>> By the way, I suspect (but do not have a recipe right now) that the 
>> same bug exists with the :width and :slant attributes (namely, that 
>> setting the font for the default face with an "exotic" value leads to 
>> suboptimal/wrong font choices for other faces), and that they should 
>> also be set to nil in font_find_for_lface.
>
> IIUC the way we manage choose fonts is that we ask for a list of fonts 
> matching a particular pattern.  If that list is empty we make the 
> pattern more coarse and try again.  And when the list is not empty we 
> choose the best fit based on a score.
>

Yes, that's roughly speaking what is happening indeed.  Except that if we 
constraint, when we build it, the list of candidates to a specific weight 
and/or slant and/or width, scoring the candidates becomes pointless, 
because all candidates in the list will (by definition) be perfect 
matches.

>
> So in essence, what you're saying is that we should rely more on 
> scoring, and start with a coarser pattern right from the beginning?
>

What I'm saying (now) is that we should rely on scoring, not that we 
should rely _more_ on scoring.

New (again) patch attached.