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.
