Thanks so much for taking the time to review this!

> I'd rather not completely replace the old with a brand new code
> because that makes it hard for me to see what's really changed.

I thought this would be ok, since the existing version is a complete
rewrite of the original (so there's precedent for complete rewrite of
a function in a commit to fix a bug), and if there were tests showing
the behavior to be the same as before, except where this bug is
fixed. (Though I see the tests are currently broken).

Based on your feedback, and some help from #emacs, I made a patch
that is very minimal to the existing code, with better commit
message, and attached it here.

The patch removes the =(unless (bolp) ...)= guarding inserting CRLF.
The RFC says the "boundary delimiter MUST occur at the beginning of a
line, i.e., following a CRLF".  =(bolp)= is not enough to guarantee
the boundary is preceded by CRLF.  It can be true when the point
is after just "\n".

Because CRLF is inserted unconditionally after the =cond=, the code
does not include the boundary's CRLF in each branch of the =cond=.

> when `filedata` is an empty string, this add an additional \r\n
> compared to the current code.  This seems right to me

Me too, and all the other clients I tested.

> I expect the decoding software will skip the \r\n at the of the
> header and then look for \r\n<BOUNDARY>, so it's important to have
> two \r\n

 What you said is true.  In addition, they also accept

"HEADER\r\nfile content\n--BOUNDARY"

as the content "file content", and consider the last "\n" as attached
to the boundary.  That's where the file's final "\n" gets lost if the
file's content was initially "file content\n".

> There remain some questions on this patch:

While fixing this bug, I found a few more problems in addition to the
two that you mention here.  I was not addressing them yet, since I
thought I should fix one bug per patch.

> I suspect we can also simply this code by moving the first (insert
> "--" boundary "\r\n") before the loop, and the second into the loop
> so we can make it insert "\r\n--" boundary "\r\n" (and thus remove
> \r\n from the end of each of the preceding branches).

Almost, but not quite, or at least not without some awkward (to my
eye) repositioning of inserting boundaries, "--", and "\r\n".  The
final boundary complicates things.  It is different from all the
others, it is "--BOUNDARY--" instead of "--BOUNDARY"

Here's what I ended up with when I tried that,

https://git.sr.ht/~ozzloy/emacs-bug-63941/tree/simplify-insert-boundaries-and---/item/mm-url.el#L397

This passes the tests in =emacs/tests/lisp/gnus/mm-url-tests.el=.