bug(#830): watch next-gen — P0/P1 issues found in post-merge review #834

Post-Merge Review Findings for PR #830

PR #830 (Watch Next-Gen) merged with strong architecture but several issues that need follow-up fixes.

P0 — Must Fix

1. CLI flags not wired to config — --notify-level, --overnight-start/end, --sentinel-file\ are parsed from args but never passed to \loadWatchConfig(). These flags have no effect at runtime. Fix: add them to the \cliOverrides\ object.

2. \isProcessAlive()\ EPERM bug — On Unix, \process.kill(pid, 0)\ throws EPERM when the process exists but caller lacks permission. Current code treats any error as 'dead'. Fix: check \�rror.code === 'EPERM'\ → return true (alive).

P1 — Should Fix

3. Path traversal in WorktreeBackend —
   elativePath\ parameter in
   ead()/\write()\ is not validated for ..\ segments. Fix: validate or scope the FSStorageProvider.

4. Auth status stderr parsing — \gh auth status\ writes to stderr even on success. \�xecFileSync\ only returns stdout → auth drift detection broken. Fix: use \spawnSync\ or \gh api user -q .login.

5. *\invokeFleet()\ ignores --agent-cmd* — Fleet dispatch hardcodes 'copilot'\ instead of using \config.agentCmd. Breaks environments using wrapper commands.

6. --state-backend\ not validated — Invalid values are saved to config and silently fall back at runtime. Fix: validate upfront against allowed values.

7. SDK console.log — \github.ts:249\ logs directly to stdout. SDK should return status; CLI decides how to log.

P2 — Nice to Have

8. BOM characters in cleanup.ts and cleanup-ceremony.md
9. Changeset docs say PID in .squad/\ but implementation uses \os.tmpdir()\
10. No integration tests for end-to-end watch flow
11. Fleet dispatch has no tests

Copilot Comment Validity

11/12 Copilot comments were valid — very high signal-to-noise ratio on this PR.

Priority: P0 items should be fixed ASAP — users setting CLI flags will get no behavior change.