After asking core engineers, the mechanism highlighted above is to allow relays to send messages in priority to high-stake participation nodes, in case of DOS attack. This way, we are sure that high-stake participation nodes receive consensus messages and are still able to reach consensus even if a lot of “fake” nodes are connecting to relays and trying to saturate them.
See go-algorand/localTemplate.go at e6172bc8c415e18b4afe3ae87255e498259fdc2c · algorand/go-algorand · GitHub and go-algorand/localTemplate.go at e6172bc8c415e18b4afe3ae87255e498259fdc2c · algorand/go-algorand · GitHub
This mechanism is not for nodes to select relays. Relays do not have stake. The selection of relays from nodes is a completely different mechanism that rely on performance metrics gathered by the node. See go-algorand/wsNetwork.go at 4e4ee2cd71727b10b9190e551b327c1d9c146961 · algorand/go-algorand · GitHub