An upstream Clang change (<https://reviews.llvm.org/D110596>) has been merged to make symlinking /usr/lib/cuda/include to /usr/include work. The shim header approach would still have the benefit of being smaller and thus Clang being able to parse the version from it faster.