Let me share my thinking out loud thoughts...

I think selecting relays depends on the purpose / goal in the client. For some clients a hardcoded list will work as the best solution (when it's a microservice client trying one thing at best with a one-kind specific relay) and for other clients which are doing a lot of different stuff around multiple event kinds they would use a dynamic relay selection (and this is where it gets complicated and you inititally define the biggest relays which are transmitting all the events).

I would argue a relay selection should be context-aware. I'm sure this topic is already often discussed in the NIPs repo? (which I should follow more tbh)