At first I was thinking "Yay computation!", because computation is fire.

Then I built a machine with quintessence like a fool.

Finally I remembered "Computation is fire", so I deleted that and used triplex.

Four pieces of state here:
1. The template element (if saltlike) - parks next to berlo.  Note that this is not actually the original template!
2. "Is it a saltlike?" - a wand that arm 12 uses to pull saltlikes off the ring.  Note that this is always fire, and starts as the template.
3. The number of quicksilver needed to make the template metal (if metal) - on hexarm 2.
4. "Is it a metal?" - a wand that arm 8 uses to pull lead off the ring.

Pseudoperiod 5 is forced on the metal side because if you have an inconsistent number of elements on a wheel, you cannot wand off another wheel with period 4.  If you do, sometimes the wheel steals your wand.  This is a problem since I wanted to store the metal value in a hexarm.

Period 28 is a nice compromise.  It lets the element side operate at P4, while the metal side can do P5 with a 3-cycle reset (to skip the empty spot in the hexarm).

Really glad to finally have rate in computation; I've always wanted that to be the metric instead of cycles.  I'm glad this puzzle was simple enough at its core to be manageable as rate.

A bit sad that the weighting on rate is not high enough to encourage really crazy fast machines.  This metric is more summy.