This solve is a testament to both iteration and research.  

*******

My first strategy followed my first instinct: prebuilding air-earth elbows on the left, then passing them through the input to construct products on the right.  I was able to optimize that idea to 119c.  I liked the fact that the machine could prebuild at input speed...but suppression was finicky, and the overall Rate suffered as a result.  First output dropped at cycle 54, and 13R thereafter.

*******

But I almost never stop at my first idea for tourney puzzles. Moreover, my personal hallmark as an alchemist is stealing other people's ideas.  

After some gif-staring and a good night's sleep, it occurred to me that this puzzle shares many similarities with Eyedrops of Irritation, from the 2023 Weeklies.  Height metric, elbow-shaped input, and element manipulation finishing over geometry as a central theme.  So I fired up the VOD to see what had worked well there.

The first top-shelf stealable idea I saw was one implemented by Team42 and Bambi.  By putting a biarm inside the curve of the input, they were able to calcify and duplicate on one side while simultaneously suppressing and replacing the input with a recolored one on the other.  Turns out the idea is just as powerful here!  The machine I built with this strategy had a slower overall Rate than my first effort, at 17, but the first output dropped much earlier (cycle 30), for a modestly-better 115c.  The most telling indicator of its simplicity and efficiency, though, was the arm count--only five, plus Berlo!

*******

But there was one more stolen idea to test.  You see, this simple, low-latency technique did not quite win the day in Eyedrops of Irritation; Bambi only placed 2nd with it.  The winner was Team PsyJonDras, with mR_P0zzel filling out the podium, both teams using different takes on a conveyor belt strategy.  Those sacrificed some latency, but made up for it with excellent Rates.  How might that work in this puzzle?

Fast conveyor belts are still a bit beyond me, but brick surgery is something I've done a few times and understand pretty well.  My first effort was a big improvement off the bat--102c. Iteration brought that down to 97, 95, 93, and finally the solve you see here: a lengthy 51 cycles to the first product, but a blistering Rate 8 after that.  91c total, much better than either the prebuild or low-latency approaches!

*******

Based on my work, I expect the fastest solves to be heavily Rate-centric. I'm delighted to have gotten cycles down to double digits, and feel like this is my best HC effort to date.  Can't wait to see how low it goes!