refactor(tools): read test code with the TypeScript parser in the async test-cleanup invariant

[mrgoonie]

Fixes #634. Refs #629, #632, #563.

What changed

tools/invariants/test-cleanup-retries-asynchronously.mjs no longer reads test code with a hand-written scanner. It parses each file with the TypeScript parser (the typescript dev dependency the repo already pins), choosing TypeScript, TSX or JavaScript with JSX by the file's extension.

• Calls are found in the syntax tree: a call expression whose callee is rmSync or a name the file gives it, x.rmSync, x["rmSync"] / x[`rmSync`], optional calls, calls through parentheses or type assertions, and fs.rmSync.call(...) / .apply(...). The call asks for retries when maxRetries is named among its arguments, including as a quoted key.
• Aliases come from the tree too: import specifiers, object destructuring, const name = ... and name = .... A mock factory's { rmSync: vi.fn() } object literal no longer counts as an alias.
• The exemption marker still has to be in a comment on the call's line or the line above. Comments are found by walking every token's leading trivia. JSX text is never read as trivia, so a marker written as JSX text does not count. commentsOf(source, path?) exposes this classification.
• Removed export: blankCommentsAndStrings is gone. Nothing used it since calls are found on the tree, and nothing outside the spec imported it.
• Unparsable files: a file that does not parse is named in a warning: note, because error recovery may read code after the error differently from how it was written.
• A fast path: files that never spell both rmSync and maxRetries are not parsed.

Skipped is reported as skipped

• check(name) results now carry skipped.
  • formatReport() (in tools/invariants/context.mjs, used by tools/check-invariants.mjs) prints SKIP for a skipped check.
  • The summary then reads 14 invariant check(s) passed, 1 skipped instead of all 15 invariant checks passed.
  • A check with failures is FAIL even if it also said it skipped.
• When the check skips: only when dependencies are not installed at all, meaning the error is ERR_MODULE_NOT_FOUND for typescript itself and there is no node_modules at the repository root.
• When it fails: a missing typescript with node_modules present, any other missing module, or any other load error. The failure carries the error message.
• CI: .github/workflows/ci.yml gets a second step, Repository invariants after install, right after Install and under the same if:. The pre-install step stays for prose-only changes.

A fresh worktree with no install was run end to end:

SKIP  test-cleanup-retries-asynchronously
      · dependencies are not installed, so the TypeScript parser is not there; CI runs this check again after install
14 invariant check(s) passed, 1 skipped

The four cases from the #632 review, each with a unit test

Case	main	this PR
Regex literal holding a backtick and a quote (/say: `Đã /u, /it's "(/g), then a call	call missed	found
String right after a keyword (return'(', case'(', typeof'(')	calls missed	found
Bracket access across a line break (fs[\n"rmSync"], fs?.[\n`rmSync`\n])	missed	found
Marker written as JSX text (<p>// invariant-allow: ...</p>), or inside a regex	honoured	not honoured; {/* ... */} inside JSX still is

There are also tests for:

• reading a .ts file as TypeScript and a .jsx file as JSX;
• (await import("node:fs")).rmSync(...);
• .call/.apply;
• the parse warning.

Agreement with the parser

The test agrees with the TypeScript parser on where the comments are, in every file it reads covers every file the invariant scans (708 today). Comments are what the marker depends on.

• What it compares: commentsOf character by character against an independent AST view. That view takes getLeadingCommentRanges/getTrailingCommentRanges at every node and node-list edge, and excludes JSX text.
• Result: 0 disagreements.
• Mutation check: a deliberately broken comment walk, one that ignores comments before }, is caught.
• Runtime: about 3.2 s, down from about 5–6 s for the earlier whole-code comparison. It keeps an explicit 60 s timeout.

The repository-wide test finds no retrying synchronous removal in the repository's test code also stays.

For the record, from round 1: the earlier whole-code version of this comparison, run against main's scanner, found real code blanked in 32 test-scoped files. Among them were apps/runtime/test/action-speech.spec.ts:94, packages/contracts/test/activity-timeline.spec.ts:198 and tools/smoke-widget-tooling.mjs:55.

Detection compared with main, as a one-off script over all 1528 source files under apps packages packs tools examples: 0 files differ in the reported lines. There are no new false positives.

Performance

node tools/check-invariants.mjs, whole tree, Windows 11, three runs each:

	run 1	run 2	run 3
main	1630 ms	1473 ms	1585 ms
this PR	1306 ms	1193 ms	1217 ms

Verification

• tools/test/test-cleanup-retries.spec.ts + tools/test/invariants-context.spec.ts: 33/33 pass.
  • The new tests for the four cases, extensions and comment agreement fail against main's module.
  • The skip and fail tests run a copy of the check in a separate Node in three states: no node_modules, an empty node_modules, and a typescript package that throws on load. The formatReport tests cover the SKIP, FAIL and summary lines. Neither the exports nor the skipped field exist on main.
• pnpm typecheck: exit 0.
• eslint on the 5 changed JS/TS files: exit 0.
• pnpm invariants: all 15 pass with dependencies installed. In a fresh worktree with no install: 14 passed, 1 skipped, exit 0.
• .github/workflows/ci.yml parses with yaml with no errors. The verify job's steps are Repository invariants, Install [if], then Repository invariants after install [if].
• pnpm verify: exit 0, 7797 tests passed.

Docs impact

None. This is an internal invariant and its runner. No doc in docs/ or on the official site describes them.

[mrgoonie]

Review attestation: ready to merge at eca2c89892012982e3fb1c9425259ce902f39aec, reviewed by agent:code-reviewer.

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