fix(make): MacPorts libomp built every macOS engine single-threaded

[crichalchemist]

The libomp probe in c/Makefile derives -I and -L from one prefix and only knows Homebrew's layout. MacPorts (port install libomp) puts the header at include/libomp/omp.h and the library at lib/libomp/libomp.dylib, one level deeper on both sides, so no prefix can ever satisfy the check. On such a Mac every build prints libomp not found: building single-threaded and means it — the OpenMP pragmas are silently discarded across all nine engines, with nothing else to tell the user the warning is wrong.

Found while forward-merging #1338 on an x86_64 Mac with MacPorts libomp 21.1.7; kept out of that PR because it is unrelated to the Vulkan tier.

c/Makefile — the probe resolves separate OMPINC / OMPLIB directories. The Homebrew prefix check is unchanged and still comes first (including the CLT-only fallback to /opt/homebrew/opt/libomp and /usr/local/opt/libomp); MacPorts' two directories are probed directly when it fails — but only when OMPDIR was not the user's choice. An explicit OMPDIR, from the command line or the environment, is taken as given: with libomp inside it is used, without it the build is single-threaded, which tests/test_makefile_platform.py pins (OMPDIR=/colibri-test/missing-libomp must not yield -fopenmp). An environment OMPDIR used to be overwritten by the probe's own := assignment; its origin is now read before that assignment, so it counts as the choice it is. The warning names both install spellings.

c/Makefile.deepseek-v4 — carries its own copy of the probe and had already drifted from the parent block its comment claims to mirror: it lacked the CLT-only fallback. Both blocks are identical again, modulo each file's spacing.

c/qwen38_core.h — workers is read only by num_threads(workers); without OpenMP the pragma is gone and clang reports it unused. Discharged in the #else of the existing _OPENMP guard; the OpenMP build is untouched.

c/tests/test_olmoe_matmul_q.c — HAVE_FAST_DOT_I8 is defined under __ARM_NEON or __AVX2__, and the test binaries compile without -march, so on x86_64 the SKIP branch is taken with failures declared and never read. Discharged on that branch. Pre-existing and unrelated to OpenMP — simply the other warning an x86_64 Mac meets on the way to the 0-warning bar.

docs/quickstart.md — the MacPorts spelling next to the Homebrew one.

Verified on a 2017 iMac (x86_64, macOS 13, Apple clang, MacPorts libomp 21.1.7), from a clean binary each time:

check	result
plain make qwen38 (no flags)	probe warning gone, 0 warnings, otool -L shows /opt/local/lib/libomp/libomp.dylib
Homebrew-shaped prefix as OMPDIR, command line and environment (dry runs)	-I$(OMPDIR)/include -L$(OMPDIR)/lib -lomp — the primary branch is unchanged, and the environment is honoured
OMPDIR=/nonexistent, command line and environment (dry runs)	libomp not found, no -fopenmp — the explicit choice wins
forced no-OpenMP build, make qwen38 OMPC= OMPL=	0 warnings — the workers fix holds
Makefile.deepseek-v4 with UNAME_M=arm64, plain and with a bogus environment OMPDIR (dry runs)	same flags and same refusal as the parent; the mirrored block behaves identically
python3 -m unittest tests.test_makefile_platform	3/3 OK
make check	exit 0, 0 compiler warnings, Python suite 970 tests OK (105 skipped), OpenMP active on 43 compile lines

One thing this changed on purpose during review: my first cut let a bogus OMPDIR fall through to MacPorts, and test_darwin_portable_build_does_not_force_x86_architecture caught it. The test is right — the fallback is now gated on $(origin OMPDIR).

Not exercised here: a real Homebrew Mac (the macOS CI jobs brew install libomp, so that branch is covered there) and a linked arm64 deepseek_v4 — the file refuses x86_64 macOS, and CI deliberately does not build V4 on macOS either, so the mirrored block is verified by parse and by being identical logic to the parent's, which is verified in full.

🤖 Generated with Claude Code

https://claude.ai/code/session_01NJcbVYkasdfXg5DEsHdSCt

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.