[red-knot] Handle possibly-unbound instance members

[sharkdp]

Summary

Adds support for possibly-unbound/undeclared instance members.

Test Plan

New MD tests.

[sharkdp]

We came to the same conclusion in an earlier PR. This is also mentioned higher up in this test suite. And mypy/pyright don't emit a diagnostic for this either (which is also mentioned higher up). This test is only added here for easier reference / completeness.

[AlexWaygood]

Looks good! A couple of minor questions

[carljm]

The handling of inherited attributes here reminds me that we will also need to add eager Liskov checking for inherited classes (once we have callable types and can do callable subtype checks). For a mutable attribute (which all are by default) a strict Liskov interpretation would require invariance, which would make the union handling here irrelevant.

[sharkdp]

which would make the union handling here irrelevant

Mainly curious: is this true even in the presence of gradual types? Could a subclass refine a base-class Any attribute to a more static type T? And would it then make sense to treat a possibly-unbound-on-child-class attribute as being of type Any | T?

[carljm]

Mainly curious: is this true even in the presence of gradual types? Could a subclass refine a base-class Any attribute to a more static type T? And would it then make sense to treat a possibly-unbound-on-child-class attribute as being of type Any | T?

That's a good point: I think that would make sense.

[carljm]

Probably the diagnostic would be raised by an eager Liskov check at the inheritance site, rather than here? But it doesn't hurt to have a TODO here until we do that.