feat: rolling-origin event-aware evaluation harness for model variants

One fold per monsoon season (train <= 30 Apr, test Jun-Nov, 2021-2025)
replaces the single fixed holdout that contained only ~4 warning events.
Metrics are what matters operationally: sustained first-alert lead vs
each observed 3.70m crossing (two consecutive alerting samples required;
lookback floored at the previous event's end so multi-peak floods can't
launder lead credit), peak error from the prediction actually issued 24h
before the peak (3h match tolerance, null on outages), false-alarm
episodes (12h gap tolerance), MAE / flood-regime MAE, and a Brier score
on warning exceedance — included because sigma cancels algebraically in
any p>=0.5 alert metric, so lead times compare predictors while Brier
compares uncertainty models.

Variants: baseline_abs (current), rise (target = future max - current
level), rise_weighted (flood-regime sample weights 1x->5x), and
rise_quantile (q50/q90 heads, spread-implied sigma). Harness verified by
a 3-agent adversarial review (features bit-identical across fold
cutoffs; three metric flaws found and fixed before first use).

Also: features.build_labels/build_matrix gain stats_end so the rescue
quantile is computed from pre-cutoff data only, closing the label-
construction leak flagged in the earlier ML review.

src/ml/features.py

@@ -272,9 +272,17 @@ def _future_window_stats(col: pd.Series, horizon_h: int) -> Tuple[pd.Series, pd.
 
 
 def build_labels(
-    grid: HourlyGrid, station: str, horizons: Tuple[int, ...] = (6, 12, 24)
+    grid: HourlyGrid,
+    station: str,
+    horizons: Tuple[int, ...] = (6, 12, 24),
+    stats_end: Optional[str] = None,
 ) -> pd.DataFrame:
-    """Build max-level and threshold-exceedance labels for one target station."""
+    """Build max-level and threshold-exceedance labels for one target station.
+
+    ``stats_end`` bounds the data used for label-construction statistics (the
+    rescue quantile below): pass the training cutoff during evaluation so
+    test-period extremes cannot influence which training rows receive labels.
+    """
     idx = grid.observed.index
     observed_level = _series(grid.observed, station, "water_level", idx)
     warn_thr, danger_thr = get_thresholds(station)
@@ -283,10 +291,11 @@ def build_labels(
     # NOT to warn_thr: coupling it to the configurable threshold made raising a
     # station's threshold silently shrink its regression training set (P.5 lost
     # 34% of rows and +46% MAE when its warning went 3.0 -> 4.55).
+    stats_level = (
+        observed_level.loc[: pd.Timestamp(stats_end)] if stats_end else observed_level
+    )
     rescue_thr = (
-        float(observed_level.quantile(0.975))
-        if observed_level.notna().any()
-        else np.inf
+        float(stats_level.quantile(0.975)) if stats_level.notna().any() else np.inf
     )
 
     out: Dict[str, pd.Series] = {}
@@ -321,11 +330,12 @@ def build_matrix(
     df_long: pd.DataFrame,
     station: str,
     horizons: Tuple[int, ...] = (6, 12, 24),
+    stats_end: Optional[str] = None,
 ) -> Tuple[pd.DataFrame, pd.DataFrame, dict]:
     """Build (X, Y, meta) training/inference matrices for one station."""
     grid = make_hourly_grid(df_long)
     X = build_features(grid, station)
-    Y = build_labels(grid, station, horizons)
+    Y = build_labels(grid, station, horizons, stats_end=stats_end)
 
     keep = X["obs_age_h"].notna()
     train_start = TRAIN_START.get(station)