According to mypy_primer, this change has no effect on the checked open source code. 🤖🎉

A bit of churn above. It seems that Awaitable is required as a base class because it allows asyncio.Future to sometimes behave covariantly even though it's not typed as covariant itself.

I tested locally and could see that the mypy-primer results cleared up if the class was Future(Generic[_T_co]), so that definitely seems like the cause of the blow-up seen in #12818. With no way to make Future covariant directly, I believe Awaitable needs to stay.

Is it normal to have this kind of indirectly-covariant typing? I was surprised to find that it works that way, but I don't have a deep understanding of variance.

Yes, IIUC you would break the transitivity of subtyping. See also:

class A: ...
class B(A): ...
from typing import Sequence
def takes_seq_a(x: Sequence[A]): ...
def main(x: list[B]):
    takes_seq_a(x)

asyncio.Future is mutable, so cannot be covariant: python/mypy#13689 (comment)

(and if you want to break your brain a little bit, check out #8781 , which surfaced unsoundness and was patched in both pyright and mypy, see python/mypy#13714)