>
> > Try to clear out one of :post-read-conversion and :pre-write-conversion
> to see which of the two is the culprit. (While you're at it, try to clear
> out both first just to make sure we're barking up the right tree.)
>
> Bingo.  I had a look at the function and it’s the `insert’ into the
> temporary buffer that is causing this.
>

Great news!


> ;; Pre-write conversion for `utf-8-hfs'.
> (defun ucs-normalize-hfs-nfd-pre-write-conversion (from to)
>   (let ((old-buf (current-buffer)))
>     (set-buffer (generate-new-buffer " *temp*"))
>     (if (stringp from)
>         (insert from)
>       (insert-buffer-substring old-buf from to))
>     (ucs-normalize-HFS-NFD-region (point-min) (point-max))
>     nil))
>
>
> old-buf here is *code-conversion-work*.
>
> I don’t understand this pre-write-conversion function.  According to the
> documentation of   `define-coding-system’, why would it make a new buffer
> and switch to it?
>
> > VALUE must be a function to call after all functions in
> > ‘write-region-annotate-functions’ and ‘buffer-file-format’ are
> > called, and before the text is encoded by the coding system
> > itself.  This function should convert the whole text in the
> > current buffer.  For backward compatibility, this function is
> > passed two arguments which can be ignored.
>
>
It looks like this function needs to be modernized. (In my december rewrite
I simply started using it, assuming that it was fully functional.) Given
the documentation, it *should* be enough to implement it like this (with a
big reservation -- this is all new to me):

(defun ucs-normalize-hfs-nfd-pre-write-conversion (from to)
  (ucs-normalize-HFS-NFD-region (point-min) (point-max))
  nil)

If I remember correctly, the rewrite tried to correct two problems:

* When deleting a character like "ä" (which in the HFS file system is
decomposed into "a" and "¨") the full character should be deleted. Earlier,
when pressing backspace, "ä" was converted to an "a".

* Completion: If a directory contains files like "åäö.txt" and "aao.txt",
Emacs used to say that "a" was the common start to both sequences.

See the discussion in bug 22169 for more information.

    -- Anders