CI/build hardening: caching, make -j, header deps, header-repo CI, reusable workflows

[balbasty]

Closing — this umbrella is done. Checking off against the actual state:

• D2 (ccache): done — 4 repos wired this session, 7.5x–36x measured speedup.
• D3 (header dependency tracking): done.
• D4 (CI on header-only repos): done — kernels/cpu-impl/cuda-impl all have test-via-cpu-lib/build-via-cuda-lib smoke workflows.
• D5 (reusable workflows): done — python-test.yml/python-lint.yml in fastfields/.github, plus the new cpp-submodule-staleness.yml/cpp-submodule-bump.yml.
• D6 (C++ lint/format/codespell): done — .clang-format/.codespellrc/lint.yaml across all 6 C++/CUDA repos (clang-format non-blocking via continue-on-error, codespell blocking).
• Lower — timeout-minutes/concurrency: done except one loose end already tracked separately as fastfields-lib#54 (this repo's own test.yaml still missing the block the others have).
• Lower — auditwheel/delvewheel: done — confirmed already wired in the reusable release workflow; delvewheel deliberately skipped (no Windows/macOS compiled-wheel leg exists to need it).
• Lower — unpinned fastfields-dlpack dep: done — now ==0.* across the 4 wrapper repos.

Workstream: claude-jitfields-to-fastfields

[balbasty]

Progress on this epic (all merged):

• D4 — CI on the three header-only source repos. ✅ Added a Test workflow to fastfields-kernels, fastfields-cpu-impl, and fastfields-cuda-impl. Each builds the change through its downstream lib (kernels/cpu-impl → fastfields-cpu-lib make test; cuda-impl → fastfields-cuda-lib compile+link with nvcc). cpu-impl/cuda-impl use a two-checkout layout (lib without submodules, then this repo recursively into the lib's impl slot) so the nested kernels submodule populates reliably. (cpu-impl also had a kernels gitlink that a prior commit had clobbered with the dev-tree symlink — restored.)
• D5 — reusable workflows. ✅ Extracted the byte-identical Test/Lint workflows into fastfields/.github (python-test.yml, python-lint.yml, workflow_call); numpy/torch/cupy/bind-py now call them via thin stubs. secrets: inherit forwards CI_SUBMODULE_TOKEN for bind-py's C++ chain; the dlpack-install step self-skips for bind-py via a repo-name guard.
• D3 — header dependency tracking. ✅ fastfields-cpu-lib now compiles both test and library objects with -MMD -MP and -includes the *.d, so editing a kernel/impl header rebuilds the affected object. Also restructured the test build into shared per-module objects (compiled once, .SECONDARY), which both speeds the gate and makes it ccache-friendly.
• D2 (partial) + Lower. ✅ setup-python pip caching and a repo-scoped cancel-in-progress concurrency group added once in the reusable Python workflows (benefits all 5 packages). make -jN was already present in the C++ gates.

Still open:

• D2 ccache for the C++ builds (cpu-lib/cuda-lib + the D4 header-repo CIs that rebuild them). Now unblocked by D3's header-dep tracking; deferred as the highest-effort/-risk piece (a misconfigured cache can produce false greens).
• D6 — no C++ .clang-format/clang-tidy/codespell.
• Lower — concurrency/timeout-minutes on the C++ workflows; packaging auditwheel/delvewheel; unpinned fastfields-dlpack dep in the wrappers.

---

Generated by Claude Code

[balbasty]

Lower — concurrency/timeout-minutes: done. Every C++ gate now has a concurrency group with cancel-in-progress (a new push to a PR cancels its superseded run), added via: cpu-lib #31, kernels #24, cpu-impl #25, cuda-impl #7, cuda-lib #15. The reusable Python workflows got the same in fastfields/.github#2. timeout-minutes is now set on every gate (30 min for the CPU/make test gates; the nvcc gates already had 60).

Still open on this issue: D2 ccache for the C++ builds (deferred — highest effort/risk), D6 (C++ .clang-format/clang-tidy/codespell), and the packaging bits (auditwheel/delvewheel; the wrappers' fastfields-dlpack @ …@main dep is intentionally unpinned during active development).

---

Generated by Claude Code

[balbasty]

Re-triaged this before assigning the remainder — most of it is already done:

• D2 (caching + make -j): done — .github's reusable workflows have pip caching (97b17d4), and fastfields-cpu-lib/.github/workflows/test.yaml runs make -j2 -C . test CXX=clang++.
• D3 (header dependency tracking, -MMD -MP): done.
• D4 (CI on the header-only repos): done.
• D5 (dedupe reusable workflows): done (17e7697).
• Lower — timeout-minutes: done (cpu-lib's test job: timeout-minutes: 30).
• Lower — concurrency: group: done (group: ${{ github.workflow }}-${{ github.ref }}, cancel-in-progress: true).

Confirmed still open:

• D6 — no C++ lint/format/codespell anywhere: zero .clang-format files across kernels/cpu-impl/cuda-impl/cpu-lib/cuda-lib/lib.
• Lower — packaging: no auditwheel/delvewheel step in any wrapper repo's release workflow.
• Lower — unpinned dep: fastfields-dlpack appears as a bare, unversioned dependency in fastfields-numpy/fastfields-torch/fastfields-cupy/fastfields's pyproject.toml.

Assigning an agent for just this narrowed remainder (D6 + the two packaging/pinning items) rather than re-doing what's already landed.

Workstream: claude-jitfields-to-fastfields

---

Generated by Claude Code