Of these the hardest one to deal with is route lookup caching and reuse w/o connect(2). Obviously the UDP connected TCB can cache that, but if you don't want a "connected" socket fd... then there's nowhere else to cache it except ancillary data, so ancillary data it would have to be. But getting return-to-sender ancillary data on every read (so as to be able to copy it to any sends back to the same peer) adds overhead, so that's not good.

A system call to get that ancillary data adds overhead that can be amortized by having the application cache it, so that's probably the right design, and if it could be combined with sending (so a new flavor of sendto(2)) that would be even better, and it all has to be uring-friendly.