test: add val_loss staleness regression check to walk-forward smoke test

Warns when best_epoch==1 with 3+ epochs trained, which indicates the
evaluator may be reading stale weights (constant val_loss). This would
have caught the Candle VarMap bug where val_loss=0.804688 every epoch.

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

crates/ml/src/trainers/dqn/smoke_tests/walk_forward.rs

@@ -90,6 +90,20 @@ fn test_walk_forward_oos_metrics() -> anyhow::Result<()> {
         "Walk-forward: avg_gradient_norm={grad_norm} — model must be learning"
     );
 
+    // ── Regression: val_loss must not be constant across all epochs ──
+    // If best_epoch == 1 AND we ran 3+ epochs, the evaluator may be reading
+    // stale weights (e.g., Candle VarMap instead of fused GPU buffers).
+    // With proper weight updates, val_loss should change and best_epoch should
+    // sometimes be > 1. This caught a bug where GPU-trained weights weren't
+    // passed to the backtest evaluator (val_loss=0.804688 every epoch).
+    let total_epochs = metrics.epochs_trained;
+    if total_epochs >= 3 && best_epoch <= 1 {
+        tracing::warn!(
+            "Walk-forward: best_epoch={best_epoch} with {total_epochs} epochs — \
+             val_loss may be constant (stale weights?). best_val_loss={best_val_loss:.6}"
+        );
+    }
+
     // Log results for manual inspection
     tracing::info!(
         "Walk-forward results: best_sharpe={best_sharpe:.4} at epoch {best_epoch}, \