so, cycles solves -

If you make one input to one output, N=6 L=10 (4xdebond, 4xreject, +1 for grab last Hg, bond). That gives you 23.

However!
If you take the seventh input, debond a cardinal, use it to replace one cardinal from the sixth, and reject twice to get 2Hg that you use instead of the Hg from the sixth, then the QE from the 6+7 is produced one cycle early - fastest it can be included in a complete output is now L=8 (3xdebond, unify, tribond, tribond).
What of the Hg from 6+7? Well, the second Hg from 6 has latency of 7 (3xdebond, 2xreject, grab, bond), the second Hg from 7 has latency of 7 (2 cycle penalty for coming from 7, debond, 2xreject, grab, bond).
So the 6th output can come out at cycle 21, and is quintessence-bound.

What of the 5th? Currently it'll collide because that's also coming out at cycle 21. Solution: we can grab the first Hg produced from input 6 because it has a spare latency; it comes out the same time as the second from input 5, so we have time to use that and the first 3 Hg from 5 to build it. Will this help?
QE has a latency of 9, so yes. The fifth output comes out one cycle early, the sixth two cycles early, and I'm pretty sure 21 is the limit.

End result is, I think this is latency +21.


(after listening to stream):

OK, so I missed a regrab on the QE that makes it latency 10. If we grab input 6 by cardinal, we fix that by getting -1 on QE - it is now actually latency 8, but the mercury from input 6 is now delayed as well. Input 6 therefore no longer has Hg to spare.

To fix input 5, however, is trickier. If we grab it by the cardinal, we now have L=9 on QE but L=11 on Hg. One of those Hg can be pulled from input 7 - that still has one coming out the same time as the first Hg from input 5 for free (would be second, but extra latency on 6), and for the other we can use input 8 (6-cycle penalty, reject, grab, bond = 9).

It's horrifying, but it is technically possible to do 21.