Deploy reviewed operational cloud-sea model

[TIMtechnology]

Summary

• reconstruct Five-Woman Mountain training inputs from exact ECMWF IFS HRES 00Z single runs available before the decision cutoff
• replace executable Pickle deployment with a strict, point-bound JSON logistic model and fail-closed loader
• add chronological holdout gates, complete rule/terrain/water baseline, immutable candidate workflow, and production ML smoke checks
• bake the reviewed model and Wunvshan terrain snapshot into the image

Evidence

• 72 usable days: 22 positive / 50 negative
• final 15-day holdout: balanced accuracy 0.65, recall 1.00, Brier 0.209
• rule baseline: 0.50 / 0.60 / 0.250
• bash scripts/validate-release.sh: 39 passed; pip-audit 0; npm audit 0; frontend build and bundle budgets passed

Rollout

• merge after CI
• back up production DB, compose, secrets and previous image
• import the exact-run forecast archive for future candidate retraining
• enable only the reviewed Wunvshan model; all other points remain rule-only
• smoke test and monitor with automatic image rollback on health failure