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

> [ It'd have been better to file a new bug report for this one, FWIW.  ]
>
> I installed the patch below, which seems safe, but is probably
> not optimal.  Matthias?

Wouldn't it be safer to simply disable prettification of the "diff
header" when diff-buffer-type isn't equal to git?

In case of the output of the command diff-buffers, I don't see what
would be a usefull prettification of that header.

For other cases, back in 2018 you wrote that "This has only been tested
with Git's diff output."  (and I extended diff--font-lock-prettify
without testing other outputs as Git's).  Note also that prettification
was already broken with emacs 27.1.  See the attached screenshot where
--- #<buffer *Messages*> has one of its minus sign in the fringe when
diff-font-lock-prettify is t.  (Your patch fixed that!)