Researchers on the Grid Project at MIT's Parallel and Distributed Operating Systems group have found that throughput can vary wildly. There are more details at http://www.pdos.lcs.mit.edu/grid/pubs.html, but I seem to recall reading that the conventional wisdom with ad-hoc mesh routing is that bandwidth varies with the inverse of the number of hops between nodes.
So ad-hoc routing may not be very useful for networks with an average width of more than about three hops. Cliff Skolnick suggests that mesh routing might be a part of an integrated solution for covering the distance between fixed point, long distance aggregation nodes and the hotspot in your house: the last five hundred feet or so, as it were.
Also, there's got to be better ways to pick out addresses on an ad-hoc mesh network than choosing them ahead of time. A Zeroconf-life solution comes to mind (pick an address deterministically, then ping to see if anyone's already on it, with a definite method for resolving contention) but no doubt the Grid folks already have three better ways of doing it already working in the lab...