Positional arguments to TypedDict constructor

The spec says "The constructor takes only keyword arguments, corresponding to the items of the TypedDict." This derives from wording in PEP 589 "It can be used as a callable object with keyword arguments corresponding to the TypedDict items. Non-keyword arguments are not allowed."

But at runtime, the TypedDict constructor also takes positional arguments. This works:

from typing import TypedDict
class TD1(TypedDict):
    a: int
class TD2(TD1):
    b: str

def f(x: TD2) -> TD1:
    return TD1(x)

Pyright allows this program; mypy does not and reports Expected keyword arguments, {...}, or dict(...) in TypedDict constructor.

The spec should say whether or not this should be allowed. (My opinion: it should be allowed, since it works at runtime and doesn't appear to break any type safety rule.)

At runtime, even more is allowed; the TypedDict constructor actually allows whatever dict() allows. So you could do TD1([("a", 1)]), or TD1(TD2, a=3). Pyright allows some but not all of that. I'm inclined to prohibit anything other than calls with keyword arguments and calls with a single positional argument of TypedDict type; anything else is rarely useful and prone to unsound behavior.

Start with the linked TypedDict constructor specification and the corresponding PEP 589 section, then compare the documented rules with the runtime examples and the differing mypy and Pyright behavior described here. Done means reaching a decision on permitted positional arguments and updating the specification with clear normative wording.