Yep will look into this at the weekend. Reminds of the n things I need to do also, I did external (in beam) work on canonical labelling myself. It would be good to integrate this all under one umbrella at some point.

• saucy-like partition sub-linear partition refinement, with this you can do Titin (~350k atoms) in ~100 ms (see slide 17 https://www.slideshare.net/NextMoveSoftware/cinf-1-generating-canonical-identifiers-for-glycoproteins-and-other-chemically-modified-biopolymers).
• McKay's Lemma 2.25 is really useful, asked him for the proof that I still need to understand
• stereochemistry refinement/leaf comparison handling
• randomized permutation group (since I wanted canon no need to be complete)

Would it be possible to make the public classes package-private and encapsulate as much as possible behind a stable public API? Code looks fine but I don't like the API breakages, we can minimise this in future if the classes are package-private.

Ok, given it a go. Now split into public interfaces, and a factory for creating refiners:

AtomContainerDiscretePartitionRefiner refiner = PartitionRefinement.forAtoms().create();

which is not exactly wonderful, but I couldn't think of a better mechanism.

All good thanks for patches - please try to keep public method signatures intact in future but otherwise okay.

I'll remove the print to standard out (AtomContainerPrinter) from the Azulene test.