Heal stale active rows when a test runner dies mid-run

[soulgalore]

When a test runner is killed mid-job (docker restart, OOM kill, host reboot), nothing flipped its DB row off status='active'. /search
kept showing a ghost active entry indefinitely, and /result rendered "Streaming logs · running" over the dead worker's frozen log
output — because the queue still keeps the job in its active set until its periodic stall check moves it back to wait (~30 s
default), and getState() can't distinguish a real worker from an orphaned lock. Operators saw two jobs apparently running on one test
runner after a restart even though per-worker concurrency=1 was intact: one was real, the other was the abandoned ghost.

Three fixes close the gap from different sides. A global:stalled listener mirrors the queue back to the DB the moment a stalled job
goes back to wait. A 60-second reconcile pass picks up rows the listener missed because the server itself was also restarting when
the stall fired. And /result/:id peeks at the job's lock key in Redis so an orphaned-active job renders as "stalled, retrying"
immediately instead of waiting 30 seconds for the stall check to catch up. Stalled rows are mirrored as waiting rather than failed,
so a successful retry doesn't leave a phantom failure in the 24 h health pill.

Co-authored-by: Claude noreply@anthropic.com