Flaky test: "terminates the whole managed tree" fails on Windows and macOS runners #74

Summary

The unit test "terminates the whole managed tree (grandchild included) before resolving" (packages/core/test/unit/mcp/mcp-remote-runtime.test.ts:1195 on main, c36d50f) fails now and then on CI. It fails in a different way on each OS, and reruns pass.

Failures

• Windows, on main (run 33529800770, 2026-09-01, test (windows-latest)):

  FAIL: terminates the whole managed tree (grandchild included) before resolving
    error:   Unexpected end of JSON input
  

• macOS, on a docs PR (run 34455726328, 2026-09-10, test (macos-latest)):

  The input did not match the regular expression /timed out after \d+s/. Input:
  'npm install of mcp-remote@0.1.38 timed out and its process tree could not be confirmed stopped (process group check failed (EPERM)). ...'
  

Likely causes

• Windows: the fixture is killed before it writes its pids. The test gives the fake npm timeoutMs: 300 (:1212). In that time the fixture has to start Node.js, spawn the grandchild, and writeFileSync the pids file. On a slow runner the kill can land after the file is created but before it is written, so the test reads an empty file. If the kill lands earlier, the read fails with ENOENT instead.
• macOS: EPERM from the group check. After SIGKILL, mcp-runtime-runner.ts polls process.kill(-pgid, 0) (:254) and treats anything but ESRCH as unconfirmed (:258). macOS can answer EPERM while a member of the group is a zombie that hasn't been reaped yet: the grandchild, reparented to launchd once the fake npm dies. Not confirmed.

Suggested fix

• Test: give the fixture time to write before the timeout fires. The fixture never exits, so something like timeoutMs: 3000 still exercises the timeout. Write the pids file atomically (write a temp file, then rename it), so a kill can't leave it empty.
• Runner: on EPERM, keep polling until the deadline instead of failing at once. Only report "could not be confirmed stopped" if the group still answers at the deadline.