GeezerVison

Since this is a rollback of the original fix plus a new, debatably dodgy fix on top, can we just do a regular rollback of #124648 first? I think we need more time for the proper fix and I'd rather release 3.13 (and 3.12, really) with known issues than introduce wholly new ones.

You should rollback GH-124648, it was not as safe a change as I thought it was. I thought it only kicks into action of you are using "basic single-phase init" extensions and at that point you are already on shaky ground. However it actually kicks in if your sub-interpreter is allowed to use those extensions, i.e Py_RTFLAGS_MULTI_INTERP_EXTENSIONS is set. I didn't realize that detail until investigating the gh-124785.

With GH-124648 rolled back, you can make a separate decision if you want to merge this PR into a RC. It fixes a use-after-free segfault bug. However, those bugs only get triggered if you are on the shaky ground as explained above (embedding case with legacy sub-interpreters). So, I think you would be well justified to say you don't want to merge this given we are already into RC2. If the ABI change is a problem, I think it can be re-worked to avoid that. Then it could go into 3.13.1 as just a regular bugfix.

Closing, for the 'main' branch I think GH-124808 is a better fix. For 3.13 (and 3.12) I'll work on either a backported version of GH-124808 or potentially a re-worked version of this PR.