> I'd like to elaborate on my comments: the thought was that instead of
> extending the scope of the search for the first random Dired window
> from the selected frame to all frames, would it be better to improve the
> heuristic of finding the window that the user really meant to use
> (remember that the user option name contains the word "DWIM").
>
> The proposed heuristic was to use `get-mru-window' to get
> the most recently used window from all frames, and even
> better way is to traverse all windows ordered by their visiting
> recency on all frames to find the window with Dired mode buffer.

I can't find an existing function that would sort windows by recency,
but fortunately the implementation is straightforward:

  (sort (window-list-1)
        (lambda (a b)
          (> (window-use-time a)
             (window-use-time b))))

using `>' gives the mru order, `<' - lru order.

BTW, while looking at windows walking functions, I noticed
an opportunity for simplification.

Martin, could you please confirm if I'm not mistaken with this patch: