>> I'm more and more thinking that we should really bite the bullet and go 
>> for a "really buffer-local" setup, such as in the patch below.
>> 
>> I'd be surprised if it doesn't introduce backward incompatibilities, 
>> but at least it's "The Right Thing" to do.
>
> Well, that's what my patch does, too, except that it delays the backward 
> incompatilibites during one Emacs major release...  (I did not include 
> minibuffer-completion-predicate and minibuffer-completion-confirm 
> because they haven't been discussed yet, but they are easy to add.)
>

And (to show that they are indeed easy to add) here is the updated patch, 
which also takes care of minibuffer-completion-predicate and 
minibuffer-completion-confirm.