Fix setup wheel-build

I fixed the likely wheel-build failures in the release workflow.

Changed:
- Updated `.github/workflows/python-publish.yml`
  - `cibuildwheel==2.21.3` -> `cibuildwheel>=3.0,<4`
  - Added `cp314-*` wheels to match Python 3.14 support
  - Removed musllinux wheel builds for now to reduce release surface
  - Replaced forced Linux `CFLAGS/LDFLAGS` with `METBIT_PORTABLE_BUILD=1`
  - Fixed artifact downloads by downloading `sdist` and `wheels-*` separately

- Updated `setup.py`
  - Avoids `-march=native` during portable wheel builds
  - Uses Windows-safe compile flags (`/O2`) instead of Unix-only flags
  - Adds Python 3.14 classifier

Verification:
- Built a portable Python 3.14 macOS arm64 wheel successfully:
  `metbit-9.0.0-cp314-cp314-macosx_26_0_arm64.whl`
- Installed that wheel into a fresh venv successfully
- Smoke test passed:
  `import metbit`, `pearson_columns(np.eye(5))`, `backend_info()`
- `twine check` passes with only the existing long-description warnings

Local cibuildwheel limitations:
- macOS cibuildwheel could select `cp314`, but cannot complete outside CI without python.org framework Pythons.
- Linux cibuildwheel could select `cp314-manylinux_x86_64`, but Docker is not running locally, so I couldn’t complete the container build here.

Fixed the release wheel failures from the CI log.

What changed:
- [python-publish.yml](/Users/kawa/work/kawa-technology/GitHub/metbit/.github/workflows/python-publish.yml:135): upgraded to `cibuildwheel>=3.0,<4` so Python 3.14 wheel selectors work.
- [python-publish.yml](/Users/kawa/work/kawa-technology/GitHub/metbit/.github/workflows/python-publish.yml:140): added `cp314-*`, skipped musllinux, and set `METBIT_PORTABLE_BUILD=1`.
- [python-publish.yml](/Users/kawa/work/kawa-technology/GitHub/metbit/.github/workflows/python-publish.yml:148): replaced the multiline Windows-breaking `python -c` command with a single-line command and assert `native_c` is present.
- [setup.py](/Users/kawa/work/kawa-technology/GitHub/metbit/setup.py:77): removed `-march=native` from portable wheel builds and uses `/O2` on Windows.
- [metbit/_native_backend.c](/Users/kawa/work/kawa-technology/GitHub/metbit/metbit/_native_backend.c:344): fixed the OpenMP/GIL macro bug that caused Linux C compilation failure.
- [setup.cfg](/Users/kawa/work/kawa-technology/GitHub/metbit/setup.cfg:2): changed deprecated `description-file` to `description_file`.

CI root causes found:
- macOS: `-march=native` on Apple M3 leaked into x86_64 wheel builds, extension skipped, then `delocate` failed.
- Linux: C extension failed to compile due nested `Py_END_ALLOW_THREADS`, extension skipped, then `auditwheel` failed because no shared library existed.
- Windows: workflow test command had leading indentation inside `python -c`, causing `IndentationError`.

Verification:
- YAML parse: passed.
- Portable Python 3.14 wheel build: passed, native `.so` included.
- Fresh wheel smoke install: passed with `native_c: True`.
- OpenMP branch syntax parse: passed.
- `twine check`: passed with existing long-description warnings.
- `tests/test_native_backend.py --no-cov`: `12 passed`.

Author: aeiwz

.github/workflows/python-publish.yml

@@ -106,7 +106,7 @@ jobs:
 
   # ---------------------------------------------------------------------------
   # 3. Binary wheels via cibuildwheel
-  #    Linux:   manylinux_2_28 x86_64 + aarch64 (QEMU), musllinux x86_64
+  #    Linux:   manylinux_2_28 x86_64 + aarch64 (QEMU)
   #    macOS:   x86_64 + arm64
   #    Windows: AMD64
   # ---------------------------------------------------------------------------
@@ -133,28 +133,19 @@ jobs:
           python-version: "3.12"
 
       - name: Install cibuildwheel
-        run: pip install cibuildwheel==2.21.3
+        run: pip install "cibuildwheel>=3.0,<4"
 
       - name: Build wheels
         env:
-          CIBW_BUILD: "cp310-* cp311-* cp312-* cp313-*"
-          CIBW_SKIP: "*-win32 *-manylinux_i686 pp*"
+          CIBW_BUILD: "cp310-* cp311-* cp312-* cp313-* cp314-*"
+          CIBW_SKIP: "*-win32 *-manylinux_i686 *-musllinux_* pp*"
           CIBW_MANYLINUX_X86_64_IMAGE: manylinux_2_28
           CIBW_MANYLINUX_AARCH64_IMAGE: manylinux_2_28
           CIBW_ARCHS_LINUX: "x86_64 aarch64"
           CIBW_ARCHS_MACOS: "x86_64 arm64"
           CIBW_ARCHS_WINDOWS: "AMD64"
-          CIBW_ENVIRONMENT_LINUX: "CFLAGS='-fopenmp' LDFLAGS='-lgomp'"
-          CIBW_TEST_COMMAND: >
-            python -c "
-            import metbit;
-            assert metbit.__version__, 'version missing';
-            from metbit._native import pearson_columns, backend_info;
-            import numpy as np;
-            r = pearson_columns(np.eye(5), anchor_index=0);
-            assert abs(r[0] - 1.0) < 1e-12, r[0];
-            print('OK - backend:', backend_info())
-            "
+          CIBW_ENVIRONMENT: "METBIT_PORTABLE_BUILD=1"
+          CIBW_TEST_COMMAND: 'python -c "import metbit; assert metbit.__version__, \"version missing\"; from metbit._native import pearson_columns, backend_info; import numpy as np; info = backend_info(); assert info[\"native_c\"], info; r = pearson_columns(np.eye(5), anchor_index=0); assert abs(r[0] - 1.0) < 1e-12, r[0]; print(\"OK - backend:\", info)"'
           CIBW_TEST_REQUIRES: "numpy scipy scikit-learn"
           CIBW_BUILD_VERBOSITY: 1
         run: cibuildwheel --output-dir wheelhouse
@@ -175,10 +166,16 @@ jobs:
     steps:
       - uses: actions/checkout@v4
 
-      - name: Download all built artifacts
+      - name: Download source distribution
         uses: actions/download-artifact@v4
         with:
-          pattern: "sdist wheels-*"
+          name: sdist
+          path: dist/
+
+      - name: Download wheels
+        uses: actions/download-artifact@v4
+        with:
+          pattern: "wheels-*"
           path: dist/
           merge-multiple: true
 
@@ -210,10 +207,16 @@ jobs:
       url: https://pypi.org/project/metbit/${{ env.RELEASE_TAG }}
 
     steps:
-      - name: Download all built artifacts
+      - name: Download source distribution
         uses: actions/download-artifact@v4
         with:
-          pattern: "sdist wheels-*"
+          name: sdist
+          path: dist/
+
+      - name: Download wheels
+        uses: actions/download-artifact@v4
+        with:
+          pattern: "wheels-*"
           path: dist/
           merge-multiple: true
 
@@ -250,10 +253,16 @@ jobs:
         env:
           RELEASE_TAG: ${{ env.RELEASE_TAG }}
 
-      - name: Download all built artifacts
+      - name: Download source distribution
+        uses: actions/download-artifact@v4
+        with:
+          name: sdist
+          path: release-assets/
+
+      - name: Download wheels
         uses: actions/download-artifact@v4
         with:
-          pattern: "sdist wheels-*"
+          pattern: "wheels-*"
           path: release-assets/
           merge-multiple: true
 

metbit/_native_backend.c

@@ -341,6 +341,7 @@ pearson_columns_f32(PyObject *self, PyObject *args)
 
     const float *data = (const float *)buf.buf;
     double anchor_sq = 0.0;
+    int oom_f32 = 0;
 
     Py_BEGIN_ALLOW_THREADS
 
@@ -355,7 +356,6 @@ pearson_columns_f32(PyObject *self, PyObject *args)
 
     /* Pass 2: covariance and sum-of-squares (parallel when OpenMP available) */
 #ifdef _OPENMP
-    int oom_f32 = 0;
     #pragma omp parallel reduction(+:anchor_sq)
     {
         double *lc  = (double *)calloc((size_t)columns, sizeof(double));
@@ -386,12 +386,6 @@ pearson_columns_f32(PyObject *self, PyObject *args)
             free(lc); free(ls);
         }
     }
-    if (oom_f32) {
-        Py_END_ALLOW_THREADS
-        PyMem_Free(means); PyMem_Free(cov); PyMem_Free(col_sq);
-        PyBuffer_Release(&buf); Py_DECREF(output);
-        return PyErr_NoMemory();
-    }
 #else
     for (Py_ssize_t r = 0; r < rows; ++r) {
         const float *row = data + r * columns;
@@ -405,17 +399,23 @@ pearson_columns_f32(PyObject *self, PyObject *args)
     }
 #endif
 
-    for (Py_ssize_t c = 0; c < columns; ++c) {
-        double denom = sqrt(anchor_sq * col_sq[c]);
-        if (denom == 0.0) { corr[c] = 0.0 / 0.0; continue; }
-        double r = cov[c] / denom;
-        corr[c] = (r > 1.0) ? 1.0 : (r < -1.0) ? -1.0 : r;
+    if (!oom_f32) {
+        for (Py_ssize_t c = 0; c < columns; ++c) {
+            double denom = sqrt(anchor_sq * col_sq[c]);
+            if (denom == 0.0) { corr[c] = 0.0 / 0.0; continue; }
+            double r = cov[c] / denom;
+            corr[c] = (r > 1.0) ? 1.0 : (r < -1.0) ? -1.0 : r;
+        }
     }
 
     Py_END_ALLOW_THREADS
 
     PyMem_Free(means); PyMem_Free(cov); PyMem_Free(col_sq);
     PyBuffer_Release(&buf);
+    if (oom_f32) {
+        Py_DECREF(output);
+        return PyErr_NoMemory();
+    }
     return output;
 }
 
@@ -461,6 +461,7 @@ column_variances(PyObject *self, PyObject *args)
     memset(var, 0, (size_t)columns * sizeof(double));
 
     const double *data = (const double *)buf.buf;
+    int oom_var = 0;
 
     Py_BEGIN_ALLOW_THREADS
 
@@ -475,7 +476,6 @@ column_variances(PyObject *self, PyObject *args)
 
     /* Pass 2: sum of squared deviations - OpenMP parallel over rows */
 #ifdef _OPENMP
-    int oom_var = 0;
     #pragma omp parallel
     {
         double *lvar = (double *)calloc((size_t)columns, sizeof(double));
@@ -519,6 +519,12 @@ column_variances(PyObject *self, PyObject *args)
 
     PyMem_Free(means);
     PyBuffer_Release(&buf);
+#ifdef _OPENMP
+    if (oom_var) {
+        Py_DECREF(output);
+        return PyErr_NoMemory();
+    }
+#endif
     return output;
 }
 
@@ -565,6 +571,7 @@ column_variances_f32(PyObject *self, PyObject *args)
     memset(var, 0, (size_t)columns * sizeof(double));
 
     const float *data = (const float *)buf.buf;
+    int oom_vf = 0;
 
     Py_BEGIN_ALLOW_THREADS
 
@@ -577,7 +584,6 @@ column_variances_f32(PyObject *self, PyObject *args)
         means[c] /= (double)rows;
 
 #ifdef _OPENMP
-    int oom_vf = 0;
     #pragma omp parallel
     {
         double *lv = (double *)calloc((size_t)columns, sizeof(double));
@@ -621,6 +627,12 @@ column_variances_f32(PyObject *self, PyObject *args)
 
     PyMem_Free(means);
     PyBuffer_Release(&buf);
+#ifdef _OPENMP
+    if (oom_vf) {
+        Py_DECREF(output);
+        return PyErr_NoMemory();
+    }
+#endif
     return output;
 }
 

setup.cfg

@@ -1,2 +1,2 @@
 [metadata]
-description-file = README.rst
+description_file = README.rst

setup.py

@@ -74,6 +74,20 @@ def _detect_openmp():
 _OMP_COMPILE, _OMP_LINK = _detect_openmp()
 
 
+def _native_compile_args():
+    """Return optimization flags suitable for local and portable wheel builds."""
+    if sys.platform == "win32":
+        return ["/O2"]
+
+    args = ["-O3", "-ffast-math"]
+    portable_build = os.environ.get("METBIT_PORTABLE_BUILD", "").lower() in {
+        "1", "true", "yes",
+    }
+    if not portable_build:
+        args.insert(1, "-march=native")
+    return args
+
+
 # ---------------------------------------------------------------------------
 # Custom build_ext: tolerate compiler errors so a missing C toolchain never
 # prevents a pure-Python install.
@@ -99,7 +113,7 @@ def build_extension(self, ext):
 _native_ext = Extension(
     "metbit._native_backend",
     sources=["metbit/_native_backend.c"],
-    extra_compile_args=["-O3", "-march=native", "-ffast-math"] + _OMP_COMPILE,
+    extra_compile_args=_native_compile_args() + _OMP_COMPILE,
     extra_link_args=_OMP_LINK,
     optional=True,
 )
@@ -149,5 +163,6 @@ def build_extension(self, ext):
         "Programming Language :: Python :: 3.11",
         "Programming Language :: Python :: 3.12",
         "Programming Language :: Python :: 3.13",
+        "Programming Language :: Python :: 3.14",
     ],
 )