This is the crucial part of the change, which implements the new type narrowing behavior. It's the same as isinstance(), a little way up in the same method.

I propose type-narrower-not-subtype, right now this error code can be misleading for readers: type is not a subtype?

It would be consistent with:

TYPE_NARROWER_NOT_SUBTYPE: Final = ErrorMessage(
    "Narrowed type {} is not a subtype of input type {}", codes.TYPE_NARROWER_NOT_SUBTYPE
)

Maybe narrowed-type-not-subtype? Your suggestion sounds like TypeNarrower, which was a rejected name for teh feature.

Can you please add a comment on why this is marked as xfail?

I copied this from the TypeGuard test. Seems like the xfail is just because we generate a name with an extra 1 at the end, which looks like a typo:

Expected:
  main:16: note: Revealed type is "__main__.<subclass of "Foo" and "Bar">"
  main:21: note: Revealed type is "__main__.<subclass of "Foo" and "Bar">" (diff)
Actual:
  main:16: note: Revealed type is "__main__.<subclass of "Foo" and "Bar">"
  main:21: note: Revealed type is "__main__.<subclass of "Foo" and "Bar">1" (diff)

I will remove the xfail and make it pass.

T is not defined, why does this test pass?

This is a general issue with test cases that I've seen in other tests as well. I'm not sure if T is coming from the builtins stubs or whatever else. If you are really interested I could quickly try and search where this has happened.

Obviously still good to point it out and change it here.

I think the root cause of this is that a normal builtins.pyi in typeshed defined _T (which does not get exported) wheres the test fixtures (like dict.pyi) typically use stuff like T (which does get exported).

It would probably be a good idea to change all fixtures, but that might be quite a bit of work.

Added an explicit definition of T in this test. Agree that ideally we should fix this throughout the test suite.

Nice, thanks! I think this should fix a thing I ran into while testing #16939