@polycode-projects/the-mechanical-code-talker 1.9.2 → 1.10.0

package/src/syllogise.mjs

@@ -1,105 +1,28 @@
-// syllogise.mjs — tmct's speculative-inference engine (ROADMAP Phase 9 /
-// archive/PLAN_SPECULATIVE_INFERENCE.md, growing toward tier-5 "the Syllogist"
-// per PLAN_INFERENCE_TESTING.md). Offline, $0, deterministic, HARD-bounded idle
-// CPU: forward-chain entailments over the OWL-labelled memory graph so a future
-// query-time MISS becomes a lookup. The MATERIALIZING pass (`syllogise()`,
-// below) is NEVER on the chat hot path — it only ever runs as the explicit
-// `npx tmct syllogise` batch job. The PURE kernels it's built from
-// (`deriveSubClassClosure`/`deriveTypePropagation`) are plain, I/O-free
-// functions, so a caller may also reuse them for a small, bounded, READ-ONLY
-// live check (e.g. chat.mjs's ISA proof chase, PLAN_INFERENCE_TESTING.md
-// INF-A2) without that being "the batch pass on the hot path" — nothing is
-// written to memory unless `syllogise()` itself is called.
+// syllogise.mjs — tmct's speculative-inference engine, growing toward
+// tier-5 "the Syllogist". Offline, deterministic: forward-chains entailments
+// over the OWL-labelled memory graph so a future query-time MISS becomes a
+// lookup. `syllogise()` (the materializing pass) only runs as the explicit
+// `npx tmct syllogise` batch job; the pure kernels below may also be reused
+// for a small, bounded, read-only live check without writing anything.
 //
-// Deliberately narrow (the plan's kill-criterion discipline): FIVE rules —
-//   - scm-sco: rdfs:subClassOf transitivity — (a ⊑ b), (b ⊑ c) ⊨ (a ⊑ c).
-//     The simplest OWL 2 RL rule, exactly what the ACE grammar's pattern 1
-//     emits (`deriveSubClassClosure`).
-//   - cax-sco: rdf:type propagation across a subClassOf chain — (x rdf:type
-//     C), (C ⊑ … ⊑ D) ⊨ (x rdf:type D). The OWL 2 RL class-axiom rule that
-//     completes scm-sco for INSTANCE membership (`deriveTypePropagation`).
-//   - cax-dw: disjointness violation — (x rdf:type C1), (C1 owl:disjointWith
-//     C2) ⊨ x is NOT of type C2, checked over C1's FULL ⊑-ancestor closure
-//     (the ⊑-lift, PLAN_INFERENCE_TESTING.md §1 footnote²) — the first rule on
-//     this ladder to produce a PROVABLE "no" rather than only "yes"/"unproven"
-//     (`deriveDisjointViolations`, PLAN_INFERENCE_TESTING.md §4 stage 3).
-//   - cls-svf1: someValuesFrom application — (x P y), (y rdf:type C2), (R
-//     owl:onProperty P), (R owl:someValuesFrom C2) ⊨ (x rdf:type R) — anyone
-//     who P's something of type C2 is of type R, the restriction CLASS
-//     itself (OWL 2 RL Table 8's cls-svf1, W3C OWL 2 RL profile). Joins over
-//     exactly the triple shape the ACE grammar's pattern 4 ("every N1 that
-//     VERBs a N2 is a N3") already emits for its restriction node `R` (`R
-//     rdf:type owl:Restriction`, `R owl:onProperty P`, `R owl:someValuesFrom
-//     N2` — `src/grammar/ace.mjs`'s `parseRestriction`) — no representational
-//     gap, this rule was simply never implemented. Deliberately scoped to
-//     JUST cls-svf1 (the restriction-membership half): it does NOT chase the
-//     further `owl:intersectionOf`/cls-int1 step pattern 4 ALSO emits (`(N1 ⊓
-//     R) ⊑ N3`) — concluding the ORIGINAL worked example's "chat.mjs is a
-//     suite" needs x typed in BOTH N1 and R, then intersection-membership,
-//     then cax-sco across `⊑ N3`, three more rules deep. That composition is
-//     a documented follow-up (PLAN_INFERENCE_TESTING.md §4 stage 4's
-//     remaining scm-svf row), not attempted here (`deriveSomeValuesFromApplication`).
-//   - scm-svf1: someValuesFrom restriction SUBSUMPTION — two INDEPENDENTLY
-//     declared restrictions over the SAME property, whose filler classes are
-//     ⊑-related, entail the restriction NODES are themselves ⊑-related (W3C
-//     OWL 2 RL Table 9's scm-svf1 — confirmed, against the real downloaded
-//     spec, a DISTINCT rule from scm-svf2, which needs `rdfs:subPropertyOf`
-//     instead; tmct's ACE grammar has no way to teach property subsumption at
-//     all, so scm-svf2 is out of scope for a real reason, not laziness). Joined
-//     the batch pass in a follow-up build (this session): originally deferred
-//     (INFBENCH's two drive points never touched the batch pass, so nothing
-//     was lost by deferring), closed once a positive INFBENCH fixture case
-//     (`c1ScmSvfApply`) existed to measure it against
-//     (`deriveSomeValuesFromSubsumption`).
+// Five rules:
+//   - scm-sco: (a ⊑ b), (b ⊑ c) ⊨ (a ⊑ c) — subClassOf transitivity (`deriveSubClassClosure`).
+//   - cax-sco: (x rdf:type C), (C ⊑ … ⊑ D) ⊨ (x rdf:type D) — type propagates across a ⊑ chain (`deriveTypePropagation`).
+//   - cax-dw: (x rdf:type C1), (C1 owl:disjointWith C2) ⊨ x is NOT of type C2, checked over C1's full ⊑-ancestor closure — the first rule that proves a "no" (`deriveDisjointViolations`).
+//   - cls-svf1: (x P y), (y rdf:type C2), (R owl:onProperty P), (R owl:someValuesFrom C2) ⊨ (x rdf:type R) — someValuesFrom application; stops at restriction membership, doesn't chase the further intersectionOf/cls-int1 step (`deriveSomeValuesFromApplication`).
+//   - scm-svf1: two independently declared someValuesFrom restrictions over the same property, with ⊑-related filler classes, entail the restriction nodes are themselves ⊑-related (`deriveSomeValuesFromSubsumption`).
 //
-// The whole safety story is four guards, all mechanical, shared by all five rules:
-//   - BUDGET — at most `budget` NEW derivations per pass (default 50), and at
-//     most `depth` fixpoint rounds (scm-sco only — cax-sco/cax-dw need no
-//     fixpoint, see their own doc comments). The pass stops at whichever bites
-//     first, deterministically (candidates are sorted before truncation).
-//   - FOCUS — when a focus set is given (the fold hands us the just-touched class
-//     footprint), a derivation is admitted only if it touches focus (a, the pivot
-//     b, or c ∈ focus — "focus-connected, one step out"). No focus ⇒ whole graph
-//     (the explicit `npx tmct syllogise` batch pass).
-//   - SCREENS — tautology (a ⊑ a / x:x is never written) and dedup (a derivation
-//     that already exists, stated OR previously entailed, is skipped via its
-//     content key — the same novelty test appendFact's content-hash id enforces).
+// Four safety guards, shared by all five: BUDGET (max new derivations per
+// pass, and max fixpoint rounds for scm-sco); FOCUS (an optional touched-
+// class footprint scopes a derivation to what's relevant — no focus means
+// whole-graph); SCREENS (tautology and dedup against facts already stated or
+// entailed); and TRUST (every derived fact writes under a first-class
+// `entailed:*` Source, prior 0.3, so it never outranks a stated fact and is
+// retractable by provenance).
 //
-// Every derived fact is written via appendFacts with `entailed:subClassOf`
-// (scm-sco), `entailed:type` (cax-sco), `entailed:disjointWith` (cax-dw),
-// `entailed:someValuesFrom` (cls-svf1), or `entailed:someValuesFromSubsumption`
-// (scm-svf1) provenance (a first-class entailed Source, trust prior 0.3 in
-// memory/trust.mjs) so it is LOW trust, NEVER outranks a stated fact, and is
-// fully RETRACTABLE by provenance when the source graph moves. cax-dw's,
-// cls-svf1's, and scm-svf1's conclusions additionally ride trust.mjs's
-// entailed hook when their OWN premises are resolvable in the pre-pass
-// snapshot — premise-derived (`min(premiseTrusts) × ruleConfidence`), still
-// always strictly below its weakest premise (see syllogise()'s own doc
-// comment) — rather than the bare entailed floor; scm-sco/cax-sco do not
-// (yet) engage that hook — see syllogise()'s doc comment for why that
-// specific extension is non-trivial.
-//
-// TWO MORE capabilities live below (PLAN_INFERENCE_TESTING.md §4 stage 4's
-// remainder, INF-C1), both LIVE-CHASE ONLY — never added to `syllogise()`'s own
-// materializing batch pass (documented as a deliberate design choice at each
-// export's own doc comment, not an oversight — see each one for the specific
-// reason: neither is an enumerable "derive every new fact of this shape"
-// closure the way the five rules above are):
-//   - cardinality monotonicity: a class's OWN declared exactly/min cardinality
-//     restriction (n) proves "at least m" for any QUERIED m ≤ n. Confirmed
-//     OUTSIDE OWL 2 RL's own decidable profile (the spec's own `cls-*` rule
-//     table has no rule comparing exactly/min/max cardinalities to each
-//     other, and the profile's syntactic restriction limits cardinality
-//     expressions to 0 or 1 only) — genuinely LIVE-CHASE ONLY, never a
-//     candidate for the batch pass at all, not merely deferred
-//     (`proveCardinalityAtLeast`).
-//   - cax-maxc0: max-cardinality-0 as encoded negation. Grounded in W3C OWL 2
-//     RL's real `cls-maxc1` rule (an ABox contradiction: asserting a specific
-//     individual has a value for a max-0-restricted property is
-//     inconsistent) via a one-step universal generalization — since ANY
-//     witnessed individual would be a contradiction, no true witness can
-//     exist, so a general "no" is provable without needing one
-//     (`proveMaxCardinalityZeroDenial`).
+// Two further capabilities below are LIVE-CHASE ONLY, never part of the
+// batch pass: cardinality monotonicity (`proveCardinalityAtLeast`) and
+// max-cardinality-0 as encoded negation (`proveMaxCardinalityZeroDenial`).
 
 import { loadMemory, appendFacts, readFactRows, normFactTerm, factIdForTriple, removeFacts } from "./memory/core.mjs";
 
@@ -110,52 +33,30 @@ export const SYLLOGISE_RULE = "subClassOf";
 export const ENTAILED_PROVENANCE = `entailed:${SYLLOGISE_RULE}`;
 
 /** cax-sco: the type-propagation rule, and the provenance tag its conclusions
- *  carry (PLAN_INFERENCE_TESTING.md §4 stage 1). */
+ *  carry. */
 export const TYPE_PREDICATE = "rdf:type";
 export const CAX_SCO_RULE = "type";
 export const ENTAILED_TYPE_PROVENANCE = `entailed:${CAX_SCO_RULE}`;
 
-/** cax-dw: x rdf:type C1, C1 owl:disjointWith C2 |= x is NOT of type C2 - a
- *  provable "no", never a guessed one (PLAN_INFERENCE_TESTING.md S4 stage 3,
- *  INF-B1). The conclusion is materialized on the SAME owl:disjointWith
- *  predicate the ACE grammar's pattern 6 ("no N1 is a N2") already emits
- *  class-to-class - chat.mjs's FACT_PREDICATE_PHRASES already renders it "is
- *  not a" for either shape, so an instance-level disjointWith fact reads
- *  correctly ("redis.mjs is not a queue") with no new phrase table entry
- *  needed. */
+/** cax-dw: x rdf:type C1, C1 owl:disjointWith C2 |= x is NOT of type C2 — a
+ *  provable "no", never a guessed one. */
 export const DISJOINT_PREDICATE = "owl:disjointWith";
 export const CAX_DW_RULE = "disjointWith";
 export const ENTAILED_DISJOINT_PROVENANCE = `entailed:${CAX_DW_RULE}`;
-/** trust.mjs's entailed hook's rule-confidence for cax-dw, deliberately < 1
- *  (see syllogise()'s toWrite mapping): keeps a premise-derived conclusion
- *  STRICTLY below its weakest premise's trust, every time, honouring this
- *  module's "never outranks a stated fact" invariant while still riding well
- *  above the bare entailed prior (memory/trust.mjs SOURCE_PRIOR.entailed). */
+/** Rule-confidence < 1 so a premise-derived conclusion stays strictly below
+ *  its weakest premise's trust, never a stated fact's equal. */
 export const CAX_DW_RULE_CONFIDENCE = 0.95;
 
 /** cls-svf1: x P y, y rdf:type C2, R owl:onProperty P, R owl:someValuesFrom
- *  C2 |= x rdf:type R — anyone who P's something of type C2 is of type R,
- *  the restriction CLASS itself (OWL 2 RL Table 8's cls-svf1, PLAN_INFERENCE
- *  _TESTING.md S1 INF-B2, S4 stage 4). Joins over the SAME `owl:Restriction`/
- *  `owl:onProperty`/`owl:someValuesFrom` triple shape the ACE grammar's
- *  pattern 4 already emits for "every N1 that VERBs a N2 is a N3"
- *  (`src/grammar/ace.mjs`'s `parseRestriction`) - no new representational
- *  work, this rule simply consumes triples the grammar was already writing.
- *  `owl:onProperty`'s object and a taught property edge's PREDICATE are
- *  stored in two different casings today (a stored fact's predicate keeps
- *  its raw vocabulary spelling, e.g. "tmct:imports"; a triple's OBJECT slot -
- *  which is what `owl:onProperty`'s value occupies - is normFactTerm'd down
- *  to "imports" at write time, `memory/core.mjs` `appendFact`) - the pure
- *  kernel below normFactTerm's the raw predicate itself before comparing, so
- *  both sides converge on the same spelling without a special case. */
+ *  C2 |= x rdf:type R (OWL 2 RL Table 8). `owl:onProperty`'s stored value and
+ *  a taught property edge's predicate differ in casing (normFactTerm'd vs.
+ *  raw vocabulary spelling), so the kernel below normalizes both before
+ *  comparing. */
 export const ON_PROPERTY_PREDICATE = "owl:onProperty";
 export const SOME_VALUES_FROM_PREDICATE = "owl:someValuesFrom";
 export const CLS_SVF1_RULE = "someValuesFrom";
 export const ENTAILED_SVF1_PROVENANCE = `entailed:${CLS_SVF1_RULE}`;
-/** trust.mjs's entailed hook's rule-confidence for cls-svf1 - the same
- *  sub-1 discount cax-dw uses and for the identical reason (see
- *  CAX_DW_RULE_CONFIDENCE's own comment): keeps a premise-derived conclusion
- *  STRICTLY below its weakest premise's trust, every time. */
+/** Same sub-1 discount as CAX_DW_RULE_CONFIDENCE, same reason. */
 export const CLS_SVF1_RULE_CONFIDENCE = 0.95;
 
 const SEP = "␟"; // an in-key separator no fact term can contain
@@ -174,27 +75,13 @@ const RESERVED_PREDICATES = new Set([
   "rdfs:subclassof", "rdf:type", "owl:disjointwith", "owl:onproperty", "owl:somevaluesfrom", "owl:intersectionof",
 ]);
 
-/**
- * PURE `min(premiseTrusts) × ruleConfidence` — the entailed hook's own core
- * arithmetic (memory/trust.mjs `computeTrust`), exposed here so a caller that
- * has NO Fact/appendFacts call to delegate to (a LIVE, read-only chat proof
- * chase — chat.mjs's scm-svf1/cardinality-monotonicity/cax-maxc0 call sites)
- * can still compute the SAME premise-derived figure `computeTrust` would land
- * on for a materialized fact (recency sits at ~1.0 for a freshly-derived
- * conclusion, so the two agree). `syllogise()`'s own `appendFacts` mapping
- * does NOT call this — it hands `premiseTrusts`/`ruleConfidence` straight to
- * `appendFacts` → `computeTrust`, which ALSO applies the recency nudge; this
- * helper is for the read-only case that has no Fact to nudge. Returns `null`
- * when no numeric premise trust was supplied (nothing to compute from — never
- * a magic default), else clamped to [0,1]. Pure, no I/O.
- */
+/** PURE `min(premiseTrusts) × ruleConfidence`, clamped to [0,1] — for a live,
+ *  read-only chase with no Fact to hand to appendFacts/computeTrust. Returns
+ *  `null` when no numeric premise trust was supplied. */
 export function entailedTrustFrom(premiseTrusts, ruleConfidence = 1) {
   const nums = (Array.isArray(premiseTrusts) ? premiseTrusts : []).filter((t) => typeof t === "number");
   if (!nums.length) return null;
   const clamped = Math.max(0, Math.min(1, Math.min(...nums) * ruleConfidence));
-  // rounded to 6dp — mirrors memory/trust.mjs's own `round(n, 6)` convention,
-  // so a live-chase figure reads identically to what computeTrust would have
-  // stored had this conclusion been persisted.
   return Number(clamped.toFixed(6));
 }
 
@@ -274,8 +161,8 @@ export function deriveSubClassClosure(edges, { depth = 32, budget = 50, focus =
  * call, cached per class so repeated queries against the same edge set never
  * re-walk. Factored out so every rule that needs "the whole taught ⊑-chain a
  * class sits in" (`deriveTypePropagation`'s cax-sco, `deriveDisjointViolations`'s
- * cax-dw ⊑-lift, PLAN_INFERENCE_TESTING.md §1 footnote²) shares ONE closure
- * walk instead of three near-identical ones. Pure, no I/O.
+ * cax-dw ⊑-lift) shares ONE closure walk instead of three near-identical ones.
+ * Pure, no I/O.
  */
 function buildAncestorCloser(subClassEdges) {
   const succ = new Map(); // class -> Set(direct superclass): the subClassOf relation to close
@@ -302,18 +189,11 @@ function buildAncestorCloser(subClassEdges) {
 
 /**
  * PURE cax-sco: rdf:type propagation across a subClassOf chain — (x rdf:type
- * C), (C ⊑ … ⊑ D) ⊨ (x rdf:type D). `typeEdges` ([[x,C], …]) and
- * `subClassEdges` ([[a,b], …]) are both already-normalized term pairs —
- * exactly `deriveSubClassClosure`'s edge shape, just filtered by predicate.
- * `subClassEdges` is closed HERE (a full ancestor walk per class, memoized —
- * `buildAncestorCloser`, above), so a single call sees the WHOLE taught
- * ⊑-chain a class sits in, not just its immediate superclass — no fixpoint
- * rounds are needed (unlike `deriveSubClassClosure`, which must re-scan as
- * ITS OWN relation grows; `subClassEdges` here is a fixed input, never
- * mutated by this function).
- * Returns ONLY new `{ subject, object, via }` conclusions (`via` = the
- * subject's directly-taught type), bounded by `budget`, focus-filtered,
- * tautology- and dedup-screened, deterministic order. No I/O.
+ * C), (C ⊑ … ⊑ D) ⊨ (x rdf:type D). `subClassEdges` is a fixed input (unlike
+ * `deriveSubClassClosure`'s own growing relation), so one ancestor walk per
+ * class (`buildAncestorCloser`) covers the whole chain — no fixpoint rounds
+ * needed. Returns ONLY new `{ subject, object, via }` conclusions, bounded by
+ * `budget`, focus-filtered, tautology- and dedup-screened, deterministic order.
  */
 export function deriveTypePropagation(typeEdges, subClassEdges, { budget = 50, focus = null } = {}) {
   const ancestorsOf = buildAncestorCloser(subClassEdges);
@@ -351,42 +231,22 @@ export function deriveTypePropagation(typeEdges, subClassEdges, { budget = 50, f
 }
 
 /**
- * PURE cax-dw: x rdf:type C1, C1 owl:disjointWith C2 |= x is NOT of type C2 -
- * a provable "no" (PLAN_INFERENCE_TESTING.md S1 INF-B1, S4 stage 3), never a
- * guessed one: a pair this rule cannot connect through a stated disjointness
- * is simply not returned (never asserted "no" by absence - that "cannot be
- * proven" answer shape is the CALLER's job, chat.mjs, not this kernel's).
- *
- * Includes the S1 footnote2 lift ("B1's hardest cell"): x in mock, mock ⊑
- * fixture, fixture disjointWith test |= x not-in test - disjointness is
- * checked over `c`'s FULL ⊑-ancestor closure (`buildAncestorCloser`, shared
- * with `deriveTypePropagation` - reused, not reimplemented), not merely `x`'s
- * direct stated type. `disjointEdges` ([[c1,c2], …]) is the OWL-symmetric
- * relation as taught ("no cache is a queue" stores one directed [cache,
- * queue] row - disjointness has no preferred direction, so both orderings
- * are treated as the same fact here) and also doubles as this rule's OWN
- * idempotency ledger: a caller that re-feeds a prior pass's own entailed
- * instance-level disjointWith rows back in through `disjointEdges` (exactly
- * how `syllogise()` re-reads `readFactRows` every call) gets them skipped by
- * the same dedup/novelty screen below, with no separate "already derived"
- * bookkeeping needed.
- *
- * Returns ONLY new `{ subject, object, viaType, viaClass }` conclusions -
- * `viaType` is `x`'s directly-taught type, `viaClass` is the specific class
- * in that type's ⊑-closure (itself, for a direct hit, or an ancestor, for the
- * lift) the disjointness was actually asserted against - bounded by `budget`,
- * focus-filtered, tautology- and dedup-screened, deterministic order. No I/O.
+ * PURE cax-dw: x rdf:type C1, C1 owl:disjointWith C2 |= x is NOT of type C2 —
+ * a provable "no", never a guessed one; a pair this rule can't connect is
+ * simply not returned. Includes the ⊑-lift: disjointness is checked over
+ * `c`'s FULL ⊑-ancestor closure, not merely `x`'s direct stated type.
+ * `disjointEdges` is the OWL-symmetric relation as taught (one directed row
+ * covers both orderings). Returns ONLY new `{ subject, object, viaType,
+ * viaClass }` conclusions, bounded by `budget`, focus-filtered, tautology-
+ * and dedup-screened, deterministic order.
  */
 export function deriveDisjointViolations(typeEdges, subClassEdges, disjointEdges, { budget = 50, focus = null } = {}) {
   const ancestorsOf = buildAncestorCloser(subClassEdges);
 
-  // disjointWith is symmetric (OWL semantics: C1 disjointWith C2 |= C2
-  // disjointWith C1), so both directions of every taught/prior-entailed pair
-  // are indexed. `presentPairs` doubles as the dedup/novelty screen (see
-  // above) - it is deliberately NOT split by "class-class taught" vs
-  // "instance-class entailed": the predicate is the same either way, and a
-  // synthetic individual term never collides with a class-noun term in this
-  // domain (CODE_REF individuals always contain one of `. / \ # : @`).
+  // disjointWith is symmetric, so both directions of every pair are indexed.
+  // A synthetic individual term never collides with a class-noun term here
+  // (CODE_REF individuals always contain one of `. / \ # : @`), so class- and
+  // instance-level pairs share one map safely.
   const disjointOf = new Map(); // term -> Set(disjoint partner terms)
   const presentPairs = new Set(); // "a\0b" for every disjointWith row already known (either order)
   for (const [a, b] of disjointEdges || []) {
@@ -411,7 +271,7 @@ export function deriveDisjointViolations(typeEdges, subClassEdges, disjointEdges
     if (seenTypeEdge.has(tk)) continue;
     seenTypeEdge.add(tk);
     // the ⊑-lift: x's own class closure is {c} ∪ ancestorsOf(c) - a direct
-    // hit needs no lift (d === c), the S1 footnote2 case needs one hop or more.
+    // hit needs no lift (d === c), an inherited one needs one hop or more.
     for (const d of [c, ...ancestorsOf(c)]) {
       const partners = disjointOf.get(d);
       if (!partners) continue;
@@ -438,41 +298,15 @@ export function deriveDisjointViolations(typeEdges, subClassEdges, disjointEdges
 
 /**
  * PURE cls-svf1: x P y, y rdf:type C2, R owl:onProperty P, R
- * owl:someValuesFrom C2 |= x rdf:type R (OWL 2 RL Table 8's cls-svf1,
- * PLAN_INFERENCE_TESTING.md S1 INF-B2, S4 stage 4) - see this module's
- * header comment for the deliberate scope line (stops at restriction
- * membership, does not chase the further owl:intersectionOf/cls-int1 step
- * pattern 4 also emits).
- *
- * `propertyEdges` ([[x, predicate, y], …], `predicate` the RAW, un-normalized
- * vocabulary spelling exactly as a stored Fact's predicate reads, e.g.
- * "tmct:imports") is every taught/prior-entailed object-property assertion —
- * `syllogise()` builds this from every stored row whose predicate is NOT one
- * of the other three rules' reserved predicates (RESERVED_PREDICATES, above),
- * so ANY declared object property is a candidate, not just one hard-coded
- * verb. `typeEdges`/`subClassEdges` are the same shape `deriveTypePropagation`/
- * `deriveDisjointViolations` take (already-normalized [x,C] / [a,b] pairs);
- * `y`'s type is lifted through its FULL ⊑-ancestor closure (`buildAncestorCloser`,
- * shared machinery, same ⊑-lift discipline as cax-dw's own footnote2 case) so
+ * owl:someValuesFrom C2 |= x rdf:type R (OWL 2 RL Table 8). `propertyEdges`
+ * is every taught/prior-entailed object-property assertion (raw predicate
+ * spelling); `restrictionEdges` is each restriction node's (property, target)
+ * declaration. `y`'s type is lifted through its full ⊑-ancestor closure, so
  * "y is a mock" still satisfies a restriction declared over "fixture" when
- * mock⊑fixture is taught. `restrictionEdges` ([{ restriction, property,
- * target }, …], `property`/`target` already normFactTerm-normalized — the
- * spelling `owl:onProperty`/`owl:someValuesFrom` rows store their OBJECT in)
- * is every restriction node's (P, C2) declaration, reconstructed by joining
- * a restriction's `owl:onProperty` row with its `owl:someValuesFrom` row on
- * the restriction's own subject — exactly `syllogise()`'s own join, exposed
- * here as a plain parameter so the pure kernel stays I/O-free and unit-
- * testable without a memory store.
- *
- * Returns ONLY new `{ subject, object, viaProperty, viaPropertyKey, viaValue,
- * viaType, viaTarget }` conclusions — `object` is the restriction node R
- * itself (the newly-entailed rdf:type value), `viaProperty` the RAW predicate
- * matched, `viaPropertyKey` its normalized form (R's own `owl:onProperty`
- * value), `viaValue` the property's object `y`, `viaType` the specific class
- * `y` was directly taught as, `viaTarget` the class in that type's
- * ⊑-closure the restriction was actually declared against (itself, for a
- * direct hit, or an ancestor, for the lift) — bounded by `budget`,
- * focus-filtered, tautology- and dedup-screened, deterministic order. No I/O.
+ * mock⊑fixture is taught. Returns ONLY new `{ subject, object, viaProperty,
+ * viaPropertyKey, viaValue, viaType, viaTarget }` conclusions (`object` is
+ * the restriction node R), bounded by `budget`, focus-filtered, tautology-
+ * and dedup-screened, deterministic order.
  */
 export function deriveSomeValuesFromApplication(propertyEdges, typeEdges, subClassEdges, restrictionEdges, { budget = 50, focus = null } = {}) {
   const ancestorsOf = buildAncestorCloser(subClassEdges);
@@ -543,34 +377,20 @@ export function deriveSomeValuesFromApplication(propertyEdges, typeEdges, subCla
 }
 
 // ---- shared cardinality-restriction reconstruction (pattern-5, parseCardinality) ----
-// "every N1 has exactly n N2s" already stores `{N1, rdfs:subClassOf, r}` as a
-// plain row (`r = tmct:exactly-N-n2`) plus the restriction node r's own
-// scaffolding: `r owl:onProperty tmct:has`, `r <kind> "n"` (kind one of
-// owl:cardinality/owl:minCardinality/owl:maxCardinality), `r owl:onClass N2`
-// (`src/grammar/ace.mjs`'s `parseCardinality`, ~lines 234-257) — the SAME
-// per-fact storage discipline `deriveSomeValuesFromApplication`'s
-// `restrictionEdges` already reconstructs for someValuesFrom restrictions,
-// applied to pattern 5's shape instead.
-const HAS_PROPERTY_KEY = "has"; // the fixed synthetic marker property parseCardinality always mints (ace.mjs ~line 252) — never a real taught verb, so it doubles as this reconstruction's own defensive filter (see buildCardinalityRestrictions below): a someValuesFrom restriction's onProperty is always a REAL taught verb, never this literal marker.
+// "every N1 has exactly n N2s" stores `{N1, rdfs:subClassOf, r}` plus the
+// restriction node r's own scaffolding rows (owl:onProperty/cardinality
+// kind/owl:onClass) — reconstructed here the same way
+// deriveSomeValuesFromApplication reconstructs someValuesFrom restrictions.
+const HAS_PROPERTY_KEY = "has"; // synthetic marker parseCardinality always mints, never a real taught verb
 const CARDINALITY_KIND_OF = { "owl:cardinality": "exactly", "owl:mincardinality": "min", "owl:maxcardinality": "max" };
 export const ON_CLASS_PREDICATE = "owl:onClass";
 
 /** Reconstructs pattern-5 cardinality restriction records from raw stored
- *  rows touching a restriction node — same reconstruction discipline
- *  `deriveSomeValuesFromApplication` already uses for someValuesFrom
- *  restrictions, applied to pattern 5's shape instead. `rows` is the same
- *  `[{subject,predicate,object}, …]` shape `readFactRows`/`syllogise()`'s own
- *  `rows` param takes (raw predicate casing, already-normalized subject/
- *  object) — a caller may hand it EVERY stored row (this function ignores
- *  anything that isn't one of the four predicates it cares about) or a
- *  pre-filtered subset. A restriction is only admitted when its OWN
- *  `owl:onProperty` row resolves to `HAS_PROPERTY_KEY` — the defensive belt
- *  that keeps a someValuesFrom restriction's scaffolding (which ALSO uses
- *  `owl:onProperty`, just with a real verb) from ever being mistaken for a
- *  cardinality restriction when both kinds' rows are scanned together (e.g.
- *  chat.mjs's live wiring, which reads the whole taught-fact set at once).
- *  Returns `[{ restriction, kind, n, onClass }, …]`, deterministic order
- *  (sorted by restriction id). Pure, no I/O. */
+ *  rows touching a restriction node. A restriction is only admitted when its
+ *  own `owl:onProperty` row resolves to `HAS_PROPERTY_KEY` — keeps a
+ *  someValuesFrom restriction's scaffolding (which also uses `owl:onProperty`,
+ *  with a real verb) from being mistaken for a cardinality restriction.
+ *  Returns `[{ restriction, kind, n, onClass }, …]`, sorted by restriction id. */
 export function buildCardinalityRestrictions(rows) {
   const onPropertyOf = new Map(); // restriction -> owl:onProperty's object
   const kindOf = new Map();       // restriction -> { kind, n }
@@ -597,43 +417,21 @@ export function buildCardinalityRestrictions(rows) {
   return restrictions;
 }
 
-// ---- scm-svf1: someValuesFrom restriction subsumption (W3C OWL 2 RL Table 9,
-// scm-svf1 — confirmed distinct from scm-svf2, which needs property
-// subsumption tmct can't teach yet, see this file's header comment) ----
+// ---- scm-svf1: someValuesFrom restriction subsumption (W3C OWL 2 RL Table 9) ----
 export const SCM_SVF_RULE = "someValuesFromSubsumption";
 export const ENTAILED_SCM_SVF_PROVENANCE = `entailed:${SCM_SVF_RULE}`;
-/** trust.mjs's entailed hook's rule-confidence for scm-svf1 — the same sub-1
- *  discount cax-dw/cls-svf1 use and for the identical reason (see
- *  CAX_DW_RULE_CONFIDENCE's own comment): keeps a premise-derived conclusion
- *  STRICTLY below its weakest premise's trust, every time. Wired into
- *  `syllogise()`'s own materializing pass (below) — scm-svf1 joined the batch
- *  pass in a follow-up build, see this file's header comment. */
+/** Same sub-1 discount as CAX_DW_RULE_CONFIDENCE, same reason. */
 export const SCM_SVF_RULE_CONFIDENCE = 0.95;
 
 /**
  * PURE scm-svf1: c1 someValuesFrom y1, c1 onProperty p, c2 someValuesFrom y2,
- * c2 onProperty p, y1 ⊑ y2 (lifted through y1's FULL ⊑-ancestor closure, same
- * lift discipline as cax-dw/cls-svf1) |= c1 ⊑ c2 — a schema-level fact about
- * the restriction NODES themselves, TWO independently-declared restrictions
- * being required to compare (unlike cardinality monotonicity/cax-maxc0 below,
- * each sufficient from a SINGLE declared restriction). `restrictionEdges` is
- * the SAME `[{ restriction, property, target }, …]` shape
- * `deriveSomeValuesFromApplication` already takes (`property`/`target`
- * already normFactTerm-normalized); `subClassEdges` is the ordinary
- * `[[a,b], …]` shape every other rule in this file takes, a FIXED input never
- * mutated by this function (so no fixpoint rounds are needed, same reasoning
- * as `deriveTypePropagation`'s own doc comment). Restrictions are grouped by
- * their (normalized) property — only restrictions sharing the SAME property
- * are ever compared, matching the rule's own premise shape (`c1 onProperty p`,
- * `c2 onProperty p`, the SAME p).
- *
- * Deliberately LIVE-CHASE ONLY for this build (see this file's header
- * comment): never added to `syllogise()`'s materializing batch pass.
- *
- * Returns ONLY new `{ subject, object, viaY1, viaY2 }` conclusions (`subject`/
- * `object` are the two restriction node ids, `viaY1`/`viaY2` the specific
- * filler classes whose ⊑-relation licensed it), bounded by `budget`,
- * focus-filtered, tautology- and dedup-screened, deterministic order. No I/O.
+ * c2 onProperty p, y1 ⊑ y2 (lifted through y1's full ⊑-ancestor closure) |=
+ * c1 ⊑ c2 — a schema-level fact about the restriction NODES themselves,
+ * requiring two independently-declared restrictions over the SAME property
+ * to compare. Returns ONLY new `{ subject, object, viaY1, viaY2 }`
+ * conclusions (the two restriction node ids, and the filler classes whose
+ * ⊑-relation licensed it), bounded by `budget`, focus-filtered, tautology-
+ * and dedup-screened, deterministic order.
  */
 export function deriveSomeValuesFromSubsumption(restrictionEdges, subClassEdges, { budget = 50, focus = null } = {}) {
   const ancestorsOf = buildAncestorCloser(subClassEdges);
@@ -680,22 +478,14 @@ export function deriveSomeValuesFromSubsumption(restrictionEdges, subClassEdges,
 }
 
 // ---- shared machinery for cardinality monotonicity / cax-maxc0: both are
-// single-premise-sufficient (no SECOND independently-taught restriction
-// needed, unlike scm-svf1 above) — a class's OWN declared cardinality
-// restriction, walked through its FULL ⊑-ancestor closure (same lift
-// discipline as every other rule in this file). ----
-
-/** Shared bounded proof search for `proveCardinalityAtLeast`/
- *  `proveMaxCardinalityZeroDenial`: walks `subject`'s OWN ⊑-ancestor closure
- *  (itself first, then ancestors — the FULL lift, same discipline as cax-dw/
- *  cls-svf1) looking for a class with a DIRECTLY declared cardinality
- *  restriction satisfying `matches(record)`. `cardinalityRestrictionEdges` is
- *  `buildCardinalityRestrictions`'s own output shape. Returns the first
- *  `{ viaClass, viaRestriction, record }` found (deterministic — the ancestor
- *  walk's order is fixed for a given edge set) or null. `budget` bounds how
- *  many candidate classes (subject + ancestors) are examined, a QUERY-rooted
- *  proof search in the same spirit as `findIsaChain`'s `maxHops`, not a
- *  batch-derivation cap. Pure, no I/O. */
+// single-premise-sufficient — a class's OWN declared cardinality restriction,
+// walked through its FULL ⊑-ancestor closure. ----
+
+/** Shared bounded proof search: walks `subject`'s own ⊑-ancestor closure
+ *  looking for a class with a directly declared cardinality restriction
+ *  satisfying `matches(record)`. Returns the first `{ viaClass,
+ *  viaRestriction, record }` found (deterministic) or null. `budget` bounds
+ *  how many candidate classes are examined. */
 function findOwnCardinalityRestriction(subClassEdges, cardinalityRestrictionEdges, subject, matches, { budget = 20, focus = null } = {}) {
   if (!subject) return null;
   const ancestorsOf = buildAncestorCloser(subClassEdges);
@@ -721,40 +511,23 @@ function findOwnCardinalityRestriction(subClassEdges, cardinalityRestrictionEdge
   return null;
 }
 
-// ---- cardinality monotonicity (confirmed OUTSIDE OWL 2 RL's own decidable
-// profile, see this file's header comment) ----
+// ---- cardinality monotonicity (outside OWL 2 RL's own decidable profile) ----
 export const SCM_CARD_RULE = "cardinalityMonotonicity";
-/** trust.mjs's entailed hook's rule-confidence for cardinality monotonicity —
- *  the same sub-1 discount every other rule on this ladder uses (see
- *  CAX_DW_RULE_CONFIDENCE's own comment). Unlike CAX_DW_RULE_CONFIDENCE/
- *  CLS_SVF1_RULE_CONFIDENCE/SCM_SVF_RULE_CONFIDENCE, this constant has no
- *  `syllogise()`/`appendFacts` call site to feed: `proveCardinalityAtLeast` is
- *  QUERY-rooted (see its own doc comment) and never produces an enumerable
- *  Fact for the entailed hook to score — there is no `mgx:trustScore` for this
- *  rule's answer to carry. Defined here anyway, for the same reason every
- *  other rule-confidence constant is a named export rather than an inline
- *  literal: chat.mjs's LIVE proof chase (the only caller) computes and
- *  attaches a `min(premiseTrusts) × ruleConfidence` figure to its OWN answer
- *  for auditability (`entailedTrustFrom`, below), even though today's answer
- *  plumbing does not yet surface it past that one function's return value. */
+/** Same sub-1 discount as CAX_DW_RULE_CONFIDENCE. No `syllogise()` call site
+ *  (this rule is query-rooted, never an enumerable Fact) — defined anyway so
+ *  chat.mjs's live proof chase can attach an auditable confidence figure
+ *  (`entailedTrustFrom`, below) to its own answer. */
 export const CARDINALITY_RULE_CONFIDENCE = 0.95;
 
 /**
  * PURE: given one class `subject`'s OWN declared cardinality restriction
- * (kind ∈ {exactly,min}, n, onClass — lifted through `subject`'s FULL
- * ⊑-ancestor closure, so an inherited restriction counts too) and a QUERIED
- * (`onClass`, `m`), proves "`subject` has at least `m` `onClass`" whenever
- * `onClass` matches and `n ≥ m`. A bounded, QUERY-rooted proof (there is no
- * fixed enumerable "new fact" to write — `m` is query-specific, a different
- * shape than every derivation-producing rule above) — genuinely LIVE-CHASE
- * ONLY (see this file's header comment: this is outside OWL 2 RL's own
- * profile, and unlike scm-svf1, this one has no enumerable fact shape to ever
- * join the batch pass — not merely deferred).
+ * (kind ∈ {exactly,min}, n, onClass — lifted through its full ⊑-ancestor
+ * closure) and a QUERIED (`onClass`, `m`), proves "`subject` has at least `m`
+ * `onClass`" whenever `onClass` matches and `n ≥ m`. Query-rooted, not a
+ * batch derivation — `m` is query-specific.
  *
  * Returns the witnessing `{ subject, object: onClass, m, n, kind, viaClass,
- * viaRestriction }` or null (`viaClass` is the specific class in `subject`'s
- * ⊑-closure the restriction was actually declared against — itself, for a
- * direct hit, or an ancestor, for the lift). No I/O.
+ * viaRestriction }` or null.
  */
 export function proveCardinalityAtLeast(subClassEdges, cardinalityRestrictionEdges, subject, onClass, m, opts = {}) {
   if (!onClass || !Number.isFinite(m)) return null;
@@ -772,31 +545,19 @@ export function proveCardinalityAtLeast(subClassEdges, cardinalityRestrictionEdg
 // ladder's "produces a provable no" naming convention, same epistemic status
 // as cax-dw) ----
 export const CAX_MAXC0_RULE = "maxCardinalityZero";
-/** trust.mjs's entailed hook's rule-confidence for cax-maxc0 — same sub-1
- *  discount, same reason (CAX_DW_RULE_CONFIDENCE's own comment). Same caveat
- *  as CARDINALITY_RULE_CONFIDENCE just above: `proveMaxCardinalityZeroDenial`
- *  is QUERY-rooted and never produces an enumerable Fact either, so there is
- *  no `mgx:trustScore` for this rule to carry — chat.mjs's LIVE proof chase
- *  computes and attaches the `min(premiseTrusts) × ruleConfidence` figure to
- *  its own answer for auditability (`entailedTrustFrom`, below). */
+/** Same sub-1 discount and query-rooted caveat as CARDINALITY_RULE_CONFIDENCE. */
 export const CAX_MAXC0_RULE_CONFIDENCE = 0.95;
 
 /**
- * PURE: `subject` ⊑ r (lifted through `subject`'s FULL ⊑-ancestor closure), r
- * a maxCardinality-0 restriction (property `has`, onClass `onClass`) |= "no
- * `subject` has a `onClass`" — the universal-generalization bridge from
- * `cls-maxc1`'s per-individual ABox contradiction (asserting a witnessed
- * individual would be inconsistent) to a class-level provable negative: since
- * NO witness can exist without contradiction, the general "no" is sound.
- * Same query-rooted, LIVE-CHASE-ONLY scope as `proveCardinalityAtLeast`
- * (never `syllogise()`'s batch pass). NEVER infers "no" from absence — a
- * subject with no declared max-0 restriction at all simply returns null
- * (matching cax-dw's own discipline, `deriveDisjointViolations`'s doc
- * comment above).
+ * PURE: `subject` ⊑ r (lifted through its full ⊑-ancestor closure), r a
+ * maxCardinality-0 restriction (property `has`, onClass `onClass`) |= "no
+ * `subject` has a `onClass`" — a universal-generalization bridge from
+ * `cls-maxc1`'s per-individual ABox contradiction to a class-level provable
+ * negative: since no witness can exist without contradiction, the general
+ * "no" is sound. Never infers "no" from absence — no declared restriction
+ * simply returns null.
  *
- * Returns `{ subject, object: onClass, viaClass, viaRestriction }` or null
- * (`viaClass` — itself, for a direct hit, or an ancestor, for the lift). No
- * I/O.
+ * Returns `{ subject, object: onClass, viaClass, viaRestriction }` or null.
  */
 export function proveMaxCardinalityZeroDenial(subClassEdges, cardinalityRestrictionEdges, subject, onClass, opts = {}) {
   if (!onClass) return null;
@@ -809,26 +570,17 @@ export function proveMaxCardinalityZeroDenial(subClassEdges, cardinalityRestrict
 }
 
 /**
- * PURE consistency checker (PLAN_INFERENCE_TESTING.md S1 INF-C2, S4 stage 5):
- * detects when a SINGLE subject's own already-asserted types contradict each
- * other — x rdf:type C1, x rdf:type C2, C1 owl:disjointWith C2 (checked over
- * BOTH types' FULL ⊑-ancestor closures, the same lift `deriveDisjointViolations`
- * uses) — a REFUSE-worthy clash, not a "no" to derive and move on from. This is
- * a fundamentally different shape than cax-dw: cax-dw asks "can I derive a NO
- * for an UNASSERTED type"; this asks "do X's OWN taught/entailed types already
- * clash with each other" — every stored belief about a contradictory subject is
- * suspect, not just the one pair being queried, so the caller's job (chat.mjs)
- * is to REFUSE to answer from that subject's memory at all, not to keep
- * answering everything except the one clashing pair.
+ * PURE consistency checker: detects when a SINGLE subject's own already-
+ * asserted types contradict each other — x rdf:type C1, x rdf:type C2, C1
+ * owl:disjointWith C2 (checked over both types' full ⊑-ancestor closures) —
+ * a refuse-worthy clash, not a "no" to derive and move on from: unlike cax-dw
+ * (deriving a no for an unasserted type), every stored belief about a
+ * contradictory subject is suspect, so the caller should refuse to answer
+ * from that subject's memory at all.
  *
  * Returns ONLY the clashes found — `{ subject, classA, classB, viaA, viaB }`
- * (`viaA`/`viaB` are the specific taught types whose ⊑-closures actually
- * licensed the clash — itself, for a direct hit, or an ancestor, for the
- * lift) — bounded by `budget`, focus-filtered, deterministic order, DEDUPED
- * so a subject with N mutually-clashing types reports each unordered pair
- * once. No I/O; nothing is written — same read-only discipline as
- * `deriveDisjointViolations`'s own live chat-side use (chat.mjs's INF-B1
- * cax-dw chase).
+ * — bounded by `budget`, focus-filtered, deduped so a subject with N
+ * mutually-clashing types reports each unordered pair once.
  */
 export function findConsistencyViolations(typeEdges, subClassEdges, disjointEdges, { budget = 50, focus = null } = {}) {
   const ancestorsOf = buildAncestorCloser(subClassEdges);
@@ -891,41 +643,21 @@ export function findConsistencyViolations(typeEdges, subClassEdges, disjointEdge
 }
 
 /**
- * Run one bounded speculative pass over the memory graph under `repoDir`
- * (the repo dir whose .tmct/memory/graph.json appendFact/loadMemory manage).
- * Reads the stored subClassOf, rdf:type, owl:disjointWith AND (for cls-svf1/
- * scm-svf1) owl:onProperty/owl:someValuesFrom + every other object-property
- * fact, forward-chains FIVE rules — scm-sco (⊑-transitivity) then cax-sco
- * (type propagation, seeing THIS pass's own scm-sco conclusions too, so a
- * fresh two-hop taught chain and its type propagation both materialize in one
- * call) then cax-dw (disjointness violations, seeing THIS pass's own scm-sco
- * AND cax-sco conclusions too) then cls-svf1 (someValuesFrom restriction
- * membership, also seeing the enlarged subClassOf set for its own ⊑-lift)
- * then scm-svf1 (restriction-to-restriction subsumption, seeing the SAME
- * enlarged subClassOf set — its own ⊑-lift over the two restrictions' filler
- * classes) — and materializes each NEW conclusion via `appendFacts` with
- * `entailed:subClassOf`/`entailed:type`/`entailed:disjointWith`/
- * `entailed:someValuesFrom`/`entailed:someValuesFromSubsumption` provenance +
- * trust (PLAN_INFERENCE_TESTING.md S4 stage 2's entailed hook:
- * `min(premiseTrusts) x ruleConfidence` when the conclusion's OWN premises
- * are resolvable in the pre-pass snapshot, falling back to the bare entailed
- * prior — memory/trust.mjs's SOURCE_PRIOR floor — when they are not, e.g. a
- * premise itself only exists because THIS SAME pass just derived it a round
- * earlier; still low, still never outranks a stated fact, just less
- * precisely premise-derived for that one case).
+ * Run one bounded speculative pass over the memory graph under `repoDir`.
+ * Forward-chains the five rules in order — scm-sco, cax-sco, cax-dw,
+ * cls-svf1, scm-svf1 — each seeing the prior rules' conclusions from this
+ * same pass, and materializes each new conclusion via `appendFacts` with its
+ * `entailed:*` provenance. Trust rides the entailed hook
+ * (`min(premiseTrusts) x ruleConfidence`) when premises are resolvable in the
+ * pre-pass snapshot, else falls back to the bare entailed prior.
  *
- * opts:
- *   - depth   max fixpoint rounds (scm-sco chain growth), default 32
- *   - budget  max NEW derivations written this pass, SHARED across all five
- *             rules (scm-sco, then cax-sco, then cax-dw, then cls-svf1, then
- *             scm-svf1), default 50
- *   - focus   Set|array of class terms; when given, only derivations touching
- *             focus (subject, pivot, or object ∈ focus) are admitted. Omit for a
- *             whole-graph batch pass.
+ * opts: `depth` (max fixpoint rounds, default 32), `budget` (max new
+ * derivations this pass, shared across all five rules, default 50), `focus`
+ * (Set|array of class terms scoping derivations to what touches it — omit
+ * for a whole-graph pass).
  *
  * Returns { derived: [{ id, subject, object, via, rule }], count, budget,
- * depth, truncated } — `truncated` flags that the budget may have capped the
- * pass. Deterministic, offline, side-effects only in .tmct/memory.
+ * depth, truncated }.
  */
 export async function syllogise(repoDir, { depth = 32, budget = 50, focus = null } = {}) {
   const memory = await loadMemory(repoDir);
@@ -933,13 +665,10 @@ export async function syllogise(repoDir, { depth = 32, budget = 50, focus = null
   const subClassEdges = rows.filter((r) => isSubClassOf(r.predicate)).map((r) => [r.subject, r.object]);
   const typeEdges = rows.filter((r) => isType(r.predicate)).map((r) => [r.subject, r.object]);
   const disjointEdges = rows.filter((r) => isDisjoint(r.predicate)).map((r) => [r.subject, r.object]);
-  // cls-svf1's own join inputs: a restriction's owl:onProperty and
-  // owl:someValuesFrom rows, keyed by the restriction's OWN subject so the
-  // two can be paired without a second graph pass; `propertyEdges` is every
-  // OTHER stored fact (any predicate not one of the three rules' reserved
-  // predicates above) — a taught object-property assertion is a candidate
-  // premise for whichever restriction (if any) was declared over its
-  // predicate, never hard-coded to one verb.
+  // cls-svf1's join inputs: a restriction's onProperty/someValuesFrom rows,
+  // keyed by the restriction's own subject; propertyEdges is every other
+  // stored fact, a candidate premise for whichever restriction was declared
+  // over its predicate.
   const onPropertyOf = new Map();      // restriction -> owl:onProperty's (normalized) object
   const someValuesFromOf = new Map();  // restriction -> owl:someValuesFrom's (normalized) object
   const propertyEdges = [];            // [[x, rawPredicate, y], …]
@@ -956,23 +685,12 @@ export async function syllogise(repoDir, { depth = 32, budget = 50, focus = null
   }
   const normalizedFocus = normalizeFocus(focus);
 
-  // Pre-pass trust snapshot, keyed by exact (subject, predicate, object) —
-  // the entailed hook's premiseTrusts lookup, wired for cax-dw ONLY (see the
-  // dwDerived mapping below). scm-sco/cax-sco deliberately stay on the bare
-  // entailed prior here: a spike wiring premiseTrusts through them the same
-  // way broke three PINNED tests (test/syllogise.test.mjs's own "entailed
-  // trust is low"/"never outranks its stated premise" assertions,
-  // test/memory-fold.test.mjs's "entailed facts carry speculative trust") —
-  // with `ruleConfidence` defaulting to 1, `min(premiseTrusts) × 1` can EQUAL
-  // a stated premise's trust (e.g. two operator-taught 1.0 premises → 1.0),
-  // which ties or outranks the very premise it was derived from, violating
-  // this module's own "NEVER outranks a stated fact" invariant (its own
-  // header comment, and the KILL CRITERION test). Fixing that for scm-sco/
-  // cax-sco needs the same sub-1 `ruleConfidence` discount cax-dw uses below
-  // (or an equivalent design decision) — real, but a separate, deliberate
-  // follow-up, not a trivial addition (PLAN_INFERENCE_TESTING.md §4 stage 2
-  // note left for the next pass). cax-dw is new — no pinned floor-trust tests
-  // exist for it — so it is designed with the sub-1 discount from the start.
+  // Pre-pass trust snapshot for the entailed hook's premiseTrusts lookup,
+  // wired for cax-dw/cls-svf1/scm-svf1 only: with ruleConfidence defaulting
+  // to 1, min(premiseTrusts) x 1 can EQUAL a stated premise's trust (e.g. two
+  // 1.0 premises), tying or outranking the very premise it derived from —
+  // scm-sco/cax-sco stay on the bare entailed prior until they get the same
+  // sub-1 discount cax-dw uses below.
   const trustByTriple = new Map();
   for (const r of rows) trustByTriple.set(`${r.subject}${SEP}${r.predicate}${SEP}${r.object}`, r.trust);
   const premiseTrust = (s, p, o) => trustByTriple.get(`${s}${SEP}${p}${SEP}${o}`);
@@ -986,63 +704,39 @@ export async function syllogise(repoDir, { depth = 32, budget = 50, focus = null
   const caxDerived = remainingBudget > 0
     ? deriveTypePropagation(typeEdges, enlargedSubClassEdges, { budget: remainingBudget, focus: normalizedFocus })
     : [];
-  // cax-dw sees the SAME enlarged subClassOf set (so its own ⊑-lift, S1
-  // footnote2, reaches a chain scm-sco just grew this pass) — it does not
-  // need the enlarged TYPE edge set too: it walks each direct type's full
-  // ⊑-ancestor closure itself (deriveDisjointViolations' own doc comment).
+  // cax-dw sees the SAME enlarged subClassOf set (so its own ⊑-lift reaches a
+  // chain scm-sco just grew this pass) — it doesn't need the enlarged TYPE
+  // edge set too, since it walks each direct type's own ⊑-ancestor closure.
   const remainingBudgetDw = Math.max(0, budget - scmDerived.length - caxDerived.length);
   const dwDerived = remainingBudgetDw > 0
     ? deriveDisjointViolations(typeEdges, enlargedSubClassEdges, disjointEdges, { budget: remainingBudgetDw, focus: normalizedFocus })
     : [];
-  // cls-svf1 sees the SAME enlarged subClassOf set (its own ⊑-lift, mirroring
-  // cax-dw's) — it deliberately does NOT see the enlarged type edge set: a
-  // direct taught type on the property's VALUE is the common case (the B2
-  // worked example), and enlarging risks a same-pass cax-sco conclusion on
-  // `y` being consumed before a human can audit it; a documented scope line,
-  // not an oversight (mirrors cax-dw's own "does not need the enlarged TYPE
-  // edge set" choice, just for a different reason here).
+  // cls-svf1 sees the SAME enlarged subClassOf set (its own ⊑-lift) but NOT
+  // the enlarged type edge set, so a same-pass cax-sco conclusion on `y`
+  // can't be consumed before a human can audit it.
   const remainingBudgetSvf1 = Math.max(0, budget - scmDerived.length - caxDerived.length - dwDerived.length);
   const svf1Derived = remainingBudgetSvf1 > 0 && restrictionEdges.length
     ? deriveSomeValuesFromApplication(propertyEdges, typeEdges, enlargedSubClassEdges, restrictionEdges, { budget: remainingBudgetSvf1, focus: normalizedFocus })
     : [];
-  // scm-svf1 sees the SAME enlarged subClassOf set (its own ⊑-lift over the
-  // two restrictions' filler classes) and reuses the SAME restrictionEdges
-  // just built for cls-svf1 above — needs at least two independently-declared
-  // restrictions over one property to have anything to compare (the kernel's
-  // own guard, `deriveSomeValuesFromSubsumption`'s doc comment).
+  // scm-svf1 reuses the SAME restrictionEdges built for cls-svf1 above —
+  // needs at least two restrictions over one property to compare.
   const remainingBudgetScmSvf = Math.max(0, budget - scmDerived.length - caxDerived.length - dwDerived.length - svf1Derived.length);
   const scmSvfDerived = remainingBudgetScmSvf > 0 && restrictionEdges.length > 1
     ? deriveSomeValuesFromSubsumption(restrictionEdges, enlargedSubClassEdges, { budget: remainingBudgetScmSvf, focus: normalizedFocus })
     : [];
-  // scm-svf1's own two structural premises per restriction (owl:onProperty /
-  // owl:someValuesFrom) are looked up by restriction id — restrictionEdges
-  // already carries each restriction's (property, target) pair, keyed the
-  // same way `deriveSomeValuesFromSubsumption`'s own output does.
   const restrictionByRid = new Map(restrictionEdges.map((r) => [r.restriction, r]));
 
-  // Batched write: ONE mutateMemory pass for the whole pass's conclusions
-  // (all five rules), not one appendFact per derived fact — appendFacts (the
-  // appendUtterances-precedent batch path, memory/core.mjs) does the same
-  // normalize+prose-tokenize+upsert work per fact but a SINGLE read-mutate-
-  // write, so a pass with many derivations no longer pays per-fact I/O.
-  // subject/object here are already-normalized terms straight off the stored
-  // graph (readFactRows/deriveSubClassClosure/deriveTypePropagation/
-  // deriveDisjointViolations never hand back an empty term), and predicate is
-  // always one of the three fixed constants above, so appendFacts never skips
-  // one of these — `ids` comes back exactly one-per-input, in order, safe to
-  // zip positionally below.
+  // Batched write: ONE mutateMemory pass for all five rules' conclusions via
+  // appendFacts, not one appendFact per derived fact.
   const toWrite = [
     ...scmDerived.map((d) => ({
       subject: d.subject, predicate: SUBCLASS_PREDICATE, object: d.object,
       provenance: ENTAILED_PROVENANCE,
-      // PLAN_SYLLOGIST.md §3's persisted-justification step, scm-sco only: the
-      // two premise fact ids THIS conclusion actually rode (a⊑b, b⊑c) — ids
-      // are content-addressed (factIdForTriple/memory/core.mjs), so this works
-      // whether the premise is a stated fact or another entailment this SAME
-      // pass just derived a round earlier (its id is predictable before it's
-      // even written). Read back by retractSubClassOf (below) to find every
-      // entailment a retracted premise could have supported, without a
-      // whole-graph re-scan.
+      // Persisted justification, scm-sco only: the two premise fact ids this
+      // conclusion rode (a⊑b, b⊑c) — content-addressed ids work even when a
+      // premise is itself an entailment this same pass just derived. Read
+      // back by retractSubClassOf (below) to find every entailment a
+      // retracted premise could have supported.
       justification: [
         factIdForTriple(d.subject, SUBCLASS_PREDICATE, d.via),
         factIdForTriple(d.via, SUBCLASS_PREDICATE, d.object),
@@ -1067,22 +761,10 @@ export async function syllogise(repoDir, { depth = 32, budget = 50, focus = null
       return {
         subject: d.subject, predicate: DISJOINT_PREDICATE, object: d.object,
         provenance: ENTAILED_DISJOINT_PROVENANCE,
-        // ruleConfidence < 1 (CAX_DW_RULE_CONFIDENCE) is deliberate, not a
-        // magic number: with the hook's default confidence of 1,
-        // min(premiseTrusts) × 1 can EQUAL a premise's own trust (e.g. two
-        // operator-taught 1.0 premises), tying/outranking the very premise it
-        // came from — this module's invariant is "never outranks a stated
-        // fact" (header comment), so cax-dw's conclusion is discounted
-        // strictly below its weakest premise, always, while still riding
-        // FAR above the bare 0.3 floor for a well-sourced premise pair.
         ...(premiseTrusts.length ? { premiseTrusts, ruleConfidence: CAX_DW_RULE_CONFIDENCE } : {}),
       };
     }),
     ...svf1Derived.map((d) => {
-      // the restriction's own two structural premises (owl:onProperty /
-      // owl:someValuesFrom rows) are stored with the restriction node as
-      // SUBJECT and the (already-normalized) property/target as OBJECT —
-      // premiseTrust's exact-triple lookup, so no extra normalization here.
       const premiseTrusts = numericOnly([
         premiseTrust(d.subject, d.viaProperty, d.viaValue),
         premiseTrust(d.viaValue, TYPE_PREDICATE, d.viaType),
@@ -1100,12 +782,6 @@ export async function syllogise(repoDir, { depth = 32, budget = 50, focus = null
       };
     }),
     ...scmSvfDerived.map((d) => {
-      // each restriction's own two structural premises (owl:onProperty /
-      // owl:someValuesFrom rows), for BOTH restrictions being compared, plus
-      // the y1⊑y2 subClassOf premise that licensed the comparison — always
-      // present here (unlike cax-dw/cls-svf1's optional lift premise): the
-      // kernel's own tautology screen guarantees viaY1 !== viaY2 for every
-      // derived scm-svf1 fact (`deriveSomeValuesFromSubsumption`'s doc comment).
       const r1 = restrictionByRid.get(d.subject);
       const r2 = restrictionByRid.get(d.object);
       const premiseTrusts = numericOnly([
@@ -1155,72 +831,31 @@ export async function syllogise(repoDir, { depth = 32, budget = 50, focus = null
  *  taught (same (s,p,o) → same id → provenance union, appendFact's own upsert
  *  contract), is NOT purely entailed any more — `retractSubClassOf`'s cascade
  *  must never delete it just because its now-stale justification broke; the
- *  taught half is a real, independent reason to keep believing it (the trust-
- *  tier concern PLAN_SYLLOGIST.md §3 names: "must never touch a higher-trust
- *  taught-only derivation"). */
+ *  taught half is a real, independent reason to keep believing it (must never
+ *  touch a higher-trust taught-only derivation). */
 function isPurelyEntailed(provenance) {
   const tags = String(provenance || "").split(" | ").filter(Boolean);
   return tags.length > 0 && tags.every((t) => t.startsWith("entailed:"));
 }
 
 /**
- * PLAN_SYLLOGIST.md §3's first real, scoped retraction slice: JTMS-style
- * dependency-directed removal, for scm-sco ONLY (subClassOf transitivity —
- * this file's simplest rule, and §3's own worked example: "a premise later
- * disappears, what else must be un-believed?"). Deliberately NOT the fuller
- * ATMS §3 also sketches (tracking every alternate premise-SET per fact, "a
- * further, NOT-currently-planned step") — this tracks exactly ONE
- * justification per entailed fact (the premise pair it actually rode, set at
- * write time — syllogise()'s own toWrite mapping, above): the JTMS-shaped
- * step §3 names as missing today ("a JTMS-shaped single justification per
- * fact in spirit, though not yet a persisted, walkable one").
- *
- * Retracting `subject ⊑ object` (a STATED or a previously-ENTAILED fact —
- * either may be retracted) proceeds in bounded rounds:
- *   1. Remove the named fact.
- *   2. Scan stored entailed scm-sco facts (purely-entailed ones only —
- *      `isPurelyEntailed`, above) for any whose persisted justification cites
- *      an id removed so far — candidates.
- *   3. VERIFY, never assume: a candidate is removed only if `subject ⊑
- *      object` is NO LONGER reachable over the SURVIVING subClassOf edge set
- *      (a full ⊑-ancestor walk, `buildAncestorCloser` — the SAME shared
- *      machinery `deriveTypePropagation`/`deriveDisjointViolations` already
- *      reuse, not reimplemented here). A fact with a SECOND, independent
- *      derivation path survives — a real possibility scm-sco's transitive
- *      closure allows (a⊑b⊑d AND a⊑c⊑d both license a⊑d) — exactly the
- *      failure mode a bare "delete anything citing the retracted id" JTMS
- *      walk gets wrong, and precisely why de Kleer's ATMS exists at all (§3's
- *      own citation). This VERIFY step is this slice's cheap, bounded answer
- *      to that known JTMS over-retraction limitation: one local graph walk
- *      per candidate, never a full alternate-justification enumeration.
- *   4. Repeat: a fact confirmed-removed this round becomes a new cascade
- *      source for the next round (removing a mid-chain link can ripple).
- *
- * Bounded by `budget` (max facts examined+removed, default 50 — the SAME
- * default every other rule in this file uses) and `depth` (max cascade
- * rounds, default 32, mirroring `deriveSubClassClosure`'s own fixpoint cap).
- * `truncated` flags the cascade may have been cut short before reaching a
- * fixpoint (candidates still pending when budget/depth ran out) — the SAME
- * honest-signal discipline `syllogise()`'s own `truncated` flag follows: a
- * caller must not read a truncated cascade's survivors as "provably still
- * consistent," only as "not yet shown inconsistent within budget."
+ * A scoped retraction slice: JTMS-style dependency-directed removal, for
+ * scm-sco ONLY. Retracting `subject ⊑ object` removes the fact, then cascades
+ * to any purely-entailed scm-sco fact whose persisted justification cites a
+ * removed id — but each candidate is VERIFIED (re-derivable over the
+ * surviving subClassOf edge set, not just "cited a removed id") before it is
+ * actually removed, since a fact can have a second, independent derivation
+ * path (a⊑b⊑d AND a⊑c⊑d both license a⊑d) that a bare delete-by-justification
+ * walk would wrongly discard. Repeats in rounds — a removed mid-chain link
+ * can ripple — bounded by `budget` (max facts examined+removed) and `depth`
+ * (max cascade rounds).
  *
- * Known, DELIBERATE scope limit (not a bug): this only ever touches scm-sco's
- * own entailed subClassOf facts. The other four rules (cax-sco/cax-dw/
- * cls-svf1/scm-svf1) do not yet persist a justification (none call
- * `factIdForTriple`/write `justification` — only `syllogise()`'s scmDerived
- * mapping does, above), so a type/disjointWith/someValuesFrom conclusion that
- * ALSO went stale when this same premise was retracted is not cascaded here.
- * Extending justification-tracking to the other four rules is mechanical
- * (each already computes a `via`/`viaX` pivot) but is a separate follow-up,
- * not attempted in this slice.
+ * Scope limit: only scm-sco persists a justification today, so a
+ * type/disjointWith/someValuesFrom conclusion that also went stale is not
+ * cascaded here (mechanical to extend, not attempted in this slice).
  *
- * Returns { retracted, count, budget, depth, truncated, found } — `retracted`
- * is every id actually removed (target first, then cascade order); `found`
- * is false (nothing else meaningful) when `subject ⊑ object` was never a
- * stored fact at all — an honest no-op, matching this module's "never guess"
- * discipline. No I/O beyond the one `removeFacts` call (skipped entirely when
- * `found` is false).
+ * Returns { retracted, count, budget, depth, truncated, found } — `found` is
+ * false when `subject ⊑ object` was never a stored fact.
  */
 export async function retractSubClassOf(repoDir, subject, object, { budget = 50, depth = 32 } = {}) {
   const s = normFactTerm(subject);
@@ -1236,11 +871,7 @@ export async function retractSubClassOf(repoDir, subject, object, { budget = 50,
   // removed id's own edge is excluded from that round's walk onward.
   const scRows = rows.filter((r) => isSubClassOf(r.predicate));
   const edgeOf = new Map(scRows.map((r) => [r.id, [r.subject, r.object]]));
-  // Only a purely-entailed scm-sco fact ever carries a walkable justification
-  // (see syllogise()'s toWrite mapping + isPurelyEntailed, above) — every
-  // other row's justification is [] (or the fact is also independently
-  // taught, so it is EXCLUDED here even if it happens to carry a stale one),
-  // so this candidate pool is naturally, correctly scoped.
+  // Only a purely-entailed scm-sco fact ever carries a walkable justification.
   const entailedScRows = scRows.filter((r) => r.justification.length && isPurelyEntailed(r.provenance));
 
   const removed = new Set([targetId]);
@@ -1253,17 +884,10 @@ export async function retractSubClassOf(repoDir, subject, object, { budget = 50,
       .sort((a, b) => a.subject.localeCompare(b.subject) || a.object.localeCompare(b.object));
     if (!candidates.length) break; // fixpoint — nothing left to (re-)check
 
-    // The surviving edge set for THIS round's verify walk — DRed's own
-    // "delete a superset, then selectively rederive" discipline (§3's own
-    // citation, Gupta/Mumick/Subrahmanian 1993): every candidate's OWN edge
-    // is excluded too, alongside every OTHER candidate under suspicion this
-    // SAME round, not just `removed` — otherwise a candidate would trivially
-    // "reach itself" through its own not-yet-deleted edge (or lean on a
-    // sibling candidate that is itself only standing on the same broken
-    // premise), understating what actually still needs re-verifying. A
-    // candidate that reaches its target through some OTHER, untouched edge
-    // (a genuinely independent derivation path this fact's single persisted
-    // justification never recorded) correctly survives.
+    // The surviving edge set for THIS round's verify walk excludes every
+    // candidate's own edge too, not just `removed` — otherwise a candidate
+    // could trivially "reach itself" through its own not-yet-deleted edge, or
+    // lean on a sibling candidate standing on the same broken premise.
     const candidateIds = new Set(candidates.map((c) => c.id));
     const survivingEdges = [...edgeOf.entries()]
       .filter(([id]) => !removed.has(id) && !candidateIds.has(id))
@@ -1297,27 +921,16 @@ export async function retractSubClassOf(repoDir, subject, object, { budget = 50,
 }
 
 /**
- * PROOF SEARCH (not a third rule — a bounded ROOTED chase that composes the
- * two rules above for a single "does `subj` reach one of `targets`?" query,
- * PLAN_INFERENCE_TESTING.md §4 stage 2's proof-chain substrate). Unlike
- * `deriveSubClassClosure`/`deriveTypePropagation` — which each compute a
- * WHOLE-GRAPH closure and then focus/budget-filter the result (so an
- * unrelated derivation that merely TOUCHES the focus term as a pivot or
- * object can still fill the budget before the one the caller actually wants)
- * — this walks OUTWARD from `subj` only, breadth-first, stopping the instant
- * a target is reached. That makes it safe to call live, per query, even over
- * a large (e.g. corpus-seeded) fact store: cost is bounded by `subj`'s own
- * reachable set and `maxHops`, never by how many OTHER classes the store
- * happens to know about.
+ * PROOF SEARCH (not a third rule — a bounded rooted chase for a single "does
+ * `subj` reach one of `targets`?" query). Walks OUTWARD from `subj` only,
+ * breadth-first, stopping the instant a target is reached — cost is bounded
+ * by `subj`'s own reachable set and `maxHops`, not by the whole graph, unlike
+ * `deriveSubClassClosure`/`deriveTypePropagation`'s whole-graph closures.
  *
- * The first hop may be EITHER a taught type edge (cax-sco: `subj rdf:type
- * C`) OR a taught subClassOf edge (scm-sco: `subj ⊑ C`) — `subj` may be an
- * individual or a class. Every hop after the first is subClassOf-only (once
- * "in" a class, propagation continues up the class hierarchy). Returns the
- * shortest chain as an ordered `[{ subject, predicate, object }, …]` premise
- * list (each already a stored fact — the caller cites its provenance), or
- * null when no chain reaches `targets` within `maxHops`. Pure, no I/O,
- * deterministic given the same edge lists.
+ * The first hop may be a taught type edge (cax-sco) or subClassOf edge
+ * (scm-sco); every hop after is subClassOf-only. Returns the shortest chain
+ * as an ordered `[{ subject, predicate, object }, …]` premise list, or null
+ * when no chain reaches `targets` within `maxHops`.
  */
 export function findIsaChain(subj, targets, typeEdges, subClassEdges, { maxHops = 6 } = {}) {
   const targetSet = targets instanceof Set ? targets : new Set(targets || []);
@@ -1339,8 +952,7 @@ export function findIsaChain(subj, targets, typeEdges, subClassEdges, { maxHops
   // hop counts the LENGTH of the paths currently in `frontier` (1 at the
   // first check). Check-then-extend, and never extend past maxHops — the
   // frontier is checked AT every length up to and including maxHops, never
-  // one hop beyond it (an earlier version's separate post-loop check did
-  // exactly that off-by-one — fixed here, regression-tested below).
+  // one hop beyond it.
   const seen = new Set([subj]);
   for (let hop = 1; hop <= maxHops && frontier.length; hop += 1) {
     for (const { node, path } of frontier) if (targetSet.has(node)) return path;