This is trickier than I thought.  The primary reason of the bug is that
record_insert combines two adjacent insertions thus messing up my
comparison after format decoding wrt the start of the inserted region.
Unless someone has a better idea, I'd handle this by inserting an undo
boundary for `insert-file-contents'.  The second problem is that
`insert-file-contents' does not consider an insertion a buffer change
when the buffer is empty.  This is obviously TRT in the file-visiting
case, but doesn't seem quite right in the non-visiting case.  Both
issues are involved in the buggy behaviors observed.  Please try another
time with the attached patch (I'm afraid it won't be the last).