Fix isinstance on namedtuple type aliases

[paulorochaoliveira]

Summary

Fix #21100
Fix a false positive where isinstance(x, Alias) reports
Parameterized generics cannot be used with class or instance checks
when Alias is a TypeAlias to a class inheriting from NamedTuple.

Changes

• Exempt namedtuple-backed tuple aliases from the generic-alias isinstance rejection path
• Add a regression test for TypeAlias = Foo where Foo is a NamedTuple

Testing

• ./venv/bin/pytest -n0 mypy/test/testcheck.py -k "testIsinstanceTypeAliasToNamedTuple or testIsinstanceTypeArgsAliases or testAliasNonGeneric"
• ./venv/bin/pytest -n0 mypy/test/testcheck.py -k check-isinstance.test

[github-actions]

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

[A5rocks]

I feel like no_args should be True in this case? In which case this if statement is fine.

I'd rather that, unless there's a good reason for existing behavior...

[paulorochaoliveira]

@A5rocks, I considered that, but no_args currently seems to mean more than “this alias was written without subscripting”.

In particular, it appears to assume an Instance target in a few places:

• mypy/types.py (TypeAliasType._expand_once)
• mypy/typeanal.py (instantiate_type_alias)
• the alias creation logic in mypy/semanal.py

For aliases to classes inheriting from NamedTuple, the target is a TupleType with a namedtuple fallback rather than an Instance, so making no_args=True here would likely require broadening that invariant first.

Because of that, I went with the narrower fix in the isinstance check path. If you’d prefer, I can rework this to make no_args support namedtuple-backed aliases more generally, but that seemed like a larger behavioral change than needed for this bug.

[A5rocks]

That's interesting! I had assumed this already worked for tuples but actually they don't work at runtime! I guess I agree that no_args is the wrong approach then. This still feels a bit hacky though so I'll think a bit more... But as a first comment, could you add a test to show that you didn't modify the following output?:

from typing import Any, TypeAlias

Alias: TypeAlias = tuple[int, int]

def is_foo(x: Any) -> bool:
    return isinstance(x, Alias)

(your code shouldn't have changed that, but unless that's already a test it's better to prevent people like me from modifying this without knowing the runtime behavior!)

[A5rocks]

Actually I read the context and this is already what we do for unions, so I'm happy with this.