Windows npm test hangs — no suite timeout, job still running after 60+ minutes #481

`test (windows-latest)` in `.github/workflows/test.yml` never finishes. `npm test` starts and is still running an hour later. The job has no `timeout-minutes` (smoke has 15; this one inherits the 6h Actions default), so a hang just sits there and blocks the matrix.

macOS `npm test` on the same commits is ~4 minutes.

Evidence

All three current Windows-matrix runs are stuck on the Run npm test step. Logs are not downloadable until the job ends (`BlobNotFound`).

This is the first time `npm test` has been asked to run on `windows-latest` (Windows job added with #449 / hardening). It has never completed.

What `npm test` is on win32

`test:core && test:renderer && test:electron && test:memory`.

• renderer has `--test-timeout=20000`.
• `scripts/test-electron.js` on win32 runs a subset (WSL/doctor/fake-bin + several adapter files via `writeFakeBin`). It does not pass `--test-timeout`.
• Those adapter tests spawn `.cmd` fakes through `claude.js` `detached: true`. Some fakes intentionally never exit (`setInterval(() => {}, 10000)` in `writeSlowClaude` in `session-record.test.js` / `memory-record.test.js`). Cleanup is `runner.stopAll()` → `killTree` → `process.kill(-pid)` which `proc.js` already documents as falling back on Windows.

If stop does not reap the node grandchild behind the `.cmd`, the `node --test` process keeps an open stdio handle and never exits. That matches an infinite `npm test` with no failing assertion.

Same class of hole as #480 (Windows smoke pass C never sees Claude fake stdout). This one is worse for CI: the job does not fail, it just occupies a runner.

Fix bar

1. Put `timeout-minutes: 20` (or similar) on the `test` job so a hang is a red X, not a 6-hour checkout.
2. Pass `--test-timeout` in `scripts/test-electron.js` so a single wedged `waitFor` cannot hold the process.
3. Prove `stopAll` / `killTree` actually kills a `writeFakeBin` `.cmd` → node grandchild on win32 (the slow-claude `setInterval` fake is the fixture).
4. Do not "fix" this by dropping the Windows test job. The job exists so a dead win32 path cannot go silent.

Related: #480 (smoke pass C), #442 / #449 (Windows dispatch + CI job), #143 (Stop must kill the tree).