[EthanThatOneKid]

What

Two changes to applyInference (src/wiki/infer.ts). Each leaves the inferred closure identical.

1. engine.load() gets only the ontology. load() is the engine's preprocessing pass: it closes the rules over the vocabulary it's given, then infer() runs the data through. We passed the whole graph as the vocabulary, so the instance data was reasoned over twice. Now isOntologyTriple selects RDFS/OWL axioms, class and property declarations, and list cells. Annotations, sameAs, and differentFrom stay in the data.
2. The nine rules whose premise reads owl:sameAs are dropped when no sameAs can arise. Eyeling evaluates owl:sameAs as a builtin that scans every fact (evalOwlSameAsBuiltin), and eq-rep-s/p/o call it once per triple, so inference was O(n²) even on graphs that derive nothing. mayDeriveSameAs keeps those rules whenever an asserted triple mentions a term that some rule needs in order to conclude owl:sameAs: sameAs, NamedIndividual (the ruleset declares sameAs reflexive via prp-rp), Nothing (closed under NamedIndividual), functional or inverse-functional properties, hasKey, or max (qualified) cardinality. No rule concludes any of these terms, which I checked by scanning every rule conclusion in the baked ruleset. The rules are found by shape, and the count is checked at load, the same way dt-eq is, so a re-bake that changes them fails loudly.

Parity

The closure matches main triple for triple (the harness is in the session record, not committed):

Corpus	Asserted	Closure	Equality rules	main	This PR
Synthetic OWL 2 RL (restrictions, chains, hasKey, functional props, sameAs on classes, a disjointness inconsistency)	76	350	kept	375 ms	129 ms
Same, with every sameAs trigger removed	59	263	dropped	124 ms	60 ms
tests/fixtures/micro	38	159	dropped	136 ms	49 ms
docs wiki (docs/wiki.yml)	717	832	dropped	1.9 s	183 ms
wazootech/memory	1,074	1,189	dropped	1.7 s	171 ms
bench:shacl fixture, 10³	1,041	1,156	dropped	771 ms	59 ms
bench:shacl fixture, 10⁴	10,040	10,155	dropped	41.2 s	410 ms

A scratch patch of Eyeling's builtin, using its existing predicate index, also made infer() linear (8,005 quads: 6.9 s → 0.56 s) with identical output once the per-run inconsistency-report IRIs are normalized. That confirms the diagnosis. It isn't part of this PR; we don't own Eyeling (see wazootech/reasoner#1).

Benchmark (deno task bench:shacl, #416), full wiki check

Quads	main	This PR	Peak RSS
1,007	2.2 s	0.7 s	203 MB
10,006	45.1 s	1.6 s	441 MB
100,007	timeout (> 20 min)	8.8 s	1,058 MB
1,000,017	not run	92.7 s	3,656 MB

At 10⁶, loadGraph takes 61 s and SHACL validation 28 s. SHACL result counts match the validator-only cases at every size (4 / 24 / 194 / 2,078). Measured on Windows with 1.4–3.5 GB of RAM free.

Not fixed

Graphs that do mention one of those terms still go through Eyeling's quadratic sameAs builtin. The fix for that is our own reasoner (wazootech/reasoner#1).

Checks

deno task check, lint, fmt:check, test (581 passed, 1 ignored), and the docs wiki's fmt --check, lint --strict, check --strict, and render --check all pass. The docs' embedded SPARQL results are unchanged. New: tests/infer_test.ts covers ontology classification, the sameAs trigger terms, sameAs substitution with an asserted sameAs, sameAs derived from a functional property, and axioms applying with the equality rules dropped.

Refs #414

🤖 Generated with Claude Code