diff --git a/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs b/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs
index aad9eca7c..df32d48dc 100644
--- a/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs
+++ b/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs
@@ -2079,6 +2079,13 @@ impl GpuDqnTrainer {
         Ok(())
     }
 
+    /// ISV signal update — safe to call inside graph capture.
+    /// Reads pinned device-mapped scalars (loss, grad_norm, Q-mean)
+    /// written by Adam, updates ISV signal vector [12] + history.
+    pub(crate) fn submit_isv_signal_update(&self) -> Result<(), MLError> {
+        self.update_isv_signals()
+    }
+
     /// ISV encoder MLP forward: temporal decay → MLP → branch gate + gamma mod.
     /// Reads ISV signals/history (pinned), ISV encoder weights from param buffer.
     /// Writes isv_embedding_buf [ISV_EMB_DIM=8], branch_gate_buf [4], gamma_mod_buf [1].
@@ -7883,6 +7890,16 @@ impl GpuDqnTrainer {
             self.launch_trade_plan_forward(batch_size)?;
         }
 
+        // ── 1c. Temporal pipeline (enriches h_s2 before loss) ──
+        {
+            let batch_size = self.config.batch_size;
+            self.mamba2_step(batch_size)?;
+            self.compute_predictive_coding_loss(batch_size)?;
+            self.apply_regime_dropout(batch_size, true)?;
+            self.launch_isv_temporal_route()?;
+            self.risk_budget_forward(batch_size)?;
+        }
+
         // ── 2+3. Loss + gradient (blended MSE + C51 via c51_alpha ramp) ─
         self.launch_curiosity_inference()?;
 
diff --git a/crates/ml/src/trainers/dqn/fused_training.rs b/crates/ml/src/trainers/dqn/fused_training.rs
index dae4fd7e4..ad65239b3 100644
--- a/crates/ml/src/trainers/dqn/fused_training.rs
+++ b/crates/ml/src/trainers/dqn/fused_training.rs
@@ -950,11 +950,8 @@ impl FusedTrainingCtx {
             }
 
             // ── Step 2b: Temporal pipeline ────
-            // TODO: temporal ops (mamba2, regime_dropout, isv_temporal_route,
-            // risk_budget, predictive_coding) corrupt graph_aux replay when run
-            // ungraphed between graph_forward and graph_aux. Will be addressed
-            // in the cuBLAS aux rewrite — either inside graph_temporal or
-            // by eliminating graph_aux dependency.
+            // Temporal ops now run inside submit_forward_ops_main() (step 1c),
+            // captured in the cuBLAS training graph.
 
             // ── Step 2c: HER donor computation (outside graph_aux) ───────────
             if let Some(ref mut her) = self.gpu_her {