I think half a bar is way too much to be accounted for by unreported latency. Manual Offset's usually on the order of 20-40 samples. More likely this is the metronome count-in thing that comes up every once in a while. Are you running a 2-beat count-in?
In any case try disabling count-in, and see if it makes a difference. If it does, I would call Cakewalk, and explain the situation to see if they can confirm it's a known issue, and whether they have a solution. As rare as this is, it seems to be a hardware/driver-specific thing. In my memory, no one has ever reported back how they ultimately resolved it other than by not using count-in.
If it's not the metronome count-in , I dunno...