Trying to fix CI issues

- gcc 13.3.0 seems to have an issue with math.h on macOS. It spews a
  bunch of errors about the API_DEPRECATED macro not being used
  correctly, which is not an issue on our side. gcc 13.2.0 does not
  seem to have this issue, but since we cannot select gcc with such
  granularity we downgrade to gcc 12 and hope this fixes the issue for
  now.

- numpy 2.0.0 has come out which causes runtime issues. Fortunately the
  fix is easy, just build with pybind11 >=2.12.0. However this is still
  an issue for the pdfo compatibility test since pdfo does not support
  numpy 2.0.0, and sometimes the tests install 2.0.0 because it happens
  to be in the pip cache. So we upgrade pybind11 (the new version is
  compatible with 2.0 and 1.0 numpy) and skip the pdfo test if we
  happened to pick up numpy 2.

- macOS builds for building Python wheels started to complain when running
  the "delocate" step that libquadmath, libstdc++, and libgfortran have
  minimum targets of 11.0, whereas the default was 10.9, hence we added
  the relevant env variable for CIBW.

- Notes have been made in build_python.yml for failures related to
  Ubuntu 24.04.

- A couple tests were failing on macOS 14, so I modified the rhoend and
  rhobeg values to try to get it to converge. The test was surprisingly
  sensitive to these values.

.github/actions/spelling/allow.txt

@@ -2261,3 +2261,4 @@ TESTQUAD
 WAITALL
 xcodebuild
 ICX
+sw_vers

.github/workflows/build_python.yml

@@ -26,7 +26,11 @@ jobs:
       fail-fast: false
       matrix:
         # As of 20240501, macos-11/12/13 are AMD64, and macOS-14 is ARM64.
-        os: [ubuntu-20.04, ubuntu-22.04, ubuntu-24.04, windows-2019, windows-2022, macos-11, macos-12, macos-13, macos-14]
+        # As of 20240616, ubuntu 24.04 is failing to copy files into the quay.io/pypa/manylinux2014_i686:2024-05-13-0983f6f container
+        # It fails with "tar: ./todo: Cannot utime: Function not implemented" for every file in the repo.
+        # This is strange since it worked for the quay.io/pypa/manylinux2014_x86_64:2024-05-13-0983f6f container.
+        # I suggest trying ubuntu-24.04 again in a month.
+        os: [ubuntu-20.04, ubuntu-22.04, windows-2019, windows-2022, macos-11, macos-12, macos-13, macos-14]
 
     steps:
       - name: Clone Repository (Latest)
@@ -46,11 +50,20 @@ jobs:
       - name: Checkout pybind11 submodule
         run: git submodule update --init python/pybind11
 
+      - name: Set the MACOSX_DEPLOYMENT_TARGET
+        if: ${{ runner.os == 'macOS' }}
+        run: |
+          MACOSX_DEPLOYMENT_TARGET=$(sw_vers -productVersion | cut -d'.' -f 1)
+          echo "MACOSX_DEPLOYMENT_TARGET is $MACOSX_DEPLOYMENT_TARGET"
+          echo "MACOSX_DEPLOYMENT_TARGET=$MACOSX_DEPLOYMENT_TARGET" >> $GITHUB_ENV
+
+
       - name: Set up Fortran
         uses: fortran-lang/setup-fortran@main
         if: ${{ runner.os == 'macOS' }}
         with:
           compiler: gcc
+          version: 12
 
       # Copied from https://github.com/scipy/scipy/blob/main/.github/workflows/wheels.yml
       # For rtools, see https://github.com/r-windows/rtools-installer/releases, which has been

.github/workflows/cmake.yml

@@ -42,7 +42,8 @@ jobs:
         toolchain:
           - {compiler: gcc, version: 11, cflags: '-Wall -Wextra -Wpedantic -Werror', fflags: '-Wall -Wextra -Wpedantic -Werror -fimplicit-none -fcheck=all -fstack-check -Wno-function-elimination'}
           - {compiler: gcc, version: 12, cflags: '-Wall -Wextra -Wpedantic -Werror', fflags: '-Wall -Wextra -Wpedantic -Werror -fimplicit-none -fcheck=all -fstack-check -Wno-function-elimination'}
-          - {compiler: gcc, version: 13, cflags: '-Wall -Wextra -Wpedantic -Werror', fflags: '-Wall -Wextra -Wpedantic -Werror -fimplicit-none -fcheck=all -fstack-check -Wno-function-elimination'}
+          # As of 20240616 gcc13 has issues importing math.h on macOS.
+          # - {compiler: gcc, version: 13, cflags: '-Wall -Wextra -Wpedantic -Werror', fflags: '-Wall -Wextra -Wpedantic -Werror -fimplicit-none -fcheck=all -fstack-check -Wno-function-elimination'}
           - {compiler: intel-classic, version: '2021.8', cflags: '-diag-disable=10441 -Wall -w3 -Werror-all', fflags: '-warn all -debug extended -fimplicit-none -standard-semantics'}
           - {compiler: intel-classic, version: '2021.9', cflags: '-diag-disable=10441 -Wall -w3 -Werror-all', fflags: '-warn all -debug extended -fimplicit-none -standard-semantics'}
           - {compiler: intel-classic, version: '2021.10', cflags: '-diag-disable=10441 -Wall -w3 -Werror-all', fflags: '-warn all -debug extended -fimplicit-none -standard-semantics'}

python/tests/test_combining_constraints.py

@@ -29,7 +29,8 @@ def test_providing_bounds_and_nonlinear_constraints():
     nlc = prima_NLC(lambda x: x[0]**2, lb=[25], ub=[100])
     bounds = prima_Bounds([None, 1], [None, 1])
     x0 = [0, 0]
-    res = prima_minimize(fun, x0, constraints=nlc, bounds=bounds)
+    # We start with a higher rhobeg to help the optimizer find the optimal solution
+    res = prima_minimize(fun, x0, constraints=nlc, bounds=bounds, options={'rhobeg': 1.5})
     assert np.isclose(res.x[0], 5, atol=1e-6, rtol=1e-6)
     assert np.isclose(res.x[1], 1, atol=1e-6, rtol=1e-6)
     assert np.isclose(res.fun, 9, atol=1e-6, rtol=1e-6)
@@ -48,16 +49,17 @@ def newfun(x):
     lc = LC(np.array([1,1,1]), lb=10, ub=15)
     x0 = [0, 0, 0]
     # macOS seems to stop just short of the optimal solution, so we help it along by
-    # taking a larger initial trust region radius and requiring a smaller final radius
+    # taking a larger initial trust region radius and requiring a larger final radius
     # before stopping. The different packages have different names for these options.
+    # This test seems surprisingly sensitive to the initial and final trust region radius.
     if package == 'prima':
-        options = {'rhobeg': 2, 'rhoend': 1e-8}
+        options = {'rhobeg': 2.2, 'rhoend': 1e-5}
         method = None
     elif package == 'pdfo':
-        options = {'radius_init': 2, 'radius_final': 1e-8}
+        options = {'radius_init': 2.2, 'radius_final': 1e-5}
         method = None
     elif package == 'scipy':
-        options = {'rhobeg': 2, 'tol': 1e-8}
+        options = {'rhobeg': 2.2, 'tol': 1e-5}
         # PDFO and PRIMA will select COBYLA but SciPy may select something else, so we tell it to select COBYLA
         method = 'COBYLA'
     else:

python/tests/test_compatibility_pdfo.py

@@ -20,6 +20,9 @@ def test_pdfo():
     scipy = pytest.importorskip("scipy")
     if version.parse(scipy.__version__) < version.parse("1.11.0"):
         pytest.skip("scipy version too old for this test (its version of COBYLA does not accept bounds)")
+    numpy = pytest.importorskip("numpy")
+    if version.parse(numpy.__version__) >= version.parse("2.0.0"):
+        pytest.skip("numpy version too new for this test (pdfo does not yet support numpy v2)")
 
     from pdfo import pdfo
     from scipy.optimize import NonlinearConstraint as NLC, LinearConstraint as LC, Bounds