Fix 1.20 regression: type narrowing lost on reassignment with generator/ternary expressions

[parinporecha]

Summary

PR #20622 introduced a binder_version guard to prevent false negatives when walrus operators (:=) interact with type inference fallback logic. However, this guard was too broad: any expression that bumps the binder version (including generator expressions with ternary operators) would disable the union fallback path, causing type narrowing to be lost on reassignment.

Problem

from typing import Iterable

def foo(args: Iterable[str | int] | str | int) -> Iterable[str]:
    if isinstance(args, (str, int)):
        args = (args,)
    args = (
        arg if isinstance(arg, str) else str(arg)
        for arg in args
    )
    return args  # Error: Generator[str | int, ...] not compatible with Iterable[str]

This worked in mypy 1.19 but regressed in 1.20.

Fix

Replace the binder_version == self.binder.version check with an explicit check for AssignmentExpr (walrus :=) in the expression tree, following the suggestion from @ilevkivskyi in the issue.

The walrus operator is the specific construct that causes problematic binder mutations during r.h.s. acceptance. Other constructs that bump the binder version (ternary in generators, etc.) are benign and should not disable the fallback.

Changes

• Add _AssignmentExprSeeker (TraverserVisitor) to detect walrus in expression trees
• Add TypeChecker._has_assignment_expr() static method
• Replace binder_version == self.binder.version with not has_walrus in union_fallback guard

Testing

• All 60 check-python38 tests pass, including testAssignToOptionalTupleWalrus and testReturnTupleOptionalWalrus
• Full testcheck suite running locally

Issue 21273

[github-actions]

Diff from mypy_primer, showing the effect of this PR on open source code:

graphql-core (https://github.com/graphql-python/graphql-core)
+ src/graphql/execution/execute.py:1793: error: Unused "type: ignore" comment  [unused-ignore]

[ilevkivskyi]

This will have horrible performance implications. Since the decision is made post-visit, a simple flag set in visit_assignment_expression() would be sufficient. Moreover, that flag can be set only if the binder was actually used.