fix(worker): Only publish occupancy between first task and drain

[enochtangg]

Pods that are not handling traffic still publish occupancy, which pulls the KEDA pool average down during rollouts and blocks scale-up. This limits publishing to the window where a pod's slots actually reflect pool load.

Change

A pod now waits until it accepts its first task before it publishes occupancy. A push worker can only receive a task after it is SERVING and the broker is routing to it. So publishing starts once the pod is actually doing work, not when its first child finishes warming up.

A pod stops publishing occupancy as soon as it starts draining. Instead of leaving the last value in place, the worker removes the Prometheus series, so Prometheus marks it stale and it drops out of the average. Only the metrics thread writes the gauge, and it removes the series on every flush once draining has started.

Draining starts as soon as the worker sees the file /tmp/taskworker-draining. The pod's preStop hook creates this file when the pod is deleted, which currently is set to 15 seconds before SIGTERM arrives. That change is in a separate ops PR.

A new metric, taskworker.worker.occupancy.withheld, counts the times a pod skips publishing, tagged with the reason (no_task_yet or stopped). This makes the new behaviour easy to check during a rollout.

[linear-code]

STREAM-1843

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

Want reviews to match your repository better? Bugbot Learning can learn team-specific rules from PR activity. A team admin can enable Learning in the Cursor dashboard.

^{Reviewed by Cursor Bugbot for commit ef07cd0. Configure here.}