No. There would be no NAT box holding IP-port mappings in its internal memory, with the related timeouts, flakiness, port clobbering etc. and no packet re-writing. All routing decisions would be static, based on information in the IP header: the legacy outside routers would just examine the legacy part of the IP address and packet, while the internal IPv4.1 would use the extended bits. So just like any packet routing and without translation.

Critically, this solves the cold start and connectability problem of NAT: if you get a packet addressed to your outside IP, to a port that has no memorized mapping, to what internal IP do you send it to? Lacking a static or UPnP port assignment, it can only be dropped. The extended packet format would provide this information for every packet, the upgraded outside host would tell you what upgraded internal host it wants to talk to.

Not really, since in this proposal you'd still have the end-to-end routability that NAT prevents