feat(ci): boot-time sweep requeues runs orphaned by the deploy restart #5468

Run #257 exposed the class: the runner executes inside the app container, so a push-to-main verify-deploy run is killed by the very deploy it verifies, ~40s in — leaving a zombie "running" row the hourly reaper eventually calls "abandoned". Under reaping alone that workflow could never go green on push (this is also what happened to stuck run #253).

On boot, any run still running from before this process started is dead with certainty (runs execute in-process; one container by design). The sweep fails each honestly as runner_restarted and enqueues one fresh attempt for recent victims, which then executes against the post-deploy world.

Guards:

• one retry per (workflow, commit) — a workflow whose execution restarts its own container cannot loop
• 15-minute requeue window — stale zombies are finalised but not re-run ahead of live work on the serial drain
• bounded (limit 50) and non-throwing, same contract as the reaper

Tests extend workflow-run-abandonment.test.ts: sweep ordering (sweep → reap → drain), one-shot flag set before the await, loop guard, window, bound, catch. This PR's own deploy is the live test: its push-triggered verify run will be killed by its deploy, then requeued by the new boot sweep and go green.

🤖 Generated with Claude Code