pi-mega-compact 0.17.1 → 0.18.1

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

@@ -0,0 +1,104 @@
+/**
+ * vector-cortex/encoder/lexical.ts — VC2B mode C: token/phrase lexical encoder.
+ *
+ * Lexical C is a token/phrase lexical feature generator used when both mode A
+ * (learned asset) and mode B (trigram) are unavailable or fail. It is continuity,
+ * NOT semantic completeness: C operates on exact current tokens/phrases only and
+ * MUST state that it has lost old semantic context (task 4 + TRIAD_RESILIENCE.
+ * "C is continuity, not semantic completeness: it may omit old context and must
+ * report that limitation").
+ *
+ * C never imports the learned asset or learned calibration (task 4): it is a
+ * pure token/phrase lexical projection (token counts + phrase hashes) computed
+ * from the exact input. It is independently implemented from B (which hashes
+ * byte-level trigrams) — C works at the token/phrase level, B at the byte-ngram
+ * level, so the two share no algorithm.
+ *
+ * Authority outage freezes derived high-water: C never advances any derived
+ * frontier; it is purely a local reconstruction from the exact present tokens.
+ *
+ * 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 lexical C. */
+export const ENCODER_LEXICAL_WIDTH = 256;
+/** The documented limitation lexical C reports (continuity, not semantics). */
+export const ENCODER_LEXICAL_LIMITATION = "lexical C: token/phrase-level continuity only; old semantic context is omitted";
+export function tokenizeLexical(text) {
+    // Split into lowercase token/phrase units on non-alphanumeric boundaries.
+    return text.toLowerCase().split(/[^a-z0-9_]+/).filter((t) => t.length > 0);
+}
+/**
+ * Normalize a single token/phrase unit the same way `tokenizeLexical` does for
+ * the string form: lowercase and strip leading/trailing non-alphanumeric runs.
+ * Applied to array-form tokens so both accepted input forms of `embedLexical`
+ * hash identical conceptual content to identical buckets, regardless of how the
+ * caller chose to pass it (code-review Q03).
+ */
+function normalizeToken(tok) {
+    return tok.toLowerCase().split(/[^a-z0-9_]+/).join("");
+}
+/**
+ * Resolve either accepted input form to a normalized token sequence: the string
+ * form routes through `tokenizeLexical` (split on non-alphanumeric boundaries,
+ * fragmented tokens discarded); the array form applies the same per-token
+ * lowercase/strip normalization and drops tokens that normalize to empty. Both
+ * paths therefore agree on hash buckets for the same conceptual content.
+ */
+function resolveTokens(tokensOrText) {
+    if (typeof tokensOrText === "string")
+        return tokenizeLexical(tokensOrText);
+    const out = [];
+    for (const raw of tokensOrText) {
+        const norm = normalizeToken(raw);
+        if (norm.length > 0)
+            out.push(norm);
+    }
+    return out;
+}
+/**
+ * Encode a token/phrase sequence into a 256-dim L2-normalized lexical vector
+ * (all-zero on empty input). Features: exact token count + token id-hash sums +
+ * phrase-adjacency hashes. Deterministic (repeat drift == 0). Pure local compute.
+ */
+export function embedLexical(tokensOrText) {
+    const width = ENCODER_LEXICAL_WIDTH;
+    const out = new Float32Array(width);
+    const tokens = resolveTokens(tokensOrText);
+    for (let i = 0; i < tokens.length; i++) {
+        const tok = tokens[i];
+        const h = createHash("sha256").update(`t:${tok}`).digest();
+        const bucket = h.readUInt32BE(0) % width;
+        const weight = (h.readUInt32BE(4) / 4294967295) * 2 - 1;
+        out[bucket] += weight;
+        // Phrase adjacency: bigram hash blended in so semantic-free ordering matters.
+        if (i > 0) {
+            const pair = createHash("sha256").update(`p:${tokens[i - 1]}:${tok}`).digest();
+            const pb = pair.readUInt32BE(0) % width;
+            const pw = (pair.readUInt32BE(4) / 4294967295) * 2 - 1;
+            out[pb] += pw;
+        }
+    }
+    return l2Normalize(out);
+}
+/** The documented limitation string, surfaced when lexical C is selected. */
+export function selectLexicalC(options = {}) {
+    const reporter = options.reporter ?? createEncoderHeadsReporter();
+    const selection = {
+        ok: true,
+        mode: "C",
+        dim: ENCODER_LEXICAL_WIDTH,
+        width: ENCODER_LEXICAL_WIDTH,
+        limitation: ENCODER_LEXICAL_LIMITATION,
+    };
+    reporter.fallbackSelected({
+        mode: selection.mode,
+        dim: selection.dim,
+        width: selection.width,
+        limitation: selection.limitation,
+    });
+    return selection;
+}
+export { l2Normalize };

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

@@ -0,0 +1,115 @@
+/**
+ * vector-cortex/encoder/router.ts — VC2B encode-or-fallback router (S2).
+ *
+ * The single production seam that catches a real VC2A `load()` failure and
+ * hands off to the independently initialized VC2B fallbacks:
+ *
+ *   - mode A: the learned multi-head projection — an EncoderRuntime that
+ *     verifies+loads a local qualified ONNX (VC2A) plus `encodeVectorSet`
+ *     producing VectorSetV1, emitting `vector_cortex_encoder_heads_emitted`;
+ *   - when that `load()` fails for ANY reason (removed model ->
+ *     ENC_ASSET_UNREADABLE, digest mismatch -> ENC_DIGEST_MISMATCH, missing
+ *     manifest -> ENC_MANIFEST_INVALID, unsupported platform, RSS over budget,
+ *     ...) the router catches the real failure and selects the independently
+ *     initialized asset-free trigram B (`selectTrigramBFallback`) — or lexical C
+ *     when the runtime reports B-unavailable / the caller forces mode C. The
+ *     `vector_cortex_encoder_fallback_selected` event fires from the REAL
+ *     producer seam, not test wiring (task 5 + code-review S1).
+ *
+ * Best-effort and non-fatal: every branch returns an explicit verdict and never
+ * throws across the boundary into the agent loop. Flag-OFF parity: with
+ * `MEGACOMPACT_VC2B=0` the VC2B reporter is a no-op (zero emissions), and the
+ * mode-A path is governed by the VC2A runtime itself — the router only adds the
+ * fallback handoff and changes no producer bytes.
+ *
+ * Pi-agnostic, zero network (PREVENT-PI-004), no `any` (PREVENT-011).
+ */
+import { createEncoderRuntime } from "./runtime.js";
+import { encodeVectorSet } from "./heads.js";
+import { embedTrigram512, selectTrigramBFallback } from "./trigram.js";
+import { embedLexical, selectLexicalC } from "./lexical.js";
+import { createEncoderHeadsReporter } from "./emit-vc2b.js";
+import { ENC_FAIL, } from "./types.js";
+/** Deterministic text derived from an int token sequence so the asset-free
+ *  fallback producers operate on the same authority the learned path encoded. */
+function textFromTokens(tokens) {
+    return tokens.join("-");
+}
+/**
+ * Try to produce an encoding for an input token sequence. Mode A: verify+load
+ * the local learned asset via the EncoderRuntime; on a real `load()` failure
+ * (removed model, digest mismatch, missing manifest, ...) the router catches it
+ * and selects the independently initialized trigram B or lexical C, emitting
+ * `vector_cortex_encoder_fallback_selected` (and `heads_emitted` when A wins).
+ */
+export function encodeOrFallback(input, assetDir, options = {}) {
+    const reporter = options.reporter ?? createEncoderHeadsReporter();
+    const tokens = Array.isArray(input?.tokens) ? input.tokens : [];
+    const runtime = options.runtime ?? createEncoderRuntime();
+    // A caller that explicitly forces a fallback mode wins even over the empty-
+    // input degenerate case: the forced-mode contract must hold for ANY input, so
+    // handle forceFallback before the empty-input selection below (Q02). A forced
+    // B/C is NOT a rollback and NOT a demotion — it is an intentional selection
+    // for capacity/testing — so the verdict carries `code: null`, never
+    // ENC_FAIL.ROLLBACK, so a consumer that reads `verdict.code` as "what
+    // triggered the fallback" won't misread a deliberately forced mode as a
+    // rollback and take rollback-specific action (code-review Q04).
+    if (options.forceFallback !== undefined) {
+        if (options.forceFallback === "C") {
+            const sel = selectLexicalC({ reporter });
+            const vector = embedLexical(textFromTokens(tokens));
+            return { ok: true, mode: "C", vector, width: sel.width, limitation: sel.limitation, code: null };
+        }
+        const sel = selectTrigramBFallback({ reporter });
+        const vector = embedTrigram512(textFromTokens(tokens));
+        return { ok: true, mode: "B", vector, width: sel.width, limitation: null, code: null };
+    }
+    // Empty input yields the asset-free fallback (finite, deterministic zero
+    // vector, ENC-ZERO-002). This is legitimate degenerate behavior, NOT a shape
+    // failure, so the verdict carries no failure code (code === null) — a consumer
+    // that interprets `code` as "what went wrong" must not misread a valid all-zero
+    // B vector as a shape rejection (Q05). Only reached when no mode is forced, so
+    // a forced B/C is never subverted by empty tokens.
+    if (tokens.length === 0) {
+        const sel = selectTrigramBFallback({ reporter });
+        const vector = embedTrigram512("");
+        return { ok: true, mode: "B", vector, width: sel.width, limitation: null, code: null };
+    }
+    // Mode A attempt — reached only when no mode is forced (the forceFallback
+    // guard at the top already returned), so `options.forceFallback` is undefined here.
+    const loaded = runtime.load(assetDir);
+    if (!loaded.ok) {
+        return fallbackFromLoad(loaded, reporter, tokens);
+    }
+    // Q01/Q03: a qualified mode-A load is not enough — the verified per-manifest
+    // token capacity (maxTokens, <= global 512) must also be enforced before we
+    // produce a VectorSetV1. run inference over the input; an over-cap sequence
+    // (e.g. 100 tokens against a verified maxTokens=64 manifest) or an
+    // over-budget inference is rejected here and routed to the B/C fallback with
+    // the real failure code, rather than silently emitting an ok:true mode-A
+    // set whose inputTokens breach the model's declared capacity.
+    const inferred = runtime.infer({ tokens });
+    if (!inferred.ok) {
+        return fallbackFromLoad({ ok: false, mode: "B", code: inferred.code }, reporter, tokens);
+    }
+    const vectorSet = encodeVectorSet(tokens, { reporter, seed: options.seed });
+    return { ok: true, mode: "A", vectorSet, code: null };
+}
+/**
+ * Catch a (real or forced) A load failure (ok === false only) and select the
+ * B/C fallback that emits `vector_cortex_encoder_fallback_selected` from the
+ * production seam. The parameter is narrowed to the failed-load variant because
+ * the router hands off here only on a non-A/failed load — a qualified mode-A
+ * success never reaches this function (Q04).
+ */
+function fallbackFromLoad(loaded, reporter, tokens) {
+    if (loaded.mode === "C") {
+        const sel = selectLexicalC({ reporter });
+        const vector = embedLexical(textFromTokens(tokens));
+        return { ok: true, mode: "C", vector, width: sel.width, limitation: sel.limitation, code: loaded.code };
+    }
+    const sel = selectTrigramBFallback({ reporter });
+    const vector = embedTrigram512(textFromTokens(tokens));
+    return { ok: true, mode: "B", vector, width: sel.width, limitation: null, code: loaded.code };
+}
+export { ENC_FAIL };

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

@@ -0,0 +1,228 @@
+/**
+ * vector-cortex/encoder/runtime.ts — VC2A EncoderRuntime (task 3).
+ *
+ * Allocates (prepares an inference session) ONLY after manifest verification;
+ * rejects any non (batch 1, tokens <= maxTokens, <=512) input with
+ * ENC_SHAPE_INVALID; caps the encoder's MARGINAL footprint at 150 MiB
+ * (ENC_RSS_BUDGET_EXCEEDED -> mode B); and yields a deterministic mode-A
+ * inference over the verified asset (the trained weights are substituted in
+ * VC2C — the contract, shape gating and budgets all land here).
+ *
+ * MEMORY BUDGET (Q01/Q02): the 150 MiB cap measures the encoder's INCREMENTAL
+ * footprint — an in-process allocation counter (`selfAllocated`) plus any
+ * externally staged asset working set (`host.allocatedBytes()`) — NOT the
+ * whole-process RSS. In a live pi extension the process baseline (node:sqlite
+ * DatabaseSync + dashboard + loaded context) routinely exceeds 150 MiB, so an
+ * absolute-RSS cap would permanently demote a qualified asset to mode B and
+ * make mode A unreachable in production. Bounding the marginal footprint keeps
+ * mode A reachable while still enforcing the budget. `selfAllocated` models a
+ * single REUSABLE 384-float projection buffer (first inference allocates it,
+ * every later inference reuses it), so it is capped at `SEMANTIC_BUFFER_BYTES`
+ * — the marginal footprint can never grow without bound (Q01), and a long-lived
+ * runtime cannot drift over budget from healthy operation. The check runs
+ * BEFORE the allocation on both the load and the inference path
+ * (cap-before-allocation, task 3), and an over-budget inference demotes the
+ * runtime to mode B just as an over-budget load does (consistent demotion per
+ * ENC_FAIL.RSS_BUDGET_EXCEEDED).
+ *
+ * TOKEN CAPACITY (Q03): the per-manifest `maxTokens` (<= 512) is stored at load
+ * and enforced at inference — an input longer than the verified manifest's
+ * declared capacity is rejected with ENC_SHAPE_INVALID, honoring the model
+ * contract rather than a global 512 ceiling.
+ *
+ * FLAG GATING (Q04): the default factory consults `MEGACOMPACT_VC2A`; when the
+ * flag is OFF the runtime is fixed at mode C (rollback, byte-identical to the
+ * predecessor — no asset is read or verified). `forcedMode: "C"` is the
+ * explicit override for the same rollback path.
+ *
+ * Triad: A = qualified local ONNX (verified); B = asset-free trigram (forced by
+ * a missing/unsupported/digest-bad asset, no remote fetch); C = lexical forced
+ * when A verification fails AND B initialization itself fails. Demotions always
+ * select B/C locally and never attempt a network fetch (PREVENT-PI-004).
+ *
+ * Pi-agnostic. No `any` (PREVENT-011). Emits the two VC2A events via the
+ * reporter (non-fatal).
+ */
+import { detectPlatform, readEncoderManifest, verifyEncoderAsset, } from "./asset.js";
+import { createEncoderReporter } from "./emit.js";
+import { VC2A_ENABLED } from "../../config/vector-cortex.js";
+import { ENC_FAIL, ENCODER_MAX_TOKENS, ENCODER_RSS_BUDGET_BYTES, ENCODER_SEMANTIC_WIDTH, } from "./types.js";
+/** Bytes a single encoder-owned projection buffer commits to the marginal
+ *  footprint (Float32Array, 4 bytes per element). */
+const SEMANTIC_BUFFER_BYTES = ENCODER_SEMANTIC_WIDTH * 4;
+const DEFAULT_HOST = {
+    allocatedBytes: () => 0,
+    allocatorFails: () => false,
+    nowMs: () => Date.now(),
+};
+function mergeHost(partial) {
+    return { ...DEFAULT_HOST, ...partial };
+}
+/** A deterministic seeded projection so the mode-A inference path is testable
+ *  end-to-end without onnxruntime (real weights + execution are VC2C). */
+function projectSemantic(seed, n) {
+    const out = new Float32Array(n);
+    let state = (seed >>> 0) ^ 0x9e3779b9;
+    let sum = 0;
+    for (let i = 0; i < n; i++) {
+        state = (state * 1664525 + 1013904223) >>> 0;
+        out[i] = (state / 4294967296) * 2 - 1;
+        sum += out[i] * out[i];
+    }
+    const norm = Math.sqrt(sum) || 1;
+    for (let i = 0; i < n; i++)
+        out[i] = out[i] / norm;
+    return out;
+}
+/** Deterministic token seed derived from the verified asset bytes count. */
+function seedFromBytes(embeddedBytes) {
+    return (embeddedBytes * 2654435761) >>> 0;
+}
+function modeLabel(mode) {
+    return mode === "A" ? "qualified-onnx" : mode === "B" ? "trigram" : "lexical";
+}
+export function createEncoderRuntime(options = {}) {
+    const reporter = options.reporter ?? createEncoderReporter();
+    const host = mergeHost(options.host);
+    const forced = options.forcedMode;
+    const plat = options.platform ?? detectPlatform;
+    // Q04: rollback contract — MEGACOMPACT_VC2A=0 selects mode C (byte-identical
+    // to the predecessor: no asset read/verify, no learned infer). An explicit
+    // forcedMode "C" takes precedence; otherwise the flag gates the default.
+    const rolledBack = forced === "C" || !VC2A_ENABLED();
+    let mode = rolledBack ? "C" : "C";
+    let embeddedBytes = 0;
+    let verified = false;
+    /** Per-manifest token capacity (<= 512) from the verified asset; enforced at
+     *  inference (Q03). Defaults to the global ceiling before a load. */
+    let maxTokens = ENCODER_MAX_TOKENS;
+    /** Bytes this runtime itself has allocated. This models a SINGLE reusable
+     *  384-float projection buffer: the first inference allocates it (1536
+     *  bytes), every later inference reuses it, so the counter is capped at
+     *  `SEMANTIC_BUFFER_BYTES` and never grows without bound (Q01). Combined
+     *  with `host.allocatedBytes()` it drives the 150 MiB marginal budget (Q02),
+     *  never whole-process RSS. */
+    let selfAllocated = 0;
+    /** The encoder's marginal working-set footprint, in bytes. */
+    const footprint = () => selfAllocated + host.allocatedBytes();
+    const demoteTo = (rmode, code) => {
+        mode = rmode;
+        verified = false;
+        reporter.runtimeDemoted({ reason: code, mode: rmode, platform: plat()?.toString() ?? "unsupported" });
+    };
+    const runtime = {
+        schema: "encoder-runtime-v1",
+        // Live getter so `mode` always reflects the latest load/demote outcome
+        // (a plain property would freeze at its construction-time value forever).
+        get mode() {
+            return mode;
+        },
+        load(assetDir) {
+            if (rolledBack) {
+                // Rollback path (forcedMode "C" or MEGACOMPACT_VC2A=0): mode C restores
+                // the prior derived pointer; no asset is read or verified; no emission.
+                // Q04: report the rollback with its own code, not MANIFEST_INVALID, so a
+                // correctly-shaped, digest-correct asset is not mis-read as corrupted.
+                mode = "C";
+                verified = false;
+                return { ok: false, mode: "C", code: ENC_FAIL.ROLLBACK };
+            }
+            // Attempt A: verify the local qualified ONNX asset (never a remote fetch).
+            const manifest = readEncoderManifest(assetDir);
+            let verify;
+            if (manifest === null) {
+                verify = { ok: false, code: ENC_FAIL.MANIFEST_INVALID };
+            }
+            else {
+                verify = verifyEncoderAsset(assetDir, manifest, plat());
+            }
+            if (!verify.ok) {
+                // A failed -> B, unless B init itself fails (allocator) -> C.
+                if (host.allocatorFails()) {
+                    demoteTo("C", ENC_FAIL.ASSET_UNREADABLE);
+                    return { ok: false, mode: "C", code: ENC_FAIL.ASSET_UNREADABLE };
+                }
+                demoteTo("B", verify.code);
+                return { ok: false, mode: "B", code: verify.code };
+            }
+            // Allocate only after verification (task 3). Simulate allocator failure.
+            if (host.allocatorFails()) {
+                demoteTo("B", ENC_FAIL.ASSET_UNREADABLE);
+                return { ok: false, mode: "B", code: ENC_FAIL.ASSET_UNREADABLE };
+            }
+            // Cap the encoder's MARGINAL footprint at 150 MiB (task 3, Q01). This
+            // bounds the encoder's incremental allocation, so a healthy process with
+            // a large baseline RSS still reaches mode A.
+            if (footprint() > ENCODER_RSS_BUDGET_BYTES) {
+                demoteTo("B", ENC_FAIL.RSS_BUDGET_EXCEEDED);
+                return { ok: false, mode: "B", code: ENC_FAIL.RSS_BUDGET_EXCEEDED };
+            }
+            embeddedBytes = verify.embeddedBytes;
+            // Q03: record the verified manifest's token capacity so inference can
+            // enforce the model's declared maximum, not just the global 512 ceiling.
+            maxTokens = verify.maxTokens;
+            verified = true;
+            mode = "A";
+            reporter.assetVerified({
+                mode: "A",
+                embeddedBytes: verify.embeddedBytes,
+                onnxDigest: verify.onnxDigest.slice(0, 12),
+            });
+            return {
+                ok: true,
+                mode: "A",
+                embeddedBytes: verify.embeddedBytes,
+                rssBytes: footprint(),
+                sessionId: `enc-${seedFromBytes(verify.embeddedBytes).toString(16)}`,
+            };
+        },
+        infer(input) {
+            if (!verified || mode !== "A") {
+                // Only batch1/max512 verified assets reach inference (mode B/C do not).
+                return {
+                    ok: false,
+                    code: ENC_FAIL.SHAPE_INVALID,
+                    shapeError: "no verified learned asset; mode is " + modeLabel(mode),
+                };
+            }
+            if (!input || !Array.isArray(input.tokens)) {
+                return { ok: false, code: ENC_FAIL.SHAPE_INVALID, shapeError: "missing tokens array" };
+            }
+            const n = input.tokens.length;
+            // Q03: enforce the per-manifest maxTokens (<= global 512 ceiling), so an
+            // over-cap request against a low-cap verified asset is rejected rather
+            // than silently exceeding the model's declared capacity.
+            if (n < 1 || n > maxTokens) {
+                return {
+                    ok: false,
+                    code: ENC_FAIL.SHAPE_INVALID,
+                    shapeError: `token count ${n} outside 1..${maxTokens} (manifest cap)`,
+                };
+            }
+            // Q03: cap-before-allocation on the inference path too. Check the
+            // marginal footprint BEFORE allocating the projection buffer; an
+            // over-budget inference demotes to mode B consistently with load() (the
+            // ENC_FAIL.RSS_BUDGET_EXCEEDED model: "measured RSS over 150 MiB -> B"),
+            // so a subsequent infer no longer attempts allocation in a stale mode A.
+            if (footprint() > ENCODER_RSS_BUDGET_BYTES) {
+                demoteTo("B", ENC_FAIL.RSS_BUDGET_EXCEEDED);
+                return {
+                    ok: false,
+                    code: ENC_FAIL.RSS_BUDGET_EXCEEDED,
+                    shapeError: "encoder footprint over budget during inference",
+                };
+            }
+            const start = host.nowMs();
+            // Batch is always 1 (single request); shape is (1, n) for n in 1..maxTokens.
+            const semantic = projectSemantic(seedFromBytes(embeddedBytes) ^ n, ENCODER_SEMANTIC_WIDTH);
+            // Q01: the projection buffer is a single reusable 384-float array; the
+            // marginal footprint is a fixed SEMANTIC_BUFFER_BYTES once it exists, so
+            // selfAllocated is SET (never accumulated) — bounded regardless of how
+            // many inferences run on a long-lived runtime.
+            selfAllocated = SEMANTIC_BUFFER_BYTES;
+            const latencyMs = host.nowMs() - start;
+            return { ok: true, semantic, rssBytes: footprint(), latencyMs, shapeError: null };
+        },
+    };
+    return runtime;
+}

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

@@ -0,0 +1,138 @@
+/**
+ * vector-cortex/encoder/types.ts — VC2A contract (ModelManifestV1 /
+ * EncoderRuntime).
+ *
+ * The offline encoder runtime owns the learned-asset path (triad mode A: a
+ * qualified local ONNX). MODEL_ASSET.md is the normative target. This
+ * sprint (VC2A) ships the manifest + verification + shaped-inference contract;
+ * the trained weights are packaged in VC2C (MODEL_ASSET: "package.json changes
+ * occur only in VC2C"), but the verification, digest-before-load, platform
+ * demotion, shape rejection and RSS/latency budget all land here so a later
+ * sprint only substitutes real weights.
+ *
+ * Contract-first (ENGINEERING_PRACTICES §3): this types file is the reviewed
+ * gate; implementations import from it; consumers import only types + factory.
+ *
+ * Pi-agnostic and dependency-free (PREVENT-PI-004 — local assets only, the
+ * runtime never fetches). No `any` (PREVENT-011).
+ */
+/** Supported matrix from MODEL_ASSET.md §qualification. */
+export const ENCODER_SUPPORTED_PLATFORMS = [
+    "linux-x64",
+    "linux-arm64",
+    "darwin-x64",
+    "darwin-arm64",
+    "win32-x64",
+];
+/** ONNX opset required by the normative v1 target (opset 17). */
+export const ENCODER_OPSET = 17;
+/** Batch must be exactly 1 (single-request inference). */
+export const ENCODER_BATCH = 1;
+/** Maximum accepted token count (WordPiece, deterministic truncation). */
+export const ENCODER_MAX_TOKENS = 512;
+/** Caps the encoder's MARGINAL footprint (bytes) at 150 MiB (MODEL_ASSET
+ *  §qualification). The budget bounds the encoder's own incremental allocation
+ *  (a reusable projection buffer + any externally staged asset working set),
+ *  NOT the whole-process RSS — in a live pi extension the process baseline
+ *  routinely exceeds 150 MiB, so measuring absolute RSS would make mode A
+ *  unreachable in production. This is the "RSS" figure the acceptance metric
+ *  and ENC_FAIL.RSS_BUDGET_EXCEEDED refer to: it is the encoder's marginal
+ *  footprint, never the process RSS (code-review Q01/Q02). */
+export const ENCODER_RSS_BUDGET_BYTES = 150 * 1024 * 1024;
+/** p95 inference budget in milliseconds (MODEL_ASSET §qualification). */
+export const ENCODER_LATENCY_P95_MS = 40;
+/** Semantic projection head width (MODEL_ASSET: 384 float32 L2-normalized). */
+export const ENCODER_SEMANTIC_WIDTH = 384;
+/** Exact VC2A failure codes (returned, never thrown across the boundary). */
+export const ENC_FAIL = {
+    /** opset != 17. */
+    OPSET_INVALID: "ENC_OPSET_INVALID",
+    /** batch != 1. */
+    BATCH_INVALID: "ENC_BATCH_INVALID",
+    /** maxTokens > 512. */
+    TOKENS_EXCEEDED: "ENC_TOKENS_EXCEEDED",
+    /** input token count > declared maxTokens / 512, or not batch 1. */
+    SHAPE_INVALID: "ENC_SHAPE_INVALID",
+    /** asset file unreadable (truncated during digest read, allocator failure). */
+    ASSET_UNREADABLE: "ENC_ASSET_UNREADABLE",
+    /** on-disk digest does not match the manifest (one-byte mutation). */
+    DIGEST_MISMATCH: "ENC_DIGEST_MISMATCH",
+    /** platform not in the supported matrix (selects trigram B). */
+    PLATFORM_UNSUPPORTED: "ENC_PLATFORM_UNSUPPORTED",
+    /** manifest missing/invalid (selects trigram B). */
+    MANIFEST_INVALID: "ENC_MANIFEST_INVALID",
+    /** encoder MARGINAL footprint over the 150 MiB budget (selects trigram B).
+     *  This is the encoder's own incremental allocation (a reusable projection
+     *  buffer + any externally staged asset working set), NOT whole-process RSS
+     *  — see ENCODER_RSS_BUDGET_BYTES. */
+    RSS_BUDGET_EXCEEDED: "ENC_RSS_BUDGET_EXCEEDED",
+    /** mode C forced by the rollback path (MEGACOMPACT_VC2A=0 / forcedMode "C").
+     *  Distinct from MANIFEST_INVALID so a non-corrupt, correctly-shaped asset
+     *  present on disk is not mis-reported as "manifest invalid" when the runtime
+     *  is simply rolled back to the predecessor path (code-review Q04). */
+    ROLLBACK: "ENC_ROLLBACK_ACTIVE",
+};
+/** The 8 registered VC2A conformance IDs (task 1: "register ENC-001..008"). */
+export const ENC_IDS = [
+    "ENC-001",
+    "ENC-002",
+    "ENC-003",
+    "ENC-004",
+    "ENC-005",
+    "ENC-006",
+    "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",
+];