It was using -a (alternatives) rather than -A (architecture) so the test wasn't checking the exact error branch.