Interaction of Unpack and a TypeVar bound to a tuple

[Enegg]

This seems similar to #7317, however in their code pyright did show errors.
The following MRE typechecks without issues:

from collections import abc
from typing import NamedTuple, Unpack

class Foo(NamedTuple):
    x: str
    y: int

def func1[TupleT: tuple[object, ...]](p: type[TupleT]) -> abc.Callable[[Unpack[TupleT]], None]: ...

f1 = func1(tuple[str, int])
f1(("abc", 123))  # Unpack seemingly ignored
f1foo = func1(Foo)
f1(Foo("abc", 123))

def func2[TupleT: tuple[object, ...]](p: type[TupleT]) -> abc.Callable[[*tuple[*TupleT]], None]: ...

f2 = func2(tuple[str, int])
f2("abc", 123)  # two parameters, as expected
f2foo = func2(Foo)
f2foo(Foo("abc", 123))  # single parameter, otherwise errors

In the definition of func1, the Unpack in the return type seems allowed, but is ignored. When replaced with *TupleT, pyright reports:

Unpack operator not allowed in type expression

In the definition of func2, both Unpack and * of TupleT is allowed within a tuple; the returned callable has an inconsistent parameter count depending on the accepted tuple type being a parametrized tuple, or a NamedTuple (of same length).

I believe the intended behavior is to error in all of the cases above due to TupleT not being a valid type to unpack?

[erictraut]

You've discovered an experimental (and undocumented) feature that I implemented in support of my recent explorations related to tensor shape type checking.

You're correct that the code above is not allowed by the typing spec currently. You'll see that mypy produces errors here.

I could have added this support under the reportExperimentalFeatures flag, but I didn't bother doing this because I figured it was unlikely that anyone would try to use this feature since it's undocumented. In any case, please don't rely on this working because it's quite likely that the support will go away.