This would be handy for a number of reasons:
It would make Tor better able to handle new protocols like VoIP.
It could solve the whole need to socksify applications.
Exit relays would also not need to allocate a lot of file descriptors for all the exit connections.
We're heading in this direction. Some of the hard problems are:
IP packets reveal OS characteristics.
We would still need to do IP-level packet normalization, to stop things like TCP fingerprinting attacks.
Given the diversity and complexity of TCP stacks, along with device fingerprinting attacks, it looks like our best bet is shipping our own user-space TCP stack.
Application-level streams still need scrubbing.
We will still need user-side applications like Torbutton.
So it won't become just a matter of capturing packets and anonymizing them at the IP layer.
Certain protocols will still leak information.
For example, we must rewrite DNS requests so they are delivered to an unlinkable DNS server rather than the DNS server at a user's ISP; thus, we must understand the protocols we are transporting.
DTLS (datagram TLS) basically has no users, and IPsec sure is big.
Once we've picked a transport mechanism, we need to design a new end-to-end Tor protocol for avoiding tagging attacks and other potential anonymity and integrity issues now that we allow drops, resends, et cetera.
Exit policies for arbitrary IP packets mean building a secure Intrusion Detection System (IDS).
Our node operators tell us that exit policies are one of the main reasons they're willing to run Tor.
Adding an IDS to handle exit policies would increase the security complexity of Tor, and would likely not work anyway, as evidenced by the entire field of IDS and counter-IDS papers.
Many potential abuse issues are resolved by the fact that Tor only transports valid TCP streams (as opposed to arbitrary IP including malformed packets and IP floods.)
Exit policies become even more important as we become able to transport IP packets.
We also need to compactly describe exit policies in the Tor directory, so clients can predict which nodes will allow their packets to exit.
Clients also need to predict all the packets they will want to send in a session before picking their exit node!
The Tor-internal name spaces would need to be redesigned.
We support onion service ".onion" addresses by intercepting the addresses when they are passed to the Tor client.
Doing so at the IP level will require a more complex interface between Tor and the local DNS resolver.