diag: eprintln per-step in training loop

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

@@ -1175,6 +1175,7 @@ impl DQNTrainer {
         }
 
         for _step in 0..num_training_steps {
+            if _step < 3 { eprintln!("H100_LOOP: step {_step}/{num_training_steps} — sampling PER"); }
             let sample_start = std::time::Instant::now();
             let (batch, vaccine_batch) = {
                 let agent = self.agent.read().await;