mypy doesn't take value restrictions into account when solving type variables

[hauntsaninja]

This came up in python/typeshed#6762

from typing import Any, TypeVar, Union

T = TypeVar("T")
V = TypeVar("V", str, bytes)

def check_t(x: T | list[T]) -> T: ...
def check_v(x: V | list[V]) -> V: ...

x_str: str
x_list_any: list[Any]
x_str_list_any: str | list[Any]

# mypy does the correct thing here

reveal_type(check_t(x_str))  # str
reveal_type(check_t(x_list_any))  # list[Any] | Any
reveal_type(check_t(x_str_list_any))  # str | list[Any] | Any
reveal_type(check_v(x_str))  # str

# but maybe not here

# E: Value of type variable "V" of "check_v" cannot be "Union[List[Any], Any]"
# N: Revealed type is "Union[builtins.list[Any], Any]"
reveal_type(check_v(x_list_any))
# E: Value of type variable "V" of "check_v" cannot be "Union[str, List[Any], Any]"
# N: Revealed type is "Union[builtins.str, builtins.list[Any], Any]"
reveal_type(check_v(x_str_list_any))

Looking at the output, my guess is that mypy isn't taking value restrictions into account when solving constraints and only using value restriction as a check later. That is, list[Any] cannot match the T part of T | list[T] because it would violate the value restriction.

[hauntsaninja]

Playing around with the following as a potential fix, but I think there are some other issues with how mypy handles this kind of thing.

diff --git a/mypy/constraints.py b/mypy/constraints.py
index 5a78cdb94..9ca16bcaf 100644
--- a/mypy/constraints.py
+++ b/mypy/constraints.py
@@ -199,6 +199,12 @@ def infer_constraints_if_possible(template: Type, actual: Type,
         # This is not caught by the above branch because of the erase_typevars() call,
         # that would return 'Any' for a type variable.
         return None
+    if (
+        isinstance(template, TypeVarType) and
+        template.values and
+        all(not mypy.subtypes.is_equivalent(actual, erase_typevars(t)) for t in template.values)
+    ):
+        return None
     return infer_constraints(template, actual, direction)

[JelleZijlstra]

#3644 is a similar issue with T | Cls[T].