test_runner: --test-force-exit does not end a run held open by a leaked child with inherited stdio (Linux)

[thejebo]

Version

v22.23.3, v24.21.0

Platform

Linux x64: official node:22-bookworm and node:24-bookworm Docker images. Not reproducible on Windows.

Subsystem

test_runner

What steps will reproduce the bug?

// leak.test.mjs
import { test } from 'node:test';
import { spawn } from 'node:child_process';

test('leaks a child with inherited stdio', () => {
  spawn(process.execPath, ['-e', 'setTimeout(() => {}, 20000)'], { stdio: 'inherit' });
});

time node --test --test-force-exit leak.test.mjs

How often does it reproduce?

Every run we measured.

What is the expected behavior?

The run ends shortly after the test passes, as it does on Windows (about 1 s). --test-force-exit is documented to exit the process once all known tests have finished, even if the event loop would otherwise stay active.

What do you see instead?

The test passes immediately, but the run does not end until the leaked child exits: 20 s for a child that lives 20 s. That happened on both versions above, with --test-force-exit and with or without a preload that makes stdout blocking.

Our reading: the per-file test process does exit, but the child still holds a copy of that process's stdout pipe (inherited), and the parent runner waits for the pipe to close. Nothing inside the test process can close another process's copy of the pipe, so the test file itself cannot work around it.

Additional information

• The timings come from a test that leaks a child this way, in a larger project suite (thejebo/orbi-dev#373). The snippet above is that pattern cut down to its minimum; the reduced file itself was not re-run separately.
• On Windows, the same case with --test-timeout=15000 --test-force-exit exits in about 1.1 s.
• Related but distinct: test_runner: --test-force-exit with concurrency silently loses test verdicts (parent reports fewer tests than ran, exit 0) #64833 (--test-force-exit dropping queued test reports on POSIX asynchronous stdout pipes). A blocking-stdout preload fixes that one, but not this hang.