Static CUDA seems to be broken anyway, the upstream build system is awkward and uncooperative, and it’s simpler to just patch it to use dynamic libraries.