Using both sides of the Van-Berlos wheel for two inputs felt very slow, so I settled on finding a solution using a template instead.

I settled on what I wanted the geometry of the final input swing to be and this locked a lot of the rest of it in place. I started with a 15-latency solve, then struggled to get it down to 13. I think the theoretical maximum with this setup is 11, due to an extra move needed to translate the second tail molecule into place, but I don't think it's possible. I think 12 might be doable, though I couldn't get it to work.