Instead of relying on the very brittle "X-track-desc-*" keys to figure
out where things on XR are, we now have some reference data to take
care of this for us.
Hopefully, this should make this generated model considerably more
accurate (though note that it does miss out basically any abnormal
move).