>> >>   (if (and
>> >>        ;; When `inhibit-message' is non-nil, the intention was to just
>> >>        ;; log the message to the *Messages* buffer using `message'.
>> >>        (null inhibit-message)
>> >>        (window-live-p (active-minibuffer-window))
>> >>        (window-live-p (old-selected-window))
>> >>        (bufferp (window-buffer (old-selected-window)))
>> >>        (minibufferp (window-buffer (old-selected-window))))
>> >
>> > Btw, can you explain why every part of this condition is needed?  IOW,
>> > why isn't just the below enough?
>> >
>> >    (window-live-p (active-minibuffer-window))
>> >
>> > (I do understand the reason for the test of inhibit-message).
>> >
>> > Maybe the other conditions need a comment to explain them?
>> 
>> Unfortunately, I forgot why they were added, i.e. during testing
>> I added them one by one when noticed that some cases don't work.
>> Now I'll try to reproduce these cases by removing conditions
>> and checking which part doesn't work.

Now I recall why just (window-live-p (active-minibuffer-window))
is not enough.  Because it returns non-nil even when the
current buffer is not the minibuffer, but the minibuffer was
activated earlier.  Test case:

0. emacs -Q
1. M-x   ;; activate the minibuffer
2. C-x o ;; switch back to *scratch*
3. Eval in *scratch* buffer:

   (window-live-p (active-minibuffer-window))
   => t

A message overlay should not be added to the *scratch* buffer, so it's
important to check if old-selected-window is a minibuffer window
(i.e. the current buffer is the minibuffer).

>> But if implementation would be straightforward and if you see no problems
>> then why not.
>
> The implementation looks straightforward to me, since you already
> implemented almost all of it in that ELPA package.  What's left is
> minor details.

Let's iron out the details.  A new patch attached works well
in all cases I tested (dabbrev, icomplete, etc.)  But I'm sure
it could be improved further because I might have made wrong
assumptions on the C side, or something.