The color-graph example was a tricky and perhaps unfair benchmark comparison. In this example, due to Michael Ballantyne, Orchid, and Greg Rosenblatt, they used higher-order goals in order to adjust a local part of the search to force a non-standard search strategy.
This gave a dramatic improvement. It is also very much like using a local DFS, and so does not (at least not directly) show the impact of associating conjunctions.
There’s the possibility that the Zzzs are making an important impact--cf our version w/Orchid’s. One more axis for comparison. But I suspect not.
There’s the possibility that the delays in the cdr are a big impact. One more axis for comparison.
Also need to run tests where we shuffle the ordering of edges and nodes and left-right of the edge pairs, to kinda avoid around that.
See also separate repo benchmarking orchid’s results w/various options