CI should finish in roughly 6 minutes compared to 24 minutes in master.

This is an outstanding improvement 🎉

What's the status of this, is it safe to merge? I see "PoC" in the description, but it seems to be working? (Same with cmake)

That's very interesting indeed!
I once thought, couldn't we have a compiler that escapes the container's emulation and runs on the host arch, while being pre-configured for cross-compilation to the target arch?, but put it aside as a wild dream.
But if I understood this correctly, this is actually possible, and shown here?
I'd be really intrigued if that approach could be used more widely across the Python ecosystem, or if there are any drawbacks that make adoption difficult?

What's the status of this, is it safe to merge?

yes I think it is since we run the ninja test suite and, since the PR was first introduced, the manylinux images have moved to using the same trick using clang meaning I'm committed to keep clang up-to-date.

I see the PoC in the description, but it seems to be working?

It's a working proof-of-concept from the Python ecosystem point of view.
I'm almost sure using clang is not an issue at all when building executables and I don't think it would be an issue with Python extensions (although not tested).
If it's acceptable as-is by other maintainers, I'd like to reference this PR (and probably the one in CMake as well) as an example for this trick.
If it proves useful enough then the install-static-clang.sh script will most likely end-up in manylinux itself.

I'd be really intrigued if that approach could be used more widely across the Python ecosystem, or if there are any drawbacks that make adoption difficult?

There are definitely some differences, I mostly encountered some issues with the linker either here, in scikit-build/cmake-python-distributions#632 or manylinux. Those were easy enough to fix in any case.