PR #1417 — Daemon Orphan Reap and Start Lock (#349)
The Broader Goal at the End of the Previous PR
Recent work hardened onboarding, heal/TPM autonomy, and ephemeral swarm execution. The broader goal remained containerized and OS-level agent sandboxing: credential masking and containment across every dispatch. That arc assumes the daemon is a trustworthy supervisor — if it dies, the workspace must still know which billed agent processes are alive and must not accidentally run two supervisors at once.
Strategic Shifts in This PR
No strategic shift. This is a correctness fix for two daemon-lifecycle gaps filed as gh:#349: orphaned children that survive a daemon crash with nobody left to reap them, and a classic check-then-act race on daemon start / watch start that can double-dispatch the same queued job.
What This PR Shipped
Gap 1 — orphaned child after daemon death. Daemon-launched jobs use subprocess.Popen(..., start_new_session=True), so the agent process is not in the daemon's process group. Reconciliation (_reconcile_daemon_jobs()) previously ran only inside the daemon poll loop. If the daemon itself died, running rows stayed status='running' forever and the live billed process disappeared from synlynk status / daemon status.
Fix: _synlynk_daemon_child_main() now calls SynlynkDaemon._reconcile_orphans_on_startup() before entering _run_loop(). That path calls the shared _reconcile_daemon_jobs() logic (not a parallel judgment): still-alive PIDs stay running (adopted/monitored); dead PIDs settle through the existing GTV path (timed_out / failed_unverified / etc.). Parallel work on gh:#136 that extends the same reconciler therefore stays compatible.
Gap 2 — concurrent start TOCTOU. WatchDaemon.start() / SynlynkDaemon.start() checked _is_running() then later re-exec'd and wrote the pidfile. Two near-simultaneous starts could both pass the check.
Fix: exclusive fcntl.flock(LOCK_EX | LOCK_NB) on a sibling lockfile (<pidfile>.lock, still resolved via _daemon_state_path() / _repo_common_dir()). The parent holds the lock across spawn until the child publishes a live pidfile (or a short timeout); the child then holds the lock for its lifetime. If the lock cannot be acquired, the existing "already running" message is printed — same UX, race closed.
Tests cover: alive orphan adopted on child startup; dead orphan reaped to timed_out; concurrent start() threads produce a single spawn; direct lock-helper exclusivity.
Brainstorm Visuals Used
None.
What This Achieved on the Path to Autonomy
A crashed or OOM-killed daemon no longer leaves billed agent processes invisible until a human finds them in ps, and laptop wake / launchd RunAtLoad races can no longer start two job-dispatch loops against the same queue. That is table-stakes supervisor honesty for any sandboxing story that depends on the daemon as the containment boundary.
Strategic Note: The Goal at the End of This PR
Same active goal — containerized and OS-level agent execution sandboxing — with one fewer way for the supervisor itself to lie about what is running. Keep gh:#136 (git-state cross-check in the same reconciler) and gh:#355 (HTTP handler) merges narrowly scoped so they do not fight this change.