There's an isolation leak somewhere. It may be just in the _xxsubinterpreters module, but I suspect it's not.

See#99114 (comment).

Reproducers:

• https://gist.github.com/tonybaloney/262986212e1061b97908657a53a605d6
• https://gist.github.com/tonybaloney/504be6d1541610f50963c32a3f1c379d

FYI, I see crashes on this fairly infrequently.

Linked PRs

• gh-105699: Use a Thread-Local Variable for PKGCONTEXT #105740
• [3.12] gh-105699: Use a Thread-Local Variable for PKGCONTEXT (gh-105740) #105765
• gh-105699: Use a Linked List for the PyModuleDef Cache #106899
• gh-105699: Fix a Crasher Related to a Deprecated Global Variable #106923
• gh-105699: Fix an Interned Strings Crasher #106930
• [3.12] gh-105699: Fix an Interned Strings Crasher (gh-106930) #106963
• [3.12] gh-105699: Fix a Crasher Related to a Deprecated Global Variable (gh-106923) #106964
• gh-105699: Add Some Stress Tests for Subinterpreter Creation #106966
• gh-105699: Use a _Py_hashtable_t for the PyModuleDef Cache #106974
• [3.12] gh-105699: Add some stress tests for subinterpreter creation (GH-106966) #107012
• gh-105699: Disable the Interpreters Stress Tests #107354
• [3.12] gh-105699: Disable the Interpreters Stress Tests (gh-107354) #107357
• [3.12] gh-105699: Use a _Py_hashtable_t for the PyModuleDef Cache (gh-106974) #107412
• gh-105699: Re-enable the Multiple-Interpreters Stress Tests #107572
• [3.12] gh-105699: Re-enable the Multiple-Interpreters Stress Tests (gh-107572) #107783
• [3.12] gh-105699: Re-enable the Multiple-Interpreters Stress Tests (gh-107572) #112483