Stale inuse.<pid>.lock files left behind on unclean exit (SIGKILL / crashes)

## Summary

The agency CLI / Copilot CLI does not reliably unlink its `inuse..lock` file on every exit path. When the process is sent SIGKILL, crashes, or exits via certain unclean paths, stale lock files accumulate in `~/.copilot/session-state//`. Future attempts to resume the affected sessions then incorrectly believe the session is still in use.

## Symptom (downstream)

We hit this in [copilot-ide](https://github.com/shachaf-ashkenazi/copilot-ide), an Electron wrapper around `agency copilot --acp` / `agency copilot`. After any unclean exit, opening an old session shows a "Force resume?" prompt because the IDE sees a leftover `inuse..lock` and (correctly) refuses to overwrite a session that *might* still be open elsewhere.

User-visible bug we filed downstream: shachaf-ashkenazi/copilot-ide#18.

Reproduction is trivial: start a session (`agency copilot &`), then SIGKILL the resulting PID (signal 9). The `inuse..lock` file in the session dir survives the dead process indefinitely.

## Current downstream workaround

We do PID-liveness probing in the IDE before trusting any lock file:
[`src/main/sessionLocks.ts`](https://github.com/shachaf-ashkenazi/copilot-ide/blob/master/src/main/sessionLocks.ts).

It parses the PID out of the filename, calls `process.kill(pid, 0)` (POSIX no-op probe) to check if it's alive, then runs `ps -p -o command=` to verify the live PID actually looks like the Copilot CLI (kernel reuses PIDs aggressively — a multi-day-old lock will frequently point at an unrelated `login` shell or `node` process). Locks owned by dead OR live-but-unrelated PIDs are unlinked.

Empirically: ~5–10 ms per lock probed, only paid when the PID is alive. Catches the bug but is a workaround for behavior the CLI should own.

## Proposed fix (upstream)

Any one of these would close the gap:

1. **Trap `SIGTERM` / `SIGINT` and unlink the lock before exit.** Covers the common "user closes terminal" / IDE-shutdown case. Doesn't help SIGKILL or hard crashes.
2. **Use OS-reclaimed advisory locks (`flock(2)` / `fcntl(F_SETLK)`)** instead of (or in addition to) the on-disk marker file. The kernel releases these when the holding process dies, regardless of how. This is the most robust option.
3. **Bake the PID-liveness reclaim logic into the CLI itself**, the same approach our `sessionLocks.ts` takes. Cheap, no semantic change, defends against pre-existing stale locks created by older versions.

(1) + (3) is probably the easiest path that handles all exit modes. (2) is the cleanest if you're willing to change the lock format.

## Notes

- We'll keep our PID-liveness reclaim in the IDE either way as defense in depth. Removing it would only be safe once we can assume all CLI versions in the wild handle this correctly, which won't be true for years.
- This affects both layouts of the binary that write into `~/.copilot/session-state/`: `~/.copilot-cli//copilot` (the agency-bundled inner binary) and `/opt/homebrew/Caskroom/copilot-cli//copilot` (the standalone Copilot CLI brew Cask). Both need the fix.

## Environment

- agency 2026.5.11.1 (wrapping Copilot CLI 1.0.45)
- macOS (aarch64)

## Related

- Downstream user-visible bug: shachaf-ashkenazi/copilot-ide#18
- Downstream tracker: shachaf-ashkenazi/copilot-ide#37

透過啟動 `agency copilot`、傳送 SIGKILL 並檢查產生的工作階段目錄來重現問題；比較 upstream 行為與 `src/main/sessionLocks.ts` 中的 downstream workaround。找出 CLI 的鎖定實作，並決定適當的回收或清理方式；成功的定義是，在兩種受支援的配置中，過時的鎖定在當機或強制終止後不再阻止工作階段恢復。

由索引模型根據 Issue 內容生成。