Expressing the purpose of attachments

The following question was raised in today's B call:

How does a recipient know the purpose served by each attachment in a message? If Alice sends two pictures to Bob, how does Bob's agent know which is the cat picture and which is the dog picture?

The only way I can think of is to make use of appended attachments and reference them via message attributes with semantic meaning. This idea has limited reuse because it cannot be implemented on messages belonging to existing protocols.

Another idea is to submit both objects in an envelope structure inside a single attachment. The problem here is you might end up with double-encoding: you'd base64-encode each picture and put them in the envelope, then base64-encode *that* and set it as the attachment's contents.

cc @TelegramSam @swcurran @dhh1128