Regression related to forward reference and nested classes

[JukkaL]

This example started generating a false positive about an undefined name (simplified from a real-world generated protobuf stubs):

from typing import NewType

class W[T]: pass

class R:
    class M(W[Action.V], type):  # Name "Action.V" is not defined
        FOO = R.Action.V(0)
    class Action(metaclass=M):
        V = NewType('V', int)

class Action:
    pass

This was a side effect of #18625.

There may be some confusion between R.Action and Action classes.

This looks like a release blocker. This is happening in generated protobuf stubs code which can't be edited easily.

If this is non-trivial to fix, we could temporarily revert the original PR until a fix is available.

cc @ilevkivskyi as the author of the above PR

[ilevkivskyi]

A possible quick fix may be to reduce precedence of global forward reference (although this may further complicate precedence rules for forward references). I can make a PR later today and check what is the fallout (likely there will be none, so it is more a question of documenting this clearly at some point).

[sterliakov]

Line 6 in your example produces a runtime NameError (verified on 3.12, 3.13, and my common sense), so if anything is to be changed here, that should only apply to stubs?..

[JukkaL]

There could also be a string literal forward reference in a non-stub file. The main thing is maintaining backward compatibility.

[ilevkivskyi]

Yeah, I am going to go now with a minimal change to preserve backwards compatibility in this case. If there will be further complains from real code, we can further adjust the rules (and sorry for a delay, was distracted by unrelated things).

[ilevkivskyi]

Here is a possible fix #19000