Summary

We don't emit a diagnostic on this snippet... but we should, because _T_co is not defined 😆

class SupportsNext:
    def __next__(self) -> _T_co:
        raise NotImplementedError

We infer _T_co as being available here because of its definition in builtins.pyi here. But that definition should be treated as private-to-the-stub, I think.

Possibly we should just filter out typevar definitions when falling back to the builtin scope in name lookup? The same could also be said for type-alias definitions and protocol definitions in builtins.pyi, but they're arguably both more useful and less confusing to have available (you might want to use them quoted in type annotations, e.g. x: "_PositiveInteger" wouldn't be an unreasonable type annotation.

Version

No response