A host in the Internet that wants to talk to a host in China would require an update to parse new DNS double-A (realm, address) records to encapsulate the p acket IP-in-IP, outer src= 240.0.0.1 outer dest=240.0.0.2. The router that ser ves the shaft at level 1 attracts 240.0.0.0/8 within realm 1 and routes up the elevator for more specific (host) routes within that prefix. The router that serves the shaft at level 2 attracts 240.0.0.2/32 inside the shaft; upon the s aid packet it would swap the inner and outer destination and the packet would reach the Chinese address with classical routing within realm 2.
Routers serving the shaft need an update, but then, only those do. Obviously t he host in China can only reply if its stack is updated to understand the form at. But all the other hosts and routers in China can be classical IPv4 as we k now them long as their traffic stays in China. To migrate to IPv6 what you can do is map the elevator shaft prefix in, say, 400::/3 (sadly cannot use F00/3 that would map 240 neatly but is already assigned).
The current internet would own 400:1::/32, China would own 400:2::/32, etc... You encode the double-A of the host in the prefix, reserve a well known suffix for IPv4 mapped double-A, and you have an IPv6 address that can be mapped bot h ways statelessly. When migrating to v6, each IPv4 node that owns a public IP v4 address in one realm gets a full IPv6 /64 for free. "
Somehow this sounds a lot like 6to4: packets get routed to special devices in the network and ISPs have little control over this. Not a popular architecture. Or another way to look at it is the resemblance with the ill fated 'Provider-Based Global Unicast Addresses' (RFC 1884, Section 2.4.7). This was not very popular either.