Health-check fix batch (CREST_AI 3dca271)

V30 operations: episode final products (consolidated stats + event-scoped nowcast verification replace the running event on end), CPU rung demoted to last resort behind worker claims, provenance stamps in manifests, /api/health self-monitoring, runner-split plumbing (EVENT_TICK / EVENT_RUNNER_URL), event-score labels + no Model-options panel in nowcast mode

V29 fixed episode windows: anchored at trigger-12h (no moving window), USGS-obs-based episode end, frame carry-forward on republish, warm-start IC from previous depth frame, in-flight queue shown in the event panel

queue sweep: stale = spec queued-stamp age + no fresh claim (id age alone swept re-enqueued episodes and killed a live claim mid-solve)

event ladder: delegated worker hand-off frees the runner (no more result-wait/CPU while claimed); new detections outrank episode re-sims; fresh claims survive re-enqueue; event_run gains t0 backdating + defer for refills

Nowcast MRMS rain-radar overlay (24h/48h/7d anim) + persistent 2D depth legend + queue worker wake-ping

V27 P1: event job queue (shadow mode) — bundle EF5 grids + spec for GPU workers