I still think that we should disallow this in some future versions.

I still think that we should disallow this in some future versions.

I agree that the current behaviour doesn't make any kind of "sense". Personally, I would be happy to disallow this behaviour if:

1. We're happy to start adding __class_getitem__ to stdlib collections.namedtuples if typeshed types them as generic typing.NamedTuples. (This is currently unnecessary, since they inherit __class_getitem__ from tuple, but it would be necessary if we made the change you suggest.)
2. We're happy to start backporting PRs that add __class_getitem__ to stdlib classes. Historically we've treated these as enhancements, and haven't backported them. I understand the rationale for that policy. But it means that it generally takes years before users can stop doing ugly if TYPE_CHECKING blocks of code to work around discrepancies between the runtime and the stubs.