fix(runner): run file teardown after the file's tests in parallel

[Chemaclass]

🤔 Background

Related #1320

Under --parallel the runner dispatched a file's tests as a background worker and ran tear_down_after_script in the parent on the next line, so the hook released the fixture its own tests were still reading. The same file passed sequentially and failed in parallel.

💡 Changes

• Move the file's teardown into the worker, after the tests it wraps. call_test_functions already waits for its per-test workers, which is what makes the ordering hold.
• Publish a hook failure as a .result payload so the parent still counts it. A failure recorded inside a worker dies with the subshell, and without this the hook error printed while the run reported All tests passed.
• Drop the docs/ai-agents.md caveat that told authors not to release a shared fixture in the hook. It existed only to work around this.
• Waiting on the worker from the parent was the other option. Measured here it cost 12s of 52s, all of it two slow acceptance files that define the hook and stopped overlapping. The worker-side hook keeps the suite at 52.8s.