Ignore plugin-generated members when inferring PEP 695 variance

[allanlewis]

Important

AI / LLM disclosure. This pull request was produced by an AI coding agent — Anthropic's Claude Code (model Claude Opus 4.8). The diff, the added tests, and this description were generated in their entirety by the model. I have reviewed the change and take responsibility for it, but want to be fully transparent that I did not write it by hand.

I'm aware of the CONTRIBUTING.md guidance on LLM-assisted contributions, in particular that mypy discourages use of LLMs by new contributors and that PRs from new contributors that are mostly LLM-generated may be closed. As a first-time contributor I didn't want to spring this on anyone unannounced, so I opened a Gitter thread to discuss it first, and I'm filing this as a draft: https://matrix.to/#/!FjKyUUyKpKFUvCNnMz:gitter.im/$NQqpiczcc57aIWYrrmkPXWl6GFA7cD4gZhgtKPslYS0?via=gitter.im
If this isn't a welcome way to contribute, please feel free to close it.

What

PEP 695 variance inference treats generic @attrs.define / @attrs.frozen classes as invariant — and empty ones as contravariant — even when the type variable is used only covariantly. Removing the decorator makes the same class infer covariant, as expected.

@attrs.frozen
class Box[T]:
    def get(self) -> T: ...

bi: Box[int]
bo: Box[object] = bi  # error today; should be allowed (Box is covariant)

Why

infer_variance (mypy/subtypes.py) iterates over every member from all_non_object_members, including members synthesized by plugins. Two attrs-generated members reuse the class's own type and skew inference:

Synthesized member	Type after self-binding	Effect on T
__lt__ / __le__ / __gt__ / __ge__	(self, other: Self[T]) -> bool	contravariant T
__attrs_attrs__ (with a T-typed field)	ClassVar[tuple[Attribute[T], ...]]; Attribute is invariant	invariant T

The dataclass plugin's __replace__ (synthesized on Python 3.13+) has the same effect on frozen dataclasses with fields.

This is the same class of bug as #17783, which special-cased the dataclass plugin's internal __mypy-replace; attrs uses different member names, so that fix never covered it.

Fix

Skip members flagged plugin_generated during variance inference. A user's own fields and methods are never plugin_generated, so they continue to drive the result; only synthesized members are ignored. This subsumes the existing __mypy-replace special case (kept in place for clarity).

sym = info.get(member)
if sym is not None and sym.plugin_generated:
    continue

Tests

• testAttrsPEP695InferVariance (check-plugin-attrs.test) — a covariant attrs class infers covariant; a mutable-field class stays invariant; a class with a user-written T-parameter method stays contravariant (confirming only synthesized members are skipped).
• testPEP695InferVarianceInFrozenDataclassPython313 (check-python312.test) — a frozen dataclass stays covariant under --python-version 3.13.

Both fail without the change and pass with it; the full testcheck suite passes locally.

Relates to #17623.

[github-actions]

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