pi-mega-compact 0.20.70 → 0.20.72

package/extensions/dashboard-server/routes-vector-cortex-economics.ts

@@ -28,18 +28,25 @@ import type { RouteContext } from "./routes-core.js";
 import { VC7B_ENABLED } from "../../src/config.js";
 import { sendJson } from "./routes-vector-cortex-shared.js";
 import { deriveVcStatus } from "./vc-status.js";
+import { readLivewireSnapshot } from "../../src/vector-cortex/livewire/livewire-registry.js";
 import type { VectorCortexEconomicsView } from "./api-contracts/vector-cortex-economics.js";
 
+// The flag-off deferred reason, kept byte-identical to the pre-LIVEWIRE
+// predecessor (VC7B: `economics_not_computed_v0_20_23`).
+const DEFERRED_REASON = "economics_not_computed_v0_20_23";
+
 /**
  * Reader-only GET /api/vector-cortex/cache-economics (VC7B).
  *
- * Counts, profile tallies, and ECON_* codes only — a static reader-only
- * aggregate seam with the same shape as the VC7A crystals handler.
+ * Counts, profile tallies, and ECON_* codes only. With the flag ON it surfaces
+ * the LIVEWIRE-provided profile economics tallies + the runtime `computed` bit;
+ * with the flag OFF it returns the byte-identical legacy deferred view (mode C,
+ * deferredReason present).
  */
 export function handleVectorCortexEconomics(
   req: IncomingMessage,
   res: ServerResponse,
-  _ctx: RouteContext,
+  ctx: RouteContext,
 ): boolean {
   const url = req.url ?? "";
   const path = url.split("?")[0] ?? url;
@@ -50,25 +57,33 @@ export function handleVectorCortexEconomics(
   }
 
   const enabled = VC7B_ENABLED();
-  // Flag-off routes to mode C: with VC7B off no cache economics are served, which
-  // is exactly the spec's "economics bypass" outcome. Reporting A (cached render
-  // priced) or B (fresh render priced) would imply an economics path that is not
-  // wired at all. Mirrors how VC6C/VC7A OFF views report the mode they take.
-  const mode: "A" | "B" | "C" = enabled ? "A" : "C";
+  if (!enabled) {
+    // Flag-off parity: byte-identical to the predecessor (mode C + deferred).
+    const body: VectorCortexEconomicsView = {
+      enabled: false,
+      mode: "C",
+      profileCount: 0,
+      provenExclusions: 0,
+      unprovenExclusions: 0,
+      lastFailure: null,
+      updatedAt: new Date().toISOString(),
+      deferredReason: DEFERRED_REASON,
+      status: deriveVcStatus({ enabled: false, hasData: false }),
+    };
+    sendJson(res, 200, body);
+    return true;
+  }
+
+  const econ = readLivewireSnapshot(ctx.stateDir).economics;
   const body: VectorCortexEconomicsView = {
-    enabled,
-    mode,
-    profileCount: 0,
-    provenExclusions: 0,
-    unprovenExclusions: 0,
-    lastFailure: null,
+    enabled: true,
+    mode: "A",
+    profileCount: econ.profileCount,
+    provenExclusions: econ.provenExclusions,
+    unprovenExclusions: econ.unprovenExclusions,
+    lastFailure: econ.lastFailure,
     updatedAt: new Date().toISOString(),
-    deferredReason: "economics_not_computed_v0_20_23",
-    status: deriveVcStatus({
-      enabled,
-      deferredReason: "economics_not_computed_v0_20_23",
-      hasData: false,
-    }),
+    status: deriveVcStatus({ enabled: true, hasData: econ.computed }),
   };
   sendJson(res, 200, body);
   return true;

package/extensions/dashboard-server/routes-vector-cortex-helpers.ts

@@ -108,3 +108,87 @@ export async function seedEval(dir: string): Promise<void> {
     { session: "s1", seq: 4, event: "encode", value: 12, unit: "ms", mode: "A" },
   ]);
 }
+
+/**
+ * Seed a LIVEWIRE aggregate snapshot directly on disk so the spawned dashboard
+ * server (a separate process) rehydrates real subsystem counts and its reader-only
+ * routes report LIVE (non-deferred) data. This is how the LIVEWIRE route tests
+ * prove the wiring: the runtimes' persisted counts are exactly what a reader
+ * process reconstructs. Counts + codes + triad mode only (SECURITY_PRIVACY).
+ */
+export async function seedLivewireSnapshot(
+  dir: string,
+  overrides?: {
+    crystals?: Partial<Record<
+      "crystalCount" | "totalBytes" | "hits" | "misses" | "hitBytes" |
+      "writes" | "duplicateWrites" | "collisions",
+      number
+    >>;
+    diagnostics?: Partial<Record<
+      "profileMisses" | "rangeMisses" | "dependencyMisses" | "requestMisses" |
+      "generationMisses" | "unknownMisses" | "serveBlocked",
+      number
+    >> & { breakerState?: string };
+    economics?: {
+      computed?: boolean;
+      profileCount?: number;
+      provenExclusions?: number;
+      unprovenExclusions?: number;
+    };
+    policy?: {
+      shadowDecisions?: number;
+      clampedDecisions?: number;
+      rejectedInputs?: number;
+      pressureVersion?: number;
+      liveMutations?: number;
+    };
+  },
+): Promise<void> {
+  const { saveLivewireSnapshot } = await import(
+    "../../src/vector-cortex/livewire/livewire-snapshot.js"
+  );
+  const c = overrides?.crystals ?? {};
+  const d = overrides?.diagnostics ?? {};
+  const e = overrides?.economics ?? {};
+  const p = overrides?.policy ?? {};
+  saveLivewireSnapshot(dir, {
+    schema: "vector-cortex-livewire-v1",
+    crystals: {
+      mode: "A",
+      crystalCount: c.crystalCount ?? 3,
+      totalBytes: c.totalBytes ?? 1024,
+      hits: c.hits ?? 0,
+      misses: c.misses ?? 0,
+      hitBytes: c.hitBytes ?? 0,
+      writes: c.writes ?? 0,
+      duplicateWrites: c.duplicateWrites ?? 0,
+      collisions: c.collisions ?? 0,
+    },
+    diagnostics: {
+      profileMisses: d.profileMisses ?? 0,
+      rangeMisses: d.rangeMisses ?? 0,
+      dependencyMisses: d.dependencyMisses ?? 0,
+      requestMisses: d.requestMisses ?? 0,
+      generationMisses: d.generationMisses ?? 0,
+      unknownMisses: d.unknownMisses ?? 0,
+      serveBlocked: d.serveBlocked ?? 0,
+      breakerState: d.breakerState ?? "CLOSED_A",
+      lastFailure: null,
+    },
+    economics: {
+      profileCount: e.profileCount ?? 4,
+      provenExclusions: e.provenExclusions ?? 1,
+      unprovenExclusions: e.unprovenExclusions ?? 0,
+      computed: e.computed ?? true,
+      lastFailure: null,
+    },
+    policy: {
+      shadowDecisions: p.shadowDecisions ?? 0,
+      clampedDecisions: p.clampedDecisions ?? 0,
+      rejectedInputs: p.rejectedInputs ?? 0,
+      liveMutations: p.liveMutations ?? 0,
+      pressureVersion: (p.pressureVersion ?? 1) as 1 | 2,
+      lastFailure: null,
+    },
+  });
+}

package/extensions/dashboard-server/routes-vector-cortex-policy.ts

@@ -21,13 +21,18 @@ import type { RouteContext } from "./routes-core.js";
 import { VC8B_ENABLED } from "../../src/config.js";
 import { sendJson } from "./routes-vector-cortex-shared.js";
 import { deriveVcStatus } from "./vc-status.js";
+import { readLivewireSnapshot } from "../../src/vector-cortex/livewire/livewire-registry.js";
 import type { VectorCortexPolicyView } from "./api-contracts/vector-cortex-policy.js";
 
+// The flag-off deferred reason, kept byte-identical to the pre-LIVEWIRE
+// predecessor (VC8B: `shadow_controller_not_instantiated_v0_20_23`).
+const DEFERRED_REASON = "shadow_controller_not_instantiated_v0_20_23";
+
 /** GET /api/vector-cortex/policy — reader-only policy + shadow aggregate (VC8B). */
 export function handleVectorCortexPolicy(
   req: IncomingMessage,
   res: ServerResponse,
-  _ctx: RouteContext,
+  ctx: RouteContext,
 ): boolean {
   const url = req.url ?? "";
   const path = url.split("?")[0] ?? url;
@@ -39,23 +44,37 @@ export function handleVectorCortexPolicy(
   }
 
   const enabled = VC8B_ENABLED();
-  const mode: VectorCortexPolicyView["mode"] = enabled ? "A" : "C";
+  if (!enabled) {
+    // Flag-off parity: byte-identical to the predecessor (mode C + deferred).
+    const body: VectorCortexPolicyView = {
+      enabled: false,
+      mode: "C",
+      shadowDecisions: 0,
+      clampedDecisions: 0,
+      rejectedInputs: 0,
+      liveMutations: 0,
+      pressureVersion: 1,
+      lastFailure: null,
+      updatedAt: new Date().toISOString(),
+      deferredReason: DEFERRED_REASON,
+      status: deriveVcStatus({ enabled: false, hasData: false }),
+    };
+    sendJson(res, 200, body);
+    return true;
+  }
+
+  const policy = readLivewireSnapshot(ctx.stateDir).policy;
   const body: VectorCortexPolicyView = {
-    enabled,
-    mode,
-    shadowDecisions: 0,
-    clampedDecisions: 0,
-    rejectedInputs: 0,
-    liveMutations: 0,
-    pressureVersion: 1,
-    lastFailure: null,
+    enabled: true,
+    mode: "A",
+    shadowDecisions: policy.shadowDecisions,
+    clampedDecisions: policy.clampedDecisions,
+    rejectedInputs: policy.rejectedInputs,
+    liveMutations: policy.liveMutations,
+    pressureVersion: policy.pressureVersion,
+    lastFailure: policy.lastFailure,
     updatedAt: new Date().toISOString(),
-    deferredReason: "shadow_controller_not_instantiated_v0_20_23",
-    status: deriveVcStatus({
-      enabled,
-      deferredReason: "shadow_controller_not_instantiated_v0_20_23",
-      hasData: false,
-    }),
+    status: deriveVcStatus({ enabled: true, hasData: policy.shadowDecisions > 0 }),
   };
   sendJson(res, 200, body);
   return true;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-mega-compact",
-  "version": "0.20.70",
+  "version": "0.20.72",
   "description": "Layered, local, vector-backed context compressor for pi — supersede/collapse/cluster compaction with deduped inline recall.",
   "type": "module",
   "license": "BSD-3-Clause",

package/src/vector-cortex/cache/store.ts

@@ -78,6 +78,11 @@ export class CrystalStore {
   private writes = 0;
   private duplicateWrites = 0;
   private collisions = 0;
+  /** Restart-survival crystal count/bytes (LIVEWIRE rehydrate). The `committed`
+   *  map is empty in a fresh process; these offsets carry the persisted totals
+   *  so the dashboard does not reset to zero on restart. */
+  private rehydratedCrystalCount = 0;
+  private rehydratedTotalBytes = 0;
 
   /** Freeze a crystal object for a key/bytes pair (digest computed here). */
   static freeze(keyDigest: string, bytes: Uint8Array, key: CrystalV1["key"]): CrystalV1 {
@@ -196,11 +201,11 @@ export class CrystalStore {
 
   /** Reader-only aggregate for the dashboard seam — counts and bytes only. */
   stats(): CrystalStoreStats {
-    let totalBytes = 0;
+    let totalBytes = this.rehydratedTotalBytes;
     for (const c of this.committed.values()) totalBytes += c.byteCount;
     return {
       mode: this.mode(),
-      crystalCount: this.committed.size,
+      crystalCount: this.rehydratedCrystalCount + this.committed.size,
       totalBytes,
       hits: this.hits,
       misses: this.misses,
@@ -210,4 +215,23 @@ export class CrystalStore {
       collisions: this.collisions,
     };
   }
+
+  /**
+   * Restart survival: seed the CUMULATIVE counters from a previously-persisted
+   * aggregate (LIVEWIRE). The in-memory `committed` map is naturally empty in a
+   * fresh process — this only restores the running totals the dashboard reports,
+   * so a process restart does not reset the dashboard to zero. It deliberately
+   * does NOT reconstruct crystals: frozen bytes are never persisted to the
+   * reader aggregate, only the counts.
+   */
+  rehydrate(stats: CrystalStoreStats): void {
+    this.rehydratedCrystalCount = stats.crystalCount;
+    this.rehydratedTotalBytes = stats.totalBytes;
+    this.hits = stats.hits;
+    this.misses = stats.misses;
+    this.hitBytes = stats.hitBytes;
+    this.writes = stats.writes;
+    this.duplicateWrites = stats.duplicateWrites;
+    this.collisions = stats.collisions;
+  }
 }

package/src/vector-cortex/encoder/asset.ts

@@ -26,7 +26,19 @@ import {
 } from "./types.js";
 
 export type AssetVerifyResult =
-  | { ok: true; embeddedBytes: number; maxTokens: number; onnxDigest: string; tokenizerDigest: string }
+  | {
+      ok: true;
+      embeddedBytes: number;
+      maxTokens: number;
+      onnxDigest: string;
+      tokenizerDigest: string;
+      /**
+       * ML5-A (VC2B-2): digest of the manifest-pinned `trained-heads.json`
+       * sibling when the manifest declares `headWeights`; null when the
+       * manifest ships no head weights (the committed placeholder bundle).
+       */
+      headWeightsDigest: string | null;
+    }
   | { ok: false; code: string };
 
 /** True when `p` is a single basename: non-empty, no path separators, no "..",
@@ -80,7 +92,14 @@ function isManifest(m: unknown): m is ModelManifestV1 {
     typeof o.onnx.path === "string" &&
     typeof o.onnx.sha256 === "string" &&
     typeof o.tokenizer.path === "string" &&
-    typeof o.tokenizer.sha256 === "string"
+    typeof o.tokenizer.sha256 === "string" &&
+    // ML5-A (VC2B-2): optional headWeights must, when present, be a valid
+    // ManifestAssetFile (path basename + sha256 + bytes).
+    (o.headWeights === undefined ||
+      (!!o.headWeights &&
+        typeof o.headWeights.path === "string" &&
+        typeof o.headWeights.sha256 === "string" &&
+        typeof o.headWeights.bytes === "number"))
   );
 }
 
@@ -122,6 +141,11 @@ export function verifyEncoderAsset(
   if (!isBasename(manifest.onnx.path) || !isBasename(manifest.tokenizer.path)) {
     return { ok: false, code: ENC_FAIL.MANIFEST_INVALID };
   }
+  // ML5-A (VC2B-2): a manifest-declared headWeights path must also be a bare
+  // basename (no traversal into arbitrary paths off the asset dir).
+  if (manifest.headWeights !== undefined && !isBasename(manifest.headWeights.path)) {
+    return { ok: false, code: ENC_FAIL.MANIFEST_INVALID };
+  }
 
   const onnxPath = join(assetDir, manifest.onnx.path);
   const onnxDigest = digestFile(onnxPath);
@@ -133,13 +157,32 @@ export function verifyEncoderAsset(
   if (tokDigest === null) return { ok: false, code: ENC_FAIL.ASSET_UNREADABLE };
   if (tokDigest !== manifest.tokenizer.sha256) return { ok: false, code: ENC_FAIL.DIGEST_MISMATCH };
 
+  // ML5-A (VC2B-2): verify the manifest-pinned trained-heads sibling when the
+  // manifest declares it. Absent declaration -> headWeightsDigest: null and no
+  // failure (committed placeholder bundle ships no trained weights).
+  let headWeightsDigest: string | null = null;
+  if (manifest.headWeights !== undefined) {
+    const hwPath = join(assetDir, manifest.headWeights.path);
+    const hwDigest = digestFile(hwPath);
+    if (hwDigest === null) return { ok: false, code: ENC_FAIL.ASSET_UNREADABLE };
+    if (hwDigest !== manifest.headWeights.sha256) return { ok: false, code: ENC_FAIL.DIGEST_MISMATCH };
+    headWeightsDigest = hwDigest;
+  }
+
   let embeddedBytes = 0;
   try {
     embeddedBytes = statSync(onnxPath).size + statSync(tokPath).size;
   } catch {
     return { ok: false, code: ENC_FAIL.ASSET_UNREADABLE };
   }
-  return { ok: true, embeddedBytes, maxTokens: manifest.maxTokens, onnxDigest, tokenizerDigest: tokDigest };
+  return {
+    ok: true,
+    embeddedBytes,
+    maxTokens: manifest.maxTokens,
+    onnxDigest,
+    tokenizerDigest: tokDigest,
+    headWeightsDigest,
+  };
 }
 
 /**

package/src/vector-cortex/encoder/router.ts

@@ -26,15 +26,24 @@
  */
 
 import { createEncoderRuntime } from "./runtime.js";
-import { encodeVectorSet, type HeadProjectionOptions } from "./heads.js";
+import {
+  encodeVectorSet,
+  loadHeadProjections,
+  projectHeadFromTrunk,
+  type HeadProjectionOptions,
+  type HeadProjectionTable,
+} from "./heads.js";
 import { embedTrigram512, selectTrigramBFallback } from "./trigram.js";
 import { embedLexical, selectLexicalC } from "./lexical.js";
 import { createEncoderHeadsReporter, type EncoderHeadsReporter } from "./emit-vc2b.js";
+import { ML5A_ENABLED } from "../../config/vector-cortex.js";
 import {
   ENC_FAIL,
+  ENCODER_HEAD_ORDER,
   type EncoderInput,
   type EncoderLoadResult,
   type EncoderRuntime,
+  type HeadVector,
   type VectorSetV1,
 } from "./types.js";
 
@@ -61,6 +70,15 @@ export interface RouterOptions extends HeadProjectionOptions {
   /** Forced fallback: skips the A load and selects the named VC2B fallback
    *  (used to exercise C when A and B are both disabled). Optional. */
   readonly forceFallback?: "B" | "C";
+  /**
+   * ML5-A (VC2B-2): path to the `trained-heads-v1` artifact. When supplied AND
+   * the MEGACOMPACT_ML5_A gate is on AND the artifact loads (correct seed +
+   * shape), the mode-A VectorSet is produced by projecting the real trained head
+   * matrices over the trunk embedding (`projectHeadFromTrunk`). When absent /
+   * flag-off / unloadable, the mode-A producer falls back to the deterministic
+   * LCG placeholder `encodeVectorSet` (byte-identical predecessor).
+   */
+  readonly trainedHeadsPath?: string;
 }
 
 /** Deterministic text derived from an int token sequence so the asset-free
@@ -133,10 +151,46 @@ export function encodeOrFallback(
   if (!inferred.ok) {
     return fallbackFromLoad({ ok: false, mode: "B", code: inferred.code }, reporter, tokens);
   }
-  const vectorSet = encodeVectorSet(tokens, { reporter, seed: options.seed });
+  const vectorSet = produceVectorSet(inferred.semantic, tokens, options, reporter);
   return { ok: true, mode: "A", vectorSet, code: null };
 }
 
+/**
+ * Produce a mode-A `VectorSetV1` from the [1,384] trunk embedding (the
+ * `runtime.infer` result). VC2B-2 ML5-A: when real trained heads are loaded
+ * (`MEGACOMPACT_ML5_A` on + a `trainedHeadsPath` that loads), the multi-head
+ * output is projected through the real trained projection matrices via
+ * `projectHeadFromTrunk`. Otherwise (flag-off / absent / unloadable artifact)
+ * the deterministic LCG placeholder `encodeVectorSet` serves mode A —
+ * byte-identical to the VC2B predecessor. Non-fatal: an unloadable artifact
+ * degrades to the placeholder, never a throw.
+ */
+function produceVectorSet(
+  trunkEmbedding: Float32Array,
+  tokens: readonly number[],
+  options: RouterOptions,
+  reporter: EncoderHeadsReporter,
+): VectorSetV1 {
+  const table: HeadProjectionTable | null =
+    ML5A_ENABLED() && options.trainedHeadsPath !== undefined
+      ? loadHeadProjections(options.trainedHeadsPath)
+      : null;
+  if (table !== null) {
+    const heads: HeadVector[] = ENCODER_HEAD_ORDER.map((h) =>
+      projectHeadFromTrunk(h, trunkEmbedding, table),
+    );
+    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 };
+  }
+  // ML5-A placeholder fallback (byte-identical predecessor).
+  return encodeVectorSet(tokens, { reporter, seed: options.seed });
+}
+
 /**
  * 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

package/src/vector-cortex/encoder/types.ts

@@ -92,6 +92,9 @@ export interface ModelManifestV1 {
   readonly heads: EncoderHeads;
   readonly onnx: ManifestAssetFile;
   readonly tokenizer: ManifestAssetFile;
+  /** ML5-A (VC2B-2): optional sibling `trained-heads.json` (`trained-heads-v1`) shipping
+   *  REAL trained head weights; absent in the placeholder bundle, non-fatal when missing. */
+  readonly headWeights?: ManifestAssetFile;
   readonly totalBytes: number;
   readonly trainingManifestDigest: string;
 }