test: remove spec directories with the retrying async helper, since rmSync never retries on Windows

[mrgoonie]

Fixes #563. Refs #535, #543, #561.

Why

On Windows, rmSync(path, { recursive: true, force: true, maxRetries }) reports a held directory as EPERM
(Node 24) or EBUSY (Node 22) at once and never runs the retries. fs.promises.rm with the same options does retry. Measured on this
machine (Windows 11, Node v24.11.0). The directory was held as the working directory of a child process for 1.5 s:

Form	Result
rmSync(..., { maxRetries: 10, retryDelay: 100 })	threw EPERM after 0 ms
await rm(..., { maxRetries: 10, retryDelay: 100 })	removed the directory after 1536 ms, once the child exited

A held directory is also why many spec cleanups asked for retries in the first place.

What changes

• Every retrying rmSync in test code now awaits removeTestDirectory() from tools/test-cleanup.ts (added by
  test: wait for a timed-out body before cleanup, and retry removal where Windows retries #561). This covers 58 spec files and the apps/runtime/test/live-nodes.ts helper: 64 call sites in total. Each
  afterEach/afterAll hook that cleans up became async, and rmSync was dropped from imports that no longer use it.
  Calls that were already in an async context were changed in place:
  • the finally blocks in coding-journey and service-container-engine;
  • the cleanup list in task-dispatch-scoped, whose steps are already awaited;
  • liveNodes().stopAll();
  • the image-generator reference app's spec, which already imports repo sources.
• Two widget-dev tests keep a synchronous rmSync, now without maxRetries. They delete a watched folder and make
  it again within one turn of the event loop, and that synchronous step is what they test. A comment says why. The
  other widget-dev deletions use the helper.
• New invariant test-cleanup-retries-asynchronously (tools/invariants/, wired into pnpm invariants). It fails
  when test code calls rmSync( with maxRetries. Test code means specs, plus the files under test/ and e2e/
  folders in apps, packages, packs, tools and examples. The check finds each call's closing parenthesis by
  counting brackets, so an options object that spans several lines is still read whole. The failure message names the
  file and line and points to the helper. On main, it reports all 64 call sites. On this branch it passes.
• New spec tools/test/test-cleanup-retries.spec.ts:
  • unit tests for the invariant's matcher: a call that spans several lines, the promise form, calls without retries,
    and which paths count as test code;
  • a Windows-only test of the real behaviour, using a child process that keeps the directory as its working directory
    until it is released. In that test:
    • the synchronous form with maxRetries: 10 throws EPERM/EBUSY in under 500 ms, and the directory remains;
    • removeTestDirectory() succeeds only after the holder is released 700 ms later, which shows that the retries ran.

Changes after review (round 1)

• Merged main. fix(browser): close the browser when its page cannot be opened #587's packs/browser-playwright/test/new-page-failure.spec.ts is converted: its afterAll is async and awaits removeTestDirectory(dir). Without this change, the merged tree failed the new invariant.
• tools/smoke-widget-tooling.mjs is converted, and the invariant now covers it, because CI runs it on windows-latest. Its cleanUp() was already async, and it now awaits rm from node:fs/promises with the same retries.
• The invariant reads code, not comments or strings.
  • It blanks comments and the text of string literals before it matches. A call named in a comment or string is no longer flagged, and a ( inside a string no longer makes it read to the end of the file. Code inside a template literal's ${...} is still read.
  • It follows rmSync under a name the file gives it: rmSync as x, { rmSync: x }, or const x = rmSync.
  • A call that shows the failing form on purpose carries // invariant-allow: sync-rm-retries. Only the Windows test in the new spec does.
  • Known limits, documented in the module: options passed in a variable, and a regular expression literal that holds a quote.
• New matcher unit tests, one per case: a call in a comment, a parenthesis in a string, a call inside a string, template-literal expressions, each alias form, the allow marker, and the smoke script's scope.
• The doc comments no longer say "Under Node 24": Node 22 throws at once too.

Production code is out of scope and tracked in #626: apps/runtime/src/worker-process.ts:344 and apps/runtime/src/application/widget-dev-sessions.ts:645 still call rmSync with maxRetries.

Verification (Windows 11, Node v24.11.0, before review)

Check	Result
pnpm invariants	15/15 pass. On main, the new check fails with 64 call sites.
pnpm typecheck	exit 0
pnpm lint	exit 0
The 63 specs this touches, plus the 4 live-nodes.ts users, 3 runs in a row on the final head	63/63 files and 1054 tests passed on each run (an earlier set of runs had failures, explained below)
tools/test/test-cleanup-retries.spec.ts, 5 runs in a row	5/5 each run
pnpm test (full suite)	569 files passed, 7665 tests passed, exit 0

An earlier set of three batch runs had failures, and none came from the converted cleanup:

• In two of those runs, the first version of the Windows test failed. It used a child process with a fixed lifetime,
  and under load the synchronous removal succeeded, so the child was not holding the directory at that moment. The
  child now reports that it is running and holds the directory until it is released. The test then passed 5 runs in a
  row on its own and in all 3 later batch runs.
• In one of those runs, two tests in peer-delegation.spec.ts failed. It then passed 5 runs in a row on its own, and
  in all 3 later batch runs.

After review, at the new head (Windows 11, Node v24.11.0):

Check	Result
tools/test/test-cleanup-retries.spec.ts, 3 runs in a row	10/10 each run
packs/browser-playwright/test/new-page-failure.spec.ts	1/1
pnpm invariants	15/15 pass
pnpm typecheck	exit 0
eslint on the changed files	exit 0
node --check tools/smoke-widget-tooling.mjs	exit 0

Overlap with open PRs

• fix(widgets): re-arm a dev session on a changed folder id and bound persistent stat errors #614 (fix(widgets): re-arm a dev session on a changed folder id and bound persistent stat errors #611) changes the same "remade in one tick" tests in packages/core/test/widget-dev-engine.spec.ts and apps/runtime/test/widget-dev-sessions.spec.ts. fix(widgets): re-arm a dev session on a changed folder id and bound persistent stat errors #614 keeps rmSync(root, { recursive: true, force: true, maxRetries: 5 }) there, so whichever PR lands second must drop maxRetries while it keeps fix(widgets): re-arm a dev session on a changed folder id and bound persistent stat errors #614's test bodies. Otherwise the new invariant fails.
• feat(voice): hand a sentence naming no offered label to Clark's turn with the focused widget's actions #600 is merged. Its import-block conflict in apps/runtime/test/widget-perform.spec.ts is resolved here by keeping both imports. That spec and the cleanup spec pass (90/90), along with invariants, typecheck and eslint, on the merged head.
• fix(browser): close the browser when its page cannot be opened #587 is merged, and its spec is converted here.
• feat(widgets): let the person choose a project folder for widget dev sessions #612 and test(runtime): close every connection when a live node stops #601 merge without conflicts (checked with git merge-tree).

Follow-ups (not in this PR)

• fix(runtime): production cleanup uses rmSync with maxRetries, which never retries on Windows #626: the production sites worker-process.ts:344 and widget-dev-sessions.ts:645.

[mrgoonie]

Review attestation: ready to merge at 5c89d6d6ac23b99ad23e0c73b77cc6bcf430fa24, reviewed by agent:code-reviewer.

A push to this PR makes this attestation stale; the new head needs its own review.