mypy incorrectly flags a type-var error when a descriptor overrides a protocol property in a different file.

[thomas-mckay]

Bug Report

To Reproduce

• Create an app.py file with this inside:

from __future__ import annotations

from typing import Protocol

# PROTOCOL
class ServiceProto(Protocol):
    @property
    def value(self) -> int: ...

# DESCRIPTOR
class Descriptor:
    def __get__(self, instance: ServiceImpl, owner: type[ServiceImpl]) -> int:
        return 5

# IMPLEMENTATION
class ServiceImpl:
    value = Descriptor()

class App[ServiceT: ServiceProto]:
    def __init__(self, service: ServiceT) -> None:
        self.service = service

• create a second file (important) with this inside (mine is named run.py):

from __future__ import annotations

from .app import App, ServiceImpl

def bootstrap() -> App[ServiceImpl]:
    return App(ServiceImpl())

Expected Behavior
No errors

Actual Behavior

mypy src/
run.py:5: error: Type argument "ServiceImpl" of "App" must be a subtype of "ServiceProto"  [type-var]
run.py:6: error: Value of type variable "ServiceT" of "App" cannot be "ServiceImpl"  [type-var]
Found 2 errors in 1 file (checked 3 source files)

Your Environment

• Mypy version used: mypy 1.18.2 (compiled: yes)
• Mypy command-line flags: none
• Mypy configuration options from mypy.ini (and other config files): none
• Python version used: Python 3.13.7 (via uv)

Weird Behavior

Merge both files, remove redundant imports. No more errors.

PS: Feel free to edit the title. This was a hard one to describe in one sentence without knowing the root cause.
PPS: This is a toy example with integers to illustrate the bug. My real use case is more complex: the descriptor is a lazy instantiator that works like cached_property (the descriptor is given a class object on instantiation, it instantiates and caches an instance of that class on the owning service instance, then serves the cached instance on further accesses).

[ilevkivskyi]

Yeah, this is a known limitation, for technical reasons descriptors/protocols combo doesn't work during type variable bound/values checking (because it is performed during semantic analysis phase, before actual type checking starts), see second bullet point in #19299. To set correct expectations, this is not easy to fix properly, but we want to do this (mostly for other benefits, like lazy deserialization of symbols), probable somewhere in early 2026.