Find my comments inline;
Yes, I think the overhead in the FaultMediator is rather low. It already handles a lot of other application protocol specific stuff. The only thing which is not nice is that the way to detect the Hessian message is making assumptions on the transport used (content-type of http transport header as a decision criteria). But there are obviously other alternatives to implement the isHessianMessage() method (e.g. letting the builder write an info about the application protocol used in a defined place within the message context or even something smarter?).
Yes, there are limitations regarding message transformations changing the application protocol, this is true. On the other side this would be a relatively hard job. Either reimplementing the whole protocol or integrating a Hessian library (many library versions are incompatible amongst each other). Once we really do this, the effort to change a few lines in the FaultMediator can be neglected.