I've attached a patch below that's implements the `final face
realization filtering' function that bob wanted.  This is just a
first-cut, so I'd appreciate any comments people have.

The basic interface is a variable, `realize-face-filter':

   realize-face-filter's value is nil

   If non-nil, a function called to perturb faces before final realization.
   It is passed a lisp-vector containing all the attributes of the
   fully-specified face, and can change any that it wishes.

For instance the following will completely turn off bold fonts:

   (defun unboldify-lface (lface)
     (aset lface 4 'normal))
   (setq realize-face-filter 'unboldify-lface)
   (clear-face-cache)

It seems to work fine.

If I instrument the filter function a bit, like:

   (defvar ubf-count 0)
   (defun unboldify-lface (lface)
     (setq ubf-count (1+ ubf-count))
     (aset lface 4 'normal))
   (setq realize-face-filter 'unboldify-lface)
   (clear-face-cache)

`ubf-count' seems to continually incremented (if I keep checking in
*scratch* it goes up 2 or 3 each time, but if I display a complicated
new buffer, it can go up by 100).  It doesn't seem noticably slow to me,
but I have a reasonably fast cpu (1GHz).

I guess I'll need to see exactly what it's getting called on, but this
suggests that maybe more places should be checking the face cache
(or that the face-cache checking should be moved into realize_face).
[Gerd, are you listening...?]

Also, perhaps `realize-face-filter' should be a list of functions
instead, e.g., `realize-face-filter-functions', as there might be
circumstances when two entities want to add a filter -- but maybe this
is something that will only ever be changed by the end-user, who can
make his own filter function do everything he wants.

Any comments?

Thanks,

-Miles

Patch: