martin rudalics <rudalics@gmx.at> writes:

> This is quite a weakness of the present mechanism and I think 
> you got
> it right.  To summarize your approach:
>
> - When we have to replace a buffer in a side window, that 
> window's
>   dedicated status is 'side', and some other buffer is found 
>   that was
>   shown there before (it's on the list of that window's 
>   _previous_
>   buffers) we show that other buffer in the window and make sure 
>   to
>   restore the window's dedicated status to 'side'.
>
> - Otherwise delete the window.  Deleting the window is always 
> possible
>   and we have to make sure one thing - never show in a side 
>   window a
>   buffer that has not been shown before in that window.  This 
>   rule
>   should take care of the DOOM workaround.
>
> And users who override the behavior sketched here by setting the 
> side
> window's dedicated status to t should have the window deleted 
> (since
> that is always possible).
>
> Have I got it right?

Yes.
Maybe it is a step toward implementing the `soft' dedication that 
is mentioned in the comments.
Thanks you for the explanations.

> We have to document it in the Elisp manual.

Here another draft with the info manual changes:

From what I have tested so far, there were more functions that 
needed to preserve the side dedication. I put in my modeline a 
token for the window dedication and it was quite useful to spot 
them (maybe not the clever way but worked). I also grepped into 
window.el to see if I was missing something without more success.

There is a change in the indentation of `quit-restore-window' and 
I don't know if there is convention in such cases.