Backend matrix

#	Backend	CMake options (in addition to -DCMAKE_BUILD_TYPE=Release)	Local build?	CI?	Validated where
1	Kokkos Serial	-DPOPS_USE_KOKKOS=ON + Kokkos Serial	Yes	Yes (ci-fast, PR gate)	CI ubuntu (reference oracle)
2	Kokkos OpenMP	-DPOPS_USE_KOKKOS=ON + Kokkos OpenMP	Yes	Yes (ci-full)	CI job kokkos-openmp, 91/91 ctest
3	MPI + Kokkos Serial	-DPOPS_USE_KOKKOS=ON -DPOPS_USE_MPI=ON + Kokkos Serial	Yes	Yes (ci-full)	CI job mpi, rank-invariant np=1/2/4
4	MPI + Kokkos OpenMP	-DPOPS_USE_MPI=ON -DPOPS_USE_KOKKOS=ON + Kokkos OpenMP	Yes	No (never MPI+Kokkos in CI)	ROMEO x64cpu manual (52/57 rank-invariant)
5	Kokkos CUDA	-DPOPS_USE_KOKKOS=ON + Kokkos Cuda + nvcc_wrapper	No (no local nvcc)	No (never CUDA in CI)	ROMEO GH200 manual (python/tests/gpu/, sbatch)
6	MPI + Kokkos CUDA	-DPOPS_USE_MPI=ON -DPOPS_USE_KOKKOS=ON + Kokkos Cuda + nvcc_wrapper	No	No	ROMEO multi-GPU manual (srun --gpus-per-task=1)

Quick read:

• Configs 1-3 are covered by GitHub CI (1 on every PR; 2-3 in full mode ci-full: push to master, cron, dispatch, or ci-full label).

• Configs 4-6 are never in CI: CI joins MPI with Kokkos Serial but never with Kokkos OpenMP, and never builds CUDA. They are validated manually on ROMEO, by bit-for-bit comparison to the Serial oracle (dmax).

• Configs 5-6 have no local build: nvcc runs only on the aarch64 GPU node of ROMEO.

For the detailed coverage (each test x each backend column, with the status ci-fast / ci-full / ROMEO / self-skip / ?), the source of truth remains BACKEND_COVERAGE.md. For the GPU port phases and the detailed validation results, see GPU_RUNTIME_PORT.md.