[koleharv]

Fixes #15717

Binding a classmethod to a second name makes stubtest reject a stub that matches the runtime exactly:

# mod.py
class T:
    def a(self): pass
    b = a
    @classmethod
    def c(cls): pass
    d = c

# mod.pyi
class T:
    def a(self) -> None: ...
    b = a
    @classmethod
    def c(cls) -> None: ...
    d = c

error: mod.T.d variable differs from runtime type def () -> Any
Stub: in file mod.pyi:6
def (cls: type[mod.T])
Runtime: in file mod.py:4
def ()

The plain-method alias b is fine; only the classmethod alias errors.

Cause

Both sides are right, and the comparison is what is wrong. d = c in the stub names the classmethod object, whose type keeps the cls argument, whereas reading T.d at runtime goes through the descriptor and yields a method already bound to the class. verify_var then compares an unbound signature against a bound one and reports a difference that does not exist.

Change

When the runtime value is a method bound to a class, verify_var now compares against the stub type with its leading cls removed.

The check is deliberately narrow. It applies only when:

• the runtime object is a bound method whose __self__ is a class — an instance method bound to an instance is untouched,
• the stub type is a callable with at least one argument, and
• that first argument is a type[...], or is named cls / mcls / metacls.

So a stub that already accounts for the binding, or one whose first parameter is a genuine value parameter, keeps being compared as before. A real mismatch still fails: annotating the alias as c_alias: int against a runtime method is still an error, and there is a test for it.

Tests

Two cases added to test_static_class_method:

• a class with both a plain-method alias and a classmethod alias, matching at runtime: no error (fails before this change)
• a classmethod alias annotated as int in the stub: still an error

mypy/test/teststubtest.py passes (67 tests), self-check on mypy/stubtest.py is clean, and ruff / black are clean.

Note

The enum branch of verify_var now reads the same local variable rather than stub.type directly. The value is identical there — an enum member is never a bound method — but self-check cannot narrow stub.type to non-None once the surrounding condition tests the local.