GeezerVison

As a quick preliminary check, I would suggest running doc tests, since they indicated the issue at first and caused further findings (see #118891 for details).

Currently, it appears to me that the PR achieves the opposite of its purpose.

Doctest on main branch takes ~12-15 minutes
With PR it takes >40m (still running): https://github.com/Privat33r-dev/cpython/actions/runs/11082906901/job/30796516669?pr=4
Main (before PR): https://github.com/Privat33r-dev/cpython/actions/runs/11083066610/job/30796883045?pr=5

The test at https://github.com/Privat33r-dev/cpython/actions/runs/11082906901/job/30796516669?pr=4 was killed by the 60 minute timeout, less than half-way through the build (source files not fully read). This PR isn't running documentation tests as no files in Doc/ have changed and there's no NEWS entry currently.

A

This PR isn't running documentation tests as no files in Doc/ have changed.

That's how I triggered it in my repo. I suggest to add/remove space in Docs in this PR as well, so we can see the end result of the changes.

I probably should've added further context for clarity. I used this PR as a base to make a PR in my repository, with 1 more commit from me (space change in docs), so Doctest will start automatically. It took more than 1 hour and stopped halfway through, because the job had 1 hour limit set.
Privat33r-dev#4

At the same time, I made a control Doctest in my repo on main branch + minor doc change. It took 13 m:
Privat33r-dev#5

Closing this, as we don't need an urgent fix (3.13 goes back to generational collection)

Longer term, I expect reduced reference counting operations to make faster-cpython/ideas#693 viable.

The pre-scan to find live objects is definitely worth further investigation, though.