test: own the temp namespace in the test runners

[Astro-Han]

Summary

Test suites reach the temp directory through 933 mkdtemp(join(tmpdir(), …)) calls spread over 250 files, and nothing removes what they create. On a developer machine that runs the suite repeatedly this accumulates without bound — the tmpdir this was found on held over 47,000 stale maka-* directories. CI containers are discarded after a run, so the cost lands entirely on local development: slower tmpdir traversal, noise for anything that scans it, and enough churn that "did this run leak?" stops being measurable.

The obvious fix is to give each call site a cleanup array and an afterEach, which is the convention builtin-tools-file-worker.test.ts already follows. Applied across the repo that is the same eight lines of book-keeping in 250 files — roughly 2,000 lines whose correctness depends on nobody ever forgetting them again, which the current state shows is not a safe assumption.

So this puts the temp namespace where the process lifetime already is. runWorkspace is the single spawn point for every workspace suite: it now hands each test process its own TMPDIR/TMP/TEMP and removes the tree when the workspace finishes, however it finishes. The headless runner gets the same treatment — it already isolates HOME, the XDG directories and APPDATA into a tree it deletes on exit, and simply left the temp namespace pointing at the host. No call site changes, and nothing new for a future test author to remember.

Ownership is the whole point, so the exit paths matter: the tree is created before the try that removes it covers everything after it, including a synchronous throw from spawn, and a removal that fails warns rather than replacing whatever result the run was already reporting.

Verification

packages/runtime against the same build, once through the runner and once directly (npm run test:dist, which is what the runner invokes and therefore the no-isolation control), counting top-level maka-* directories in the host tmpdir before and after:

	tests	pass	fail	skipped	leaked dirs
direct, no isolation	3167	3158	0	9	51
through the runner	3167	3158	0	9	0

• node --test scripts/run-workspace-tests-parallel.test.mjs scripts/run-headless-tests.test.mjs — 7/7 pass.
• Both new behaviours were checked against their own removal rather than assumed: reverting the implementation file turns the spawn-throw test red, and dropping the rm turns the three namespace tests red.
• npm run test:scripts (the one npm test path that does not go through the runner) measured at zero leaked directories already, so both paths are now clean.
• npm run lint, npm run format:check — pass.

Known boundaries

• Running a suite outside the runner (node --test by hand in a package) still leaks, by design: the isolation belongs to whatever starts the test process, and a bare node --test has no such owner.
• packages/runtime-host builds its Unix socket endpoint under a hardcoded /tmp rather than os.tmpdir() (src/control/endpoint.ts), so it is unaffected either way; it already prunes its own stale directories.
• The Windows TMP/TEMP variables are set alongside TMPDIR but were not exercised — no Windows host was available.
• Not run: the full npm test across every workspace.