__slots__ and protocols don't work right for issubclass and typevar bounds

[A5rocks]

cc @sobolevn since you've been doing stuff here

Bug Report

Slots become part of a interface with protocols, they shouldn't be.

To Reproduce

import typing

@typing.runtime_checkable
class Foo(typing.Protocol):
    __slots__ = ()


class Bar:
    pass

_T = typing.TypeVar("_T", bound="Foo")


def foo(f: _T) -> _T:
    return f

print(issubclass(Bar, Foo))
print(isinstance(Bar(), Foo))
foo(Bar())

Expected Behavior

At runtime this prints:

True
True

So issubclass is allowed, and I don't see why the TypeVar bounds shouldn't be.

Actual Behavior

main.py:17: error: Only protocols that don't have non-method members can be used with issubclass()
main.py:17: note: Protocol "Foo" has non-method member(s): __slots__
main.py:19: error: Value of type variable "_T" of "foo" cannot be "Bar"
Found 2 errors in 1 file (checked 1 source file)

Your Environment

mypy playground

[sobolevn]

Thanks, added to my backlog. We can treat this one as a regression, so the priority is high 🙂

[sobolevn]

Simplier repro:

import typing

@typing.runtime_checkable
class Foo(typing.Protocol):
    __slots__ = ()


class Bar:
    pass

def foo(f: Foo) -> Foo:
    return f

foo(Bar())
# out/ex.py:14: error: Argument 1 to "foo" has incompatible type "Bar"; expected "Foo"
# out/ex.py:14: note: Following member(s) of "Bar" have conflicts:
# out/ex.py:14: note:     __slots__: expected "Tuple[]", got "Union[str, Iterable[str]]"

[sobolevn]

Related #9314

[sobolevn]

Important question: should we treat __slots__ as a part of the protocol?

Let's see how other type checkers do that 🤔

[sobolevn]

Pyright ignores __slots__ definition from protocol members. Probably, we should too.

[sobolevn]

@A5rocks reviews are welcome! 👍