> Good, point so we just turn off undo during the dance and then insert
 > the single undo-element (this is only valid for non-REPLACE, of course).

Hopefully for VISIT as well - we should not record deletions when
handling the VISIT&REPLACE case.  I see only three cases here:

(1) undo-list was t => leave it alone.

(2) nochange => restore undo-list to previous value.

(3) otherwise => reset undo-list to nil.

I also added a nochange arm for the non-decoding case - please have a
look.  BTW nochange makes an incompatible change to `revert-buffer': You
can no longer use `revert-buffer' to get rid of the undo list.  I think
this should be documented somehow.  Finally, we should document that
REPLACE doesn't handle differences in text properties and tell users to
write their own `revert-buffer-insert-file-contents-function' to handle
them.

martin, who still considers the non-VISIT&REPLACE case obscure and
cannot see any "example" in `smerge-resolve'.