Compatibility of return type with subclass upon overload

[cmp0xff]

Hi, we have the following minimal code snippet that reproduces our concern from pandas-dev/pandas-stubs#1385. In words, when the return types from different overloads are parametrised by a type variable, type checkers can work. When return types are subclasses, however, mypy does not seem to work.

from typing import Any, Generic, Never, TypeVar, overload, reveal_type

T = TypeVar("T")

class Arg(Generic[T]): ...

class Gen(Generic[T]): ...

# part 1: return type is Generic

class D0:
    @overload
    def m(self, p: Arg[Never]) -> Gen: ...

    @overload
    def m(self, p: Arg[int]) -> Gen[int]: ...

reveal_type(D0().m(Arg[Any]()))  # pyright gives Gen[Unknown], mypy gives Gen[Any], good
reveal_type(D0().m(Arg[int]()))  # Gen[int], good
reveal_type(D0().m(Arg[float]()))  # error, good

# part 2: return types are subclasses

class Base: ...

class Sub(Base): ...

class D1:
    @overload
    def m(self, p: Arg[Never]) -> Base: ...  # we want to get Base for Arg[Any]; mypy warns overload-overlap here, whereas pyright does not warn (at least without strict)

    @overload
    def m(self, p: Arg[int]) -> Sub: ...

reveal_type(D1().m(Arg[Any]()))  # pyright gives Base, but mypy gives Any
reveal_type(D1().m(Arg[int]()))  # Sub, good
reveal_type(D1().m(Arg[float]()))  # error, good

Part 1 is understood from #19525. Could someone explain part 2? Thank you 🙂

[ilevkivskyi]

When multiple overloads can be matched because of an Any argument(s), and the corresponding return types are not equivalent modulo Any, mypy intentionally infers Any as the result of the call. This is done to guarantee a fundamental rule of gradual typing: replacing a precise type with Any should not cause new errors. Consider this code:

a: Arg[Any] = Arg[int]()
x: Sub = D1().m(a)

this code would generate an error, while it shouldn't.