> It looks suspicious to me, that the mail-sources settings within
> g-s-s-m affect the global value. If I set mail-sources to nil in the
> init file, it is overwritten by the value above, so it has no
> buffer-local binding. Hm.

<some investigation about what's going on>

List elements after the server name (=defs) within an select-method
are evaluated (for nnfolder, nnml) in nnoo.el:nnoo-change-server by
these lines:

      (unless (cdr (assoc server (cddr bstate)))
	(while (setq def (pop defs))
	  (unless (assq (car def) bvariables)
	    (nconc bvariables
		   (list (cons (car def) (and (boundp (car def))
					      (symbol-value (car def)))))))
	  (if (equal server "*internal-non-initialized-backend*")
	      (set (car def) (symbol-value (cadr def)))
	    (set (car def) (cadr def)))))

'bvariables' contain a part of 'nnoo-definition-alist' that is related
to current backend. Its change will alter nnoo-definition-alist.

According to this code, if I specify

 (setq gnus-secondary-select-methods
       '((nnfolder ""
           (mail-sources ((file :plugged t) ; get mail from system mailbox
                          (directory :path "~/Mail/incoming-nnfolder"
                                     :suffix ""
                                     :plugged t))))
         (nnml "" 
           (mail-sources ((directory :path "~/Mail/incoming-nnml"
                                     :suffix ""
                                     :plugged t))))))

and mail-sources is nil in the beginning, the first call for backend
'nnfolder' will add 

    (list (cons mail-sources nil))  to bvariables.

When nnoo-change-server is then called for backend 'nnml',
mail-sources has the value that was set by the nnfolder-backend before,
therefore

    (list (cons mail-sources ((file :plugged t)(directory :path...))))

is added to the nnml-part of nnoo-definition-alist:

  ((nnfolder nil ((nnfolder-* . *)
                  (mail-sources)))
   (nnml nil     ((nnml-* . *)
                  (mail-sources (file :plugged t)
                                (directory :path "~/Mail/incoming-nnfolder"
                                           :suffix "" :plugged t))))
   (...))


This is not what has been expected.


<alternative approach>

If I understand it right, the intention of the code in
nnoo-change-server (listed above) is to provide a means to add
variable-value pairs to the list of server variables of a backend.

Would it break the design, if mail-sources was handled as server
variable too, so that -- depending on backend -- "nn*-mail-sources"
could be specified?

nndiary, for instance, is using nndiary-mail-sources instead of
mail-sources.

As an example, the following patch makes nnmail-get-new-mail use
              <backend>-mail-sources 
rather than   mail-sources:

------------------------------------------------------------>8----------
Index: nnmail.el
===================================================================
RCS file: /usr/local/cvsroot/gnus/lisp/nnmail.el,v
retrieving revision 6.74
diff -u -p -r6.74 nnmail.el
--- nnmail.el	17 Apr 2003 19:01:45 -0000	6.74
+++ nnmail.el	3 May 2003 15:30:23 -0000
@@ -1646,7 +1646,8 @@ See the Info node `(gnus)Fancy Mail Spli
 (defun nnmail-get-new-mail (method exit-func temp
 				   &optional group spool-func)
   "Read new incoming mail."
-  (let* ((sources (or mail-sources
+  (let* ((sources (or (nnmail-get-value "%s-mail-sources" method)
+		      mail-sources
 		      (if (listp nnmail-spool-file) nnmail-spool-file
 			(list nnmail-spool-file))))
 	 fetching-sources
----------8<------------------------------------------------------------


It seems to work, if I change mail-sources to nnfolder-mail-sources
resp. nnml-mail-sources in my g-s-s-m definition.

This would also allow the use of a global mail-sources with, for
instance, a different one for backend 'nnfolder'.