fix(ml): serve the forecast rain the model was trained with

build_features cut the rain series to the gauge grid before the forward sum,
and the grid ends at the newest reading, so the row every live forecast is
made from had rain_fc24 = NaN. Training and all evaluations had the real
Open-Meteo forecast there. The forward sum now runs over the grid extended by
24 h; earlier rows are unchanged.

Harness (new opt-in variant rise_rain_fc24_missing = rise_rain with
rain_fc24 blanked on test rows, five folds, --from-cache): P.1 and P.103
nearly unaffected, P.21 false alarms 10 -> 5 and MAE 0.232 -> 0.215 with the
fix. Write-up in docs/FLOOD_FORECASTING.md section 5; results in
models/eval_2026-10-18_fc24.json.

Also: HII mirror rows in one hour now resolve to the latest reading via a
stable sort on the raw time (was an unstable sort on the floored hour, i.e.
row-order dependent). Needed for an exact Rust port.

tests/test_rain_serving.py

@@ -0,0 +1,64 @@
+"""rain_fc24 at the newest grid row (the one serving predicts from)."""
+
+import numpy as np
+import pandas as pd
+
+from src.ml import features
+from src.ml.data import _prepare_hii_levels
+
+
+def _grid(hours=48):
+    idx = pd.date_range("2026-09-01", periods=hours, freq="h")
+    df = pd.DataFrame(
+        {
+            "timestamp": idx,
+            "station_code": "P.1",
+            "water_level": np.linspace(1.0, 2.0, hours),
+            "discharge": np.nan,
+        }
+    )
+    return features.make_hourly_grid(df)
+
+
+def test_forward_rain_reaches_past_the_last_reading():
+    grid = _grid()
+    idx = grid.observed.index
+    rain = pd.Series(
+        np.arange(1, len(idx) + 49, dtype=float),
+        index=pd.date_range(idx[0], periods=len(idx) + 48, freq="h"),
+    )
+    X = features.build_features(grid, "P.1", rain=rain)
+    last = idx[-1]
+    expected = rain.loc[last + pd.Timedelta(hours=1) : last + pd.Timedelta(hours=24)]
+    assert X.loc[last, "rain_fc24"] == expected.sum()
+    # rows whose window lies inside the grid are what they always were
+    old = rain.reindex(idx).shift(-1).iloc[::-1].rolling(24, min_periods=1).sum()
+    inner = idx[:-24]
+    np.testing.assert_allclose(X.loc[inner, "rain_fc24"], old.iloc[::-1].loc[inner])
+
+
+def test_forward_rain_nan_without_forecast_hours():
+    grid = _grid()
+    idx = grid.observed.index
+    rain = pd.Series(1.0, index=idx)  # nothing after the last reading
+    X = features.build_features(grid, "P.1", rain=rain)
+    assert np.isnan(X.loc[idx[-1], "rain_fc24"])
+    assert X.loc[idx[-2], "rain_fc24"] == 1.0
+
+
+def test_hii_rows_latest_reading_in_an_hour_wins():
+    raw = pd.DataFrame(
+        {
+            "timestamp": pd.to_datetime(
+                ["2026-10-01 10:40", "2026-10-01 10:10", "2026-10-01 11:00"]
+            ),
+            "station_code": "P.1",
+            "wl_msl": [301.5, 301.2, 301.7],
+            "discharge": [10.0, 9.0, 11.0],
+        }
+    )
+    out = _prepare_hii_levels(raw)
+    assert list(out["wl_msl"]) == [301.5, 301.7]
+    assert list(out["timestamp"]) == list(
+        pd.to_datetime(["2026-10-01 10:00", "2026-10-01 11:00"])
+    )