I mildly prefer header-set to multipart/foo. I prefer either to
multipart/references without the start parameter, and definitely unprefer
the start parameter.
Header-set has two advantages over multipart/foo. One, the "header" part
is known a priori to be optional, and readers that don't understand it know
that ignoring it is ok (unlike multipart/unknown-foo, which gets treated
like mixed).
Interesting notion. Are you sure that readers are justified in ignoring the
header parts in things they don't recognize?
Two, by making the header part multipart, you can associate multiple sets of
header info without repeating the data part. The drawback is that you don't
know what the aggregate is until you look inside; for my implementation, this
is no problem at all.
As it happens I'd worked out a nice SGML example that used this trick to format
a document in different ways, but I never bothered to post it. The other issue
that arises here is how to select among the headers. If multipart/alternative
semantics are applied you have to read all of them and them pick the last one
you are capable of displaying. Are these the semantics we want?
multipart/references (without start) is a bit more complex, and offers no
advantage to what *I* want to do right *now* (AppleDouble). I can see that
it's more powerful, but without some concrete examples of how that power is
going to be used, I can't say it puts the cream in my coffee.
I'm having exactly the same problem, myself.
The start parameter is trouble for me. It could force me into using temp
files, something I don't need to do for the other three alternatives.
You may find yourself doing this for some sorts of multipart headers as well.
Document format information can be pretty large.
Ned