Allow type checkers to infer more precise types for tuples

[davidhalter]

This is consisent with the latest changes where f(1) could either be an int or a Literal[1]

    def f[T](x: T) -> T: ...
    assert_type(f(1), int)
    # or
    assert_type(f(1), Literal[1])

The same is possible with tuples:

    def f[*Ts](*x: *Ts) -> tuple[*Ts]: ...
    assert_type(f(1), tuple[int])
    # or
    assert_type(f(1), tuple[Literal[1]])

Mypy fails the test generics_typevartuple_callable.py now, because it doesn't error on any of the assert_type below:

def func3(*args: * tuple[int, *Ts, T]) -> tuple[T, *Ts]:
    raise NotImplementedError

assert_type(func3(1, "", 3j, 3.4), tuple[float, str, complex])  # E[fun3]
assert_type(func3(1, "", 3j, 3.4), tuple[float, Literal[''], complex])  # E[fun3]

This is also how Zuban used to work, but I think it's better to be precise here and it should probably fail one of the two (like Eric suggested in #2178 (comment)).

Ran the automated tests again. A few changes:

• Zuban's passes a test again (due to assert_type changes)
• Picked up a pyrefly upgrade that fixes a test

[JelleZijlstra]

To avoid the two options, could we put this in a function that takes args of type int and str? Then tuple[int, str] would be unambiguously right.

[davidhalter]

Thanks Jelle, that's a really good idea, it also feels more readable this way for people that are not used to the special # E[...] comments.

I pushed this. This makes Mypy pass generics_typevartuple_callable.py again and doesn't really change anything other than Zuban passing the tests with its latest release.

[AlexWaygood]

Hmm, I weakly prefer the version prior to c1ffe96. The new version could pass even if the type checker completely fails to solve the TypeVarTuple and just falls back to *tuple[Any, ...] for its inference of *Ts. The version with the # E[] tagged errors was precise

[davidhalter]

Didn't think about that. That's obviously a problem. I'm happy to revert...

[JelleZijlstra]

I meant something like this:

def func(x: int, y: str):
    assert_type(args_to_tuple(x, y), tuple[int, str])

I agree the current version is worse.

[AlexWaygood]

@JelleZijlstra's version there works for me, yes!

[davidhalter]

@AlexWaygood @JelleZijlstra Thank you for the good feedback. Kind of stupid of me to not think of this in the first place. I think the change is now completely uncontroversial.

[AlexWaygood]

this makes sense to me, thanks!