wait_for_warmup_jobs — dedup state writes + adaptive poll backoff

Motivation

wait_for_warmup_jobs polled k8s_warmup_status on a fixed 15 s cadence and wrote the same record_task_progress + update_state combo on every tick — including when nothing changed. A long warmup (20-30 min) produced 80-120 identical Table writes; under multiple concurrent DB warmups the workload Storage Table hit throttling. The polling itself also fanned out 6 K8s GETs per call, multiplied across every active warmup.

User-facing change

Faster + cheaper warmup waits. The chip strip update latency stays unchanged for actual transitions (still pulses at poll_seconds), but quiet periods stop generating Table writes and stretch the K8s poll cadence to 2x → 4x (capped at 60 s).

API / IaC diff

• api/tasks/storage/helpers.py::wait_for_warmup_jobs
• Track a (nodes_ready, nodes_failed, nodes_active, total_jobs) signature; skip record_task_progress + update_state when the signature is unchanged from the previous tick.
• quiet_ticks counter: after 3 unchanged ticks sleep 2 * poll_seconds, after 6 ticks sleep 4 * poll_seconds, hard ceiling 60 s.

Validation

• uv run pytest -q api/tests -k warmup — 84 passed.
• uv run ruff check api/tasks/storage/helpers.py — clean.