[adam-amara]

Scales the demo to the whole journal slice, makes the repository usable on its own, and corrects two results that were wrong.

Corpus

84,805 structures, 4.69 M atomic sites, 687,048 intermolecular contacts, 17.0 M edges. Loaded with LOAD JSONL (21 M elements in 55 s) and committed chronologically as 26 commits, one per publication year, for time travel.

Two results were wrong

Periodic-net dimensionality reported ~89% of structures as 0D isolated motifs, including textbook amide catemers. The quotient graph was built from contact edges alone — but a donor and an acceptor are joined in the crystal through the molecules they belong to, an edge that is not a contact. Built that way the graph is bipartite with no path longer than one hop: no cycles, rank 0 everywhere. Including the covalent bonds contracts each molecule to a point, which is the standard formulation. Result is now 25.1 / 28.0 / 22.0 / 24.8 % across 0D–3D.

The R²₂(8) dimer query matched every hydrogen bond against itself and returned 2₁-screw catemers. A centrosymmetric dimer is its own symmetry image, so it is one edge in a quotient graph, not the reciprocal pair the query hunted. The discriminator is whether the generating operation is an involution — its own inverse, so it relates exactly two molecules — precomputed as is_involution on every contact edge.

Graph model

• Atom -[:IN_COMPONENT]-> Component — there was no path from a contact between two atoms to the molecules those atoms belong to, so synthon, coformer and polymorph questions were unanswerable rather than merely unwritten.
• Atom -[:IN_FRAGMENT]-> Fragment, 764,852 edges — records which atoms are in a functional group, not merely which molecules contain one. Without it a motif query means "an oxygen of a molecule that has a COOH somewhere", which also matches that molecule's alcohols, esters and nitro oxygens.
• net_dim / net_dim_weak and is_involution precomputed at ingest.

Corrected motif numbers

HAS_FRAGMENT was over-emitted: the dedup guard scanned only the last 400 edges, but Component nodes are deduplicated corpus-wide, so one component carried up to 23 parallel edges. Any query joining Fragment on both ends multiplied by the product.

	before	after
R²₂(8) dimers	12,820	1,354 — mean H···O 1.79 Å at 171.7°
C(4) catemers	1,686	280
Synthon, by stored edges	47.9 / 52.1	62.2 / 37.8
Synthon, by physical bonds	63.4 / 36.6	75.0 / 25.0

The earlier claim that correcting the homodimer double-count "reverses the winner" was an artefact of that duplication. The homodimer wins on either reading.

Chemistry

Weak C–H···A contacts with their own kind and no inferred variant, so every existing kind='hbond' statistic is unchanged. Inferred donors restricted to N, O and S — fluorine and chlorine cannot donate a proton — with the rest stored as close_contact, excluded from every hydrogen-bond statistic. is_centrosymmetric asked of gemmi per space group rather than summed from a hand-picked list.

Interfaces

crystal.py: facts, 14 named queries, arbitrary Cypher, time travel, CSV/JSON. The studio gains a query library grouped by subject and a Cypher box that paints the canvas from a user's own query.

Tests

Two gates execute the shipped queries against a live server. The front-end gate previously extracted 12 of ~127 and reported success; it now recovers 76 and asserts that required schema elements appear in what it extracted, failing if one does not — a scraper that stops matching can no longer report success over an empty set. Fixing it immediately surfaced two broken queries in the studio (min/max, which this dialect does not implement).

validate_chemistry.py printed RESULT: PASS on an empty corpus; all three validators now fail cleanly with no data, and none could be run as documented before. A duplicate dictionary key in solvents.py silently discarded propanol. CI added.

Robustness

RDKit's bond-order search does not always terminate and is a C++ call no Python watchdog can interrupt, so it runs in a killable child process. Non-finite floats are nulled at serialisation — json.dumps emits a bare NaN, which is not valid JSON, and one occurrence rejects an entire LOAD JSONL.

Documentation

README.md for someone deciding whether to use the graph; TUTORIAL.md for the model, the dialect's real limits, worked questions, the studio, time travel, ingest and query performance. MIT licence. Every figure is re-derived from the live graph by tests/corpus_facts.py.

Known limits are stated rather than buried: π-stacking absent, hydrogen positions not normalised, disorder handled only by an occupancy threshold, InChIKey reaching ~36% of components, no query parameters (so no injection protection — interpolation is the only route), and only count() and avg() among aggregates.