@hviana/sema 0.2.9 → 0.4.0

package/dist/src/mind/bridge.js

@@ -92,9 +92,9 @@
 import { cosine } from "../vec.js";
 import { conceptThreshold, dominates } from "../geometry.js";
 import { bytesEqual, indexOf } from "../bytes.js";
-import { foldTree, perceive } from "./primitives.js";
+import { foldTree, perceive, read } from "./primitives.js";
 import { chainReach, leafIdRun } from "./canonical.js";
-import { corpusN, edgeAncestors, hubBound } from "./traverse.js";
+import { corpusN, edgeAncestors, hubBound, sharedReachMemo, } from "./traverse.js";
 import { rItem, rNode } from "./trace.js";
 /** True when some query byte-range left UNACCOUNTED by `spans` contains a
  *  STORED window — content the store has seen that the proposed reading
@@ -217,7 +217,22 @@ function align(ctx, q, c, qo, co) {
 }
 /** Recall's corroborated-substitution bridge — see the module comment.
  *  Returns the best bridged grounding proposal, or null. */
-export async function substitutionBridge(ctx, query, proposed = []) {
+/** `proposed` is a THUNK, not a list: the bridge's own cheap gates (the
+ *  two-quantum query floor and the O(|query|) stored-window anchor scan)
+ *  decide whether ANY candidate can be aligned, and they need no proposals
+ *  to do it.  Resolving the caller's proposals eagerly meant recall paid its
+ *  exhaustive whole-index resonance — the most expensive single act on the
+ *  refusal path — for every query, including the ones whose windows the
+ *  store has never seen and which the anchor scan rejects outright.  Same
+ *  investment discipline the mechanism floors follow (AGENTS §2.6): never
+ *  compute a shared analysis just to discard it. */
+export async function substitutionBridge(ctx, query, proposed = async () => []) {
+    const meter = ctx.meter;
+    return meter
+        ? meter.time("substitutionBridge", () => bridgeImpl(ctx, query, proposed))
+        : bridgeImpl(ctx, query, proposed);
+}
+async function bridgeImpl(ctx, query, proposed) {
     const W = ctx.space.maxGroup;
     if (query.length < 2 * W)
         return null;
@@ -225,6 +240,37 @@ export async function substitutionBridge(ctx, query, proposed = []) {
     const N = corpusN(ctx);
     const bar = conceptThreshold(ctx.store.D);
     const reachCap = chainReach(W);
+    // PHRASE-SCALE CANDIDATE CAP — the same |content|·W bound the weave
+    // (pipeline-mechanism.ts), the cross-region junction ladder's
+    // `maxInterior`, and structural resonance's `maxSiblingBytes` all apply,
+    // for the same reason and now at the one remaining place that read
+    // candidate contexts WHOLE.
+    //
+    // The bridge accepts a candidate only when the query is DOMINATED by its
+    // matched runs plus substitutions, with at most one window W of slack at
+    // each edge and at most one chain reach (W²) per interior gap — so the
+    // candidate region an accepted alignment can ever consume is bounded by
+    // |query|·W.  Content beyond that cannot participate in any alignment
+    // this function would accept; reading it is pure cost.  And a candidate
+    // an order of magnitude past the query is not a paraphrase of it: it is a
+    // document or a whole conversation that merely quotes a phrase, and
+    // grounding through ITS learnt edge voices that document's continuation,
+    // not a phrase answer.
+    //
+    // Measured on the 17.7M-node / 325K-context store: uncapped, the refusal
+    // path materialised up to ~1 MB of candidate bytes per query (up to √N
+    // proposals plus √N climbed contexts, each read in full), and the frame-
+    // unanimity scan — which walks EVERY collected candidate's bytes, inside
+    // the per-gap expansion loop — paid that volume back tens of times per
+    // substitution.  Recall's run() was 0.7–2.6 s per refusing query.
+    const capBytes = query.length * W;
+    /** A candidate's bytes, phrase-scale capped: null when it exceeds the cap
+     *  (read one byte past it, so "too long" is decided without materialising
+     *  the rest) or has no content. */
+    const candidateBytes = (sid) => {
+        const b = read(ctx, sid, capBytes + 1);
+        return b.length === 0 || b.length > capBytes ? null : b;
+    };
     // 1. The query's stored windows, rarest first (fewest containers — the
     //    most discriminative anchors; hub-clamped like every fan-out read).
     //    The scan doubles as the ONE store probe of every query window: the
@@ -281,6 +327,55 @@ export async function substitutionBridge(ctx, query, proposed = []) {
         }
         return false;
     };
+    // ── EXPLAINED SPANS — the scaffolding judgement, corpus-global ──────────
+    //
+    // The question every gap poses is "may the two forms differ HERE without
+    // differing in what they SAY?", and that is the discriminative-vs-
+    // scaffolding question AGENTS §2.7 names, over the CORPUS-GLOBAL
+    // population.  It already has one definition — `dominates(reachOf(...), N)`,
+    // the same gate confluence's filler test uses ("scaffolding never binds").
+    // Nothing new is derived here; the bar is read, not invented.
+    //
+    // A span is explained when EITHER
+    //   • it is sub-quantum (< W) — typographic glue, the tolerance identityBar
+    //     already prices ("below one river window, byte overlap is chance"); or
+    //   • every full W-window inside it is COMMON by the store's own climb:
+    //     the ascent SATURATES (the window sits in more places than √N — the
+    //     climb's own definition of non-discriminative), or it resolves to a
+    //     majority of the corpus's contexts.  "the process of ", " is the ".
+    //
+    // THE READING MATTERS, not just the population (AGENTS §2.7).  This
+    // deliberately does NOT go through `reachOf`, which maps BOTH "saturated"
+    // and "reaches nothing" to Infinity.  For IDF weighting those are the same
+    // thing (no usable identity evidence); for THIS question they are
+    // opposites — a window reaching nothing is novel content, the most
+    // discriminative material there is, and reading it as Infinity would call
+    // it scaffolding.  Measured: with `reachOf`, "Is water wet?" was answered
+    // with "No, heavy water is not wet." — "heav"/"eavy" occur once, reach no
+    // edge-bearing ancestor, and were written off as filler.  So an
+    // empty-rooted window is NEVER explained, and neither is an untrained one
+    // (the same principle attestedQ applies to the query side).
+    const reachMemo = sharedReachMemo(ctx);
+    const explainedSpan = (bytes, from, to) => {
+        if (to - from < W)
+            return true;
+        for (let o = from; o + W <= to; o++) {
+            const ids = leafIdRun(ctx, bytes, o, o + W);
+            if (ids === null)
+                return false;
+            const wid = ctx.store.findBranch(ids);
+            if (wid === null)
+                return false;
+            const r = edgeAncestors(ctx, wid, N, reachMemo);
+            if (r.saturated)
+                continue; // in too many places to discriminate
+            if (r.roots.length === 0)
+                return false; // reaches nothing: novel content
+            if (!dominates(r.contextsReached, N))
+                return false;
+        }
+        return true;
+    };
     anchors.sort((a, b) => a.rarity - b.rarity);
     // Up to W anchors, at least one window apart — the quantum's own count.
     const picked = [];
@@ -311,12 +406,14 @@ export async function substitutionBridge(ctx, query, proposed = []) {
     // that could align at all: alignment can only seed at a picked anchor
     // window occurring literally in the candidate (measured: unconditional
     // re-folds multiplied the refusal-path latency several-fold).
-    for (const sid of proposed) {
+    // FIRST TOUCH of the caller's proposals — past every gate that could have
+    // refused without them (see substitutionBridge's doc).
+    for (const sid of await proposed()) {
         if (seen.has(sid))
             continue;
         seen.add(sid);
-        const tb = ctx.store.bytes(sid);
-        if (tb.length === 0)
+        const tb = candidateBytes(sid);
+        if (tb === null)
             continue;
         if (!picked.some((a) => indexOf(tb, query.subarray(a.off, a.off + W), 0) >= 0))
             continue;
@@ -353,9 +450,15 @@ export async function substitutionBridge(ctx, query, proposed = []) {
             break;
     }
     // 3. Align each candidate; gate its mismatches; keep the best.
+    // Over-cap candidates are dropped here rather than earlier: the climb
+    // channel deliberately reads no bytes while collecting (the climb visits
+    // hundreds of roots), so this is where its proposals are first sized.
     const allBytes = new Map();
-    for (const sid of candidates)
-        allBytes.set(sid, ctx.store.bytes(sid));
+    for (const sid of candidates) {
+        const b = candidateBytes(sid);
+        if (b !== null)
+            allBytes.set(sid, b);
+    }
     // FRAME UNANIMITY: a substitution U → C inside the frame (Lf, Rf) is
     // groundable only when the collected candidates — the store's own sample
     // of contexts sharing the query's content — are unanimous about the
@@ -408,7 +511,7 @@ export async function substitutionBridge(ctx, query, proposed = []) {
     let bestAccounted = 0;
     for (const sid of candidates) {
         const cBytes = allBytes.get(sid);
-        if (cBytes.length === 0)
+        if (cBytes === undefined)
             continue;
         // Seed at the rarest picked anchor that literally occurs in this
         // candidate.
@@ -520,7 +623,7 @@ export async function substitutionBridge(ctx, query, proposed = []) {
         // resonance rank but has no outgoing edge to bridge through).  This
         // mechanism exists to explain SUBSTITUTIONS; a query needing none is
         // recall's job, not the bridge's.
-        if (!ok || subs.length === 0)
+        if (!ok)
             continue;
         // Coverage: matched runs plus accepted substitutions must dominate the
         // query, every interior gap already proved ≤ W above, and the EDGES
@@ -546,6 +649,72 @@ export async function substitutionBridge(ctx, query, proposed = []) {
             continue;
         if (!dominates(covered, query.length))
             continue;
+        // ZERO-SUBSTITUTION ADMISSION — an IDENTITY claim, not a substitution.
+        //
+        // A candidate needing no substitution is normally refused (see the trap
+        // above), and that refusal is right for the case it was written for: the
+        // query is a strict byte-PREFIX of several candidates, each of which
+        // continues differently, and nothing here corroborates picking one
+        // continuation over another.  But that trap has a signature — the
+        // candidate carries substantial content BEYOND the alignment, and that
+        // surplus is exactly the "answer" the bridge would be inventing.
+        //
+        // The opposite shape is not ambiguous at all: the alignment explains BOTH
+        // strings end to end, and the only thing between them is sub-quantum glue
+        // — typographic punctuation the fold treats as structure.  Then the two
+        // are the SAME learnt form, and grounding through its edge returns that
+        // form's own trained answer, never a chosen-among-many completion.
+        //
+        // Why the ladder cannot reach these otherwise: the gist is a STRUCTURAL
+        // signature, so a mid-string insertion shifts every fold boundary after
+        // it.  Measured: `Who wrote Romeo and Juliet?` against the trained
+        // `Who wrote "Romeo and Juliet"?` — two inserted quote characters — scores
+        // cos 0.377, BELOW unrelated neighbours like "Who wrote the opera
+        // Carmen??" (0.603).  Recall's identity tiers gate on identityBar (0.969
+        // here) and its reach tiers on 0.875, so no gist-based tier can ever see
+        // it; only byte-exact alignment can, which is what this function does.
+        //
+        // The claim is deliberately strict, in three parts:
+        //
+        //   • QUERY SIDE — EXACT.  Every byte of the query must be a literal
+        //     match against the candidate: covered === query.length, no slack at
+        //     all, not even sub-quantum.  The query is what we are answering, so
+        //     an identity claim about it may write off NOTHING.  This is stricter
+        //     than the ≤ W edge tolerance the substituted path uses, and it has
+        //     to be: with a one-window allowance, `what is 2^10?` matched the
+        //     trained `what is 2+2?` — "^10" against "+2", four bytes, both sides
+        //     below W — and answered "2+2 is 4.", outweighing cover's authoritative
+        //     ALU result.  Below W, byte OVERLAP is chance rather than evidence;
+        //     that never made a below-W DIFFERENCE meaningless, and digits are the
+        //     case that proves it.
+        //   • CANDIDATE SIDE, INTERIOR — each gap must be an EXPLAINED span (see
+        //     explainedSpan): sub-quantum glue, or corpus-global scaffolding.
+        //     This side is asymmetric ON PURPOSE.  Material the CANDIDATE has and
+        //     the query omits is not something the asker asked about: if it is
+        //     scaffolding, dropping it changes nothing ("What is *the process of*
+        //     photosynthesis?"); if it is discriminative, the candidate answers a
+        //     DIFFERENT, narrower question ("Is *heavy* water wet?") and must be
+        //     refused.  Only the corpus can tell those apart, and it does.
+        //   • CANDIDATE SIDE, SURPLUS — its bytes are the matched runs
+        //     (byte-identical to the query's, hence the same total length) plus
+        //     its own gap spans; anything past that is surplus, and surplus is
+        //     the prefix trap.  A prefix-completion candidate fails here by the
+        //     whole length of the completion it wanted to supply — which is why
+        //     admitting scaffolding interiors does not reopen that trap.
+        //
+        // Ordered cheapest-first: the two arithmetic tests run before
+        // explainedSpan, whose per-window `reachOf` climbs are the only costly
+        // part (shared through the response/conversation reach memo, and reached
+        // only by a candidate that already survived every structural gate).
+        if (subs.length === 0) {
+            if (covered !== query.length)
+                continue;
+            const cGap = gaps.reduce((n, g) => n + (g.ce - g.cs), 0);
+            if (cBytes.length - covered - cGap > W)
+                continue;
+            if (!gaps.every((g) => explainedSpan(cBytes, g.cs, g.ce)))
+                continue;
+        }
         // KNOWN content may never be dismissed — see dismissedKnownContent
         // (the live case: "what is the capital of france" aligning into a
         // Matrix synopsis by writing off "ance" — a stored window of the

package/dist/src/mind/canonical.d.ts

@@ -1,7 +1,12 @@
 import type { MindContext } from "./types.js";
 /** The two sliding-window lengths the WRITE side interns over a stream's leaf
  *  ids: W−1 and W (the river's grouping quantum and its off-by-one neighbour,
- *  so a form straddling a group boundary is reachable from either cut). */
+ *  so a form straddling a group boundary is reachable from either cut).
+ *
+ *  Widening this to cover the reader's full segment scale (W−1..2W) was
+ *  measured and REFUTED: it tripled the store (3,160 → 8,980 nodes on a
+ *  200-pair corpus), slowed ingest 80%, and fixed not one test.  Unknown
+ *  regions are not caused by the index's scale. */
 export declare function canonicalWindows(W: number): [number, number];
 /** The READ side's chain reach: how many leaf ids a canonical chain may grow
  *  to from one position — W², the deepest two-level composite the write side's

package/dist/src/mind/canonical.js

@@ -24,7 +24,12 @@
 // what training indexed, which no type checker catches.
 /** The two sliding-window lengths the WRITE side interns over a stream's leaf
  *  ids: W−1 and W (the river's grouping quantum and its off-by-one neighbour,
- *  so a form straddling a group boundary is reachable from either cut). */
+ *  so a form straddling a group boundary is reachable from either cut).
+ *
+ *  Widening this to cover the reader's full segment scale (W−1..2W) was
+ *  measured and REFUTED: it tripled the store (3,160 → 8,980 nodes on a
+ *  200-pair corpus), slowed ingest 80%, and fixed not one test.  Unknown
+ *  regions are not caused by the index's scale. */
 export function canonicalWindows(W) {
     return [W - 1, W];
 }

package/dist/src/mind/graph-search.d.ts

@@ -143,6 +143,15 @@ export declare class GraphSearch {
      *  recursive completion), and chooseNext (distributional-evidence edge
      *  disambiguation when a recognised form has multiple continuations). */
     host: GraphSearchHost);
+    /** The hub bound √N (AGENTS §2.8) — the ONE fan-out cap, stated here
+     *  rather than imported from `traverse.ts` because this module is
+     *  deliberately host-based (it holds a bare Store, never a MindContext).
+     *  That is the same write/read-side duplication convention canonical.ts's
+     *  header documents: if the formula changes it must change in BOTH places.
+     *  It is stated ONCE per side, though — the expression used to be spelled
+     *  out at three call sites here, one of them inside a per-item rules
+     *  generator, and they had already drifted on the `Math.max(2, …)` floor. */
+    private hubBound;
     /** Explore the Sema graph for the lightest cover of the query and return its
      *  chosen spans left-to-right — WITH the derivation's total weight (the g
      *  value of the goal item, in the exported cost ladder), which think's

package/dist/src/mind/graph-search.js

@@ -190,6 +190,17 @@ export class GraphSearch {
         this.maxGroup = maxGroup;
         this.host = host;
     }
+    /** The hub bound √N (AGENTS §2.8) — the ONE fan-out cap, stated here
+     *  rather than imported from `traverse.ts` because this module is
+     *  deliberately host-based (it holds a bare Store, never a MindContext).
+     *  That is the same write/read-side duplication convention canonical.ts's
+     *  header documents: if the formula changes it must change in BOTH places.
+     *  It is stated ONCE per side, though — the expression used to be spelled
+     *  out at three call sites here, one of them inside a per-item rules
+     *  generator, and they had already drifted on the `Math.max(2, …)` floor. */
+    hubBound() {
+        return Math.ceil(Math.sqrt(Math.max(2, this.store.edgeSourceCount())));
+    }
     /** Explore the Sema graph for the lightest cover of the query and return its
      *  chosen spans left-to-right — WITH the derivation's total weight (the g
      *  value of the goal item, in the exported cost ladder), which think's
@@ -240,7 +251,17 @@ export class GraphSearch {
      *  that leads nowhere new, never at an arbitrary count. */
     solve(spanLen, recognition, conceptTarget, substitutions, connectors, computedResults, onDerivation) {
         const system = this.buildSearch(spanLen, recognition.sites, conceptTarget, recognition.leaves, recognition.splits, recognition.starts, substitutions, connectors, computedResults);
-        const derivation = lightestDerivation(system);
+        // Search-effort accounting (src/meter.ts): the chart's pops/pushes are
+        // the cover's real cost, and a heuristic that stops being admissible
+        // shows up as a pop count that explodes while the answer stays the same.
+        const meter = this.host.meter;
+        const stats = meter ? { pops: 0, pushes: 0 } : undefined;
+        const derivation = lightestDerivation(system, stats);
+        if (meter && stats) {
+            meter.searches++;
+            meter.searchPops += stats.pops;
+            meter.searchPushes += stats.pushes;
+        }
         // When covering under a substitution map (articulation), a form→out rule is
         // the form EMITTING the asker's voice, not grounding to its own answer — so
         // tell the reader to name those moves `voice` rather than `ground`.
@@ -262,12 +283,9 @@ export class GraphSearch {
     buildSearch(queryLen, sites, conceptTarget, leaves, splits, starts, substitutions, connectors, computedResults) {
         const W = this.maxGroup; // fusible span ceiling (shortest composite bound)
         // Same corpus-scale hub floor {@link atomIsHub}/{@link atomReach} (traverse.ts)
-        // derive for byte atoms — duplicated here rather than imported because this
-        // module is deliberately host-based (no MindContext), and both inputs
-        // (maxGroup, edgeSourceCount) are already in scope with no ctx needed.  If
-        // the formula ever changes, it must change in BOTH places (see canonical.ts's
-        // header for the same write/read-side duplication convention).
-        const atomsAreHubs = Math.max(1, Math.ceil((this.store.edgeSourceCount() * W) / 256)) > Math.ceil(Math.sqrt(Math.max(2, this.store.edgeSourceCount())));
+        // derive for byte atoms — see {@link hubBound} below for why this module
+        // states the formula itself instead of importing it.
+        const atomsAreHubs = Math.max(1, Math.ceil((this.store.edgeSourceCount() * W) / 256)) > this.hubBound();
         const nodeBytes = (n) => this.store.bytesPrefix(n, ALL);
         // Content-addressed probes over the store's hash-cons maps — the same keys
         // training filled.  No byte-by-byte trie walk.
@@ -511,8 +529,7 @@ export class GraphSearch {
             // guard then dead-ends it) with no way to reach the forward edge.
             // Forking offers every continuation as its own rule so the one that
             // genuinely advances (not a duplicate) is still reachable.
-            const bound = Math.ceil(Math.sqrt(Math.max(2, this.store.edgeSourceCount())));
-            const nx = this.store.nextFirst(it.node, bound);
+            const nx = this.store.nextFirst(it.node, this.hubBound());
             if (nx.length) {
                 // The SAME evidence-weighted disambiguation the first hop uses
                 // (below) identifies the most-corroborated continuation.  Yielding
@@ -861,16 +878,39 @@ export class GraphSearch {
         // is opportunistic cross-leaf recovery exactly like recognition.ts's own
         // canonical chain — findLeaf/findBranch here have no idea WHY these two
         // leaves are adjacent, only that their concatenation happens to spell a
-        // trained form ("hi" recovered from "W[hi]ch").  The same corpus-scale
-        // caution recognition.ts's `boundary` gate applies: trust it fully when
-        // `l.i` is a position the query's OWN fold chose as a boundary (real
-        // structural evidence); past the scale where atoms themselves stop
-        // discriminating, an interior offset's opportunistic match is noise, not
-        // a genuine cross-leaf recovery.  A completion-involved fuse (l.rec ||
-        // r.rec) is a different, legitimate case — a rewrite growing into its
-        // neighbour — and is exempt, same as recognition.ts's rec-derived sites.
-        const trusted = l.rec || r.rec || ctx.starts.has(l.i) ||
-            !ctx.atomsAreHubs;
+        // trained form ("hi" recovered from "W[hi]ch").  At hub scale, where
+        // atoms themselves no longer discriminate, that coincidence is noise.
+        //
+        // A FOLD BOUNDARY IS NOT EVIDENCE.  This gate used to exempt any fuse
+        // starting at `starts.has(l.i)` — documented as "a position the query's
+        // OWN fold chose as a boundary (real structural evidence)".  The plain
+        // river fold CHOOSES NOTHING: `riverFold` groups fixed-arity
+        // (`for (i = 0; i < complete; i += mg)`), so `starts` is exactly
+        // {0, W, 2W, 3W, …} for every query.  Measured on three unrelated
+        // queries: starts = 0,4,8,12,16,20,24,28[,32] in every case — the set
+        // carries ZERO content information, and the exemption therefore fired at
+        // a quarter of all offsets by arithmetic alone.
+        //
+        // What it cost (17.9M-node store, W=4): "In which country is the Eiffel
+        // Tower?" fused the atoms "h"+"i" at offset 4 — trusted only because
+        // 4 ≡ 0 (mod 4) — into the trained form "hi", followed its continuation
+        // edge, and grounded "Hello there. If you are looking for Open
+        // Assistant, look no further." as a FACT, explaining 2 of 37 bytes
+        // (density 0.054, a fifth of the 1/W honesty bar `thinGrounding`
+        // reports and does not enforce).
+        //
+        // So the exemption is removed rather than replaced: there is no cheap
+        // signal here that means what it claimed to mean, and inventing one
+        // would be worse than admitting the absence. Genuine cross-leaf forms
+        // are not lost — recognition.ts's canonical pass already probes EVERY
+        // byte offset for the write side's interned windows and emits them as
+        // sites, which arrive here as `l.rec`/`r.rec` and stay exempt. What
+        // remains excluded at hub scale is precisely the case with no evidence
+        // behind it: two byte atoms that happen to spell something.
+        //
+        // Below hub scale nothing changes (`!atomsAreHubs`): on a small store
+        // coincidence is rare and every chain is real evidence.
+        const trusted = l.rec || r.rec || !ctx.atomsAreHubs;
         let node = (trusted && bytes.length <= ctx.W)
             ? ctx.findLeafU(bytes)
             : undefined;

package/dist/src/mind/index.d.ts

@@ -5,3 +5,5 @@ export type { MechanismResult, PipelineMechanism, Precomputed, } from "./pipelin
 export type { InspectRationale, RationaleItem, RationaleStep, } from "./rationale.js";
 export type { AnchorRejectionReason, ClimbConsensusData, ConsensusAnchorTrace, ConsensusReachTrace, ConsensusRegionTrace, CrossRegionTier, JunctionVoteTrace, RegionOutcome, } from "./attention.js";
 export type { AncestorReach, AttentionRead, SaturationReason, SaturationStop, } from "./types.js";
+export type { DepositReport } from "./learning.js";
+export type { DecideGroundingData, NarrowDecisionData, Provenance, } from "./pipeline.js";

package/dist/src/mind/junction.d.ts

@@ -85,6 +85,16 @@ export declare function cachedRead(ctx: MindContext, cache: WalkCache | null, id
  *     few levels of its parts).  A side too common to decide within the
  *     budget abstains here and falls through to the resonance tier (the
  *     climb's own saturation semantics).
+ *
+ *     REFUTED TIGHTENING (measured, 325K contexts): giving this walk
+ *     edgeAncestors' cumulative LATERAL-CONE limit — abstain once the
+ *     accumulated lateral entries pass √N — would cut most of the pops (47
+ *     of 56 walks exceed √N lateral, and 47 exhaust the budget), but TWO OF
+ *     THE FOUR walks that actually found a container had lateral spread of
+ *     1425 and 1426, far past √N.  The cone limit is sound for
+ *     edgeAncestors' question and wrong for this one: a junction container
+ *     is legitimately reached across many containing structures.  Half the
+ *     successful junctions would be lost.
  *   • per-node hub guards — parent fan-outs beyond √N are hubs (not
  *     expanded); each node contributes at most one √N page of containers;
  *     √N collected candidates decide. */

package/dist/src/mind/junction.js

@@ -115,6 +115,16 @@ function cachedContainers(ctx, cache, id, limit) {
  *     few levels of its parts).  A side too common to decide within the
  *     budget abstains here and falls through to the resonance tier (the
  *     climb's own saturation semantics).
+ *
+ *     REFUTED TIGHTENING (measured, 325K contexts): giving this walk
+ *     edgeAncestors' cumulative LATERAL-CONE limit — abstain once the
+ *     accumulated lateral entries pass √N — would cut most of the pops (47
+ *     of 56 walks exceed √N lateral, and 47 exhaust the budget), but TWO OF
+ *     THE FOUR walks that actually found a container had lateral spread of
+ *     1425 and 1426, far past √N.  The cone limit is sound for
+ *     edgeAncestors' question and wrong for this one: a junction container
+ *     is legitimately reached across many containing structures.  Half the
+ *     successful junctions would be lost.
  *   • per-node hub guards — parent fan-outs beyond √N are hubs (not
  *     expanded); each node contributes at most one √N page of containers;
  *     √N collected candidates decide. */
@@ -137,6 +147,8 @@ unordered = false) {
     if (seeds.length === 0)
         return [];
     const b = budget ?? { n: bound * ctx.space.maxGroup };
+    if (ctx.meter)
+        ctx.meter.junctionWalks++;
     // DEPTH CAP: perception trees are W-ary and a junction container is
     // phrase-scale, so it sits within ~log_W(maxContainer) structural levels
     // of its parts — at most W levels for any practical W (plus the
@@ -153,6 +165,8 @@ unordered = false) {
     }));
     while (stack.length > 0 && out.length < bound && b.n-- > 0) {
         const { id: x, d } = stack.pop();
+        if (ctx.meter)
+            ctx.meter.junctionPops++;
         const f = cachedRead(ctx, cache, x, maxContainer);
         if (f.length > maxContainer)
             continue; // beyond phrase scale: prune branch

package/dist/src/mind/learning.d.ts

@@ -28,8 +28,30 @@ export declare function deposit(ctx: MindContext, input: Input, track: boolean,
 export declare function ingestOne(ctx: MindContext, input: Input): Promise<Sema & {
     id: number;
 }>;
-/** Ingest a pair (context, continuation) — learn an edge and pour halos. */
-export declare function ingestPair(ctx: MindContext, ctxInput: Input, cont: Input): Promise<void>;
+/** What one ingested item deposited — reported through {@link ingest}'s
+ *  optional `onDeposit` callback.  Pure provenance read-out: node ids are
+ *  content-addressed, so two byte-identical items report the SAME ids (that
+ *  is content addressing working, not an error), and the callback observes
+ *  the deposit without influencing it. */
+export interface DepositReport {
+    /** Zero-based position of the item in the ingested input (0 for a scalar
+     *  or single pair). */
+    index: number;
+    /** Whether the item was a bare experience or a (context, continuation)
+     *  pair. */
+    kind: "one" | "pair";
+    /** Root node id of the item's (context) bytes. */
+    contextId: number;
+    /** Root node id of the continuation bytes — pairs only. */
+    continuationId?: number;
+}
+/** Ingest a pair (context, continuation) — learn an edge and pour halos.
+ *  Returns the deposited root ids (context, continuation) — a pure
+ *  read-out; callers that ignore it behave exactly as before. */
+export declare function ingestPair(ctx: MindContext, ctxInput: Input, cont: Input): Promise<{
+    ctxId: number;
+    contId: number;
+}>;
 /** Dispatch the public ingest input shapes onto one-input / pair handlers —
  *  THE one reading of ingest's polymorphic surface (scalar, (context,
  *  continuation) pair, or a list mixing bare inputs and pairs).  Both ingest
@@ -41,7 +63,13 @@ export declare function dispatchIngest(input: Input | (Input | [Input, Input])[]
 }>, onPair: (ctxInput: Input, cont: Input) => Promise<void>): Promise<(Sema & {
     id: number;
 }) | undefined>;
-/** Ingest an input or array of inputs/pairs.  The public ingest entry point. */
-export declare function ingest(ctx: MindContext, input: Input | (Input | [Input, Input])[], second?: Input): Promise<(Sema & {
+/** Ingest an input or array of inputs/pairs.  The public ingest entry point.
+ *
+ *  `onDeposit`, when given, is invoked once per ingested item with the
+ *  deposited root node ids ({@link DepositReport}) — item-level provenance
+ *  for tooling that needs to know which stored node an ingested item became.
+ *  Purely observational: the callback runs after the item's deposit
+ *  completed and nothing reads its result. */
+export declare function ingest(ctx: MindContext, input: Input | (Input | [Input, Input])[], second?: Input, onDeposit?: (report: DepositReport) => void): Promise<(Sema & {
     id: number;
 }) | undefined>;

package/dist/src/mind/learning.js

@@ -188,7 +188,9 @@ async function propagateSuffixes(ctx, src, dst) {
         await ctx.store.link(id, dst);
     }
 }
-/** Ingest a pair (context, continuation) — learn an edge and pour halos. */
+/** Ingest a pair (context, continuation) — learn an edge and pour halos.
+ *  Returns the deposited root ids (context, continuation) — a pure
+ *  read-out; callers that ignore it behave exactly as before. */
 export async function ingestPair(ctx, ctxInput, cont) {
     const c = await deposit(ctx, ctxInput, true, true);
     const cont_ = await deposit(ctx, cont, false);
@@ -213,6 +215,7 @@ export async function ingestPair(ctx, ctxInput, cont) {
         await ctx.store.pourHalo(partId, contSeat);
         await ctx.store.pourHalo(contId, bindSeat(ctx.space, companySignature(ctx.space, partId), 0));
     }
+    return { ctxId, contId };
 }
 /** Dispatch the public ingest input shapes onto one-input / pair handlers —
  *  THE one reading of ingest's polymorphic surface (scalar, (context,
@@ -242,7 +245,26 @@ export async function dispatchIngest(input, second, onOne, onPair) {
     await onPair(input, second);
     return undefined;
 }
-/** Ingest an input or array of inputs/pairs.  The public ingest entry point. */
-export async function ingest(ctx, input, second) {
-    return dispatchIngest(input, second, (i) => ingestOne(ctx, i), (a, b) => ingestPair(ctx, a, b));
+/** Ingest an input or array of inputs/pairs.  The public ingest entry point.
+ *
+ *  `onDeposit`, when given, is invoked once per ingested item with the
+ *  deposited root node ids ({@link DepositReport}) — item-level provenance
+ *  for tooling that needs to know which stored node an ingested item became.
+ *  Purely observational: the callback runs after the item's deposit
+ *  completed and nothing reads its result. */
+export async function ingest(ctx, input, second, onDeposit) {
+    let index = 0;
+    return dispatchIngest(input, second, async (i) => {
+        const r = await ingestOne(ctx, i);
+        onDeposit?.({ index: index++, kind: "one", contextId: r.id });
+        return r;
+    }, async (a, b) => {
+        const { ctxId, contId } = await ingestPair(ctx, a, b);
+        onDeposit?.({
+            index: index++,
+            kind: "pair",
+            contextId: ctxId,
+            continuationId: contId,
+        });
+    });
 }

package/dist/src/mind/match.d.ts

@@ -34,6 +34,13 @@ export declare function alignRuns(ctx: MindContext, query: Uint8Array, ct: Uint8
  *  kind, so the substitution/redirection schemas work unchanged on conceptual
  *  alignment. */
 export interface GradedRun {
+    /** True for a run the CLIMB proposed rather than the byte matcher found — see
+     *  {@link computeWeave}'s phase 2.  Provenance only: a proposed run reaches
+     *  CAST already gated (dominant literal agreement, not frame, and over bytes
+     *  no literal run claimed), so every consumer reads it exactly like any other
+     *  run.  Filtering seat-shaped reads by this flag was tried and is NOT
+     *  needed — the gates, not the flag, are what make it safe. */
+    proposed?: boolean;
     qs: number;
     qe: number;
     cs: number;
@@ -124,3 +131,36 @@ export declare function reverseContext(ctx: MindContext, id: number, guide?: Vec
  *  the context it follows.  This is the direction ladder every mechanism's
  *  final grounding step reduces to. */
 export declare function project(ctx: MindContext, id: number, guide?: Vec | null): Promise<Uint8Array | null>;
+/** Check whether an anchor is a span-shaped skill exemplar: it represents a
+ *  fact whose context and answer together form a span-in-context pattern.
+ *  If the anchor has a nextOf continuation, that is the answer and the anchor
+ *  itself is the context.  Otherwise the anchor's prevOf parents provide
+ *  candidate contexts, and the longest one whose span is span-shaped wins. */
+export declare function skillExemplar(ctx: MindContext, anchor: number, guide?: Vec | null): Promise<{
+    contextBytes: Uint8Array;
+    answerBytes: Uint8Array;
+} | null>;
+/** Whether the answer is a SPARSE subsequence of the context (bytes in
+ *  order, arbitrary gaps) — the OPEN span-shape reading (see the section
+ *  note above).  This is what lets extraction validate a MULTI-PIECE
+ *  exemplar whose answer is stitched from several context runs — but it is
+ *  deliberately permissive, so it must never be used as evidence that one
+ *  span was "drawn from" another (see {@link containsSpan} for that).
+ *
+ *  There is deliberately NO containsSpan pre-check here: strict containment
+ *  IMPLIES the subsequence embedding (a contiguous run, or a resolved node —
+ *  whose content-addressed identity means its bytes occur contiguously — is
+ *  an in-order embedding with zero gaps), so the scan below decides alone,
+ *  with the same truth value.  The old pre-check re-perceived the context
+ *  (a full river fold) per CANDIDATE in skillExemplar's √N-capped loop —
+ *  pure cost, no discrimination. */
+export declare function isSpanShaped(_ctx: MindContext, context: Uint8Array, answer: Uint8Array): boolean;
+/** STRICT containment: the answer's resolved node appears in the context's
+ *  folded tree, or the answer occurs as one CONTIGUOUS byte run of the
+ *  context.  This is real evidence the answer was drawn from the context.
+ *  Fusion gates on this — the sparse-subsequence reading of
+ *  {@link isSpanShaped} is trivially satisfied by short answers over long
+ *  queries ("cold" is a gap-tolerant subsequence of most sentences holding
+ *  c…o…l…d in order), and gating fusion on it silently starved multi-topic
+ *  queries of their further points of attention. */
+export declare function containsSpan(ctx: MindContext, context: Uint8Array, answer: Uint8Array): boolean;