>
> There's a "competing" patch from Gregory which does something similar 
> but in a slightly different way (less clean but potentially a bit more 
> backward compatible), and I think in order to decide which is the best 
> approach we need to have a more concrete information about the risk of 
> breakage and what we might need to do to minimize it.
>

In the interest of fairness in the discussion and comparison, and given 
the length of the discussion in bug#45474, here is the "competing" patch.

The five major differences between the two approaches are (m-c-* stands 
for the minibuffer-completion-* variables, r-f-m for read-from-minibuffer, 
and m-w-s-h for minibuffer-with-setup-hook):

(1) SM: m-c-* variables are passed explicitly as arguments to r-f-m with a 
m-w-s-h, they are stricly buffer-local

GH: m-c-* variables are passed implicitly as arguments to r-f-m, they 
automatically become buffer-local upon entering the minibuffer

(2) SM: it is and it will remain necessary to use a m-w-s-h with 
setq-locals to have buffer-local m-c-* variables in r-f-m

GH: for Emacs 28 it is necessary to let-bind minibuffer-local-completion 
to t around the call to r-f-m (to preserve backward compatibility), for 
Emacs 29 and above minibuffer-local-completion t will be assumed (or, with 
the strong version: the only possible behavior)

(3) SM: when m-c-* variables are strictly buffer-local (i.e., when they 
are passed in a m-w-s-h to r-f-m), they cannot interfere with code outside 
of the minibuffer

GH: because m-c-* variables become buffer-local only upon entering the 
minibuffer, they can possibly interfere with code that is executed around 
the call to r-f-m, between the moment they have been let-bound and the 
moment they become buffer-local

(4) SM: the semantics of r-f-m is not changed

GH: the semantics of r-f-m is changed, in Emacs 28 the m-c-* variables 
become buffer-local when minibuffer-local-completion is t, in Emacs 29 
they become buffer-local unless minibuffer-local-completion is nil (or, 
with the strong version: they become buffer-local)

(5) SM: the patch fixes the behavior of completing-read-default

GH: the patch fixes the behavior of completing-read-default and makes, in 
the medium term, that new behavior the default one for all uses of r-f-m