I expect they would still be reported by normal type inference, no? So I'm not sure any special handling is needed here, thus maybe we don't need a TODO either?

Maybe ideally we'd add some tests for cases where diagnostics occur either in iteration, or more generally in type inference of the RHS of a self.x = ... assignment, and verify that we do surface those diagnostics.

I expect they would still be reported by normal type inference, no?

I also hoped that would be the case, but no. We emit the non-iterable diagnostic from infer_for_statement_definition, and attribute assignments are not definitions. I added a test to demonstrate this.

We have a similar problem where we do not yet identify cases where (normal) assignments to attributes conflict the declared type of that attribute. We have existing tests for this as well.

(Edit: no, I already implemented this 😄)

I'll also open a new ticket to address both of these issues.

or more generally in type inference of the RHS of a self.x = ... assignment

The general case works fine. We only fail to emit diagnostics if they originate from infer_…_definition calls, which do not apply to attribute assignments. I added the following test case:

class C:
    def __init__(self) -> None:
        # error: [too-many-positional-arguments]
        self.x: int = len(1, 2, 3)