pi-mega-compact 0.18.0 → 0.19.0

package/dist/src/vector-cortex/encoder/trigram.js

@@ -0,0 +1,75 @@
+/**
+ * vector-cortex/encoder/trigram.ts — VC2B mode B: asset-free trigram encoder.
+ *
+ * Trigram B is a deterministic, asset-free (no learned model, no manifest, no
+ * calibration) 512-dim fixed feature encoding of a token/phrase sequence. It is
+ * the mode-B fallback selected when the learned asset (mode A) is removed,
+ * missing, unsupported, or digest-bad — and it never imports the learned asset
+ * or learned calibration (task 4). It derives directly from textual authority:
+ * the same document hashed via its byte-level trigrams yields the same 512-dim
+ * vector regardless of the asset state.
+ *
+ * Width is fixed at `ENCODER_TRIGRAM_WIDTH = 512` (VC2B task 4 "trigram B at 512
+ * dimensions"). The vector is L2-normalized; a zero-norm (empty) input maps to
+ * the all-zero vector, matching the heads convention of the VectorSet.
+ *
+ * Failure-triad independence: B's algorithm/index is distinct from A (learned
+ * projections) and C (token/phrase lexical) — it is a deterministic hashed
+ * n-gram bag-of-hashes, computed purely in-process with no external asset.
+ *
+ * Pi-agnostic, zero network (PREVENT-PI-004), no `any` (PREVENT-011).
+ */
+import { createHash } from "node:crypto";
+import { l2Normalize } from "./heads.js";
+import { createEncoderHeadsReporter, } from "./emit-vc2b.js";
+/** Fixed output width of trigram B (VC2B task 4). */
+export const ENCODER_TRIGRAM_WIDTH = 512;
+/** Tokenize a phrase into byte-level trigrams (3-byte sliding windows). For a
+ *  short phrase with fewer than 3 bytes we still emit the available shingles. */
+function trigrams(text) {
+    const bytes = Buffer.from(text, "utf8");
+    if (bytes.length === 0)
+        return [];
+    const out = [];
+    const n = bytes.length;
+    // `Math.max(1, n - 2)` already emits a single whole-string shingle for 1- and
+    // 2-byte phrases (slice(0,3) covers the whole buffer), so there is NO separate
+    // short-phrase block — adding one would hash the same shingle twice (Q05).
+    for (let i = 0; i < Math.max(1, n - 2); i++) {
+        const chunk = bytes.slice(i, i + 3);
+        out.push(chunk.toString("hex"));
+    }
+    return out;
+}
+/**
+ * Encode a phrase into a 512-dim L2-normalized trigram vector (all-zero on
+ * empty input). Deterministic: the same text always yields the same vector
+ * (repeat drift == 0) — no asset, no calibration, no network.
+ */
+export function embedTrigram512(text) {
+    const width = ENCODER_TRIGRAM_WIDTH;
+    const out = new Float32Array(width);
+    // Feistel-style double hashing of each trigram into a bucket index + weight.
+    for (const tg of trigrams(text)) {
+        const h1 = createHash("sha256").update(tg).digest();
+        const bucket = h1.readUInt32BE(0) % width;
+        const weight = (h1.readUInt32BE(4) / 4294967295) * 2 - 1;
+        out[bucket] += weight;
+    }
+    return l2Normalize(out);
+}
+/**
+ * The 512-dim vector is produced even when the learned asset is absent: this is
+ * the mode-B selection point. Returns `{ ok: true, dim, width }` always — there
+ * is no asset to consult (task 4 + ENC-FALLBACK-003). Selecting mode B also
+ * emits `vector_cortex_encoder_fallback_selected` via the flag-gated reporter
+ * (task 5) — the production seam that makes the fallback event live in the
+ * runtime, not dead test-only wiring.
+ */
+export function selectTrigramBFallback(options = {}) {
+    const reporter = options.reporter ?? createEncoderHeadsReporter();
+    const selection = { ok: true, mode: "B", dim: ENCODER_TRIGRAM_WIDTH, width: ENCODER_TRIGRAM_WIDTH };
+    reporter.fallbackSelected({ mode: selection.mode, dim: selection.dim, width: selection.width });
+    return selection;
+}
+export { l2Normalize };

package/dist/src/vector-cortex/encoder/types.js

@@ -83,3 +83,87 @@ export const ENC_IDS = [
     "ENC-007",
     "ENC-008",
 ];
+// ---------------------------------------------------------------------------
+// VC2B — multi-head encoder (VectorSetV1 / HeadCalibrationDraft).
+// ---------------------------------------------------------------------------
+/** The five independent projection heads in STABLE order (MODEL_ASSET
+ *  §decision record; VC2B task 2 "stable order"). The array order is the
+ *  normative ordering consumed by consumers: semantic, dependency,
+ *  contradiction, cache-stability, payload-routing. */
+export const ENCODER_HEAD_ORDER = [
+    "semantic",
+    "dependency",
+    "contradiction",
+    "cacheStability",
+    "payloadRouting",
+];
+/** The ordered per-head output dimensions: semantic 384, dependency 128,
+ *  contradiction 128, cacheStability 64, payloadRouting 32 (VC2B task 2). */
+export const ENCODER_HEAD_DIMS = {
+    semantic: 384,
+    dependency: 128,
+    contradiction: 128,
+    cacheStability: 64,
+    payloadRouting: 32,
+};
+/** Ordered dimension list matching ENCODER_HEAD_ORDER (384/128/128/64/32). */
+export const ENCODER_HEAD_DIM_ORDER = ENCODER_HEAD_ORDER.map((h) => ENCODER_HEAD_DIMS[h]);
+/**
+ * Weighted training losses per head (MODEL_ASSET §data/losses/calibration):
+ * semantic .35, dependency .20, contradiction .20, cache .15, payload .10.
+ * These are normative (VC2B task 3: "losses exactly .35/.20/.20/.15/.10").
+ */
+export const ENCODER_HEAD_LOSS_WEIGHTS = {
+    semantic: 0.35,
+    dependency: 0.2,
+    contradiction: 0.2,
+    cacheStability: 0.15,
+    payloadRouting: 0.1,
+};
+/** Sum of the five loss weights must be exactly 1.0 (asserted in tests). */
+export const ENCODER_HEAD_LOSS_SUM = 1.0;
+/** Deterministic seed shared by Python/NumPy training and ONNX export (VC2B
+ *  task 3: "seed ... at 1729"). */
+export const ENCODER_SEED = 1729;
+/** The 16 registered VC2B conformance IDs (task 1: "register ENC-009..016"). */
+export const ENC2B_IDS = [
+    "ENC-009",
+    "ENC-010",
+    "ENC-011",
+    "ENC-012",
+    "ENC-013",
+    "ENC-014",
+    "ENC-015",
+    "ENC-016",
+];
+/** Normative per-head + asset qualification thresholds (MODEL_ASSET §qualification
+ *  + EVALUATION.md §metrics). These are the constants `select.ts` evaluates the
+ *  candidate's held-out metrics against (task 3 atomic check). */
+export const EVALUATION_THRESHOLDS = {
+    semantic: { spearman: 0.75, recallAt10: 0.9 },
+    dependency: { precision: 0.97, recall: 0.95 },
+    contradiction: { precision: 0.98, recall: 0.9, ece: 0.05 },
+    cacheStability: { precision: 0.999, recall: 0.9 },
+    payloadRouting: { macroF1: 0.97, exactAnchorRecall: 1.0 },
+    reconstruction: { dependencyClosureRecall: 1.0 },
+    asset: { maxTokens: 512, maxLatencyP95Ms: 40, maxRssDeltaMib: 150 },
+};
+/** VC2C-specific qualification failure codes (returned, never thrown). */
+export const ENC_QUALIFICATION_FAIL = {
+    /** The qualification manifest hash does not match the calibration that was fit
+     *  (corrupt qualification manifest after calibration before selection). */
+    DIGEST_MISMATCH: "ENC_QUALIFICATION_DIGEST_MISMATCH",
+    /** One or more per-head EVALUATION thresholds failed (demotes all of A). */
+    THRESHOLD_FAILED: "ENC_QUALIFICATION_THRESHOLD_FAILED",
+    /** An asset-field qualification check (asset/latency/RSS) failed. */
+    ASSET_FAILED: "ENC_QUALIFICATION_ASSET_FAILED",
+    /** Calibration was attempted using held-out labels (fit prohibition). */
+    HELD_OUT_IN_FIT: "ENC_QUALIFICATION_HELD_OUT_IN_FIT",
+};
+/** The 4 registered VC2C conformance IDs (task 1: "register ENC-017..020"). */
+export const ENC2C_IDS = [
+    "ENC-017",
+    "ENC-018",
+    "ENC-019",
+    "ENC-020",
+];

package/dist/vector-cortex/encoder/calibrate.js

@@ -0,0 +1,174 @@
+/**
+ * vector-cortex/encoder/calibrate.ts — VC2C calibration fit (task 2).
+ *
+ * Fits a `CalibrationV1` using ONLY the calibration split. Held-out
+ * (test/eval) labels are STRICTLY PROHIBITED from the fit inputs: the fit
+ * function rejects (ENC_QUALIFICATION_HELD_OUT_IN_FIT) any example whose
+ * `itemId` appears in the caller's held-out set. Ties in example score are
+ * broken deterministically by item ID (stable score/id ties), never by arrival
+ * order, so the fit is invariant to row order.
+ *
+ * The calibration split assignment is grouped by repository+session
+ * (EVALUATION.md §corpus): every label-bearing item carries a `repository` +
+ * `session` group, and the split digest is the canonical SHA-256 over the
+ * sorted group list. A single repo/session group NEVER crosses split boundaries
+ * (the caller seeds a group wholly into the calibration split or not at all).
+ *
+ * The fit itself is a deterministic seeded per-head temperature + threshold
+ * calibration over the calibration-only examples (real isotonic/Platt weights
+ * land with trained weights; the contract, split isolation, held-out
+ * prohibition, stable ties, and frozen-temperature/threshold surface are all
+ * normative here). Pi-agnostic, zero network (PREVENT-PI-004), no `any`
+ * (PREVENT-011).
+ */
+import { createHash } from "node:crypto";
+import { ENCODER_HEAD_ORDER, ENCODER_SEED, ENC_QUALIFICATION_FAIL, } from "./types.js";
+/** Canonical digests of a sorted stable representation (order-invariant). */
+function digestStrings(values) {
+    const sorted = [...values].sort();
+    return createHash("sha256").update(sorted.join("\n")).digest("hex");
+}
+/**
+ * Compute the calibration split digest for a set of (repository, session) groups
+ * present in the fit. The group list is canonicalized (sorted, deduped) so the
+ * digest is invariant to row order within the input. A caller that seeds WHOLE
+ * groups into the calibration split guarantees no group crosses a boundary.
+ */
+/**
+ * Render one (repository, session) group as an injective canonical string.
+ * Each field is length-prefixed (`<len>:<value>`), so two distinct pairs can
+ * never collide to the same rendering — e.g. `{r:"a", s:"b::c"}` renders as
+ * `1:a3:b::c` while `{r:"a::b", s:"c"}` renders as `4:a::b1:c`. A plain
+ * `repository::session` join would conflate those when identifiers happen to
+ * contain "::"; length-prefixing makes the split digest sound for arbitrary
+ * repository/session identifiers.
+ */
+function renderGroup(g) {
+    return `${g.repository.length}:${g.repository}${g.session.length}:${g.session}`;
+}
+export function calibrationSplitDigest(groups) {
+    const rendered = new Set();
+    for (const g of groups)
+        rendered.add(renderGroup(g));
+    return digestStrings([...rendered]);
+}
+/** A deterministic 32-bit LCG step (matches the heads/runtime projectors). */
+function nextState(state) {
+    return (state * 1664525 + 1013904223) >>> 0;
+}
+/**
+ * Deterministic per-head temperature in a stable, healthy range (e.g. 0.8..1.5).
+ *
+ * The caller-supplied `seed` is mixed into the LCG state through independent
+ * steps so it ALWAYS affects the temperature — even when it equals ENCODER_SEED.
+ * (A naive `ENCODER_SEED ^ head.length ^ seed` cancels the two seed terms when
+ * `seed === ENCODER_SEED`, leaving a pure function of the head — the default
+ * path would make the seed option inert. Here the seed seeds the PRNG first,
+ * then the head name is folded in, so both vary the fit independently.) The fit
+ * stays deterministic for a fixed (seed, head) across processes.
+ */
+function fitTemperature(head, seed) {
+    let state = (seed >>> 0) ^ 0x9e3779b9;
+    state = nextState(state);
+    state = (state ^ (head.length >>> 0)) >>> 0;
+    state = nextState(state);
+    state = (state ^ 0x85ebca6b) >>> 0;
+    state = nextState(state);
+    const r = (state / 4294967296) % 1;
+    return 0.8 + r * 0.7; // 0.8 .. 1.5
+}
+/**
+ * Deterministic per-head decision threshold derived from the calibration
+ * distribution. For a head with BOTH classes present the threshold is the
+ * midpoint between the highest-scoring negative (label 0) and the lowest-scoring
+ * positive (label 1) — a true between-class balance point that NEVER lands on a
+ * negative example's own score (code-review Q03): a future inference at the
+ * highest calibration negative is still classified negative, and one at the
+ * lowest calibration positive is still classified positive. Degenerate heads with
+ * a single class fall back conservatively (no positives -> just above the top
+ * observed score; no negatives -> just below the lowest observed positive) and an
+ * empty head defaults to 0.5. Scoring is order-invariant: ties in score resolve by
+ * item ID bytewise (stable score/id ties), never by arrival order, so the fit is
+ * invariant to row order. This frozen threshold is a normative placeholder (real
+ * trained weights land later).
+ */
+function fitThreshold(head, examples) {
+    const headEx = examples
+        .filter((e) => e.head === head)
+        .slice()
+        .sort((a, b) => (a.score - b.score) || (a.itemId < b.itemId ? -1 : a.itemId > b.itemId ? 1 : 0));
+    if (headEx.length === 0)
+        return 0.5;
+    let highestNeg = -Infinity;
+    let lowestPos = Infinity;
+    for (const e of headEx) {
+        if (e.label === 0)
+            highestNeg = Math.max(highestNeg, e.score);
+        else
+            lowestPos = Math.min(lowestPos, e.score);
+    }
+    if (lowestPos === Infinity) {
+        // Only negatives observed — no positive class to balance against. Set the
+        // threshold just above the top observed score so no observed negative is
+        // re-admitted (a genuine future positive must exceed all calibration negatives).
+        return Math.max(0.5, highestNeg + 0.05);
+    }
+    if (highestNeg === -Infinity) {
+        // Only positives observed — set the threshold just below the lowest observed
+        // positive so every observed positive is admitted.
+        return Math.max(0, lowestPos - 0.05);
+    }
+    // Both classes present: the midpoint strictly between the highest negative and
+    // the lowest positive is the between-class balance point (Q03).
+    return (highestNeg + lowestPos) / 2;
+}
+/**
+ * Fit `CalibrationV1` over the calibration split only (task 2).
+ *
+ *   - rejects any item whose `itemId` is in `heldOutIds` (held-out labels are
+ *     prohibited from fit inputs).
+ *   - treats `groups` as the calibration split units; the emitted split digest
+ *     covers only the groups ACTUALLY present in the fit examples (plus the
+ *     declared `groups`, when supplied — see below).
+ *   - stable score/id ties (never arrival order).
+ *
+ * Held-out labels are prohibited by construction: the caller passes the full set
+ * of held-out item IDs, and the fit fails loudly if any calibration input is
+ * actually a held-out item — the fit can never silently learn from review labels.
+ */
+export function fitCalibration(examples, options = {}) {
+    const seed = options.seed ?? ENCODER_SEED;
+    const heldOut = new Set(options.heldOutIds ?? []);
+    for (const e of examples) {
+        if (heldOut.has(e.itemId)) {
+            return {
+                ok: false,
+                code: ENC_QUALIFICATION_FAIL.HELD_OUT_IN_FIT,
+                reason: `held-out item ${e.itemId} leaked into calibration fit`,
+            };
+        }
+    }
+    // Stable order by (score, itemId) — arrival order never affects the digest or
+    // the fit; ties resolve by item ID bytewise (stable score/id ties).
+    const stable = examples
+        .slice()
+        .sort((a, b) => (a.score - b.score) || (a.itemId < b.itemId ? -1 : a.itemId > b.itemId ? 1 : 0));
+    const usedGroups = options.groups ?? stable.map((e) => ({ repository: e.repository, session: e.session }));
+    const splitDigest = calibrationSplitDigest(usedGroups);
+    const temperatures = {};
+    const thresholds = {};
+    for (const head of ENCODER_HEAD_ORDER) {
+        temperatures[head] = fitTemperature(head, seed);
+        thresholds[head] = fitThreshold(head, stable);
+    }
+    const calibration = {
+        schema: "calibration-v1",
+        headOrder: [...ENCODER_HEAD_ORDER],
+        calibrationSplitDigest: splitDigest,
+        fittedOnCalibrationOnly: true,
+        temperatures,
+        thresholds,
+        seed,
+    };
+    return { ok: true, calibration };
+}

package/dist/vector-cortex/encoder/emit-vc2b.js

@@ -0,0 +1,63 @@
+/**
+ * vector-cortex/encoder/emit-vc2b.ts — VC2B observability seam.
+ *
+ * Owns the two VC2B events (task 5), gated on `MEGACOMPACT_VC2B` so the
+ * flag-OFF path emits zero events (mode C parity, byte-identical predecessor):
+ *
+ *   vector_cortex_encoder_heads_emitted      — a multi-head VectorSetV1 produced
+ *   vector_cortex_encoder_fallback_selected  — a mode B/C fallback selected
+ *
+ * No dashboard or API change is necessary for this internal sprint (task 5).
+ * Every event is a JSON line with `ts` + `event` (ENGINEERING_PRACTICES §8); the
+ * emitters are non-fatal (never break the agent loop). Pi-agnostic, zero network
+ * (PREVENT-PI-004), no `any` (PREVENT-011).
+ */
+import { VC2B_ENABLED } from "../../config/vector-cortex.js";
+import { Logger } from "../../log.js";
+/** A flag-gated no-op reporter (zero emissions, structural no-op). */
+export const NOOP_VC2B_REPORTER = {
+    headsEmitted: () => { },
+    fallbackSelected: () => { },
+};
+/**
+ * The default emitter: routes both VC2B events into the append-only structured
+ * logger (`src/log.ts`) as JSON lines with `ts` + `event`. Supplying `emit:` to
+ * `createEncoderHeadsReporter` replaces this with a caller-provided sink (used
+ * by tests and downstream consumers). Making the default a REAL producer means a
+ * caller that just invokes the producer seam (`encodeOrFallback`, `encodeVectorSet`,
+ * `selectTrigramBFallback`, `selectLexicalC`) without injecting an emitter still
+ * yields structured telemetry instead of silently dropping every event (task 5,
+ * code-review Q01). Best-effort: the logger swallows all I/O errors.
+ */
+function defaultEmitFor(logPath) {
+    const logger = new Logger(logPath === undefined ? {} : { path: logPath });
+    return (event, fields) => {
+        logger.info(event, fields);
+    };
+}
+/**
+ * Flag-gated emit, defaulting to a real logger-backed sink. The returned
+ * reporter is itself flag-gated (`VC2B_ENABLED`), so wiring it into a producer
+ * seam yields zero emissions when `MEGACOMPACT_VC2B=0` (byte-identical to the
+ * predecessor). Pass an explicit `emit` to route elsewhere (tests, downstream
+ * consumers); omit it to emit real structured log lines (Q01: the default is a
+ * live producer, not a silent no-op). `opts.logPath` only redirects the default
+ * sink and is ignored when `emit` is supplied.
+ */
+export function createEncoderHeadsReporter(emit, opts = {}) {
+    const sink = emit ?? defaultEmitFor(opts.logPath);
+    const fire = (event, fields) => {
+        if (!VC2B_ENABLED())
+            return;
+        try {
+            sink(event, { ...fields, ts: new Date().toISOString() });
+        }
+        catch {
+            /* non-fatal observability */
+        }
+    };
+    return {
+        headsEmitted: (fields) => fire("vector_cortex_encoder_heads_emitted", fields),
+        fallbackSelected: (fields) => fire("vector_cortex_encoder_fallback_selected", fields),
+    };
+}

package/dist/vector-cortex/encoder/emit-vc2c.js

@@ -0,0 +1,59 @@
+/**
+ * vector-cortex/encoder/emit-vc2c.ts — VC2C observability seam.
+ *
+ * Owns the two VC2C events (task 5), gated on `MEGACOMPACT_VC2C` so the
+ * flag-OFF path emits zero events (mode C parity, byte-identical predecessor):
+ *
+ *   vector_cortex_encoder_qualification_passed   — a QualifiedEncoderV1 produced (A)
+ *   vector_cortex_encoder_qualification_demoted  — qualification failed; all of A demoted
+ *
+ * Every event is a JSON line with `ts` + `event` (ENGINEERING_PRACTICES §8); the
+ * `ts` is the numeric epoch-ms timestamp the `Logger` injects (LogEntry.ts is
+ * `number`), so these events are consistent with the rest of the log stream
+ * (code-review Q04 — no ISO-string ts override). The emitters are non-fatal
+ * (never break the agent loop). Pi-agnostic, zero network (PREVENT-PI-004), no
+ * `any` (PREVENT-011).
+ */
+import { VC2C_ENABLED } from "../../config/vector-cortex.js";
+import { Logger } from "../../log.js";
+/**
+ * The default emitter: routes both VC2C events into the append-only structured
+ * logger (`src/log.ts`) as JSON lines with `event` (the logger injects the
+ * numeric `ts`). Supplying `emit:` to `createEncoderQualificationReporter`
+ * replaces this with a caller-provided sink (used by tests and downstream
+ * consumers). Making the default a REAL producer means a caller that just
+ * invokes the producer seam (`selectQualifiedEncoder`) without injecting an
+ * emitter still yields structured telemetry instead of silently dropping every
+ * event (task 5). Best-effort: the logger swallows all I/O errors.
+ */
+function defaultEmitFor() {
+    const logger = new Logger();
+    return (event, fields) => {
+        logger.info(event, fields);
+    };
+}
+/**
+ * Flag-gated emit, defaulting to a real logger-backed sink. The returned
+ * reporter is itself flag-gated (`VC2C_ENABLED`), so wiring it into a producer
+ * seam yields zero emissions when `MEGACOMPACT_VC2C=0` (byte-identical to the
+ * predecessor). Pass an explicit `emit` to route elsewhere (tests, downstream
+ * consumers); omit it to emit real structured log lines. The `ts` is set by the
+ * underlying sink (numeric epoch-ms from `Logger`), not overridden here (Q04).
+ */
+export function createEncoderQualificationReporter(emit) {
+    const sink = emit ?? defaultEmitFor();
+    const fire = (event, fields) => {
+        if (!VC2C_ENABLED())
+            return;
+        try {
+            sink(event, fields);
+        }
+        catch {
+            /* non-fatal observability */
+        }
+    };
+    return {
+        qualificationPassed: (fields) => fire("vector_cortex_encoder_qualification_passed", fields),
+        qualificationDemoted: (fields) => fire("vector_cortex_encoder_qualification_demoted", fields),
+    };
+}

package/dist/vector-cortex/encoder/fallback.js

@@ -0,0 +1,62 @@
+/**
+ * vector-cortex/encoder/fallback.ts — VC2C qualification fallback selection.
+ *
+ * After a qualification demotion (qualify/select fails to produce mode A), this
+ * seam selects the mode-B/C fallback and routes the encode request to the
+ * independently initialized VC2B producers (trigram B / lexical C). It is the
+ * breaker-recovery handoff for the encoder triad:
+ *
+ *   A = fully qualified learned asset (QualifiedEncoderV1).
+ *   B = asset-free trigram, forced by ANY one failed qualification threshold
+ *       (or a qualification-manifest digest mismatch).
+ *   C = token/phrase lexical, forced when A is absent AND B itself errors
+ *       (injected B error), or when the caller forces C.
+ *
+ * Breaker recovery follows TRIAD_RESILIENCE: a demotion may probe/promote, never
+ * directly re-qualify A; the fallback only ever SELECTS B or C locally, never a
+ * remote fetch (PREVENT-PI-004). C states its loss of old semantic context.
+ *
+ * Pi-agnostic, zero network (PREVENT-PI-004), no `any` (PREVENT-011).
+ */
+import { createEncoderHeadsReporter } from "./emit-vc2b.js";
+import { embedTrigram512, selectTrigramBFallback, ENCODER_TRIGRAM_WIDTH } from "./trigram.js";
+import { embedLexical, selectLexicalC, ENCODER_LEXICAL_WIDTH, ENCODER_LEXICAL_LIMITATION } from "./lexical.js";
+/** Deterministic text derived from an int token sequence (mirrors router.ts). */
+function textFromTokens(tokens) {
+    return tokens.join("-");
+}
+/**
+ * Select the B/C fallback for a failed qualification and produce the encode
+ * vector. When `injectBError` or `forceC` is set, selection lands on lexical C
+ * (absent A + injected B error -> C); otherwise a qualification THRESHOLD or
+ * DIGEST mismatch selects trigram B. Never returns without an explicit verdict.
+ */
+export function selectQualificationFallback(qualificationCode, tokens, options = {}) {
+    const reporter = options.reporter ?? createEncoderHeadsReporter();
+    const tokensText = textFromTokens(tokens);
+    // A absent + injected/forced B error -> C (lexical) with the documented
+    // semantic-context limitation.
+    if (options.forceC || options.injectBError) {
+        selectLexicalC({ reporter });
+        const vector = embedLexical(tokensText);
+        return {
+            ok: true,
+            mode: "C",
+            vector,
+            width: ENCODER_LEXICAL_WIDTH,
+            code: options.injectBError ? "ENC_B_INJECTED_ERROR" : qualificationCode,
+            limitation: ENCODER_LEXICAL_LIMITATION,
+        };
+    }
+    // A qualification demotion selects the independently initialized trigram B.
+    selectTrigramBFallback({ reporter });
+    const vector = embedTrigram512(tokensText);
+    return {
+        ok: true,
+        mode: "B",
+        vector,
+        width: ENCODER_TRIGRAM_WIDTH,
+        code: qualificationCode,
+        limitation: null,
+    };
+}

package/dist/vector-cortex/encoder/heads.js

@@ -0,0 +1,113 @@
+/**
+ * vector-cortex/encoder/heads.ts — VC2B multi-head encoder (tasks 1–2).
+ *
+ * Produces a `VectorSetV1`: five independent L2-normalized projection heads in
+ * STABLE order — semantic 384, dependency 128, contradiction 128, cacheStability
+ * 64, payloadRouting 32 (MODEL_ASSET §decision record). Each head L2-normalizes
+ * its raw projection; a zero-norm projection maps to an all-zero vector (task 2).
+ *
+ * The raw per-head projection is a deterministic seeded compression of the input
+ * token sequence (seeded by `ENCODER_SEED` and the head's stable index), which
+ * mirrors the VC2A `projectSemantic` placeholder pattern: the contract, shape
+ * gating, normalization, zero-norm mapping, ordering and loss/seed constants are
+ * all normative here; real trained weights are substituted in VC2C. This keeps
+ * the mode-A multi-head path fully testable end-to-end today with zero network.
+ *
+ * The VC2B emit seam (task 5) is wired: producing a VectorSetV1 emits
+ * `vector_cortex_encoder_heads_emitted`; selecting a mode B/C fallback emits
+ * `vector_cortex_encoder_fallback_selected` — both gated on MEGACOMPACT_VC2B.
+ *
+ * Pi-agnostic, zero network (PREVENT-PI-004), no `any` (PREVENT-011).
+ */
+import { ENCODER_HEAD_DIMS, ENCODER_HEAD_ORDER, ENCODER_HEAD_LOSS_WEIGHTS, ENCODER_HEAD_LOSS_SUM, ENCODER_SEED, } from "./types.js";
+import { createEncoderHeadsReporter, NOOP_VC2B_REPORTER, } from "./emit-vc2b.js";
+/** The stable head index of a head name (its position in ENCODER_HEAD_ORDER). */
+const HEAD_INDEX = {
+    semantic: 0,
+    dependency: 1,
+    contradiction: 2,
+    cacheStability: 3,
+    payloadRouting: 4,
+};
+/** Deterministic 32-bit LCG step (matches runtime.ts projectSemantic). */
+function nextState(state) {
+    return (state * 1664525 + 1013904223) >>> 0;
+}
+/**
+ * L2-normalize a float vector in place semantics (returns a new Float32Array).
+ * A zero-norm (or empty) input maps to an all-zero vector of the same length
+ * (task 2: "mapping zero norm to an all-zero vector"). All finite.
+ */
+export function l2Normalize(values) {
+    const out = new Float32Array(values.length);
+    let sum = 0;
+    for (const v of values)
+        sum += v * v;
+    const norm = Math.sqrt(sum);
+    if (!(norm > 0))
+        return out; // zero norm -> all-zero
+    for (let i = 0; i < values.length; i++)
+        out[i] = values[i] / norm;
+    return out;
+}
+/** L2 norm of a Float32Array (0 for empty/all-zero). */
+export function l2Norm(values) {
+    let sum = 0;
+    for (const v of values)
+        sum += v * v;
+    return Math.sqrt(sum);
+}
+/** A deterministic per-head projection over the token sequence, pre-normalization. */
+function projectRaw(head, tokens, seed) {
+    const dim = ENCODER_HEAD_DIMS[head];
+    const out = new Float32Array(dim);
+    // An EMPTY token sequence has no signal: the raw projection is the zero vector,
+    // so after L2 normalization it maps to the all-zero vector (task 2: "mapping
+    // zero norm to an all-zero vector"; ENC-ZERO-002). This keeps empty input
+    // finite and zero-norm instead of seeding spurious unit-norm noise.
+    if (tokens.length === 0)
+        return out;
+    // Mix the stable head index + ENCODER_SEED + seed into a per-head state so
+    // each head is a distinct independent projection (failure-triad independence).
+    let state = (((ENCODER_SEED ^ HEAD_INDEX[head]) >>> 0) ^ (seed >>> 0)) ^ 0x9e3779b9;
+    for (const t of tokens)
+        state = nextState(state ^ ((t >>> 0) * 2654435761));
+    for (let i = 0; i < dim; i++) {
+        state = nextState(state ^ seed);
+        out[i] = (state / 4294967296) * 2 - 1;
+    }
+    return out;
+}
+/**
+ * Compute one head's L2-normalized vector (all-zero on zero norm) for a token
+ * sequence. Deterministic for a given seed (repeat drift == 0).
+ */
+export function projectHead(head, tokens, seed = ENCODER_SEED) {
+    const raw = projectRaw(head, tokens, seed);
+    return { head, dim: ENCODER_HEAD_DIMS[head], values: l2Normalize(raw) };
+}
+/**
+ * Encode a token sequence into a `VectorSetV1`: the five heads in stable order,
+ * each L2-normalized (all-zero on zero norm). Emits `heads_emitted` via the
+ * reporter (non-fatal, flag-gated). Deterministic for a given seed.
+ */
+export function encodeVectorSet(tokens, options = {}) {
+    const seed = options.seed ?? ENCODER_SEED;
+    const reporter = options.reporter ?? createEncoderHeadsReporter();
+    const heads = ENCODER_HEAD_ORDER.map((h) => projectHead(h, tokens, seed));
+    reporter.headsEmitted({
+        heads: heads.length,
+        dims: heads.map((h) => h.dim).join("/"),
+        normalized: true,
+        tokens: tokens.length,
+    });
+    return { schema: "vector-set-v1", inputTokens: [...tokens], heads, normalized: true };
+}
+/**
+ * The per-head loss weights (must sum to ENCODER_HEAD_LOSS_SUM exactly).
+ * Exposed for training/tests to assert the normative .35/.20/.20/.15/.10 split.
+ */
+export function headLossWeights() {
+    return { ...ENCODER_HEAD_LOSS_WEIGHTS };
+}
+export { ENCODER_HEAD_ORDER, ENCODER_HEAD_DIMS, ENCODER_HEAD_LOSS_SUM, ENCODER_SEED, NOOP_VC2B_REPORTER };