My guess is different...and I think its a bug (or feature?) in the resent code that aborts the APs that were sent to a CPU if they get resent to an Nvidia GPU that is not OpenCl "capable" (i.e. its not on a host running Boinc 7.xx)...
I have disabled AP's for GPU for now, so it can't be that.
Well then that answers it, that AP just suffered the same fate that VLAR's do.
I disabled, removed the app_info section, for GPU APs days ago, so that AP could only be sent/resent to CPU.
But your request at the time would've been for GPU work so the same mechanism kicked into action.