Type[T] -> T has a strange behaviour

This bug report is coming from https://github.com/python/mypy/issues/8946 and https://github.com/ltworf/typedload/issues/132

Basically the pattern

def f(t: Type[T]) -> T: ...

Works for some types but not others, and I can't see a pattern. For example it will work for a NamedTuple but not for a Tuple.

See this example:

from typing import *
from dataclasses import dataclass
T = TypeVar('T')

@dataclass
class P:
    a: int

class Q(NamedTuple):
    a: int

def create(t: Type[T]) -> T:    ...

# These ones work
reveal_type(create(int))
reveal_type(create(str))
reveal_type(create(List[int]))
reveal_type(create(Dict[int, int]))
reveal_type(create(Q))
reveal_type(create(P))

# These do not
reveal_type(create(Tuple[int]))
reveal_type(create(Union[int,str]))
reveal_type(create(Optional[int]))

Now in the older mypy, the ones that do not work would just be understood as Any, which is absolutely not the intended behaviour. In the latest release instead they fail with error: Argument 1 to "create" has incompatible type "object"; expected "Type[<nothing>]" (and I have no idea of what this means).

Anyway in my view, all of the examples provided should work.

In case I'm wrong, then none of them should work, and probably a way to achieve this is needed.

Background: I am working on a library to load json-like things into more typed classes, and it promises to either return an object of the wanted type, or fail with an exception.

Begin with the supplied Python reproducer and run mypy against each reveal_type call, comparing the working cases with Tuple[int], Union[int, str], and Optional[int]. Trace how Type[T] and TypeVar[T] are inferred for those calls; done means the examples have consistent, intended inferred types or the supported behavior is clearly documented and covered by regression tests.