feat(sweeper): Phase 2D — Orphan revival and delayed-retry promotion

[smallchungus]

Summary

Phase 2D: the durability loop.

• internal/sweeper.SweepOnce does two passes:
  1. Orphan revival: finds status=running jobs whose worker_id has no live heartbeat (EXISTS heartbeat:<id> returns 0), calls MarkFailed with "worker died". Status goes back to queued (or dead if max attempts hit).
  2. Delayed-retry promotion: finds status=queued AND last_error IS NOT NULL AND next_run_at <= now() jobs and pushes their IDs onto the Redis stage queue. The last_error IS NOT NULL filter avoids re-pushing newly-enqueued jobs whose producer hasn't yet pushed them.
• New queue.IsWorkerAlive(workerID) reader (paired with the Heartbeat writer from 2C).
• New store.ListRunningJobs and store.ListReadyRetryJobs plus shared scanJobs helper.
• cmd/sweeper binary: 5 s tick, signal-handled graceful shutdown, slog JSON.

6 new tests (2 queue + 2 store + 4 sweeper integration). Existing 22 integration tests still green.

Closes the durability loop. Combined with Phase 2C, a worker that crashes mid-job is recovered automatically; failed jobs back off correctly.

Test plan

• make lint (0 issues)
• make test-unit
• make test-integration (sweeper, worker, queue, store, testutil, api — all green)
• make k8s-validate (8/8 valid)
• cmd/sweeper smoke run with docker compose stack — clean start/stop on SIGTERM
• CI green