ctf: cc4s CTF Rust port

Pinned reference: https://gitlab.cc4s.org/cc4s/ctf at f69cbb46e23bc2f39cda5722ce096f56301dab4f.

This is an independent, in-progress CPU/MPI port, not yet a replacement for the full reference. No muffin-tin integration, C++ CTF linkage, Python/C++ compatibility layer, or CUDA implementation is included.

The implementation uses a single ctf crate organized by upstream responsibilities. Native calls belong in internal FFI modules. BLAS/LAPACK provide local numerical kernels initially; their internal Rust interface must remain replaceable by faer (pending confirmation of the requested name “fear”). ScaLAPACK remains the distributed decomposition implementation, independently of local kernel choice. Do not introduce a plugin framework or replace distributed algorithms with gather.

The C++ expression-template layer's automatic ordering of multi-term contraction chains (Sum_Term::estimate_time/execute, term.cxx:426,486-518) has no Rust equivalent: interface/{term,idx_tensor,fun_term} are replaced by direct operations, not ported, so a chain written as successive contract_from/sum_from calls executes in the order the caller writes it. This is a scope limit, not a pending port item; see docs/coverage.md.

Validation contract

Class R: preserve the pinned CPU algorithms, layouts and communication steps. Integer/index/layout/discrete-algebra checks are exact. Floating-point checks must preserve each upstream test's metric and tolerance. Decompositions compare orthogonality/residual/reconstruction, not individual eigenvectors.

Run the completed acceptance set once at 1, 2, 4 MPI ranks in WSL Ubuntu-26.04, with the Cargo target directory on the Linux filesystem, then Windows native. Close passing checks immediately. At most three extra diagnostic computations for numerical failures, with affected checks rerun only after actual fixes. Compilation failures do not justify numerical tolerance changes.

Completion requires all five phases and the full source/test mapping, not just the currently implemented foundations. See docs/coverage.md and docs/upstream-inventory.tsv for scope and progress.

Running the current subset

MPI ownership

The host owns an rsmpi Universe, initializes with at least mpi::Threading::Funneled, and checks the provided thread level. Construct Context::world(&universe) or Context::from_communicator(&universe, &communicator) on the MPI main thread; ctf checks both requirements and never initializes or finalizes MPI. Contexts borrow the universe and any host communicator. Drop tensors first, explicitly close() split contexts, then release host communicators and the universe. Only ctf-owned splits are freed by close(); dropping an unclosed split leaks its handle rather than performing a collective on Drop. The rsmpi dependency has default features disabled; custom reductions do not use libffi.

Platforms

scripts/acceptance-wsl.sh is host-portable: it runs on any Unix host with mpirun on PATH, including WSL Ubuntu-26.04 and macOS.

• WSL Ubuntu-26.04: install the Rust toolchain, MPI development files, libclang-dev, BLAS/LAPACK and ScaLAPACK development libraries, and use the Linux filesystem for Cargo's target directory.
• macOS: brew install open-mpi openblas scalapack. build.rs adds the Homebrew prefix (HOMEBREW_PREFIX, else brew --prefix, else /opt/homebrew) to the link search path; ScaLAPACK/BLAS link as scalapack/openblas, the same names native Windows uses (Linux keeps scalapack-openmpi/blas/lapack).
• Native Windows: scripts/acceptance-native.ps1 -MsMpiBin <path>; see docs/native-windows.md.

Both scripts read one manifest, scripts/acceptance-targets.tsv (columns target, class mpi/local/lib/excluded, ranks, sets, note); scripts/check-targets.sh fails if a Cargo test target is not triaged in it. Each target runs once per rank count, printing RUN_EXIT <target> ranks=<n> exit=<code> and a per-target log under $CTF_ACCEPTANCE_LOG_DIR (default $HOME/.cache/ctf-rs-acceptance/<timestamp>/); a failing target does not stop the run, and the script exits nonzero at the end if any target failed. CTF_ACCEPTANCE_ONLY and CTF_ACCEPTANCE_RANKS restrict a run to a subset of targets or ranks.

The manifest now includes the thirteen S1 targets, dgtog_redistribution and model_io; upstream_bench_contraction (informational) and upstream_model_trainer (needs -write) are listed as deliberate exclusions. The S1 close ran those thirteen targets at WSL and native 1/2/4 under the fixed plan.v4 contract; every target passed at every rank and no HANDOFF question remains (docs/validation.md, section "S1d and S1 close"). This record is not permission to rerun a closed numerical check; the acceptance scripts now cover the same targets going forward.

To avoid Windows-drive source I/O, run bash scripts/sync-wsl.sh from the Windows-backed repository in WSL. Build/test in /home/xylxp/ctf-rs-work after syncing; keep edits and Git commits in D:\projects\ctf-rs. The sync does not copy .git, remove files, or change the delivery repository. Passing acceptance sets need not be repeated merely because the work copy location changed.

Commit cadence

Commit each completed implementation/validation batch with its scope and evidence. Batch commits do not imply that an entire numbered phase or the full port is done. Keep native build artifacts outside the source tree; use repository-local excludes for machine-specific state. Do not push without a separate request.