Health check (liveness) — always 200 while the process is up

Render polls this, so it deliberately does not fail on a broken dependency: the API serves reads perfectly well with Redis down, and returning non-200 would get the service restarted, turning a queue outage into a total one.

Read status (ok | degraded) and alerts[] instead of the HTTP code. alerts names any queue carrying failed jobs — the signal that was missing when every media job failed silently for months behind a green check.

For something that pages, use /v1/health/ready, which 503s when a dependency is down.