Allow an attribute declared in a superclass's class body to be overridden by an attribute redeclared in a subclass's `__init__` method

[AlexWaygood]

Summary

Consider the following snippet. Ty, mypy, pyright and pyrefly all agree on the type of B1().x, but ty disagrees with the other three on the type of B2().x:

class A:
    x: int

class B1(A):
    x: bool

class B2(A):
    def __init__(self) -> None:
        self.x: bool = True


reveal_type(B1().x)  # Ty, mypy, pyright, pyrefly: bool
reveal_type(B2().x)  # Mypy, pyright, pyrefly: bool. Ty: int!

• Ty
• Pyright
• Mypy
• Pyrefly

I think for compatibility, it would be better if we followed what the other type checkers do here.

I spotted this by looking at the ecosystem report in astral-sh/ruff#20723 (comment) (all 4 new unresolved-attribute errors on aiohttp in the ecosystem report are caused by this).

Version

No response

[sharkdp]

… but we should still error on the self.x: bool = True declaration, correct?

I just recently added some tests that are very similar to this scenario, but re-declaration scenarios are not yet considered: https://github.com/astral-sh/ruff/blob/513d2996ec599d3c1990480e0edd3a7328adcd0f/crates/ty_python_semantic/resources/mdtest/attributes.md?plain=1#L821-L919

[AlexWaygood]

… but we should still error on the self.x: bool = True declaration, correct?

yes, we also need to implement the Liskov Substitution Principle, correct! (And mutable attributes should be treated as invariant.)

[AlexWaygood]

(Just noting things down for myself as much as anything else)

It also looks like this is the cause of all 5 new unresolved-attribute errors on boostedblob in astral-sh/ruff#20723 (comment)

And all the new unresolved-attribute errors on homeassistant