feat(dqn): wire gradient explosion and checkpoint save metrics into trainer

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

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

@@ -2637,6 +2637,7 @@ impl DQNTrainer {
                     warn!("  • Reducing learning rate (current: {:.2e})", self.hyperparams.learning_rate);
                     warn!("  • Enabling target network soft updates (Polyak averaging, tau=0.005)");
                     warn!("  • Adjusting reward scaling");
+                    training_metrics::record_gradient_explosion("dqn", "current");
                 }
             }
 
@@ -2799,12 +2800,15 @@ impl DQNTrainer {
                     }
                 }
 
+                let ckpt_size = checkpoint_data.len() as f64;
+                let ckpt_start = std::time::Instant::now();
                 let best_checkpoint_path = checkpoint_callback(
                     epoch + 1,
                     checkpoint_data,
                     true, // is_best flag
                 )
                 .context("Failed to save best checkpoint")?;
+                training_metrics::record_checkpoint_save("dqn", "current", ckpt_start.elapsed().as_secs_f64(), ckpt_size);
 
                 info!("Best model saved to: {}", best_checkpoint_path);
             }