MPI + Kokkos Serial

What it is. Distributed build with the Kokkos Serial execution space: comm.hpp goes through MPI_Comm_rank/size + collectives on MPI_COMM_WORLD. The domain is split into boxes distributed across ranks; the halos exchange via cross-rank fill_boundary, the reflux/mass via all_reduce_*. Single-thread CPU per rank.

Build. Kokkos installed with Kokkos_ENABLE_SERIAL=ON, plus the MPI seam:

cmake -S . -B build-mpi -DCMAKE_BUILD_TYPE=Release \
  -DPOPS_USE_KOKKOS=ON -DPOPS_USE_MPI=ON -DKokkos_ROOT="$KOKKOS_PREFIX"
cmake --build build-mpi -j

Run. The MPI targets each replay np=1/2/4 under mpirun. For -np 4 on a small machine, allow oversubscribe:

OMPI_MCA_rmaps_base_oversubscribe=true \
  ctest --test-dir build-mpi --output-on-failure

Validation: CI (ci-full job). Job mpi (ubuntu-latest / MPI, OpenMPI). Full mode only. Checks invariance to the number of ranks: the observables (parity, AMR, Krylov, mass) are bit-identical at np=1/2/4. Status ci-full on the ~21 entries of the POPS_HAS_MPI block; the non-MPI tests run at np=1 in this build (MPI-linked, single-process).