the latest math::round's nearest function does check for big-endian or little-endian then does the right thing. Otherwise, a fix in another module uses 0.500000001 to do rounding. Of course this also exists in javascript where the simple solution is to use 0.500000001.