Conformance test: make dataclass_hash.py not rely on typing.Hashable?

[grievejia]

I noticed that the conformance test currently enforces a hashability check on dataclasses:
https://github.com/python/typing/blame/6d5c186ea3f45dd80dcad9f172479b270e54966a/conformance/tests/dataclasses_hash.py

If we take the annotation of typing.Hashable and the annotation of object literally, then no type errors should be reported on that file, as typing.Hashable is a protocol that requires a def __hash__(self) -> int method, and object satisfies that protocol regardless of whether the dataclass transform creates its own __hash__ method or not.

I understand that the issue of deciding whether an object is hashable or not in type checkers is a tricky business and there was a lot of pre-existing discussions/proposals around it. This makes me wonder if it's worth doing a dedicated "hashability" section in the spec, and use separate conformance tests to establish how typing.Hashable assignability should be handled. My understanding is that the pre-existing dataclass hash tests is intended to just test about whether the __hash__ method is nullified or not under the dataclass transform, but by testing it via typing.Hashable it kinds of indirectly dictate hashable assignment behaviors as a side effect.

Concretely, what I had in mind was a refactor to dataclasses_hash.py, where we change the current assertions of the form

# This should generate an error because DC1 isn't hashable.
v: Hashable = DC(0)

into something like this:

assert_type(DC(0).__hash__, NoneType)  # OK

The new version does not depend on how typing.Hashable gets defined in typeshed, and it (arguably) aligns more directly with the intention of the tests. But I am unsure about how controversial this proposal would be and hence want to get some feedback on it first.

[erictraut]

Yeah, as you noted, statically determining the hashability of a type is tricky because the hashing mechanism in Python unfortunately doesn't follow standard inheritance rules.

I'd welcome an extension to the typing spec that standardizes static typing behaviors related to hashability.

I agree that the current conformance test implementation makes some brittle assumptions. I'm fine with your proposed change.