[dormouse-bot]

Fixes a flaky test, surface-ids.test.ts > waits out a lock a live holder keeps, and takes it once released. It failed on main in run 39457380049 with ENOTEMPTY: directory not empty, rmdir '…/surface-ids.json.lock'. The race is in the test, not in the allocator.

The test releases the other window's lock with rm(lock(), { recursive: true }), which deletes owner.json and then calls rmdir on the directory. Meanwhile, the waiting allocator keeps retrying rename(mine, lock) every few milliseconds. On Linux, a directory rename onto an empty directory replaces it. If a retry lands between the unlink and the rmdir, the allocator takes the lock and the rmdir hits its owner.json. The second ENOTEMPTY in the log, on the temp dir, follows from the first: the test threw while the allocator was still writing.

Production code already avoids this. removeLock in vscode-ext/src/surface-ids.ts moves the lock aside in a single rename before deleting it. The test now releases the lock the same way, through a releaseLock() helper. The other rm(lock()) call, in "releases only a lock it still holds", is left unchanged: it runs while the only allocator holds the lock, so nothing is waiting to race it.

Verification: I couldn't run vitest here because the CI sandbox has no network for pnpm install, so this PR's CI is the test run. In plain Node, I confirmed that renameSync(dir, emptyDir) succeeds and that the following rmdirSync throws ENOTEMPTY, matching the CI error.

[cloudflare-workers-and-pages]

Deploying mouseterm with    Cloudflare Pages

Latest commit:	bbc021f
Status:	✅  Deploy successful!
Preview URL:	https://e3009635.mouseterm.pages.dev
Branch Preview URL:	https://fix-ci-39457380049.mouseterm.pages.dev

View logs

[dormouse-bot]

CI hasn't run the test this PR fixes yet, so the fix is still unverified. Build & Test failed in pnpm test:other at its first step: spec-lint reports docs/specs/security-remote.md (7117 words against a budget of 7100) and docs/specs/security.md (2669 against 2650) over their word budgets. The && chain stopped there, before the vscode-ext tests ran. Main has the same failure and also has security-hosted.md over budget (run 39458825365). It comes from the base branch, not this diff, and #1250 re-baselines those budgets. Once main is green, a merge or rerun will give the real result for surface-ids.test.ts.

[dormouse-bot]

The fix is now verified in CI. I merged main, which is green again since #1250 landed, into this branch as bbc021f. Build & Test passed at that head (run 39484064013), and vscode-ext ran test/surface-ids.test.ts with all 17 tests passing. This clears the blocker in my earlier review. The PR is ready for a maintainer to land.