> Feb  3 19:52:14 zeus /netbsd: ed0: NIC memory corrupt - invalid
packet length 19969
> Feb  3 19:52:17 zeus /netbsd: ed0: NIC memory corrupt - invalid
packet length 19969

I seem to recall that a long time ago, there was a popular PC ethernet chip*
which had a habit of, if hammered too soon with a packet after finishing the
previous one, would forget that it had to byte-switch data (it was bigendian
internally) and drop silly-looking packets into memory.  19969, byteswapped,
is 484, a reasonable, if not obviously predictable, packet length; could this
be what the problem is?  Reports of cards working OK before might have just
been due to such packets being silently discarded.  It might be interesting
to have the driver inspect the contents of these packets more closely to see
if that is what is going on (i.e. check the ethernet address in the ethernet
header to see if it, too, is byteswapped).

* and the name "wd8003" rings a bell, albeit faintly.

------------------------------------------------------------------------------