33 -> 20

This solution feels a lot better. As someone who is vastly inexperienced with this metric I found the easiest way to optimize to be reduce arms first, reduce instructions per arm second. The free addition to period length helps a lot and I would not have gotten this solve without it.

Getting the final arm to work in a way that never grabs any other atoms is a bit tricky, and I feel like its function might be able to be combined into another arm for a save of maybe an instruction or two.

Predict 1st gets somewhere around 16, but I'm prepared to be surprised by something like a 10.