false-positive name-defined regression with dataclasses and shadowing

[The-Compiler]

Bug Report

This file:

from dataclasses import dataclass
import collections


@dataclass
class Shadow:

    collections: collections.deque
    other: collections.defaultdict


print(Shadow(collections.deque(), collections.defaultdict()))

is perhaps a bit unorthodox, but runs fine, because the collections: ... doesn't actually shadow the module when defining other.

Actual Behavior

mypy v0.950 was happy with it, but mypy v0.960 claims that:

test.py:9: error: Name "collections.defaultdict" is not defined

Introducing commit

Bisected to 03901ef ("Running dataclass transform in a later pass to fix crashes (#12762)", @JukkaL)

Your Environment

• Mypy version used: v0.960
• Mypy command-line flags: none
• Mypy configuration options from mypy.ini (and other config files): none
• Python version used: 3.10.4
• Operating system and version: Archlinux

[AlexWaygood]

Looks similar to #10640. But it's not the same, because #10640 reproduces on 0.950, whereas this doesn't. That might be because #10640 involves a from foo import bar import, whereas this is an import foo import.

[JukkaL]

I suspect that the original behavior was accidental. Now dataclass behavior is similar to ordinary classes, since #12762 made dataclasses a little less special:

import collections

class Shadow:
    collections: collections.deque
    other: collections.defaultdict  # Error here

This is still a regression. As a workaround, you can import collections with a different name:

import collections as _collections

class Shadow:
    collections: _collections.deque
    other: _collections.defaultdict  # OK

I'm going to try to find a simple fix, but I think that we can live with this regression if there is no easy fix in sight.