fix(codex): add invocation-scoped monitor lifecycle

[GigiTiti-Kai]

Summary

• add an opt-in --invocation-scope TOKEN monitor mode with a fresh app-server record and CAS lease per physical project/scope
• supervise and reap only the captured Codex TUI, app-server, and bridge-launcher jobs while preserving argv, exit status, and legacy no-scope reuse
• route hooks to the exact scoped server while keeping request, dispatcher, role, and child locks project-singleton
• make scoped dispatcher/role handoff exact-server-bounded and reject inherited scope state in nested no-scope launches
• document the lifecycle and fail-closed SIGKILL limitation

Related to #149.

Safety boundaries

• no daemon, dependency, process inventory, or worktree deletion code
• raw invocation tokens never enter paths or logs
• pidfiles are never signal authority; only direct shell jobs captured through $! are signalled
• different scopes remain isolated; a duplicate live scope fails closed
• remote/role descendant readiness remains the caller's responsibility

Verification

• focused monitor/launcher/resume/delivery: 259/259 passed (3 expected platform skips)
• scoped regression set: 5/5 passed
• spawn FD guard: 4/4 passed
• enforced-assertion baseline: 638/638
• Bash syntax, diff checks, and changed-diff gitleaks: passed
• whole local Bats aggregation reached 1644/1645; the only failure was the unrelated intermittent test_remote_engine_start_refusal.bats .config.lock race, which also reproduced in isolation (8/9). Required Linux/macOS sharded CI remains the merge gate.

Review

• task-by-task implementation and independent reviews completed
• final whole-branch review found and verified fixes for physical-path aliasing and inherited nested no-scope routing

[GigiTiti-Kai]

Ready for review. The final head is 66e13f6 and all 22 required checks are green, including all four Ubuntu and all four macOS Bats shards plus the native Windows leg.

The second CI run exposed a macOS-only fixture mismatch (/var/... vs physical /private/var/...). The production monitor was already correctly canonicalizing project identity; the fix canonicalizes the test fixture once. A symlink-path reproduction passes 24/24 monitor tests, and the native macOS shard now passes as well.

Could you review and squash-merge this when ready?