Julian pointed out that as 0.0 is a bit string of 0s (at least for IEEE),
it's an easy result to get bogusly. So use cos instead of sin, expecting 1.0
back. Has all the other nice properties, plus better error detection.