Scope CUDA link deps to their targets in meson.build

[bodono]

Follow-up cleanup to #221, addressing review feedback on that PR, plus CI coverage for the GPU build.

1. Scope CUDA link deps to their targets (meson.build)

The use_gpu block appended cuda/cublas/cusparse to the shared _deps list:

_deps += cuda_dep
_deps += cublas_dep
_deps += cusparse_dep

Because _deps is reused by every extension module defined later in the file, those CUDA libraries leaked into _scs_mkl, _scs_accelerate, and _scs_cudss whenever use_gpu=true was combined with those backends — linking them against libraries they don't use. The link_cudss block had the same anti-pattern with cuda/cudss.

This builds a per-target dependency list instead of mutating the shared _deps, mirroring the existing _mkl_deps = _deps + [mkl_rt_dep] pattern already used for the MKL target:

gpu_deps = _deps + [cuda_dep, cublas_dep, cusparse_dep]   # use_gpu
cudss_deps = _deps + [cuda_dep, cudss_dep]                # link_cudss

Also adds a comment explaining the GPU sources are listed explicitly (Meson's fs module has no glob()) and must be kept in sync with the submodule's .c files.

2. Add a build-only GPU CI job (build.yml)

Neither the use_gpu build nor the cuBLAS/cuSPARSE linking fixed in #221 was covered by CI. The new build_gpu job installs the CUDA toolkit libraries from conda-forge and builds with -Duse_gpu=true. Linking against cuBLAS/cuSPARSE requires the toolkit but not a physical GPU, so the job builds and link-checks without running the solver.

Because a shared library links successfully even with undefined symbols, the job asserts via readelf that _scs_gpu records libcublas and libcusparse as NEEDED entries — the #221 bug was a missing link that only surfaced at import time as undefined symbol: cusparseDestroySpMat.

Testing

• meson setup + ninja on the default config build cleanly locally (all targets, including the previously-affected _scs_accelerate).
• The new build_gpu CI job exercises the full use_gpu=true build + link path (this host has no CUDA toolkit, so it could not be built locally).

🤖 Generated with Claude Code