[red-knot] support dataclasses

[carljm]

[sharkdp]

What do we plan here in a first iteration? We have no support at all for data classes, but we also don't raise a lot of false positives at the moment(?). Attribute annotations work just like for normal classes:

from dataclasses import dataclass

@dataclass
class C:
    x: int
    y: str | None = None

c = C(1, "foo")  # This call is completely unchecked; so no false positives, but also no true positives

reveal_type(c.x)  # revealed: int
reveal_type(c.y)  # revealed: str | None

We currently do report false positives if there are generated special methods whose existence we do not model:

@dataclass(order=True)
class C:
    x: int

C(2) > C(1)  # False positive: Operator `>` is not supported for types `C` and `C`

[carljm]

I think the highest priority item would be correctly synthesizing the __init__ method, because if/when constructor-call-checking lands, it will cause lots of false positives on dataclasses if we don't have that.

[AlexWaygood]

Agreed that synthesising the __init__ method should definitely be the highest-priority item.

We currently do report false positives if there are generated special methods whose existence we do not model:

@DataClass(order=True)
class C:
x: int

C(2) > C(1) # False positive: Operator > is not supported for types C and C

I think the second-highest priority item might be synthesising the __lt__, __le__, __gt__ and __ge__ methods for order=True, to address this false positive here.

[sharkdp]

Initial support for dataclasses has been implemented. See astral-sh/ty#111 for a lower-priority follow-up ticket.

[AlexWaygood]

I'll close this for now so that it no longer shows up on the alpha milestone of tasks still to be done in the next fortnight. Thanks @sharkdp! 😃