Backend seams

There are no “three layers” stacked. The architecture is MPI + Kokkos:

• MPI distributes subdomains across ranks (one GPU per rank in GPU mode). Everything goes through my_rank() / n_ranks() / all_reduce_* from include/pops/parallel/comm.hpp. Without POPS_HAS_MPI, these functions return rank 0 / 1 rank: the code is serial by construction.

• Kokkos parallelizes the local compute and abstracts the hardware through its ExecutionSpace: Cuda backend for NVIDIA GPUs, Serial/OpenMP for CPU. Everything goes through for_each_cell (and for_each_cell_reduce_*) from include/pops/mesh/execution/for_each.hpp, which switches CPU <-> GPU at compile time without touching the call sites.

No CUDA kernel is written by hand: the same .cpp targets CPU and GPU depending on the Kokkos backend active at compile time. nvcc_wrapper is only the compiler required by the Kokkos Cuda backend.

The Python pops module is serial by default. The _pops extension (pybind11) is built in CI only in Kokkos Serial (without MPI). No Python test exercises the Kokkos OpenMP, Cuda or MPI paths. Multi-thread, GPU and distributed are driven from the C++ facade (System / AmrSystem), not from Python.