There
is no need to have Bitcoin transport all using a single protocol, and
we can get better robustness and feature velocity if there are a
couple protocols in use (you could just run a block-transport-protocol
daemon that connects to your local node via the classic protocol).

Although this is a somewhat appealing notion, would it really improve feature velocity? I don't think the current p2p protocol is holding anything back, and having to implement features twice in two protocols would slow things down quite a bit.

Probably the lowest hanging fruit now is fixing the 100msec sleep and just generally having tools to measure latency and queuing inside the code.