feat(models): train-time confidence stats stashed on _neurobridge_train_stats

- _compute_train_stats() captures median, std, n_train of the model's
own predict_proba on X_train. Joblib-roundtrip-safe.
- train() persists stats alongside _neurobridge_fp_cols and
_neurobridge_calibration. INFO log line now surfaces the median.
- Foundation for Day-7 T1B drift z-score in /predict/bbb.
- 2 new tests (TestTrainStatsMetadata): attribute presence + roundtrip.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

src/models/bbb_model.py

@@ -84,6 +84,26 @@ def _compute_calibration_bins(
     return bins
 
 
+def _compute_train_stats(
+    model: RandomForestClassifier,
+    X_train: np.ndarray,
+) -> dict[str, float]:
+    """Compute median + std of the model's own confidence on the training set.
+
+    Used as the reference distribution for runtime drift detection. All values
+    are floats so the dict is joblib-roundtrip-safe and JSON-serializable.
+    """
+    if len(X_train) == 0:
+        return {"median": 0.0, "std": 0.0, "n_train": 0}
+    proba = model.predict_proba(X_train)
+    confidence = proba.max(axis=1)
+    return {
+        "median": float(np.median(confidence)),
+        "std": float(np.std(confidence)),
+        "n_train": int(len(X_train)),
+    }
+
+
 def train(
     df: pd.DataFrame,
     label_col: str = "p_np",
@@ -134,11 +154,13 @@ def train(
     model._neurobridge_calibration = _compute_calibration_bins(
         model, X_test, y_test,
     )
+    model._neurobridge_train_stats = _compute_train_stats(model, X_train)
     logger.info(
         "Trained BBB classifier: n=%d, n_features=%d, classes=%s, "
-        "calibration_bins=%d",
+        "calibration_bins=%d, train_confidence_median=%.3f",
         len(y), X.shape[1], model.classes_.tolist(),
         len(model._neurobridge_calibration),
+        model._neurobridge_train_stats["median"],
     )
     return model
 

tests/models/test_bbb_model.py

@@ -161,3 +161,29 @@ class TestCalibrationMetadata:
         reloaded = bbb_model.load(artifact)
         assert hasattr(reloaded, "_neurobridge_calibration")
         assert reloaded._neurobridge_calibration == model._neurobridge_calibration
+
+
+class TestTrainStatsMetadata:
+    """Day 7 — T1A: train()-time confidence distribution stash."""
+
+    def test_train_attaches_train_stats_attribute(self, trained_model_and_features):
+        model, _ = trained_model_and_features
+        assert hasattr(model, "_neurobridge_train_stats")
+        stats = model._neurobridge_train_stats
+        assert isinstance(stats, dict)
+        for key in ("median", "std", "n_train"):
+            assert key in stats, f"missing key {key!r} in train stats"
+        assert 0.0 <= stats["median"] <= 1.0
+        assert stats["std"] >= 0.0
+        assert stats["n_train"] >= 1
+
+    def test_train_stats_survives_save_load_roundtrip(
+        self, trained_model_and_features, tmp_path: Path,
+    ):
+        from src.models import bbb_model
+        model, _ = trained_model_and_features
+        path = tmp_path / "m.joblib"
+        bbb_model.save(model, path)
+        reloaded = bbb_model.load(path)
+        assert hasattr(reloaded, "_neurobridge_train_stats")
+        assert reloaded._neurobridge_train_stats == model._neurobridge_train_stats