[stubgen] Improve dataclass init signatures

[cdce8p]

Remove generated incomplete __init__ signatures for dataclasses. Keep the field specifiers instead.

[github-actions]

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

[hamdanal]

Thanks I’ll take a look shortly. In the meantime could you check with the typeshed people that this wouldn’t crash their CI? I’m mainly concerned about pytype not accepting assignments in stubs but haven’t checked if it works correctly now. My point is if typeshed is not ready we should wait a bit here before adding this to stubgen.

[hamdanal]

Thanks, change looks good. Please check with typeshed maintainers if this is OK before merging.

[cdce8p]

In the meantime could you check with the typeshed people that this wouldn’t crash their CI? I’m mainly concerned about pytype not accepting assignments in stubs but haven’t checked if it works correctly now.

To be fair I'm not too familiar with pytype. However, I don't think these assignments should be an issue. Type checkers already need to be able to parse dataclass assignment in .py files, why shouldn't it work in .pyi files then? It works for mypy and pyright without issues.

/CC @AlexWaygood I believe you're one of the typeshed maintainers. Do you know if dataclasses.field assignments in stubs would cause issues?

[AlexWaygood]

Type checkers already need to be able to parse dataclass assignment in .py files, why shouldn't it work in .pyi files then?

I believe pytype uses a different parser for .pyi files rather than .py files; it often crashes on syntax in stubs that mypy and pyright are totally fine with. The way to find out whether it would cause issues for pytype would probably be to open a draft PR to typeshed adding some field() calls as assignment values in some dataclass classes, e.g. this one https://github.com/python/typeshed/blob/c01e731dd1fb6d6ae3c9ef00e2f19a89ab1ffb24/stdlib/pstats.pyi#L48-L61. We can see if pytype crashes on the change.

[hauntsaninja]

Seems reasonable, thank you!

[cdce8p]

Just to catch up here. We were just in the process of testing it with typeshed. Seems pytype does indeed have an issue with it, as well as flake-pyi and ruff. I don't think this is a problem though as stubgen is meant to be a starting point to iterate from.

python/typeshed#13415