Type hints for __new__ no longer work when subclassing Dict

[ShaneHarvey]

Bug Report

mypy 1.1.1 introduces a new false positive on pymongo's codebase first noticed in python/typeshed#9694 (comment)

Type hints for __new__ no longer work when subclassing Dict (or dict).

To Reproduce

I see the same behavior on 1.2 and latest:

$ cat repro.py 
from typing import Dict, Any, Type, TypeVar

_Key = TypeVar("_Key")
_Value = TypeVar("_Value")


class MyDict(Dict[_Key, _Value]):
    def __new__(cls: "Type[MyDict[_Key, _Value]]", *args: Any, **kwargs: Any) -> "MyDict[_Key, _Value]":
        return super().__new__(cls, *args, **kwargs)

Error:

$ mypy --strict repro.py           
repro.py:9: error: Value of type variable "Self" of "__new__" of "dict" cannot be "MyDict[_Key, _Value]"  [type-var]
Found 1 error in 1 file (checked 1 source file)

Current workaround is to add # type: ignore[type-var].

Your Environment

• Mypy version used: 1.1.1, 1.2.0, 1.4.0+dev.bd6ce237f62366f156f4af074e398d761fb41359
• Mypy command-line flags: --strict although I'm not sure it's needed.
• Mypy configuration options from mypy.ini (and other config files): None
• Python version used: 3.11.1

Note that in the real code, __new__ is needed to support pickle.

[finite-state-machine]

This applies to (at least) FrozenSet as well.

mypy-play.net: minimal reproduction case

As motivation, consider [mypy-play.net]:

from __future__ import annotations
from typing import *

class ModFrozenSet(FrozenSet[int]):

    some_str: str

    def __new__(cls, init: Iterable[int], some_str: str) -> ModFrozenSet:
        _ = some_str  # used by '__init__()'
        return frozenset.__new__(cls, init)

    def __init__(self, init: Iterable[int], some_str: str):
        _ = init  # used by '__new__()'
        self.some_str = some_str


# Empirical test to show that this works in practice:

def test_motivation_empirically() -> None:

    created = ModFrozenSet([1, 2, 3], "hello")

    assert type(created) is ModFrozenSet
    assert isinstance(created, frozenset)
    assert sorted(created) == [1, 2, 3]
    assert created.some_str == "hello"

[finite-state-machine]

This appears to be fixed on the master branch, but not in 1.6.1 (per the first gist in my reply above).