This seems to mix several (should-be) distinct concerns like CAPI, Kamaji and gpu-operator.

The only things I'd take from this is the kind-based bootstrapper for the management cluster and SPIFFE/TPM integration.