Fix macos build

[ognian-]

This adds a Github CI action for macOS, and fixes some building issues.

[ognian-]

@matsen Can you try if this branch fixes the build on your macOS? It's not including the BCR code, I'll make another action with libtorch if this succeeds.

[matsen]

🤖 Tested on macOS — build succeeds, no regressions.

Environment

CPU	Apple M2 Max (arm64, T6020)
OS	macOS 15.7.3 (Darwin 24.6.0)
Compiler	Apple clang 17.0.0 (clang-1700.6.3.2), target arm64-apple-darwin24.6.0
CMake	4.2.0
Generator	Ninja 1.13.2

Build

cmake -DCMAKE_BUILD_TYPE=Debug -DUSE_USHER=no -DUSE_NETAM=no .. -G Ninja
ninja -j16

All 109/109 targets compiled successfully with no errors. Only benign linker warnings:

ld: warning: ignoring duplicate libraries: '-lz'

All four executables were produced: larch-test, larch-dagutil, larch-dag2dot, bcr-larch.

Tests (non-slow, run individually)

Ran every non-slow test (89 tests) one at a time on both main and this PR branch. Results are identical — no regressions introduced by this PR.

	PR branch	main
Pass	55	55
Fail (exit 1)	8	8
Segfault (exit 139)	26	26

The 26 segfaults and 8 assertion failures are all pre-existing on main and appear to be related to running the mock (USE_USHER=no) build on macOS/arm64 — they are not caused by this PR. The exact same test IDs pass and fail on both branches.

Passing tests include: DAG loader, overlay, compact genome, sample tree, RF distance, parsimony, weight accumulator, weight counter, S5F likelihood, Sankoff (all 4), SPR (4 of 7), merge (0 of 5), netam kmer/model/likelihood, mat conversion, mat view, parallel, UA-free parsimony, and more.

[ognian-]

@matsen It should be buildable with BCR and libtorch too: easiest should be to activate the conda environment environment-dev-macos-netam.yml

If it builds for you we can merge.

[matsen]

🤖

Build test on macOS Apple Silicon (M-series, macOS 15.5 Sequoia, AppleClang 17.0)

Compilation: PASS

• Mock build (USE_USHER=no): Compiles cleanly
• Full build (with usher): Compiles cleanly
• Only warnings: ld: warning: ignoring duplicate libraries: '-lz' (benign)

Tests: 86/89 non-slow tests pass

The 3 failures are pre-existing on main (not introduced by this PR):

• [15] Load dagbin: Sample DAG (Read-only) — dagbin section_endpos assertion
• [19] Load dagbin: test_5_trees, 3 iters, seeded — dagbin/protobuf mismatch
• [36] MATConversion: Check SampleIds — leaf sample ID check failure

Protobuf caveat

Initial builds segfaulted on every protobuf parse operation. Root cause: homebrew protobuf 29.3 has a bug in TcParser::AddMessage on Apple Silicon. After upgrading to protobuf 33.4 (brew upgrade protobuf), all protobuf operations work correctly. This affects main equally — not a PR issue.

Conda environment: Does not resolve on arm64

The environment-dev-macos.yml tries to install openmpi-mpicc/openmpi-mpicxx which pull in clangxx_osx-64 — x86-only compiler packages. This needs to be fixed for Apple Silicon. The homebrew-based build works fine though.

BCR + libtorch build: Not tested

• No libtorch installed on this machine
• The conda env (which would provide pytorch-cpu) doesn't resolve on arm64
• The cmake --preset macos-arm64-bcr preset requires a CMakePresets.json that doesn't exist on this branch

Recommendations

1. The conda env files need arm64-compatible packages (openmpi-mpicc/openmpi-mpicxx pull in x86-only compilers)
2. If the macos-arm64-bcr preset was intended, CMakePresets.json is missing from the PR
3. Worth noting the protobuf ≥33.4 requirement somewhere — users with homebrew protobuf 29.3 will hit segfaults