Nice. Can you also add the removal of those parts of #2812 that are no longer needed? (But keep the tests, obviously.)

The problematic part in this PR is the reduced clarity in error messages, as can be seen in the (currently failing) tests. What do you think about it?

Perhaps it'd be better to do it in another PR.
(For now I am skipping the tests for enum, Just to get the green light back :) )

Seeing all that, it may be better to split this out:

1. Reduce PR to just your initial commit
2. Update typeshed with an EnumMeta metaclass that supports __getitem__ (it won't be activated yet because the code from Allow E[<str>] where E is an Enum type #2812 overrides it)
3. Roll back Allow E[<str>] where E is an Enum type #2812 (but keeping its tests)

You should be able to co-develop (2) and (3) and submit them concurrently, with a dependency on merging so that (2) gets merged first but only once you know that the tests for (3) pass with that included. The PR for (3) should be updated at the last moment with a typeshed sync that incorporates (2).

Does that make sense?

Regarding the poorer error message: it would be so much better if the way types are rendered distinguished between Type[X] and X. Currently it seems Type[int] and int are both rendered as Value of type "int".

(Sorry I rewrote history. Another git accident there)

What if the indexed class is e.g. Type[...], Or a type var bounded by such a thing? Or has a bound that is a metaclass?

Type[...] will be printed as "A value of type Type[...]".
The TypeVar case seems not to be handled at all - it evaluated to Any silently. Currently it is ignored in semanal.py, but handling it there does not fix the problem.

I am a bit hesitant about changing the pretty-printing since it might affect many tests.

(I am sorry but I do not have enough time right now, and I will not be available at all in the upcoming week).