Stabilize generic-not-last-base-class (PYI059)

[ntBre]

Summary

The tests were already in the right place, I just updated the documentation slightly to reflect the discussion here and on Discord.

Documentation: https://docs.astral.sh/ruff/rules/generic-not-last-base-class

Tests: https://github.com/astral-sh/ruff/blob/brent/release-0.12.0/crates/ruff_linter/src/rules/flake8_pyi/mod.rs#L51-L52

I considered mentioning the MRO explicitly and linking to this but decided that might be a bit too technical for the rule docs.

Now I see that ty links to the glossary, which would be another option.

Test Plan

Existing tests

[github-actions]

ruff-ecosystem results

Linter (stable)

ℹ️ ecosystem check detected linter changes. (+13 -0 violations, +0 -0 fixes in 5 projects; 50 projects unchanged)

apache/airflow (+4 -0 violations, +0 -0 fixes)

ruff check --no-cache --exit-zero --ignore RUF9 --no-fix --output-format concise --no-preview --select ALL

+ airflow-core/src/airflow/api_fastapi/auth/managers/base_auth_manager.py:78:22: PYI059 [*] `Generic[]` should always be the last base class
+ airflow-core/src/airflow/api_fastapi/common/exceptions.py:37:23: PYI059 [*] `Generic[]` should always be the last base class
+ airflow-core/src/airflow/api_fastapi/core_api/base.py:52:16: PYI059 [*] `Generic[]` should always be the last base class
+ airflow-core/src/airflow/api_fastapi/core_api/services/public/common.py:37:18: PYI059 [*] `Generic[]` should always be the last base class

langchain-ai/langchain (+4 -0 violations, +0 -0 fixes)

+ libs/core/langchain_core/output_parsers/base.py:31:26: PYI059 [*] `Generic[]` should always be the last base class
+ libs/core/langchain_core/prompts/base.py:45:25: PYI059 [*] `Generic[]` should always be the last base class
+ libs/core/langchain_core/runnables/base.py:111:15: PYI059 [*] `Generic[]` should always be the last base class
+ libs/core/langchain_core/stores.py:26:16: PYI059 [*] `Generic[]` should always be the last base class

latchbio/latch (+2 -0 violations, +0 -0 fixes)

+ src/latch/types/metadata.py:440:25: PYI059 [*] `Generic[]` should always be the last base class
+ src/latch/types/metadata.py:489:24: PYI059 [*] `Generic[]` should always be the last base class

python/typeshed (+1 -0 violations, +0 -0 fixes)

ruff check --no-cache --exit-zero --ignore RUF9 --no-fix --output-format concise --no-preview --select E,F,FA,I,PYI,RUF,UP,W

+ stdlib/asyncio/queues.pyi:28:12: PYI059 [*] `Generic[]` should always be the last base class

zulip/zulip (+2 -0 violations, +0 -0 fixes)

ruff check --no-cache --exit-zero --ignore RUF9 --no-fix --output-format concise --no-preview --select ALL

+ zerver/lib/notes.py:12:16: PYI059 [*] `Generic[]` should always be the last base class
+ zerver/lib/queue.py:35:18: PYI059 [*] `Generic[]` should always be the last base class

Changes by rule (1 rules affected)

code	total	+ violation	- violation	+ fix	- fix
PYI059	13	13	0	0	0

Linter (preview)

✅ ecosystem check detected no linter changes.

[ntBre]

related: #18602

[AlexWaygood]

related: #18602

IMO I think that probably does block stabilisation, but I agree with you that for now we could just not offer a fix if there are any keyword or variadic arguments, stabilise it after making that change, and open a followup issue to add a fix for classes that have keyword arguments.

Variadic arguments (things like class Foo(*args): ... and class Bar(**args): ...) are rare enough in class definitions that I think it's fine for us never to offer a fix for those. But it would be nice to offer a fix for class Foo(Generic[T], Bar, metaclass=ABCMeta): ... at some point.

[AlexWaygood]

I think I'd probably say something like "type checkers may not be able to infer an accurate MRO for for the class, which could lead to unexpected or inaccurate results when they analyze your code". Linking to the glossary sounds like a great idea!

[ntBre]

Closing this in light of the discussion on #18611 (need to move the fix to unsafe, updating the kwarg handling), but I'll copy the docs improvements over there.