ci(macos,windows): drop submodules:true, use selective submodule list

[Fedr]

Summary

Bring build-test-macos.yml and build-test-windows.yml in line with the selective-list pattern the other six workflows in this repo (build-test-emscripten, build-test-linux-vcpkg, build-test-ubuntu-arm64, build-test-ubuntu-x64, pip-build, update-docs-manual) already use, and refine the pattern in all eight to recurse only into thirdparty/mrbind/deps/cppdecl — the one nested submodule MeshLib's build actually consumes — instead of recursing into every entry's .gitmodules blindly.

After this PR, every workflow that follows the Checkout + Checkout third-party submodules two-step pattern has the same shape: parent actions/checkout without submodules: true, followed by a selective git submodule update --init --depth 1 of just the submodules that workflow consumes, plus an explicit git -C thirdparty/mrbind submodule update --init --depth 1 deps/cppdecl for the one path that needs recursion.

Why

Both workflows previously had:

- name: Checkout
  uses: actions/checkout@v6
  with:
    submodules: true             # ← inits all 29 top-level submodules

- name: Checkout third-party submodules
  run: |
    git submodule update --init --recursive --depth 1 thirdparty/mrbind

submodules: true clones every entry in .gitmodules (29 top-level submodules — including heavyweights like openvdb, GDCM, libE57Format, mbedtls, onetbb, libzip, c-blosc) regardless of whether the build will use the source. macOS gets most of those from Homebrew or builds them via build_thirdparty.sh's thirdparty CMakeLists; Windows gets them from vcpkg. Neither workflow actually consumes most of those submodules' source.

This PR drops submodules: true and lists exactly the submodules each platform actually consumes:

Platform	Submodules listed	Why
macOS (16)	imgui, eigen, parallel-hashmap, expected	header-only, consumed directly
	googletest, OpenCTM-git, libE57Format, glad, tinygltf, laz-perf	built via thirdparty/CMakeLists.txt's Apple branch
	clip	built standalone via scripts/build_thirdparty.sh → clip.sh
	fastmcpp, nlohmann-json, cpp-httplib	add_subdirectory(${MESHLIB_THIRDPARTY_DIR}/fastmcpp); nlohmann-json + cpp-httplib are fastmcpp's transitive deps
	mrbind, mrbind-pybind11	binding-generator build tools
Windows (11)	imgui, eigen, parallel-hashmap, expected	header-only
	OpenCTM-git, laz-perf	source/OpenCTM/CMakeLists.txt and source/laz-perf/CMakeLists.txt add_subdirectory them
	fastmcpp, nlohmann-json, cpp-httplib	same as macOS
	mrbind, mrbind-pybind11	binding-generator build tools

Windows lists fewer because vcpkg supplies the equivalents — except for the five that get pulled in via add_subdirectory from source/<x>/CMakeLists.txt and so need source even on the vcpkg path.

Recursion is targeted

Of the 16 submodules on the macOS list, only thirdparty/mrbind has a nested submodule MeshLib's build actually needs (deps/cppdecl, the codegen library mrbind links against). Two other submodules carry nested .gitmodules entries but neither is consumed by MeshLib:

• thirdparty/expected → cmake/tl-cmake is unreferenced by the cantonios fork's CMakeLists at the pinned commit.
• thirdparty/libE57Format → test/extern/googletest is gated behind libE57Format's own test target, which MeshLib never builds.

The other 13 have no nested .gitmodules at all — --recursive on them was a strict no-op.

So the workflow does:

- name: Checkout third-party submodules
  run: |
    # Selective init -- parent Checkout drops submodules:true.
    # https://github.com/actions/checkout/issues/1779
    git submodule update --init --depth 1 \
      thirdparty/imgui \
      thirdparty/eigen \
      ...
      thirdparty/mrbind \
      thirdparty/mrbind-pybind11
    # mrbind needs deps/cppdecl; recurse only there
    git -C thirdparty/mrbind submodule update --init --depth 1 deps/cppdecl

Bash workflows use \ continuation; the Windows workflow uses pwsh backtick (`) for the same shape. (See the Windows step.)

Measured impact

Comparing the latest run 24994826587 (764cd258) against master baseline 24987634018 (466884e1). Times = Checkout + Checkout third-party submodules summed (seconds):

Platform	Master	This PR	Δ
macOS x64 Release (macos-15-intel)	88	53	−35
macOS arm64 Debug (macos-latest)	112	47	−65
macOS arm64 Release (self-hosted, warm)	10	8	−2
Win msvc-2019 Release CMake	96	31	−65
Win msvc-2019 Debug CMake (iter-debug)	94	26	−68
Win msvc-2022 Debug MSBuild	86	30	−56
Win msvc-2022 Release CMake	112	30	−82
Sum across 7 legs	598	225	−373 (−62 %)

Per master run that's ~6.2 minutes of cumulative runner-time saved across these 7 GitHub-hosted/self-hosted legs, with ~50–80 s shaved off each Windows leg's critical path.

The self-hosted arm64 Release runner is essentially neutral — its .git/modules/ cache stays warm between runs so neither pre- nor post-PR actually clones much there. The savings show up on the GitHub-hosted ephemeral-VM legs.

Pattern uniformity

All eight workflows that follow the Checkout + Checkout third-party submodules two-step pattern now use the same shape — no submodules: true on the parent, selective list with explicit cppdecl recursion in the follow-up step. The remaining occurrences of submodules: true / submodules: recursive in this repo are in workflows that don't follow that pattern (prepare-images.yml, update-docs.yml, update-win-version.yml, the macOS- and Windows-specific jobs in pip-build.yml) — those genuinely need the full tree for unrelated reasons and are out of scope here.

Iteration history

CI on this branch surfaced changes incrementally:

• 55816785 — initial drop of submodules: true on macOS + Windows with selective lists.
• 60fc8e3f — added OpenCTM-git, laz-perf, fastmcpp, nlohmann-json, cpp-httplib after Windows configure-time errors on source/{OpenCTM,laz-perf,fastmcpp}/CMakeLists.txt add_subdirectory calls.
• d0c12915 — added clip after macOS build_thirdparty.sh failed on the standalone clip.sh step.
• 0f13185b — dropped --recursive, added explicit git -C thirdparty/mrbind submodule update --init --depth 1 deps/cppdecl. Windows used per-line separate commands at this stage.
• 764cd258 — switched Windows from per-line separate commands to a single batched call with pwsh backtick continuation, recouping ~7 s/leg of per-invocation overhead.

Replaces

Closed PR #5991, which took the opposite direction (add submodules: true everywhere, trim the second step to just thirdparty/mrbind). The selective-list approach in this PR is what the repo had been using on its other six workflows; extending it to macOS and Windows is the simpler and faster end state.

🤖 Generated with Claude Code