Prevent duplicate output in 3.14

[Erotemic]

Fix Python 3.14.2 multiprocessing shutdown behavior by ensuring only the owner process registers the atexit reporting hook.

Problem

When running the complex example with LINE_PROFILE=1 and PROFILE_TYPE=explicit on Python 3.14.2, multiprocessing using the forkserver start method can cause non-owner processes (e.g., the forkserver process / child processes) to also register the atexit hook. At interpreter shutdown, those processes may try to print/write profiling results, resulting in:

• duplicated “Wrote profile results …” messages,
• empty or truncated profile_output.txt,
• output that continues after the shell prompt appears.

Root cause

atexit.register(...) was executed in processes that should not own reporting. Even if show() includes guards, shutdown ordering and partially-torn-down globals can make those guards unreliable. The safest approach is to avoid registering the hook outside the owner process.

Fix

Move/guard atexit.register(...) so it occurs only after establishing ownership (PID/parent/process role) and only in the designated owner process. Non-owner processes never register the shutdown reporter.

---

The above is a GPT description of the patch, which I used heavily to find the problem. I believe the diagnosis is correct, but I don't understand what changed in 3.14.2 to cause the problem. Pushing up the patch as is to see if it fixes the dashboard.

[Erotemic]

I think I see the issue now. It's a combination of 2 things, which is why 3.14.0 worked, but 3.14.2 broke.

First: In 3.14 they changed the default process pool start method

Changed in version 3.14: The default process start method (see Contexts and start methods) changed away from fork. If you require the fork start method for ProcessPoolExecutor you must explicitly pass mp_context=multiprocessing.get_context("fork").

However, there was an issue that prevented forkserver from setting sys.argv correctly, and when that was fixed:

gh-143706: Fix multiprocessing forkserver so that sys.argv is correctly set before main is preloaded. Previously, sys.argv was empty during main module import in forkserver child processes. This fixes a regression introduced in 3.13.8 and 3.14.1. Root caused by Aaron Wieczorek, test provided by Thomas Watson, thanks!)

Then the failure started showing up.