Unable to bisect, as the result always comes up as the "bad" value even if setting the good revision to ~12 hours before the regression occurred. This did return to normal at some point (and even decreased below its original value even later), so it's probably safe to say whatever caused this has been fixed.