New semantic analyzer: regression with circular import and TypeVar

[mthuurne]

I found a regression when using the new semantic analyzer in mypy 0.710.

To reproduce, you need the following two modules:

# factory.py

from typing import Generic, Type

from box import BoxT

class Factory(Generic[BoxT]):
    value = 123

    def create(self, boxClass: Type[BoxT]) -> BoxT:
        return boxClass.create(self)

# box.py

from typing import TYPE_CHECKING, Type, TypeVar

if TYPE_CHECKING:
    from factory import Factory

BoxT = TypeVar('BoxT', bound='Box')

class Box:

    @classmethod
    def create(cls: Type[BoxT], f: Factory) -> BoxT:
        return cls(f.value)

    def __init__(self, value: int):
        print(value)

When mypy is run on these modules with the default arguments, no issues are reported. But when run withthe new semantic analyzer, the following happens:

$ mypy --new-semantic-analyzer box.py factory.py 
factory.py:11: error: "Type[BoxT]" has no attribute "create"

The problem does not occur if there is no circular import. It also doesn't occur if the BoxT TypeVar is defined in the factory module instead of imported from the box module.

[JukkaL]

Thanks for reporting this! This seems like an important issue to fix, in particular since the new semantic analyzer is supposed to work better with import cycles :-)