bug#46374: Regression: erronous calls to PRED switch major-mode of unrelated modified buffers

[Juri Linkov]

> Hmmm, I'm pretty sure the changes to `save-some-buffers` here
> regarding the predicate indirection caused a regression and/or
> very unexpected behavior that causes bugs in previously working
> code.

Thanks, this is a very interesting case, so Cc'ing Stefan and Tino.

> Consider
>
> `(save-some-buffers t (lambda () (derived-mode-p 'org-mode)))'
>
> called from an org-mode buffer. It runs through the part of
> `save-some-buffers' that is trying to resolve the PRED
> indirection:
>
> ``` emacs-lisp
>  ;; Allow `pred' to be a function that returns a predicate
>  ;; with lexical bindings in its original environment
>  (bug#46374).
>  (let ((pred-fun (and (functionp pred) (funcall pred))))
>    (when (functionp pred-fun)
>      (setq pred pred-fun)))
> ```
> which evaluates the predicate to check if it returns a function --
> which `(derived-mode-p)' does, as it simply returns the major-mode
> symbol on success! In this case, it would be `#'org-mode'. So then
> `org-mode' is called on any unsaved buffers as a PREDICATE,
> switching those buffers to org-mode.

Indeed, the predicate returns a function, just to add more fun:

  (funcall (lambda () (derived-mode-p 'lisp-interaction-mode)))
  => lisp-interaction-mode

This means we need to invent some ad-hoc format to distinguish between
these cases.  For example, to create a lexically-bound predicate
at the beginning, it could be called with e.g.

  (save-some-buffers t '(eval . save-some-buffers-root))

and defcustom will look like:

(defcustom save-some-buffers-default-predicate nil
  :type '(choice (const :tag "Default" nil)
                 (function :tag "Only in subdirs of root"
                           (eval . save-some-buffers-root))
                 (function :tag "Custom function"))

Quite ugly, but I see no more natural way.