gh-130870: Preserve GenericAlias subclasses in typing.get_type_hints()

[Viicos]

This is an alternative approach to #130897, that avoids recreating GenericAlias instances two times (no noticeable performance improvement but cleans up the implementation).

It also fixes the GenericAlias reconstruction in _strip_annotations(), and as such I moved the logic into a separate function.

This approach has the benefit of not losing the __unpacked__ property of the GenericAlias if no forward annotation is present in the type hint. That is:

deffunc(*args: *tuple[int]): ... get_type_hints(func) # On main:#{'args': typing.Unpack[tuple[int]]}# On this branch:#{'args': *tuple[int]} (that is, gth(func)['args'] is the same `GenericAlias` instance, with `__unpacked__ = True`)

A couple tests are failing due to this, I'm not sure how we can test this? Maybe like:

deffunc2(*args: *tuple[int, str]): passself.assertEqual(gth(func2), func2.__annotations__) self.assertTrue(gth(func2)['args'].__unpacked__)

• Issue: typing._eval_type is not preserving GenericAlias subclasses #130870

[JelleZijlstra]

This generally makes sense to me, though there's a series of issues in CI. I think this can only go into 3.15 at this point.

[ZeroIntensity]

Yeah, this feels like a feature to me.

[Viicos]

I would rather consider this as a bug fix but the risk of regressions is relatively high so it also makes sense to me to include this in 3.15 only as well.