In runtime we rely on several aspects of CTK and their discovery strategy is a bit cluttered:

• CTK shared libraries
  • Fully rely on cuda-pathfinder's documented search order, after this PR get merged
• CTK headers
  1. $DISCOVERED_CUDA_PATH/include
  2. NVIDIA Pip packages: /path/to/site-packages/nvidia/cuda_runtime/include (for CUDA 12) or /path/to/site-packages/nvidia/cuda/cu13/include (for CUDA 13)
• NVCC
  1. $NVCC env var
  2. $DISCOVERED_CUDA_PATH/bin/nvcc
  3. OS default strategy ($PATH)

For CTK headers I think it's better to migrate to cuda.pathfinder.find_nvidia_header_directory at some point.

Originally posted by @kmaehashi in #9444 (comment)