Yeah, but I mean you just have to have something on the transport layer, you can't just encapsulate application layer into network skipping network, that's not how the network stack works.
As I said you could implement it by having TCP/UDP as is, just with a fixed port number. This wouldn't be unlike the myriad of other conventions that litter IPv6, such as using /64 for a host or ULA's having a certain prefix.
Got it, makes sense. It just seems more like an architectural decision to me than something related to the network stack - that's why I got confused. You can come up with your own convention and use it within a local network. One of the downsides of this approach - it will clutter up the routing table, but that’s probably not a big deal.
I usually prefer to use some kind of demultiplexer, like a reverse proxy, to handle the conversions.