Yes, we should be able to install CNV on a ROSA HCP cluster whose worker nodes are metal instances, and deploy kubevirt VirtualMachines on them.
One thing to note, however, is in that topology some control plane configurations might be unavailable, e.g. scheduler, oauth, ocp feature gates, olmCatalogPlacement, etc.
These configurations are controlled via the HostedCluster resource, which is located on a managed cluster the customers don't have access to.
This might be an issue if we want to use a custom catalog source to install CNV from. The default setting is that the catalog sources will be located on the management cluster, not the guest cluster.
Unless ROSA+HCP will expose these configurations somehow that the customers can tweak.

Also, I'm not sure if the storage solution on the mgmt cluster will be configurable, or it will just use gp3-csi. In that case, live migration won't be possible.