I narrowed this down to this example:

c = None
a = [1]
for c in a:
    pass

reveal_type(c)  # int (should be Optional[int])

I believe that this is not about tracking whether iterables are empty, but an issue with type inference with a partial None type and a for loop.

Marking as high priority since this can lead to false negatives and cause issues with mypyc.

Any hints on how to fix this? I'd like to take a shot.

Updates: I did some tracing and summarize below:

• In checker.py -> visit_for_stmt -> analyze_index_variables, it introduces a type assignment in the conditional binder's frames, which I believe serves the purpose of correct type inference inside the loop body.

• When calling reveal_type outside the function body, checkexpr.py -> visit_name_expr -> narrow_type_from_binder gets a restriction from the conditional binder(in @JukkaL 's case, a int) and therefore narrows the type to a int

I am still going through the details but I feel that this is caused by the incorrect conditional binder+name lookup behavior as a modified version of @JukkaL's example presented:

c = "foo"
a = [1]
for c in a:
    reveal_type(c) #  Revealed type is 'builtins.str', but I expect it should be int
    pass
reveal_type(c) #  Revealed type is 'builtins.str'