nx_chem_morgan.nx

nx_chem_morgan.nx -- C2.3c milestone: Morgan extended-connectivity canonical atom ranking. Implements Weisfeiler-Lehman (1-WL) partition refinement to compute a canonical rank per atom that is invariant under graph isomorphism. Two MolGraphs representing the same molecule produce the same set of canonical ranks (modulo atom-index relabeling), so canonical SMILES emit using these ranks yields byte-identical output for graph-equivalent inputs. Algorithm (after Weininger 1989 + modern WL formulation): 1. Initial invariant per atom: packed (z, degree, charge, isotope, h_count, aromaticity) into i64. 2. Sort by initial invariant, assign DENSE class indices 0..k-1. All atoms with same packed invariant get same class. 3. Refinement step: for each atom, signature = (current_class, sorted_neighbor_classes). Sort by signature, assign new dense class indices. 4. Iterate refinement until class count stable (or MAX_ITER). 5. Final rank = position in sorted (final_class, atom_index) order. EXCEED axis hit (landscape doc): E1 -- bit-reproducible canonical form across versions/platforms. The Morgan algorithm here is content-addressed via versioned initial-invariant packing + deterministic signature sort. RDKit's BackwardsIncompatibleChanges.html documents cross-release canonical drift; this substrate's canonical is pinned to nx_chem_morgan_canonical_rank_v1 semantics (hash-pinned at the function level by source identity). Honest gaps (deferred): - Bond orders not yet used in invariants (so cyclohexane vs cyclohexene atoms could collide on initial invariant; one WL refinement step resolves it via neighbor distinctness) - Stereo not yet folded into canonicalization (C2.3d) - Tie-breaking for true graph automorphisms uses (final_class, atom_idx); not a true automorphism canonicalizer but adequate for graph-equivalent input alternatives