+            root
+        ),
+        "resolution_wh": [
+            W,
+            H,
+        ],
+
+        "samples": len(
+            rows
+        ),
+        "selected_for_review": len(
+            selected_rows
+        ),
+        "selected_pct": float(
+            100.0
+            * len(
+                selected_rows
+            )
+            / max(
+                1,
+                len(
+                    rows
+                ),
+            )
+        ),
+
+        "elapsed_seconds": float(
+            elapsed
+        ),
+        "mean_inference_ms": float(
+            infer_ms_sum
+            / max(
+                1,
+                len(
+                    rows
+                ),
+            )
+        ),
+
+        "selection": {
+            "export_mode": args.export_mode,
+            "min_suspicion_pct": float(
+                args.min_suspicion_pct
+            ),
+            "top_k_per_stratum": int(
+                args.top_k_per_stratum
+            ),
+            "max_review_per_stratum": int(
+                args.max_review_per_stratum
+            ),
+            "report_only": bool(
+                args.report_only
+            ),
+        },
+
+        "thresholds": {
+            "high_confidence": float(
+                args.high_confidence
+            ),
+            "uncertainty_margin": float(
+                args.uncertainty_margin
+            ),
+            "boundary_radius_px": int(
+                args.boundary_radius
+            ),
+            "composition_min_fraction": float(
+                args.composition_min_fraction
+            ),
+            "min_disagree_frac": float(
+                args.min_disagree_frac
+            ),
+            "min_highconf_disagree_frac": float(
+                args.min_highconf_disagree_frac
+            ),
+            "min_highconf_interior_frac": float(
+                args.min_highconf_interior_frac
+            ),
+            "min_highconf_unsafe_frac": float(
+                args.min_highconf_unsafe_frac
+            ),
+            "min_conflict_pixels": int(
+                args.min_conflict_pixels
+            ),
+            "min_status_confidence": float(
+                args.min_status_confidence
+            ),
+        },
+
+        "metrics_global": {
+            "semantic": global_seg,
+            "status": global_status,
+        },
+
+        "annotation_health": all_summary,
+
+        "groups": group_summary,
+        "statuses": status_summary,
+        "splits": split_summary,
+        "strata_group_x_status": stratum_summary,
+
+        "physical_export": {
+            "enabled": not bool(
+                args.report_only
+            ),
+            "exported": len(
+                source_manifest_rows
+            ),
+            "errors": source_export_errors,
+            "final_masks_preserved_by_clear_review": True,
+        },
+
+        "paths": {
+            "review_report": str(
+                report_path
+            ),
+            "review_candidates": str(
+                candidates_path
+            ),
+            "confusion_semantic": str(
+                report_dir
+                / "confusion_semantic.csv"
+            ),
+            "confusion_status": str(
+                report_dir
+                / "confusion_status.csv"
+            ),
+            "review_group_root": (
+                str(
+                    review_group_root
+                )
+                if not args.report_only
+                else None
+            ),
+            "review_source_manifest": (
+                str(
+                    out_root
+                    / "review_source_manifest.csv"
+                )
+                if not args.report_only
+                else None
+            ),
+        },
+    }
+
+    summary_path = (
+        report_dir
+        / "review_summary.json"
+    )
+
+    save_json(
+        summary_path,
+        summary,
+    )
+
+    readme_path = (
+        report_dir
+        / "README_REVIEW.txt"
+    )
+
+    readme_path.write_text(
+        "\n".join([
+            "AGROBOT CORRIDOR DATASET REVIEW / MODEL MINING",
+            "",
+            "IMPORTANTE:",
+            "- Discordancia modelo x GT NAO prova que a anotacao esta errada.",
+            "- Conflito de alta confianca no interior da regiao e um sinal forte para revisao.",
+            "- Falso navegavel de alta confianca recebe peso maior por risco operacional.",
+            "- A segunda cabeca (estado global) e auditada junto da mascara.",
+            "- review_score e suspicion_pct sao apenas ranking heuristico.",
+            "- O auditor usa o checkpoint PyTorch porque o ONNX de CAMPO e deliberadamente mask-only.",
+            "- --clear_review preserva final_masks/.",
+            "- --clear_review_all apaga tudo e deve ser usado com cuidado.",
+            "",
+            f"Checkpoint: {checkpoint_path}",
+            f"Dataset: {root}",
+            f"Samples: {len(rows)}",
+            f"Selecionados: {len(selected_rows)}",
+            f"Report: {report_path}",
+            f"Candidates: {candidates_path}",
+        ]),
+        encoding="utf-8",
+    )
+
+    # -------------------------------------------------------------------------
+    # Console
+    # -------------------------------------------------------------------------
+
+    print()
+    print("=" * 94)
+    print(
+        "REVISÃO CONCLUÍDA"
+    )
+    print("=" * 94)
+    print(
+        f"Samples                 : {len(rows)}"
+    )
+    print(
+        f"Selecionados            : "
+        f"{len(selected_rows)} "
+        f"({100.0 * len(selected_rows) / max(1,len(rows)):.2f}%)"
+    )
+    print(
+        f"Tempo total             : {elapsed:.1f}s"
+    )
+    print(
+        f"Inferência média        : "
+        f"{infer_ms_sum / max(1,len(rows)):.2f} ms"
+    )
+    print(
+        f"Semantic acc/mIoU       : "
+        f"{global_seg['acc']:.4f} / "
+        f"{global_seg['miou']:.4f}"
+    )
+    print(
+        f"Nav IoU / NonNav IoU    : "
+        f"{global_seg['nav_iou']:.4f} / "
+        f"{global_seg['non_nav_iou']:.4f}"
+    )
+    print(
+        f"Unsafe nav global       : "
+        f"{100.0 * global_seg['unsafe_nav_rate']:.4f}%"
+    )
+    print(
+        f"Status acc / macroF1    : "
+        f"{global_status['acc']:.4f} / "
+        f"{global_status['macro_f1']:.4f}"
+    )
+    print(
+        f"Report                  : {report_path}"
+    )
+    print(
+        f"Candidates              : {candidates_path}"
+    )
+    print(
+        f"Summary                 : {summary_path}"
+    )
+
+    if not args.report_only:
+        print(
+            f"Revisão física          : {review_group_root}"
+        )
+        print(
+            f"Source manifest         : "
+            f"{out_root / 'review_source_manifest.csv'}"
+        )
+
+    print("=" * 94)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/Python/OAK/datasets/oak-d/config.json b/Python/OAK/datasets/oak-d/config.json
new file mode 100644
index 000000000..ef85631d1
--- /dev/null
+++ b/Python/OAK/datasets/oak-d/config.json
@@ -0,0 +1,185 @@
+{
+  "camera": "oak-d",
+  "modelo": "segformer_b0",
+  "model_name": "corredores_v2",
+  "main_class_name": "navegavel",
+  "model_to_use": "geral",
+  "ckpt_test": "best_operational",
+
+  "raw_size": [1920, 1080],
+  "resolucao": [1024, 576],
+  "roi_inicio": 0.0,
+  "roi_tamanho": 1.0,
+
+  "channels": 3,
+  "backbone": "nvidia/mit-b0",
+
+  "label_classes": [
+    "Parado",
+    "EntrandoRua",
+    "CaminhandoRua",
+    "SaindoRua",
+    "Manobrando",
+    "Direcionando",
+    "RetornandoBase",
+    "Indefinido"
+  ],
+
+  "corridor_training": {
+    "data": {
+      "strict_labels": true,
+      "strict_pipeline_contract": true,
+      "preflight_max_samples": 0
+    },
+
+    "augmentation": {
+      "enabled": true,
+      "horizontal_flip_p": 0.50,
+      "affine_p": 0.60,
+      "rotate_deg": 4.0,
+      "scale_min": 0.92,
+      "scale_max": 1.08,
+      "translate_frac": 0.03,
+
+      "global_gain_p": 0.45,
+      "global_gain_min": 0.82,
+      "global_gain_max": 1.18,
+
+      "contrast_p": 0.35,
+      "contrast_min": 0.85,
+      "contrast_max": 1.15,
+
+      "gamma_p": 0.30,
+      "gamma_min": 0.82,
+      "gamma_max": 1.18,
+
+      "rgb_gain_p": 0.25,
+      "rgb_gain_min": 0.94,
+      "rgb_gain_max": 1.06,
+
+      "shadow_p": 0.35,
+      "shadow_strength_min": 0.12,
+      "shadow_strength_max": 0.42,
+
+      "noise_p": 0.18,
+      "noise_sigma_min": 0.003,
+      "noise_sigma_max": 0.018,
+
+      "blur_p": 0.12,
+      "blur_kernel": 3,
+
+      "occlusion_p": 0.00,
+      "occlusion_min_frac": 0.04,
+      "occlusion_max_frac": 0.15,
+      "ramp_epochs": 6
+    },
+
+    "sampler": {
+      "mode": "joint",
+      "joint_power": 0.30,
+      "status_power": 0.12,
+      "group_power": 0.10,
+      "max_weight_ratio": 5.0,
+      "samples_per_epoch": 0
+    },
+
+    "class_weighting": {
+      "seg_method": "log_inverse",
+      "seg_log_offset": 1.02,
+      "seg_min_weight": 0.35,
+      "seg_max_weight": 3.0,
+      "seg_max_samples": 1200,
+
+      "status_enabled": true,
+      "status_power": 0.30,
+      "status_min_weight": 0.50,
+      "status_max_weight": 3.0
+    },
+
+    "loss": {
+      "seg_ce_weight": 0.65,
+      "seg_dice_weight": 0.35,
+      "boundary_boost": 0.15,
+      "unsafe_nav_weight": 0.06,
+
+      "status_weight": 0.40,
+      "status_ramp_epochs": 8,
+      "status_label_smoothing": 0.03
+    },
+
+    "status_head": {
+      "pool_h": 3,
+      "pool_w": 4,
+      "hidden": 256,
+      "dropout": 0.20,
+      "detach_seg_summary": true,
+      "seg_summary": "probabilities"
+    },
+
+    "optimizer": {
+      "encoder_lr_mult": 0.50,
+      "decoder_lr_mult": 1.00,
+      "status_lr_mult": 2.00,
+      "freeze_encoder_epochs": 2,
+      "no_decay_bias": true,
+      "no_decay_norm": true,
+      "betas": [0.9, 0.999],
+      "eps": 1e-8
+    },
+
+    "scheduler": {
+      "mode": "poly",
+      "warmup_ratio": 0.05,
+      "warmup_start_factor": 0.10,
+      "poly_power": 1.0,
+      "min_lr_ratio": 0.02
+    },
+
+    "optimization": {
+      "grad_clip_norm": 1.0,
+      "matmul_precision": "high",
+      "cudnn_benchmark": true,
+      "persistent_workers": true,
+      "prefetch_factor": 2
+    },
+
+    "score": {
+      "nav_iou": 0.40,
+      "nav_f1": 0.20,
+      "status_macro_f1": 0.20,
+      "safety": 0.20
+    },
+
+    "operational_gate": {
+      "enabled": true,
+      "min_epoch": 5,
+
+      "min_nav_iou": 0.70,
+      "min_nav_f1": 0.80,
+      "min_non_nav_iou": 0.60,
+      "min_status_macro_f1": 0.60,
+      "max_unsafe_nav_rate": 0.03,
+
+      "group_unsafe": {
+        "enabled": true,
+        "min_non_nav_pixels": 5000,
+        "max_rate": 0.05,
+        "max_rate_by_group": {
+          "naonavegavel": 0.04,
+          "navegavel_naonavegavel": 0.05
+        }
+      },
+
+      "status_recall": {
+        "enabled": false,
+        "min_support": 5,
+        "minimums": {}
+      }
+    },
+
+    "checkpoint": {
+      "early_stop_patience": 18,
+      "early_stop_min_delta": 0.0003
+    }
+  }
+}
diff --git a/Python/OAK/datasets/oak-d/roi_seg_dataset.py b/Python/OAK/datasets/oak-d/roi_seg_dataset.py
new file mode 100644
index 000000000..8b9b2221b
--- /dev/null
+++ b/Python/OAK/datasets/oak-d/roi_seg_dataset.py
@@ -0,0 +1,148 @@
+# -*- coding: utf-8 -*-
+from PIL import Image
+import glob
+import os
+import cv2
+import numpy as np
+import torch
+from torch.utils.data import Dataset
+
+from utils import carregar_labelmap_completo, compute_roi_indices, resize_keep_width
+
+IMG_EXTS = (".jpg", ".jpeg", ".png")
+MSK_EXTS = (".png", ".jpg", ".jpeg")  # preferimos .png se existir
+
+def _is_dir(p): return os.path.isdir(p)
+def _is_file(p): return os.path.isfile(p)
+
+def _list_groups(group_root):
+    if not _is_dir(group_root): return []
+    out = []
+    for g in sorted(os.listdir(group_root)):
+        gdir = os.path.join(group_root, g)
+        if not _is_dir(gdir): 
+            continue
+        if _is_dir(os.path.join(gdir, "images")) and _is_dir(os.path.join(gdir, "masks")):
+            out.append(g)
+    return out
+
+def _mask_for_base(msk_dir, base):
+    """Encontra a máscara que casa com o base, priorizando .png."""
+    best = None
+    for ext in MSK_EXTS:
+        cand = os.path.join(msk_dir, base + ext)
+        if _is_file(cand):
+            if best is None: best = cand
+            # mantém .png se aparecer depois
+            if os.path.splitext(cand)[1].lower() == ".png":
+                return cand
+    return best
+
+def _collect_pairs_legacy(root):
+    """root/{images,masks}"""
+    img_dir = os.path.join(root, "images")
+    msk_dir = os.path.join(root, "masks")
+    imgs = []
+    msks = []
+    for p in sorted(glob.glob(os.path.join(img_dir, "*"))):
+        base, ext = os.path.splitext(os.path.basename(p))
+        if ext.lower() not in IMG_EXTS: 
+            continue
+        m = _mask_for_base(msk_dir, base)
+        if m:
+            imgs.append(p)
+            msks.append(m)
+    return imgs, msks
+
+def _collect_pairs_grouped(root):
+    """Suporta:
+       - root/group/<g>/{images,masks}
+       - root/<g>/{images,masks} (quando root já é '.../group')
+    """
+    # case A: root tem subpasta 'group'
+    group_root = os.path.join(root, "group")
+    if not _is_dir(group_root):
+        # case B: root JÁ É a pasta 'group'
+        group_root = root
+
+    groups = _list_groups(group_root)
+    imgs, msks = [], []
+    for g in groups:
+        img_dir = os.path.join(group_root, g, "images")
+        msk_dir = os.path.join(group_root, g, "masks")
+        for p in sorted(glob.glob(os.path.join(img_dir, "*"))):
+            base, ext = os.path.splitext(os.path.basename(p))
+            if ext.lower() not in IMG_EXTS:
+                continue
+            m = _mask_for_base(msk_dir, base)
+            if m:
+                imgs.append(p)
+                msks.append(m)
+    return imgs, msks
+
+class ROISegDataset(Dataset):
+    """
+    Compatível com o dataset original, mas agora aceita:
+    - root = '.../split/train'  (com 'group' dentro)
+    - root = '.../split/train/group'
+    - root = '.../split/train/<grupo>' (ainda funciona via legado se tiver images/masks)
+    - root legado = '.../split/train' com 'images' e 'masks' diretamente
+    """
+    def __init__(self, root, out_dir, zona_inicio, faixa_atuacao, input_w=384, min_input_h=96, labelmap_path="labelmap.txt",
+                 mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225)):
+
+        self.out_dir = out_dir
+        self.zona_inicio = zona_inicio
+        self.faixa_atuacao = faixa_atuacao
+        self.input_w = input_w
+        self.min_input_h = min_input_h
+        self.mean = np.array(mean, dtype=np.float32).reshape(1, 1, 3)
+        self.std = np.array(std, dtype=np.float32).reshape(1, 1, 3)
+
+        # tenta agrupar; se não encontrar, cai pro legado
+        imgs, msks = _collect_pairs_grouped(root)
+        if not imgs:
+            imgs, msks = _collect_pairs_legacy(root)
+
+        assert len(imgs) == len(msks) and len(imgs) > 0, f"Nenhuma imagem/máscara encontrada em {root}"
+        self.img_paths = imgs
+        self.msk_paths = msks
+
+        # labelmap
+        _, _, self.classes, self.ignore_rgb = carregar_labelmap_completo(labelmap_path)
+        # tipicamente ignore_rgb é (255,255,255)
+        self.ignore_id = int(self.ignore_rgb[0]) if isinstance(self.ignore_rgb, (list, tuple)) else int(self.ignore_rgb)
+
+    def __len__(self):
+        return len(self.img_paths)
+
+    def __getitem__(self, idx):
+        img_rgb = np.array(Image.open(self.img_paths[idx]).convert("RGB"))
+        # Máscara como escala de cinza (IDs já foram normalizados na etapa de normalize)
+        msk_grayscale = np.array(Image.open(self.msk_paths[idx]).convert("L"))
+
+        H, W = img_rgb.shape[:2]
+        y_fim, y_inicio = compute_roi_indices(H, self.zona_inicio, self.faixa_atuacao)
+
+        img_roi = img_rgb[y_fim:y_inicio, 0:W]
+        msk_roi = msk_grayscale[y_fim:y_inicio, 0:W]
+
+        img_in = resize_keep_width(img_roi, self.input_w, self.min_input_h, cv2.INTER_AREA)
+        msk_ids = resize_keep_width(msk_roi, self.input_w, self.min_input_h, cv2.INTER_NEAREST)
+
+        valores_validos = list(range(len(self.classes))) + [255]
+        msk_ids[np.isin(msk_ids, valores_validos, invert=True)] = self.ignore_id
+
+        if np.all(msk_ids == self.ignore_id):
+            raise ValueError(f"Máscara {self.msk_paths[idx]} está só com valor de ignore ({self.ignore_id})")
+
+        img_f = img_in.astype(np.float32) / 255.0
+        img_f = (img_f - self.mean) / self.std
+        img_chw = np.transpose(img_f, (2, 0, 1))
+
+        if idx == 0:
+            os.makedirs(self.out_dir, exist_ok=True)
+            cv2.imwrite(os.path.join(self.out_dir, "debug_roi_input.png"), cv2.cvtColor(img_roi, cv2.COLOR_RGB2BGR))
+            cv2.imwrite(os.path.join(self.out_dir, "debug_roi_mask.png"), msk_roi)
+
+        return torch.from_numpy(img_chw).float(), torch.from_numpy(msk_ids.astype(np.int64))
diff --git a/Python/OAK/datasets/oak-d/utils.py b/Python/OAK/datasets/oak-d/utils.py
new file mode 100644
index 000000000..9207e7c27
--- /dev/null
+++ b/Python/OAK/datasets/oak-d/utils.py
@@ -0,0 +1,144 @@
+import cv2
+import numpy as np
+
+# ----------------------------
+# Helpers LABELMAP
+# ----------------------------
+def carregar_labelmap_completo(caminho):
+    cor_para_id = {}
+    id_para_nome = {}
+    cores_rgb = []
+
+    with open(caminho, 'r') as arquivo:
+        idx = 0
+        for linha in arquivo:
+            if linha.startswith("#") or not linha.strip():
+                continue
+            partes = linha.strip().split(':')
+            if len(partes) >= 2:
+                nome_classe, cor_rgb_str = partes[0], partes[1]
+                r, g, b = map(int, cor_rgb_str.split(','))
+                cor_rgb = (r, g, b)
+
+                if nome_classe.lower() == "ignore":
+                    ignore_rgb = cor_rgb
+                    continue  # NÃO adiciona ignore no LUT de classes
+
+                cor_para_id[cor_rgb] = idx
+                cores_rgb.append(cor_rgb)
+                id_para_nome[idx] = nome_classe
+                idx += 1
+
+    #print(f"Mapa: {cor_para_id}")
+    #print(f"Colormap RGB: {cores_rgb}")
+    #print(f"Classes: {id_para_nome}")
+    #print(f"Ignore RGB: {ignore_rgb}")
+
+    return cor_para_id, cores_rgb, id_para_nome, ignore_rgb
+
+def converter_mask_rgb_para_ids(img_rgb, mapa_rgb, ignore_id):
+    # Cria um mapa 256^3 para IDs (usa int32 para indexar)
+    lut = np.full((256**3,), ignore_id, dtype=np.uint8)
+
+    for cor, classe_id in mapa_rgb.items():
+        r, g, b = cor
+        lut[(r << 16) + (g << 8) + b] = classe_id
+
+    # Converte RGB para índice único
+    flat_idx = (img_rgb[:,:,0].astype(np.int32) << 16) + \
+               (img_rgb[:,:,1].astype(np.int32) << 8) + \
+               img_rgb[:,:,2].astype(np.int32)
+
+    # Aplica LUT vetorizada
+    return lut[flat_idx]
+
+def converter_mask_ids_para_rgb(mask_ids: np.ndarray, colormap_rgb: list, ignore_id: int = 255) -> np.ndarray:
+    # Criar lookup table (256 cores possíveis)
+    lut = np.zeros((256, 3), dtype=np.uint8)
+    for i, color in enumerate(colormap_rgb):
+        lut[i] = color
+    lut[ignore_id] = (255, 255, 255)
+
+    # Aplicar LUT direto (vetorizado)
+    return lut[mask_ids]
+
+def converter_mask_ids_para_bgr(mask_ids: np.ndarray, colormap_rgb: list, ignore_id: int = 255) -> np.ndarray:
+    """
+    Converte máscara de IDs para imagem BGR (uint8),
+    pronta para uso com OpenCV.
+    """
+    lut = np.zeros((256, 3), dtype=np.uint8)
+
+    for i, (r, g, b) in enumerate(colormap_rgb):
+        lut[i] = (b, g, r)   # RGB -> BGR
+
+    lut[ignore_id] = (255, 255, 255)  # branco em BGR = RGB
+
+    return lut[mask_ids]
+
+def desenhar_legenda_vertical(colormap_rgb, classes, largura=200):
+    """
+    Retorna uma imagem com a legenda das classes (cor + nome)
+    """
+    nomes_classes = [classes[i] for i in range(len(classes))]
+    altura_por_classe = 30
+    altura_total = altura_por_classe * len(colormap_rgb)
+    legenda = np.ones((altura_total, largura, 3), dtype=np.uint8) * 255
+
+    for idx, (rgb, nome) in enumerate(zip(colormap_rgb, nomes_classes)):
+        y = idx * altura_por_classe
+        color = tuple(int(c) for c in rgb)
+        cv2.rectangle(legenda, (10, y + 5), (30, y + 25), color, -1)
+        cv2.putText(legenda, nome, (40, y + 20),
+                    cv2.FONT_HERSHEY_SIMPLEX, 0.6, (0, 0, 0), 1, cv2.LINE_AA)
+
+    return legenda
+
+def desenhar_legenda_horizontal(colormap_rgb, classes, altura=30, largura_por_classe=120):
+    """
+    Retorna uma imagem com a legenda das classes (cor + nome), em uma única linha horizontal
+    """
+    nomes_classes = [classes[i] for i in range(len(classes))]
+    largura_total = largura_por_classe * len(colormap_rgb)
+    legenda = np.ones((altura, largura_total, 3), dtype=np.uint8) * 255  # faixa branca
+
+    for idx, (rgb, nome) in enumerate(zip(colormap_rgb, nomes_classes)):
+        x = idx * largura_por_classe
+        color = tuple(int(c) for c in rgb)
+        # Retângulo colorido
+        cv2.rectangle(legenda, (x + 10, 5), (x + 30, 25), color, -1)
+        # Texto da classe
+        cv2.putText(legenda, nome, (x + 35, 20),
+                    cv2.FONT_HERSHEY_SIMPLEX, 0.6, (0, 0, 0), 1, cv2.LINE_AA)
+
+    return legenda
+
+def _infer_ignore_id(ignore_rgb, default_id=255):
+    import numpy as _np
+    if isinstance(ignore_rgb, (list, tuple)):
+        if len(ignore_rgb) == 1 and isinstance(ignore_rgb[0], (int, _np.integer)):
+            return int(ignore_rgb[0])
+        if len(ignore_rgb) == 3:
+            return default_id
+    if isinstance(ignore_rgb, (int, _np.integer)):
+        return int(ignore_rgb)
+    return default_id
+
+# ----------------------------
+# Helpers ROI
+# ----------------------------
+def compute_roi_indices(H: int, zona_inicio: float, faixa_atuacao: float):
+    y_inicio = int((1.0 - zona_inicio) * H)
+    y_fim = int((1.0 - (zona_inicio + faixa_atuacao)) * H)
+    y_fim = max(0, min(H, y_fim))
+    y_inicio = max(0, min(H, y_inicio))
+    if y_fim >= y_inicio:
+        y_fim = max(0, y_inicio - 1)
+    return y_fim, y_inicio
+
+def resize_keep_width(img: np.ndarray, new_w: int, min_h: int, interpolation: int) -> np.ndarray:
+    h, w = img.shape[:2]
+    new_h = int(round(new_w * (h / w)))
+    if min_h is not None and new_h < min_h:
+        new_h = min_h
+    return cv2.resize(img, (new_w, new_h), interpolation=interpolation)