I have also observed the same behavior on 3 machines, and I have been collecting logs of GPU utilization through nvidia-smi which correlate to the cudaminer log I'm also keeping. I ran cudaminer with "-P" so I've also got a lot of the network protocol. I captured this happening probably ten times last night. It occurs on machines that are also connected to different stratum endpoints. So it's not just an issue with useast, or uswest, for example.
I should also add that the observed idle periods are occasionally quite long, often 15 to 20 seconds. This seems like a very long time to me. When it occurs, no timeouts or errors are printed by cudaminer, so from its point of view there is no problem. This behavior began when wafflepool switched to the new stratum code the other day. Prior to that, I never observed this behavior before.
Any chance you have the protocol dump around when they're idling? Might let me know whats being sent/not being sent during that time...