Thanks for the patch. The root cause looks right: protected vision/audio tensors should not be downcast to FP16 in the float16 oQ path.

One gap is that this patch only changes the _should_quantize_tensor() == false fallback. The main 2D vision/audio weight tensors still go through _should_quantize_tensor() == true, then _get_predicate_bits() returns None, so they hit the existing bits is None fallback and are still cast to target_dtype.

I will handle the remaining part in a follow-up commit: apply the same protected pass-through dtype policy to both fallback paths and add a regression test for a 2D vision/audio tensor with dtype="float16".