Finding this solve was a **major** slog.  I never seriously thought that a 9[2 6] layout existed, but very much wanted to find 9[3 5]--min latency, min rate, and ideal-plus-one latency on the second product. In the end, I had to work piecewise, and explore through brute force. 

I first narrowed the space. I chose the recipe that multibonds before duplication, reasoning that this would move the clunky Berlo wheel away from the workspace.  I also opted to only consider algorithms that delivered the product by grabbing the quintessence.  I knew that this might give up on some potentially brilliant solves, but it made the challenge seem more manageable to me.

With this strategy, I started by looked for every 9-latency solve with the quint input 1 move from the multibonder.  With three days of work, I found five.  Tuning one of them to min rate got me from my stake-in-the-ground of 9[8] down to 9[4].  A good start, but I assumed many people would find this.  One glyph is too orderly, well-behaved, and sane for these wizards.

With a decent solve in hand, I pursued two options for achieving 9[3 5].  First, I explored whether any of the 9s could run their single dispersion glyph at max speed (3 cycles).  No dice.  Fortunately, I didn't spend long on this.

Next, I explored which of the input/output locations of the 9s could support 10-latency pipes in a different location.  There were quite a few of these; I stopped counting after 12.  But they were not evenly distributed--some input/output locations had notably more options than others.  This distribution suggested to me which input locations would be most promising to explore more deeply. 

Two days of stitching together the likeliest 9s with their most agreeably-shaped 10s ultimately got me a pair of dual-pipe machines capable of 9[3 6]. 
 Both had geometry issues forcing them to wait one cycle on quint grabs, and could not quite hit min rate.  I might have sat happy with this, with my ideal first two numbers and only off by one on the third...but the better of the two machines was just *so close* to not needing the pause!

Ultimately, though, it was not poking at that so-close machine that broke through.  It was when I gave up on the last day and started working with the 9-machine that had initially looked least promising.  Its available workspace was tightest of the candidates, but the geometry *just* worked out to permit a 10L pipe, and it so happened that this pair could also run at min rate.  Success!!!  9[3 5] achieved, with 18 hours to deadline.