1.6.1 > 1.7.0 and 1.7.1 regression in for "broken" ``dataclasses``

[DanielNoord]

Bug Report

When type checking the following code the output changes depending on whether you use 1.6.0 or 1.7.1.

To Reproduce

from dataclasses import dataclass


@dataclass(frozen=True)
class A:
    attr_one: int
    attr_two: int | None


@dataclass(frozen=True)
class B(A):
    attr_two = None


instance = B(attr_one=1)
print(instance.attr_two)

Playground: https://mypy-play.net/?mypy=1.7.0&python=3.11&gist=f74740e7a7ef6bf0c9f88175a988aed5

The code is likely not doing what the author expects it to do, but the new message is wrong as attr_two can be referenced and is certainly not an instance of B.

Expected Behavior

main.py:15: error: Missing positional argument "attr_two" in call to "B"  [call-arg]
Found 1 error in 1 file (checked 1 source file)

Actual Behavior

main.py:15: error: Missing positional argument "attr_two" in call to "B"  [call-arg]
main.py:16: error: "B" not callable  [operator]
Found 2 errors in 1 file (checked 1 source file)

Your Environment
master branch on Playground.

[DanielNoord]

@AlexWaygood I have now also been able to reproduce without the deamon. You might want to update the label. (I'll update the title)