Annotate TypedDict as _SpecialForm rather than object.

[rchen152]

For consistency, I think it would be slightly better to use _SpecialForm rather than object here. The comment suggests that object was used because TypedDict is non-subscriptable, but NoReturn is also non-subscriptable and is annotated as _SpecialForm.

[hauntsaninja]

If we take this approach, should we be removing __getitem__ from _SpecialForm?

[rchen152]

If we take this approach, should we be removing __getitem__ from _SpecialForm?

Hmm, I didn't look closely enough at _SpecialForm. It does have a __getitem__, both in the stubs and at runtime.

It also looks like there are some weird mypy primer errors about a type mismatch between typing_extensions._SpecialForm and something called <typing special form>.

Let me think a bit more about this PR. Tbh I might just give up on it, since the motivation for it wasn't particularly strong in the first place. It just looked odd to me to have TypedDict described as a _SpecialForm but not annotated as such.

[rchen152]

Ok, I've had a chance to think about and play around with this a bit more:

• Conceptually, TypedDict is a special form (despite not being one at runtime), so I still think it would be better to mark it as such. We're already not hewing very close to the runtime definitions of many of these typing constructs anyway.
• The mypy error occurs whenever a _SpecialForm is imported from both typing and typing_extensions in the same file. It's not specific to TypedDict. Making typing_extensions import the typing definition seems to fix it.
• Removing __getitem__ from _SpecialForm generates stubtest errors. Easy enough to suppress with an addition to the allowlist, but I'm on the fence about whether to do it. NoReturn: _SpecialForm already exists, so it doesn't seem like the __getitem__ is causing any problems.

[github-actions]

According to mypy_primer, this change has no effect on the checked open source code. 🤖🎉