[Jahongir-Qurbonov]

#22437 changed which functions get processed together with the module top level, which is what makes non-local partial None types work. It now does this only for variables whose first assignment has a literal None rvalue. The checker, however, creates a partial None type for any inferred variable whose initial type is None, including variables that have no rvalue at all. As a result this now fails:

for v in [None]: pass

def k() -> None:
    global v
    v = 1  # error: Incompatible types in assignment (expression has type "int", variable has type "None")

Before #22437 this was accepted and v was inferred as int | None. The same regression happens for y = (z := None) (for z), for w = None if c else None, and for class attributes initialized this way and later assigned via self.a = ... in a method.

This PR treats binding forms that have no rvalue as possibly None, as they were before #22437. These are for loop indices, assignment expression targets, with/match/star targets and tuple items that can't be matched to an rvalue item. It also looks through conditional expressions and assignment expressions in the rvalue. All other rvalues still count as not None, so the speedup from #22437 is kept. This leaves one case unhandled on purpose: x = f() where f returns None. That assignment is already reported as a func-returns-value error.

Added testPartialNoneTypeNonLiteralNone to check-inference.test, next to the test added in #22437. It covers conditional expressions, walrus targets and for loop indices with global, and conditional expressions and walrus rvalues for class attributes assigned via self.

[github-actions]

According to mypy_primer, this change doesn't affect type check results on a corpus of open source code. ✅