Fix the cyclotron v2 health check for long batches #44

Parent

Spec: #22 — Generate text, an LLM action node for workflows. Carried in the spec as an independent companion change.

What to build

The cyclotron v2 health check reads a poll timestamp to decide whether a pod is alive. Any batch that takes longer than roughly thirty seconds leaves that timestamp untouched for the duration, so the check reports the pod unhealthy while a Postgres heartbeat from the same pod proves it is alive and working.

Fix the check so it reflects liveness rather than poll recency. Do not widen the staleness window — that only moves the threshold at which the same wrong signal fires.

This is independent of the LLM node. It does not block it and is not blocked by it. It is filed here because the spec carries it as a companion PR and it would otherwise be lost.

Acceptance criteria

• A pod processing a single batch for longer than the current staleness window is reported healthy.
• A pod that is genuinely stuck or dead is still reported unhealthy.
• The fix changes what the check reads, not the length of the staleness window.
• A test covers the long-batch case, which is the one the current check gets wrong.

Blocked by

None — can start immediately.