Inheriting from a generic TypedDict with a T | None key inexplicably narrows the that key type to T

[daniil-berg]

Bug Report

Inheriting from a generic TypedDict with a T | None key inexplicably narrows the that key type to T.

To Reproduce

from typing import Generic, TypeVar, TypedDict

T = TypeVar("T")


class Foo(TypedDict, Generic[T], total=False):
    a: str | None
    ...


class Bar(Foo[T], total=False):
    ...


class Baz(Foo[int], total=False):
    ...


a_value: str | None

foo: Foo[int]
bar: Bar[int]
baz: Baz

reveal_type(foo)  # TypedDict('tests.Foo', {'a'?: Union[builtins.str, None]})
reveal_type(bar)  # TypedDict('tests.Bar', {'a'?: builtins.str})
reveal_type(baz)  # TypedDict('tests.Baz', {'a'?: builtins.str})

foo["a"] = a_value  # passes
bar["a"] = a_value  # error: Value of "a" has incompatible type "str | None"; expected "str"  [typeddict-item]
baz["a"] = a_value  # error: Value of "a" has incompatible type "str | None"; expected "str"  [typeddict-item]

Expected Behavior

No errors. For all three variables foo, bar, baz the revealed type of the a-key of the TypedDict should be Union[builtins.str, None].

Actual Behavior

The type of the a-key of bar and baz is revealed to be builtins.str instead of Union[builtins.str, None].

Thus, the last two assignments cause the errors:

error: Value of "a" has incompatible type "int | None"; expected "int"  [typeddict-item]
error: Value of "a" has incompatible type "int | None"; expected "int"  [typeddict-item]

Notable obsevations

• It does not matter, if any key in Foo is actually declared to be of type T.
• Declaring for example a: T | None and adjusting the a_value type accordingly makes no difference.
• As soon as you make Foo non-generic, i.e. remove Generic[T] from the bases (and adjust Bar and Baz declarations accordingly), the error disappears.
• If you declare the a-key to be of any other union type e.g. str | int (and adjust the a_value type accordingly), the error seems to disappear. (Obviously I have only tested a few combinations.)

Your Environment

• Mypy version used: 1.6.1
• Mypy command-line flags: -
• Mypy configuration options from mypy.ini (and other config files): strict = True
• Python version used: 3.11.4 (also reproduced with 3.9.17)

[JelleZijlstra]

Confirmed in mypy playground: https://mypy-play.net/?mypy=latest&python=3.10&gist=4d53c16a3b06bad9014022b24b490d08

Feels like something is going wrong with strict-optional handling.