feat: hgb-v2 — regression heads predict rise, recovering flood warning lead

Rolling-origin evaluation (5 monsoon folds x 4 variants, P.1 + P.103;
results in models/eval_variants.json) showed the absolute-level target
alerting AT the crossing on essentially every event, while the rise
target (future max - current level, level added back at serving) gives
+6h on the hard 2024 crossings, +45h in 2025, fewer false alarms than
weighted/quantile variants, and ~11% better MAE. Weighted and quantile
variants rejected: more false alarms, no Brier-score calibration gain.

Ported to production: train.py fits rise in both eval and refit passes
(sigma/metrics computed in absolute space), bundles stamped hgb-v2 with
regression_target='rise', predict.py adds the level back for v2 and
stays compatible with v1 bundles, backtest_render.py mirrors the same
math. Regenerated backtest charts: 2024 first alert 11:00 24 Sep (6h
BEFORE the 17:00 crossing, was 18h after), 2025 alert 45h ahead, and
the record-peak underprediction is gone (rise models can exceed the
training max). The >=12h acceptance gate still fails honestly at +6h —
closing that needs rainfall inputs. New P.1 MAE 5.0/7.2/9.4 cm at
6/12/24h; docs updated throughout.

src/ml/train.py

@@ -126,7 +126,8 @@ def _p_warning_series(
     """Model score if a classifier head exists, else the sigmoid-derived fallback probability."""
     if head is not None:
         return pd.Series(head.predict_proba(X)[:, 1], index=X.index)
-    predicted_max = pd.Series(reg.predict(X), index=X.index)
+    # reg predicts the RISE over current level; add the level back
+    predicted_max = pd.Series(reg.predict(X), index=X.index) + X["level"]
     return 1.0 / (1.0 + np.exp(-(predicted_max - threshold) / sigma))
 
 
@@ -240,14 +241,22 @@ def train_station(
         )
         horizon_metrics: dict = {}
 
-        # --- regression head (max level) ---
+        # --- regression head (rise to future max) ---
+        # Target = future max MINUS current level ("rise"). Rises are far more
+        # stationary than absolute stages, which softens the cannot-exceed-
+        # training-max ceiling: on the rolling-origin harness (2026-08-12) the
+        # rise target moved P.1 first-alert leads from +0h to +6/+46h and cut
+        # the 2024 record-peak underprediction. Prediction = rise + level.
         reg_labeled = eval_Y[max_col].notna()
         reg = None
         if reg_labeled.sum() >= MIN_ROWS_FOR_HEAD:
+            rise_target = (
+                eval_Y.loc[reg_labeled, max_col] - eval_X.loc[reg_labeled, "level"]
+            )
             reg = _safe_fit(
                 _make_regressor(hgb_overrides),
                 eval_X.loc[reg_labeled],
-                eval_Y.loc[reg_labeled, max_col],
+                rise_target,
                 f"max_{h}",
                 skipped_heads,
             )
@@ -259,7 +268,10 @@ def train_station(
             test_labeled = Y_test[max_col].notna()
             if test_labeled.sum() > 0:
                 y_true = Y_test.loc[test_labeled, max_col]
-                y_pred = reg.predict(X_test.loc[test_labeled])
+                y_pred = (
+                    reg.predict(X_test.loc[test_labeled])
+                    + X_test.loc[test_labeled, "level"].to_numpy()
+                )
                 residuals = y_true.to_numpy() - y_pred
                 sigma_h = max(float(np.std(residuals)), MIN_SIGMA)
                 horizon_metrics["n_test"] = int(test_labeled.sum())
@@ -367,7 +379,7 @@ def train_station(
             reg = _safe_fit(
                 _make_regressor(hgb_overrides),
                 X.loc[labeled],
-                Y.loc[labeled, max_col],
+                Y.loc[labeled, max_col] - X.loc[labeled, "level"],  # rise target
                 head_key,
                 skipped_heads,
             )
@@ -401,7 +413,10 @@ def train_station(
 
     bundle = {
         "station_code": station,
-        "model_version": f"hgb-v1+{_git_short_sha()}",
+        "model_version": f"hgb-v2+{_git_short_sha()}",
+        # v2: regression heads predict the RISE over the current level; the
+        # serving side must add the level back. Old v1 bundles lack this key.
+        "regression_target": "rise",
         "trained_at": datetime.datetime.now().isoformat(),
         "sklearn_version": sklearn.__version__,
         "feature_names": feature_names,
@@ -428,7 +443,7 @@ def train_all(
     """Train and save every requested station's models. Returns the metrics.json payload."""
     models_dir = Path(models_dir)
     models_dir.mkdir(parents=True, exist_ok=True)
-    model_version = f"hgb-v1+{_git_short_sha()}"
+    model_version = f"hgb-v2+{_git_short_sha()}"
 
     station_results: Dict[str, dict] = {}
     for station in stations: