fix: remove ungraphed vaccine batch sampling from training loop

Vaccine now uses prev_grad_buf inside maintenance_child (graphed).
The external sample() call was launching ~10 ungraphed PER kernels
on the training stream every 5th step — potential CUmodule corruption
and CPU-GPU serialization point.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

crates/ml/src/trainers/dqn/fused_training.rs

@@ -909,7 +909,9 @@ impl FusedTrainingCtx {
         // ── Graph launch or ungraphed step 0 ──────────────────────────
         if let Some(ref pg) = self.parent_graph {
             // Steps 1+: single parent graph launch replays all children.
+            tracing::debug!("parent graph launch START");
             pg.launch(cu_stream)?;
+            tracing::debug!("parent graph launch DONE (async — GPU working)");
         } else {
             // Ungraphed step 0: run everything, then capture + compose.
             // Ensure params are initialized for the first step.

crates/ml/src/trainers/dqn/trainer/training_loop.rs

@@ -1514,22 +1514,13 @@ impl DQNTrainer {
                     break;
                 }
             }
-            let sample_start = std::time::Instant::now();
-            // Vaccine batch: only sample when needed (every 5th step for better diversity)
-            let vaccine_batch = if train_step_count % 5 == 0 {
-                agent.memory_mut().sample(self.current_batch_size).ok()
-            } else {
-                None
-            };
-            sample_total_us += sample_start.elapsed().as_micros() as u64;
+            // Vaccine batch sampling removed — vaccine now uses prev_grad_buf
+            // inside maintenance_child (graphed). No ungraphed PER sampling needed.
 
             {
                 let fused_start = std::time::Instant::now();
                 let fused = self.fused_ctx.as_mut()
                     .expect("Fused CUDA training is the only production path");
-                if let Some(vb) = vaccine_batch {
-                    fused.pending_vaccine_batch = Some(vb);
-                }
                 // PER sampling is inside run_full_step (per_sample child graph).
                 let _fused_result = fused.run_full_step(&mut *agent, &self.device)
                     .context("Fused CUDA training step failed")?;