I like the idea of trying to prove that the DoS code doesn't increase the chance of network fragmentation.
The DoS countermeasures should be careful not to penalize or ban peers for any messages that the client will (or might, in another situation) relay.
For example, double-spent transactions don't trigger the DoS countermeasure code.
That should be sufficient to prevent split-the-network attacks; if an attacker wants to try to split the network, the only way the attacker would be successful is if it could somehow send messages to peers that ARE relayed and trigger disconnections elsewhere in the network.
Looking through the patch:
https://github.com/bitcoin/bitcoin/pull/517/files
... I see a couple of cases where relayed messages are penalized (block times too far off, and hitting the free transaction relay limit). To be safe, I'll remove them.
As for relaying block headers for banned peers: "banned" means "if you try to connect to me I'll simply drop the connection attempt." I feel strongly that is the correct behavior; the motivation for the DoS prevention code is to "whitelist" peer behavior, and try to prevent possible 0-day attacks like "if I send you THIS invalid transactions followed by THAT sequence of weird bytes followed by THIS corrupted block header THEN I trigger an obscure bug in the version of OpenSSL that you're running and compromise your machine...."