Both specs that import completed-worker-retirement-fixture registered a
per-file `test.afterAll(cleanupCompletedWorkerFixture)`. That helper does
`rmSync(fakeCliDir, { recursive: true })`, and fakeCliDir holds the fake
`codex` executable itself, not just the ledger. Playwright loads the
module once per worker, so with `--workers=1` whichever spec file
finished first deleted the binary out from under every later file in the
run: the next spec launched a deleted path, wrote no lifecycle events,
and timed out waiting for them. Proven with a probe at the failing poll:
fakeCliDirExists false, ledgerExists false, ledger [].
The temp dir's lifetime is the worker process's, so the module now owns
its own teardown via process.once('exit') and no spec can end it. Ledger
accumulation was NOT the cause; both specs already truncate per test.
Separately, `liveness before reveal` sampled pty:hasPty once right after
backgroundMountTab, which only waits for the pane manager to exist. Main
learns the session when the pane's deferred reattach lands, so the sample
raced it (2 failures in 10 runs, order-independent). It now polls, and
stays soft so a slow attach cannot short-circuit the tab-teardown
assertion that follows. A restarted main that never attaches stays false
for the whole 20s window, so both assertions still fail on unfixed code.
Pre-commit hook skipped (ERR_PNPM_CMD_SHIM_CHMOD on this worktree's
symlinked node_modules); oxlint and oxfmt run manually on all three files.
Jinwoo-H
50b9327c3d