jobs reap — Kill Zombie running Rows for Real

Broader goal (end of previous PR)

PR #762 windowed sentinel_crit so lifetime sentinel.md history no longer kept platform ops RED. The scoreboard could see recent failures honestly — and immediately showed the next lie: hundreds of daemon_jobs rows stuck at status=running with dead PIDs.

What moved the goalpost

A same-day ops triage reaped 482 zombies by hand (45 in the three main project DBs). STALL_NO_OUTPUT and HARNESS_INTERNAL_TIMEOUT killed processes and wrote CRITICAL sentinels but often left the row running. _reconcile_daemon_jobs also left dead PIDs as open-ended unknown (or skipped null PIDs). Manual SQL is not a product.

What this PR ships

1. synlynk jobs reap — dry-run by default; --apply marks zombies timed_out / exit_code=-9; --all-projects walks every ~/.synlynk/projects/*/state.db.
2. Auto-reap on sentinel write — when _write_sentinel_alert emits STALL_NO_OUTPUT or HARNESS_INTERNAL_TIMEOUT, parse job-* and flip still-running daemon rows immediately.
3. Hardened _reconcile_daemon_jobs — null PID and dead PID (including non-child processes) become timed_out instead of lingering running / vague unknown.
4. Helpers: _pid_is_alive, mark_daemon_job_terminal, scan_zombie_running_jobs, apply_reap_zombies, auto_reap_job_from_sentinel.
5. Tests for dry-run/apply, auto-reap via sentinel, CLI parser, and reconcile expectation update.

On the long arc

Autonomous multi-agent dispatch needs a job table that matches process reality. Windowed sentinels (#751) fixed the alert log; jobs reap (#753) fixes the queue truth that platform ops L1 rates depend on.

New goalpost

• Operators have a one-command hygiene path (jobs reap --all-projects --apply).
• Future STALL/TIMEOUT cannot recreate the 482-zombie pile without an auto flip.
• Still open: agy timeout root cause (#750), cost capture gaps (#752), full #701 epic.