The XMPP model of a "separate" http upload service integrated might not be the worst thing. Along the lines of relays, you could probably specify your preferred HTTP upload location, and have a NIP for standaradizing marking an http upload as an attachment, that a client can render or not in their preferred way. That way relay operators can opt-in or -out. Although http upload for xmpp has become not-really-optional so maybe it is just a ratchet approach rather than all at once

why is it a red flag? text only keeps the relays running well