> 3. Should we try another VM, another image, another postgres client?

To rule out (or to finger) the postgres client,
you might turn up the debug level of the postmaster
process on the server. Maybe there'll be some
clue in the debug log on why the connection was
dropped. The postgres frontend/backend protocol
is not resilient to loss of message synchronization,
but the newer version of the protocol addresses it.

As for TCP/IP issues, a few years back, running
a Java VM on a virtual Unix over MVS, we had an
issue with reappearing phantom packets. Under high
load and multiple threads, the socket read would
come up with an old buffer of data that had already
been read earlier. It was a JVM problem which was
fixed with an update.

HTH.
--yanni