CI: fix wheel test-command shadowing + conda-forge libomp for macOS

Previous run produced a green test matrix (all 12 jobs) and sdist, but
all three build-wheels jobs failed. Two independent root causes:

1. Linux and Windows wheels BUILT fine, but cibuildwheel's test-command
   failed with `ModuleNotFoundError: No module named 'wlsqm.fitter.defs'`.

   Root cause: pytest's default `prepend` import mode walks up the
   package tree from each test file looking for a `__init__.py`-less
   ancestor, and inserts that ancestor into sys.path. With
   `tests/__init__.py` present, pytest treated `tests/` as a package,
   walked up to `/project`, and inserted `/project` into sys.path.
   Python then resolved `import wlsqm` to `/project/wlsqm/` (source
   tree, no compiled .so files) instead of the installed wheel at
   `site-packages/wlsqm/`. The source-tree `__init__.py` does
   `from .fitter.simple import *` which fails because `defs.pyx` isn't
   a compiled module.

   Fix: delete `tests/__init__.py` and change the three test files
   that had `from .conftest import ...` to plain `from conftest import ...`.
   Without `tests/__init__.py`, pytest stops the walk at `tests/` and
   inserts `tests/` (not `/project`) into sys.path, which makes
   `conftest` importable by bare name and leaves the installed wheel
   visible at `site-packages/wlsqm/`. The test job is unaffected because
   meson-python's editable loader intercepts `import wlsqm` regardless
   of sys.path order.

   Verified locally: `pytest tests/` still 57/57 green.

2. macOS wheel failed at compile time with:
     `fatal error: 'omp.h' file not found`

   Root cause: Cython's `cimport openmp` in simple.pyx, impl.pyx,
   expert.pyx, and lapackdrivers.pyx unconditionally emits
   `#include <omp.h>` in the generated C. Meson's
   `dependency('openmp', required: false)` fallback only controls
   whether `-fopenmp` is passed — it does not prevent Cython from
   emitting the header include. Apple Clang ships no `omp.h`, so the
   previous commit's "drop libomp entirely, fall back to serial"
   approach was not actually a viable baseline on macOS.

   Fix: install conda-forge's `llvm-openmp` via a standalone micromamba
   in cibuildwheel's macOS before-all. conda-forge ships llvm-openmp
   rebuilt with a conservative deployment target (11.0 on arm64,
   matching cibuildwheel's default), so delocate-wheel is happy to
   bundle the resulting `libomp.dylib` into the wheel without the
   "Library dependencies do not satisfy target MacOS version" refusal.
   CPPFLAGS / LDFLAGS / PKG_CONFIG_PATH / DYLD_FALLBACK_LIBRARY_PATH
   are set in `[tool.cibuildwheel.macos].environment` so that meson
   discovers libomp during the build and delocate finds it at
   bundling time. Same pattern as scipy and numpy's macOS wheel CI.

Also add `.claude/` to .gitignore (Claude Code internals: the
scheduled_tasks lockfile was about to sneak into `git add -A`).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: Technologicat

.gitignore

@@ -17,3 +17,4 @@ dist/
 # Cython HTML annotations (generated by `cython -a`)
 wlsqm/fitter/*.html
 wlsqm/utils/*.html
+.claude/

pyproject.toml

@@ -119,22 +119,29 @@ skip = "*-win32 *-manylinux_i686 *-musllinux*"
 test-requires = ["pytest", "numpy", "scipy"]
 test-command = "pytest {project}/tests -v"
 
-# macOS wheels ship without OpenMP. We could `brew install libomp` in a
-# `[tool.cibuildwheel.macos] before-all` step and let delocate-wheel bundle
-# libomp.dylib into the wheel, but Homebrew's current libomp (22.x) has a
-# minimum macOS target (15.0) newer than cibuildwheel's default arm64 target
-# (11.0). delocate-wheel correctly refuses to bundle it because the resulting
-# wheel would lie about its compatibility.
+# macOS: ship parallel wheels by installing conda-forge's `llvm-openmp`
+# via a standalone micromamba in `before-all`. Apple Clang does not bundle
+# an OpenMP runtime, and Homebrew's current libomp (22.x) has a minimum
+# macOS target of 15.0, so `delocate-wheel` refuses to bundle it (the
+# wheel's default cibuildwheel arm64 target is 11.0 and delocate does not
+# let a wheel lie about its compatibility). conda-forge's llvm-openmp is
+# rebuilt with a conservative deployment target that matches our wheel
+# target, so delocate is happy to bundle the .dylib.
 #
-# We could either (a) bump MACOSX_DEPLOYMENT_TARGET to 15.0, shrinking the
-# supported macOS range by ~4 years, or (b) build libomp from source with a
-# lower target (complex CI). Instead, we lean on meson's
-# `dependency('openmp', required: false)` fallback: without libomp on the
-# build host, meson builds the extension in serial mode. Cython's `prange`
-# degrades to plain `range`, tests still pass, and macOS wheels are slower
-# but portable all the way back to macOS 11. Users who want parallel fits
-# on macOS can install from source: `brew install libomp` and then
-# `pip install --no-binary wlsqm wlsqm`. README.md documents this.
+# This is the same pattern scipy and numpy use on their macOS wheel CI.
+#
+# A serial-mode macOS build is NOT a usable fallback here despite meson's
+# `dependency('openmp', required: false)`, because Cython's `cimport openmp`
+# in simple/impl/expert/lapackdrivers unconditionally emits `#include <omp.h>`
+# in the generated C. Without libomp headers at build time, Apple Clang
+# errors out on the missing omp.h before meson's fallback path can take
+# effect. So we need libomp headers at build time regardless.
+[tool.cibuildwheel.macos]
+before-all = '''
+curl -Ls https://micro.mamba.pm/api/micromamba/osx-arm64/latest | tar -xvj -C /tmp bin/micromamba
+/tmp/bin/micromamba create -y -p /tmp/libomp-env -c conda-forge llvm-openmp
+'''
+environment = { CPPFLAGS = "-Xpreprocessor -fopenmp -I/tmp/libomp-env/include", LDFLAGS = "-L/tmp/libomp-env/lib -Wl,-rpath,/tmp/libomp-env/lib -lomp", DYLD_FALLBACK_LIBRARY_PATH = "/tmp/libomp-env/lib", PKG_CONFIG_PATH = "/tmp/libomp-env/lib/pkgconfig" }
 
 [dependency-groups]
 dev = [

tests/__init__.py

@@ -16,7 +16,7 @@
 
 import wlsqm
 
-from .conftest import poly2d_order2, poly3d_order2
+from conftest import poly2d_order2, poly3d_order2
 
 
 def _make_expert_2d(ncases, nk_per_case, order=2, algorithm=None, ntasks=1):

tests/test_expert.py

@@ -11,7 +11,7 @@
 
 import wlsqm
 
-from .conftest import poly1d_order2, poly2d_order2, poly3d_order2
+from conftest import poly1d_order2, poly2d_order2, poly3d_order2
 
 
 def test_interp_2d_function_value_matches_analytical(rng):

tests/test_interp.py

@@ -10,7 +10,7 @@
 
 import wlsqm
 
-from .conftest import (  # noqa: E402
+from conftest import (  # noqa: E402
     poly1d_order2,
     poly2d_order2,
     poly2d_order3,