A few followup items:

During the "hangs", the machine was still usable; only outgoing
packets appear to stack up in the queues and not make it anywhere,
though I have no real way of telling whether incoming packets are lost
this way.  Also, while these hangs were happening, even pinging the
remote end of the ppp connection fails in such a way that the
"transmit" LED on the modem doesn't blink, so it's obvious that the
data isn't making it out of the serial port.

I thought that this was kinda odd since it tended to show up during
telnet connections and not with, say, ftp connections.  It so happens
that the particular telnet I'm using sets the IP TOS field on outgoing
packets to 0x10 (IPTOS_LOWDELAY), and this seems to have something to
do with the problem.  Explicitly setting the TOS field to zero seems
to decrease the probability with which this problem appers, but it
still happens.

Also, I have only noticed this problem since updating from 1.0A to 1.2
and 1.2B.  The rate of input silo overflows has increased by a lot as
a result of this upgrade, too.

Yes, arguably I should get some decent serial ports, but I would like
to see some feedback on why this happens at all, since having a decent
serial port may only lessen the symptoms but not repair the root cause
of the problem.

---Tom

P.S. I shouldn't do kernel hacking at 4 in the morning on Christmas
Day, but, oh well...