Hello!
Now I have understood why the schema with two replicated servers and the 
coda-clients working on the same machines is even worse than with one 
client and one server on one machine. The answer is in the thread  
http://www.coda.cs.cmu.edu/maillists/codalist/codalist-2005/7558.html . 
The diagramm drawn there is somewhat different from the one we get in my 
case (it should be "cross-wired") but the idea is the same.
We starts with the two replicated servers (ViceA and ViceB) running on 
two machines. Than we have two clients (VenusA and VenusB) each one 
running on the same machine as the corresponding server (i.e. VenusA and 
ViceA are on the same host). At the beggining Venus* sees both servers 
but earlier or later it would disconnect from "its own" server. So we 
ends up with VenusA seing only ViceB and vice versa. As it was explained 
in the above-mentioned thread if we have third client than from time to 
time (once at 10 min.) both server will be synced. If during this period 
of time both clients try to modify the same file then at the next sync 
we get a server-server conflict which we have to resolve by hand.

May be the schema with two "cross-wired client/servers" can be used in 
situations than we have two groups of non-replicated volumes  one of 
which is "as a rule" accessed from the one of the server and the  second 
- from the another. For example the volumes are user's home dirs and one 
group of user works only on host with VenusA and access  only volumes 
served by ViceB (and vice versa). But this setup looks to me a little odd.
M.Kondrin