bug#74410: 31.0.50; completion-at-point doesn't work in custom buffers

[Juri Linkov <juri@linkov.net>]

>> Currently this works only when typing 'M-x completion-at-point RET',
>> not by 'M-C-i' that is bound to 'widget-complete'.  Maybe
>> 'M-C-i' should be rebound to 'completion-at-point' in widgets?
>
> Maybe it would make even more sense to just remove the `widget-complete`
> binding since M-TAB should already be bound appropriately, but the
> problem is that we don't know that the code which inserts the widget
> into a buffer has properly setup the `completion-at-point-functions`
> hook for `completion-at-point` to work.
>
> Maybe we should try and detect when that is not the case and emit
> a warning if so (with the hope to be able to obsolete `widget-complete`
> in some future version).

Or to make the binding conditional with something like this:

diff --git a/lisp/wid-edit.el b/lisp/wid-edit.el
index ba99847f488..0701442447c 100644
--- a/lisp/wid-edit.el
+++ b/lisp/wid-edit.el
@@ -1079,7 +1079,10 @@ widget-global-map
 (defvar widget-field-keymap
   (let ((map (copy-keymap widget-keymap)))
     (define-key map "\C-k" #'widget-kill-line)
-    (define-key map "\M-\t" #'widget-complete)
+    (define-key map "\M-\t" `(menu-item "" 'widget-complete
+                              :filter ,(lambda (cmd)
+                                         (unless completion-at-point-functions
+                                           cmd))))
     (define-key map "\C-m" #'widget-field-activate)
     ;; Since the widget code uses a `field' property to identify fields,
     ;; ordinary beginning-of-line does the right thing.