Peter/Phil: Given all that, it sounds like it would be very nice to
make open-rw/read/close not invoke store.  Having a new close upcall
sounds fine, but one needs to keep versions in sync anyway - not sure
if there is a call to the kernel to tell it that using the new upcall
is permissible.

I was partway through entering a bug in Netscape and hit M-q to fill
the paragraph...  It looks like the bug database has lots of old stuff
and nothing recent, so I wasn't motivated to retype things, barring
assurances that the db is used.

The real problem is with weakly connected or disconnected clients,
where this turns a read-read 'conflict' into a real conflict.  Now, if
broken Windows sw is going to actually write spuriously when no
semantic change was requested, we can't help that, but at least it is
a step in the right direction.