Reducing the IEEE layer to the hardware and physical properties is quite practical. All it would take is a little bit of co-ordination. It is essentially STARTTLS at the link layer.
* Assign one MAC address for use as a version number.
* When a pure IP device plugs into a pure IP switch, it begins by announcing itself with the reserved MAC address. If the switch returns the reserved response, the device knows it can upgrade. Otherwise it defaults back to legacy mode.
* A Pure IP session begins with a handshake, cryptographic authentication and negotiation of the link layer encryption. Because you should always encrypt when you can.
* Once the handshake is complete, packets moving from one pureIP device to another across the switch are only limited by the IP maximum size, not the 1500 or 9000 byte limits of yore.
* The switch performs ingress filtering so that a device can only source packets that it has been authorized so source. It also maintains logs of recent activity so that the source of network issues can be identified immediately. Current networks are a nightmare because one faulty process on one machine can saturate a hub.
The advantages of this would be
* Higher throughput due to maxsize frames and no ethernet frame header
* Simplified administration. Packets only route via deterministic IP tables. There are no heuristics, no broadcast packets mucking everything up. The causes of network storms are logged and traceable (by an authenticated, authorized admin).
* Security. Packets are encrypted at the link layer. This defeats traffic analysis attacks on traffic only encrypted at the transport layer or higher.