Conformance suite: Questionable tests in dataclasses_descriptors.py

Read conformance/tests/dataclasses_descriptors.py, focusing on the Desc2 and DC2 cases described in the issue. Check the asserted types against the runtime behavior and the differing pyrefly, mypy, and pyright results. Done means the conformance suite no longer requires behavior that is not specified or supported by runtime behavior, with the proposed test removal resolved by project maintainers.

The dataclasses_descriptors.py test case has a comment "This portion of the dataclass spec is under-specified in the documentation, but its behavior can be determined from the runtime implementation."
However, some of the behaviors in tests don't align with the runtime behavior.

The first part, which is using a data descriptor Desc1 and a dataclass DC1, is fine.

Then we get this, where Desc2 is a non-data descriptor (only implementing __get__):

@dataclass
class DC2:
    x: Desc2[int]
    y: Desc2[str]
    z: Desc2[str] = Desc2()


assert_type(DC2.x, list[int])
assert_type(DC2.y, list[str])
assert_type(DC2.z, list[str])

However, DC2.x and DC2.y are AttributeErrors at runtime, because there is no Desc2 object in the runtime class dictionary. I don't know if we should require type checkers to produce an error but certainly we should allow it.

Then we construct a DC2 object:

dc2 = DC2(Desc2(), Desc2(), Desc2())
assert_type(dc2.x, int)
assert_type(dc2.y, str)
assert_type(dc2.z, str)

But at runtime, dc2's three attributes are all just those Desc2 objects, not ints or strs, because those objects are stored as instance attributes, so the descriptor protocol never runs on them.

The more complicated case is the z attribute, which has a Desc2 object stored in the class. Dataclasses would conventionally treat this as the default, but it gets the default by conventional access in the class body, which calls the descriptor's __get__ with obj=None. So the runtime default for this field is actually the value you'd get for accessing DC2.z, which is list[str].

Looking at actual type checker behavior, I think pyrefly gets all this pretty much right (which means it fails the test). Mypy and pyright pass the test, but their behavior is wrong.

Since (as the test acknowledges) the behavior is not fully specified, perhaps it is best to delete this test from the conformance suite for now, until we come up with a full specification for the behavior of descriptors in dataclasses.