S41: recover stale runner job claims

[espadonne]

Summary

• add runner heartbeat active_job_ids reconciliation so abandoned running jobs assigned to an idle runner are cancelled before capacity is counted
• make shithubd-runner send an explicit empty active job set on idle heartbeats
• update runner API docs and incident/deploy runbooks with the production queue-wedge recovery path

Production finding

On 2026-05-17, production had one stale running job assigned to shithub-runner-shared-linux-1 while the runner was heartbeating idle. That consumed the runner's only capacity slot, backed up queued jobs, and eventually suppressed fresh repo-tree status dots when repo queued-run caps were reached. I manually cancelled run/job 67, verified the queue drained, and confirmed production is now idle with zero running workflow jobs.

Deploy note

Deploy web first because the heartbeat decoder rejects unknown JSON fields. Then redeploy shithubd-runner from the same trunk build so it starts sending active_job_ids.

Tests

• GOCACHE=/private/tmp/shithub-actions-stale-run-hotfix/.gocache SHITHUB_TEST_DATABASE_URL='postgres://shithub:shithub_dev@127.0.0.1:5432/postgres?sslmode=disable' go test ./internal/actions/lifecycle ./internal/web/handlers/api ./internal/runner/api ./internal/runner
• GOCACHE=/private/tmp/shithub-actions-stale-run-hotfix/.gocache make build
• git diff --check