We've narrowed down the issue in the 2.5.0 release to be that we are not using auditwheel. To fix that we would need to redesign the build system. The plan that I have now is:
- Remove the custom fallback approach for using the system spglib vs bundled version
- Move the linked CMake project that scikit-build-core uses to the python sub-project
- Have a fall-back method at the build time, where if a system spglib is not detected, source the parent project, build and install it as part of the python build. Use
$ORIGIN for the installed RPATH when in this fallback method.
- Check if we are building for PyPI, in which case, make sure the fall-back approach is taken
This is quite a refactor and I would prefer to do it after #458. In the meantime, there is an alternative workaround suggested in pypa/cibuildwheel#1929 (comment) involving wheel tags. I'll try to see how to adapt it.
We've narrowed down the issue in the 2.5.0 release to be that we are not using
auditwheel. To fix that we would need to redesign the build system. The plan that I have now is:$ORIGINfor the installedRPATHwhen in this fallback method.This is quite a refactor and I would prefer to do it after #458. In the meantime, there is an alternative workaround suggested in pypa/cibuildwheel#1929 (comment) involving
wheel tags. I'll try to see how to adapt it.