This is exactly why you should use property replication for state and not RPC’s. Join-in-progress “just works” so long as you stick to that rule and you don’t have to do anything extra. This also extends to other concepts like network relevancy and dormancy. It’s critical that state is driven through properties.
Actors are the only engine type that support any form of gameplay networking. UObject/Actor Component replication piggybacks off of its owning actor and adds additional overhead to each packet to identify it, so if anything it’d make the problem worse.
The other problem you have with using a single actor is that all entities will be always-relevant and updating at the same frequency - so you can’t benefit from concepts like relevancy or adapative replication frequency for optimisation and/or bandwidth saving.
The engines core networking model is built for and optimised around the actor/component paradigm rather than ECS approaches, so anything that wants to use it has to piggyback off of that model and is usually going to be sub-optimal. This includes mass btw.