Stefan Monnier <monnier@iro.umontreal.ca> writes:

 >> I attach a patch with a first idea (based on Stefan's suggestions),
 >> where we "protect" the parent's markers, inside widget-default-create.
 >> This is so we can try to catch almost all situations where this bug
 >> might arise.
 >
 > I can't judge whether `widget-default-create` covers all the cases, but
 > other than that, it's looking good (well, I find this "set and later 
reset"
 > to be rather ugly/brittle, but that's probably as good as it gets given
 > what we have).

Thanks for taking a look.

 >> +(defun widget--protect-parent-markers (widget)
 >> +  "Protect the :from and :to markers of the WIDGET's parent, if 
necessary.
 >> +
 >> +Usually, the :from marker has type t, while the :to marker has type 
nil.
 >> +When creating a child or a button inside a composite widget, they 
have to be
 >> +changed to nil and t respectively, so that the WIDGET's parent (if 
any),
 >> +properly contains all of its children and buttons."
 >> +  (let* ((parent (widget-get widget :parent))
 >> +         (parent-from-marker (and parent (widget-get parent :from)))
 >> +         (parent-to-marker (and parent (widget-get parent :to)))
 >> +         (pos (point)))
 >> +    (when (and parent-from-marker
 >> +               (eq pos (marker-position parent-from-marker)))
 >> +      (set-marker-insertion-type parent-from-marker nil))
 >> +    (when (and parent-to-marker
 >> +               (eq pos (marker-position parent-to-marker)))
 >> +      (set-marker-insertion-type parent-to-marker t))))
 >> +
 >> +(defun widget--unprotect-parent-markers (widget)
 >> +  "Unprotect the :from and :to markers of the WIDGET's parent, if 
necessary.
 >> +
 >> +Changes the :from marker type back to t and the :to marker type 
back to nil."
 >> +  (let* ((parent (widget-get widget :parent))
 >> +         (parent-from-marker (and parent (widget-get parent :from)))
 >> +         (parent-to-marker (and parent (widget-get parent :to))))
 >> +    (when parent-from-marker
 >> +      (set-marker-insertion-type parent-from-marker t))
 >> +    (when parent-to-marker
 >> +      (set-marker-insertion-type parent-to-marker nil))))
 >
 > I'm not a big fan of the name "protect" here, to be honest.
 > Maybe `widget--prepare-parent-for-insertion`?

Sounds a little bit better, thanks.  I went for something similar.

 >> It doesn't catch _all_, in theory, since it might happen
 >> that a grandparent widget doesn't adjust its markers, but maybe that's
 >> unusual enough that we can get away with it.
 >
 > We can probably handle that case easily by just adding to
 > `widget--protect-parent-markers` a recursive call if either the :from or
 > the :to marker had to be adjusted.  The only difficulty I see is making
 > sure we adjust all the makers in the "unprotect" case that we
 > protected earlier.  Maybe we should make
 > `widget--protect-parent-markers` return a list of (MARKER
 > . PREVIOUS-TYPE) so the `widget--unprotect-parent-markers`
 > would turn into something like (mapcar (apply-partially #'apply
 > #'set-marker-insertion-type) ...).
 > IIUC that list would be empty in the vast majority of cases.

Thanks for this idea.

I attach a patch that implements something like that, along with a new
test to check the recursive case.