Comment # 4 on bug 965542 from Neil Brown

Is the filesystem on the server ext3 or ext4?

If you use
  tune2fs -O ^dir_index /dev/sdWHATEVER

to turn off dir_index on that filesystem, does the problem go away?

If "yes" to both of those, then you are hitting a design bug in the ext3/4
directory indexing.

Is the server running a 3.3 or older kernel?  If so then an upgrade
will turn the bug from a 32bit-collision possibility into a 64bit collision
possibility, which makes it much less likely to hit (but doesn't really solve
it).

If the above doesn't seem to explain it, then I would need kernel version on
server, filesystem details on server, and a tcpdump trace (-s 0) for the "ls
-l" attempt.

The NFSv3/NFSv4 difference is probably because they fit different numbers of
entries into a reply.  The problem particularly occurs if the last name in one
reply and the first name in the next reply hash to the same value (or something
like that).