This is part of why-- when I've banned people actively Doing Evil-- I tend to do it in the form of an iptables rule that blocks only the inbound packets. I engineer my firewalls to be stateful, so this leaves existing TCP sessions intact, at least in one direction.
From the perspective of people on the game, the person just becomes quiet and eventually idles off. This tends to de-escalate nicely; everyone says their piece, gets their last word in, then tends to subside as nothing else comes back.
From the perspective of the person banned, people are still actively responding to them, though eventually they seem to not pay as much attention and finally there is a network outage of some sort. This tends to have a moderately de-escalating effect also.
In most cases, a temporary ban is quite sufficient for the person to cool off and resume responsible adult behavior. Not always, though.
I don't think implementing a separate game is the way to go; better to do it with a variation on in-game message routing, much like how BLIND rooms (e.g. for truly quiet quietrooms, freezers, etc) make everyone in one feel like they are alone. I'd recommend calling it a SHUN flag on the player object, or similar.
Ultimately, pseudonymous services of this sort really have two options: Hope for the best, or rely upon some third-party authentication source for tracking accounts. E.g. OpenID, G+, Facebook (ugh), etc. Or allowing people in only after they have been recommended by one existing player-- and hold that player accountable for their guests' behavior. This is essentially a variation on the external authentication concept.
As it is, most of the community is horrified by the idea of supplying even an email address, yet so pathetically clueless that they won't use SSL.