I'm sure there must be various ways to beat screen into submission and get it to do what people want, but on any machine where there are multiple CUDA toolkit versions laying around (I have about 6) it's best to have binaries compiled such that they don't depend on environment variables. 

Without dealing with this at linkage time, as soon as you have different binaries compiled against different CUDA toolkit versions, environment variables either have to be embedded in per-app startup scripts to ensure they pick up the right library, or you end up having various problems.  I think that starting with CUDA 5.5 and going forward, the easy way out is simply to use libcudart_static.a and avoid this problem altogether, but I realize that not everyone has an up-to-date CUDA toolkit installed, and you probably want make the cudaminer build work out-of-the-box on the principle of least user emails :-)

Cheers,
  John