Revisiting TypeIs and TypedDict and PEP 728

[Sxderp]

I originally had a discussion talking about how TypeIs creates a intersection of TypedDicts: #10251. With the accepting of PEP 728, I would expect closed=True TypedDicts to behave more like standard classes and not create "intersections". Am I misunderstanding how PEP 728 should behave?

from typing import TypedDict
from typing import TypeIs
from typing import Never

class A():
    ...

class B():
    ...

class X(TypedDict, closed=True):
    val: int
    ret: str

class Y(TypedDict, closed=True):
    key: int
    ret: int

AB = A | B
XY = X | Y

def ab() -> AB:
    ...

def xy() -> XY:
    ...

def ti_ab(real: AB) -> TypeIs[A]:
    ...

def ti_xy(real: XY) -> TypeIs[X]:
    ...

real_ab = ab()
real_xy = xy()

if ti_ab(real_ab):
    reveal_type(real_ab)
else:
    reveal_type(real_ab)

if ti_xy(real_xy):
    reveal_type(real_xy)
else:
    reveal_type(real_xy)

The output of the above is:

Type of "real_ab" is "A"
Type of "real_ab" is "B"
Type of "real_xy" is "<subclass of Y and X>"
Type of "real_xy" is "X | Y"

I would expect it to be:

Type of "real_ab" is "A"
Type of "real_ab" is "B"
Type of "real_xy" is "X"
Type of "real_xy" is "Y"

[erictraut]

The TypeIs mechanism assumes isinstance semantics, but a TypedDict class cannot be used as the second argument of isinstance. That's because isinstance uses class hierarchy information to determine the instance relationship, but a TypedDict is a structural type definition.

For more details about TypeIs, refer to this section of the typing spec.