[JukkaL]

Fixes mypyc/mypyc#1186. Fixes mypyc/mypyc#1103.

A native class that inherits from collections.defaultdict was treated like a regular dict subclass, but defaultdict instances have an extra field, so the native instance layout was too small and compiled code could segfault or behave incorrectly:

from collections import defaultdict

class DD(defaultdict[str, int]):
    def __init__(self) -> None:
        super().__init__(int)

DD()["x"]  # Segfault

Inheriting from collections.OrderedDict has the same problem (the compiled module failed to import).

The problem is more general: native classes assume that non-mypyc base classes have the layout of object (or dict/BaseException), which isn't true for many classes implemented in C (e.g. random.Random or io.BytesIO). Stubs don't describe instance layouts, so now a native class can only inherit from a non-mypyc class defined in a stub if the class is in an allowlist. Otherwise mypyc reports an error with the usual note about using @mypyc_attr(native_class=False) as a workaround. This also covers defaultdict and OrderedDict. Classes defined in Python sources are still accepted, unless they define __slots__ or are compiled separately into a C extension.

The allowlist only includes standard library classes with evidence of being subclassed in real code (at least two of the 1000 most downloaded PyPI projects, a large private codebase, or mypy itself) that are documented in the CPython docs and have a compatible runtime layout on Python 3.10-3.15, including the free-threaded build. Native subclasses of each entry were compiled and exercised on Python 3.10, 3.14 (including free-threaded) and 3.15. This keeps the list size manageable and avoids internal classes that are likely to change.

The entries are the full names mypy uses (those of the modules where typeshed defines the classes), so a test checks that they still match how mypy resolves the classes on the Python version and platform the tests run on.

Classes that keep weak references to the instance (threading.Thread, logging handlers, ...) are left out for now and listed in a comment, since weakrefs to native subclasses of interpreted classes aren't supported yet.

[p-sawicki]

should this be something like "defined in an extension module not produced by mypyc"? "separately compiled" could be confusing considering mypyc has a separate compilation mode.

[github-actions]

According to mypy_primer, this change doesn't affect type check results on a corpus of open source code. ✅