Follow-ups from #919: zombie classification edges, caller list, mail-lock test gaps #927

Follow-ups from the review of #919 (non-blocking), tracked here as one issue.

Classification edges (pidState, packages/agent/src/lib/mail-lock.ts)

1. A process whose thread-group leader called pthread_exit while other threads keep running shows Z in /proc/<pid>/stat and reads "gone". Not reachable for Bun or Node lock owners; note it in the doc comment.
2. A zombie owned by another uid stays "alive" (kill(pid, 0) returns EPERM before the stat read), so its lock is never reclaimed. Unchanged behaviour; decide whether to read /proc on EPERM too.
3. A /proc mounted from a different pid namespace can misclassify a pid (this already affects the start token).

Callers
4. The PR's caller audit missed several production callers (status.ts, launch-attestation.ts, launch-snapshots.ts, connection-state.ts and others found by git grep); the review found all of them safe with a zombie read as gone. Keep a complete list in the code or the docs.
5. office.ts (~109): pidState(pid) === "gone" now subsumes the linuxProcessExited(pid) arm on Linux; the second arm can go.

Tests
6. Add a zombie-owner case for the CLI's tryAcquireMailLock in packages/cli.
7. If the zombie helper throws after spawn (its 10 s deadline), the child is never registered for afterEach; register it before the wait loop. Add cases for the readProcStatFields branches (no ), a truncated line).
8. packages/agent/test/mail-lock.test.ts's contended synchronous-acquisition test fails intermittently with EPIPE (3 of 10 runs on unmodified main on one Linux host; passes standalone). Find the cause, and check whether CI runs the packages/agent tests at all.

Refs #713, #919.