Thanks Slawek for your quick response!

1) Are the data models of the user managed resources abstract (enough) 
from the ML2 used?
So would the composition of a router, a network, some subnets, a few 
security group and a few instances in a project just result in a 
different instantiation of packet handling components,
but be otherwise transparent to the user?

Yes, data models are the same so all networks, routers, subnets will be the same but implemented differently by different backend.
The only significant difference may be network types as OVN works mostly with Geneve tunnel networks and with LB backend You are using VXLAN IIUC your email.

That is reassuring. Yes we currently use VXLAN. But even with the same type of tunneling, I suppose the networks and their IDs will not align to form a proper layer 2 domain,
not even talking about all the other services like DHCP or metadata. See next question about my idea to at least have some gradual switchover.

2) What could be possible migration strategies?

[...] Or project by project by changing the network agents over
to nodes already running OVN?

Even if You will keep vxlan networks with OVN backend (support is kind of limited really) You will not be able to have tunnels established between nodes with different backends so there will be no connectivity between VMs on hosts with different backends.

I was more thinking to move all of a projects resources to network nodes (and hypervisors) which already run OVN. So split the cloud in two classes of machines, one set unchanged running Linuxbridge and the other
in OVN mode. To migrate "a project" all agents of that projects routers and networks will be changed over to agents running on OVN-powered nodes....
So this would be a hard cut-over, but limited to a single project. In alternative to replacing all of the network agents on all nodes and for all projects at the same time.

Wouldn't that work  - in theory - or am I missing something obvious here?