Sync from GitHub (tests passed)

app/ai_engine.py

@@ -687,7 +687,7 @@ def train_xgboost_model(
     latest_path = model_dir / f"xgb_{target_symbol.replace('=', '_')}_latest.json"
     model.save_model(str(latest_path))
     
-    # Save metrics
+    # Save metrics (including training symbols audit)
     metrics = {
         "target_symbol": target_symbol,
         "trained_at": datetime.now(timezone.utc).isoformat(),
@@ -699,6 +699,11 @@ def train_xgboost_model(
         "val_rmse": val_rmse,
         "best_iteration": model.best_iteration,
         "feature_count": len(feature_names),
+        # Audit: which symbols were used for training
+        "symbol_set_name": settings.symbol_set,
+        "training_symbols": settings.training_symbols,
+        "training_symbols_hash": settings.training_symbols_hash,
+        "training_symbols_source": settings.training_symbols_source,
     }
     
     metrics_path = model_dir / f"xgb_{target_symbol.replace('=', '_')}_latest.metrics.json"

app/features.py

@@ -236,8 +236,8 @@ def align_to_target_calendar(
         # Reindex to target calendar
         reindexed = df.reindex(target_index)
         
-        # Limited forward-fill
-        reindexed = reindexed.ffill(limit=max_ffill)
+        # Limited forward-fill (infer_objects fixes future downcasting warning)
+        reindexed = reindexed.ffill(limit=max_ffill).infer_objects(copy=False)
         
         aligned[symbol] = reindexed
     

app/inference.py

@@ -250,10 +250,11 @@ def build_features_for_prediction(
     # Get latest row
     latest = all_features.iloc[[-1]].copy()
     
-    # Ensure we have all required features
-    for feat in feature_names:
-        if feat not in latest.columns:
-            latest[feat] = 0.0
+    # Ensure we have all required features (avoid fragmented DataFrame)
+    missing_feats = {feat: 0.0 for feat in feature_names if feat not in latest.columns}
+    if missing_feats:
+        missing_df = pd.DataFrame(missing_feats, index=latest.index)
+        latest = pd.concat([latest, missing_df], axis=1)
     
     # Select only the features the model expects
     latest = latest[feature_names]