It'd be a relatively big task, but IMO, a fork of CGMiner with the CUDAminer code built into it or CUDAminer with a similar interface would be perfect, particularly for switching pools at will and being able to adjust some parameters on the fly to try and tweak the cards for best performance. The original CPUMiner was a good base to start from for this project, but maybe it's time to consider an improved interface.
it probably won't be me doing this fork. the cpuminer code is fairly elegant and minimal - this is why it is so easy to work on it.
With CGMiner you probably need to be an Einstein to get something integrated.
Christian