Agreed. ELO is a crude but simple statistical model, and its primary benefit is that it's easy to understand and reason about.

I'd love for there to be a multidimensional player skill vector (aka "embedding") that encapsulates more information as well as a more nonlinear model to go on top of it.

For example, in many games, a team usually needs a support player, but a single number isn't enough to capture that for the purposes of matchmaking.