Visual comparison in GUI testing, and a recent "horrible" regression