fix: update confusion matrix and score distribution functions to load pre-computed CV predictions

app/training/generate_paper_figures.py

@@ -164,9 +164,9 @@ def plot_roc_curves(ml: dict, dl: dict) -> None:
 def plot_confusion_matrix_lightgbm(ml: dict) -> None:
     name = "LightGBM"
     data = ml[name]
-    y_true = np.array(data["y_true"])
-    y_prob = np.array(data["y_prob"])
-    # Use Youden's J threshold if stored, else 0.5
+    # Load pre-computed CV predictions
+    y_true = np.load(MODELS / "lgbm_cv_ytrue.npy")
+    y_prob = np.load(MODELS / "lgbm_cv_probs.npy")
     threshold = data.get("optimal_threshold", 0.5)
     y_pred = (y_prob >= threshold).astype(int)
 
@@ -496,12 +496,13 @@ def plot_feature_importance(ml: dict) -> None:
 def plot_score_distribution(ml: dict) -> None:
     best = ml.get("_best_model", "LightGBM")
     data = ml.get(best, {})
-    if not data or "y_true" not in data:
-        print("  ⚠  No score distribution data")
+    y_prob_path = MODELS / "lgbm_cv_probs.npy"
+    y_true_path = MODELS / "lgbm_cv_ytrue.npy"
+    if not y_prob_path.exists():
+        print("  SKIP  No score distribution data")
         return
-
-    y_true = np.array(data["y_true"])
-    y_prob = np.array(data["y_prob"])
+    y_true = np.load(y_true_path)
+    y_prob = np.load(y_prob_path)
     threshold = data.get("optimal_threshold", 0.5)
 
     fig, ax = plt.subplots(figsize=(7, 4.5))