Mauro Aranda <maurooaranda@gmail.com> writes:

 > Noam Postavsky <npostavs@users.sourceforge.net> writes:
 >
 >> tag 3348 + confirmed
 >> severity 3348 minor
 >> found 3348 25.0.94
 >> quit
 >>
 >> Confirmed this is still an issue. Here is an elisp script for easier
 >> reproduction (run as emacs -Q -l bug-3348.el):
 >>
 >> (defconst bug-3348-test-dir
 >>   (expand-file-name "bug-3348-testdir" temporary-file-directory))
 >> (make-directory bug-3348-test-dir t)
 >> (cd bug-3348-test-dir)
 >>
 >> (make-directory "dir-a" t)
 >> (make-directory "dir-b" t)
 >>
 >> (with-temp-file "dir-a/file"
 >>   (insert "aaa"))
 >> (with-temp-file "dir-b/file"
 >>   (insert "bbb"))
 >>
 >> (ediff-directories "dir-a" "dir-b" nil)
 >> (switch-to-buffer "*Ediff Session Group Panel*")
 >>
 >> (ediff-next-meta-item 1)
 >> (ediff-mark-for-operation-at-pos nil)
 >> (ediff-collect-custom-diffs)
 >>
 >> (with-current-buffer "*Ediff Multifile Diffs*"
 >>   (write-file "foo.patch"))
 >>
 >> (with-temp-file "dir-b/file"
 >>   (insert "BBB"))
 >>
 >> (ediff-collect-custom-diffs)
 >>
 >> (switch-to-buffer "foo.patch")
 >> (diff-buffer-with-file)
 >> (message (concat "Observe foo.patch buffer is different from its 
file,\n"
 >>                  "but is marked read-only, not modified"))
 >
 > This still happens, but I'm not really sure if this should be called a
 > bug.  ediff-collect-custom-diffs explicitly marks this buffer as not
 > modified with (set-buffer-modified-p nil).  I can't find the reasons for
 > that, but that doesn't mean there isn't a good one.  And it is
 > consistent about doing that.
 >
 > The first call to ediff-collect-custom-diffs creates the buffer
 > *Ediff Multifile Diffs* and also marks it as unmodified, even though it
 > is.  So, why is that not a problem? Why is the second time a problem?
 >
 > Using write-file again is possible, so what are the expectations of the
 > second time of typing `P' (or calling ediff-collect-custom-diffs) and
 > why do they differ from the expectations for the 1st call?

After a year and a half, I think I got it :-)

After the call to write-file, ediff-meta-diff-buffer is visting a file,
and ediff assumes it never visits a file: that's why it erases, refills
it, and marks it as not modified.

So I think the attached patch fixes it.