I agree that with you that "some errors are expensive and some are not," but you don't know which category a bug falls under (cheap/expensive) until after the fact. So making that decision of what (or what not) to test isn't always so black and white.