pi-mega-compact 0.20.38 → 0.20.40

package/src/vector-cortex/encoder/runtime-emit.ts

@@ -0,0 +1,47 @@
+/**
+ * vector-cortex/encoder/runtime-emit.ts — ML5-C seller event emitter.
+ *
+ * Emits the `vector_cortex_runtime_selected` seller event to the local
+ * events.log so the dashboard Setup Cortex blockers card can surface the HG-3
+ * (install budget) / HG-4 (darwin-x64 demotion) closure state. Aggregate
+ * fields only — never payload bytes (EVAL-REDACT-002).
+ *
+ * Extracted from runtime.ts so the runtime delegate-shell stays under the
+ * 300-line soft limit after the ML5-C dispatch was added. All writes are
+ * best-effort / non-fatal; a disk-full or missing state dir never breaks the
+ * encoder loop.
+ *
+ * Pi-agnostic, dependency-free (PREVENT-PI-004 — local filesystem append only;
+ * no network). No `any` (PREVENT-011).
+ */
+
+import { appendFileSync, mkdirSync } from "node:fs";
+import { dirname } from "node:path";
+import { defaultEventsPath } from "../../monitoring.js";
+import type { RuntimeSelectionResult } from "./runtime-select.js";
+
+/**
+ * Emit the ML5-C `vector_cortex_runtime_selected` seller event (best-effort).
+ * The event carries ONLY the four aggregate fields the sprint spec pins
+ * ({backend, p95Ms, budgetOk, platform}) — never message content.
+ */
+export function emitRuntimeSelected(
+  stateDir: string,
+  result: Pick<RuntimeSelectionResult, "backend" | "p95Ms" | "budgetOk" | "platform">,
+): void {
+  try {
+    const path = defaultEventsPath(stateDir);
+    const payload = {
+      ts: Date.now(),
+      event: "vector_cortex_runtime_selected",
+      backend: result.backend,
+      p95Ms: result.p95Ms,
+      budgetOk: result.budgetOk,
+      platform: result.platform,
+    };
+    mkdirSync(dirname(path), { recursive: true });
+    appendFileSync(path, JSON.stringify(payload) + "\n", "utf8");
+  } catch {
+    /* best-effort — never break the encoder loop */
+  }
+}

package/src/vector-cortex/encoder/runtime-native.ts

@@ -0,0 +1,117 @@
+/**
+ * vector-cortex/encoder/runtime-native.ts — ML5-C native backend (Option N).
+ *
+ * Loads an `InferenceSession` from the `onnxruntime-node` native binding for
+ * the committed encoder-v1 ONNX asset. This is the CHOSEN selection when
+ * `MEGACOMPACT_ENCODER_NATIVE=1` (the native opt-in marker) is set AND the
+ * package is present — it uses the platform-specific prebuilt binary (no
+ * postinstall compilation needed; per vc2-model-prep §1 the allowScripts
+ * removal is safe because only CUDA/TensorRT downloads use it, and pi blocks
+ * all scripts anyway).
+ *
+ * The package is NOT declared in package.json dependencies — it is a lazily-
+ * resolved peer that the runtime loads ONLY when the native opt-in is set AND
+ * selected. Loading uses dynamic `import()` so the module graph compiles
+ * cleanly on hosts without the package (absent installs return null, never
+ * throw), so the ML5-C dispatch demotes to mode B trigram rather than
+ * breaking.
+ *
+ * Pi-agnostic, dependency-free (PREVENT-PI-004 — the native binary + model are
+ * committed local files). No `any` (PREVENT-011).
+ */
+
+import {
+  ENCODER_OPSET,
+  ENCODER_SEMANTIC_WIDTH,
+  ENCODER_MAX_TOKENS,
+} from "./types.js";
+
+/** The shape of the optionalImport result when onnxruntime-node is present.
+ *  Shadow-types instead of `import("onnxruntime-node")` so the module graph
+ *  builds without the package being declared in package.json (ML5-B precedent). */
+export interface OrtNativeModule {
+  InferenceSession: {
+    create(
+      path: string,
+      opts: { executionProviders: string[]; intraOpNumThreads: number },
+    ): Promise<{
+      run(
+        feeds: Record<string, Float32Array>,
+        outputNames: string[],
+      ): Promise<Record<string, { data: Float32Array }>>;
+    }>;
+  };
+}
+
+/** The backend's inference session — a thin wrapper over the real native session. */
+export interface NativeSession {
+  /** The declared ONNX opset in the loaded manifest (normative 17). */
+  readonly opset: number;
+  /** The semantic embedding width (normative 384). */
+  readonly semanticWidth: number;
+  /** The per-asset token capacity cap (normative <= 512). */
+  readonly maxTokens: number;
+  /** Run one inference over already shape-checked input tokens. */
+  infer(inputIds: Float32Array): Promise<Float32Array>;
+}
+
+/** True when `MEGACOMPACT_ENCODER_NATIVE=1` (the native opt-in operator flag). */
+export function nativeOptIn(): boolean {
+  return process.env.MEGACOMPACT_ENCODER_NATIVE === "1";
+}
+
+/** True if `onnxruntime-node` resolves on this host (loading is best-effort).
+ *  Absent installs return null (never throw) so the ML5-C dispatch can demote
+ *  to mode B trigram cleanly. */
+async function loadOrtNative(): Promise<OrtNativeModule | null> {
+  try {
+    // @ts-expect-error — optional peer; the shadow type above covers the surface
+    const mod = (await import("onnxruntime-node")) as OrtNativeModule;
+    return mod;
+  } catch {
+    return null;
+  }
+}
+
+/**
+ * Create a native-backed `NativeSession` over the committed ONNX asset, gated
+ * first on `nativeOptIn()`. Returns null (never throws) on any failure
+ * (opt-in off, absent package, unreadable asset, bad session creation) so the
+ * caller demotes to mode B trigram.
+ */
+export async function createNativeSession(
+  modelPath: string,
+  options: { threads?: number; maxTokens?: number } = {},
+): Promise<NativeSession | null> {
+  if (!nativeOptIn()) return null;
+
+  const ort = await loadOrtNative();
+  if (!ort || !ort.InferenceSession?.create) return null;
+
+  const threads = options.threads ?? 4;
+  const maxTokens = options.maxTokens ?? ENCODER_MAX_TOKENS;
+
+  try {
+    const session = await ort.InferenceSession.create(modelPath, {
+      executionProviders: ["cpu"],
+      intraOpNumThreads: threads,
+    });
+
+    return {
+      opset: ENCODER_OPSET,
+      semanticWidth: ENCODER_SEMANTIC_WIDTH,
+      maxTokens,
+      async infer(inputIds: Float32Array): Promise<Float32Array> {
+        const feeds = { input_ids: inputIds };
+        const results = await session.run(feeds, ["embedding"]);
+        const out = results["embedding"];
+        if (!out || !(out.data instanceof Float32Array)) {
+          return new Float32Array(ENCODER_SEMANTIC_WIDTH);
+        }
+        return out.data;
+      },
+    };
+  } catch {
+    return null;
+  }
+}

package/src/vector-cortex/encoder/runtime-select.ts

@@ -0,0 +1,167 @@
+/**
+ * vector-cortex/encoder/runtime-select.ts — ML5-C decision-rule dispatch.
+ *
+ * Pure function of {platform, benchRecord, nativeOptIn} → the chosen ONNX
+ * runtime backend. This is the deterministic selection that closes HG-3
+ * (install budget) and HG-4 (darwin-x64 disposition) per the ML5-C spec:
+ *
+ *   - Measured p95 at 512 tokens on 4 threads (linux-x64) <= 40 ms  →  Option W (WASM)
+ *   - Measured p95 > 40 ms or absent (degraded)                        →  Option N (native)
+ *   - Platform is darwin-x64 (Intel Mac, HG-1 deferral)               →  WASM demotion or mode B
+ *
+ * The `platform` comes from `detectPlatform()` (already in `asset.ts`); the
+ * `benchRecord` is the latest `BenchResultV1` (from ML5-B / JSONL); the
+ * `nativeOptIn` helper reads `MEGACOMPACT_ENCODER_NATIVE=1`. The selection is
+ * PURE — no side effects, no I/O, no network. The caller stamps the result on
+ * the `vector_cortex_runtime_selected` event carried to the dashboard.
+ *
+ * Pi-agnostic, dependency-free (PREVENT-PI-004 — selection is computed+in memory;
+ * no fetch/HTTP). No `any` (PREVENT-011).
+ */
+
+import {
+  ENCODER_LATENCY_P95_MS,
+  type EncoderPlatform,
+} from "./types.js";
+import type { BenchResultV1 } from "./bench-export.js";
+import { ML5C_ENABLED } from "../../config/vector-cortex.js";
+
+/** The chosen runtime backend name as it appears on the seller event. */
+export type RuntimeBackendChoice = "wasm" | "native" | "modeB";
+
+/** Inputs to the backend selection. `platform` is `${process.platform}-${process.arch}`. */
+export interface RuntimeSelectionInput {
+  /** The platform this process is running on (from detectPlatform()). */
+  readonly platform: EncoderPlatform | "unsupported";
+  /**
+   * The ML5-B bench record for this platform (linux-x64 normative). Null when
+   * no bench has run yet (degraded: the 42-byte placeholder has no real p95).
+   * Must provide a numeric p95Ms for WASM consideration; a null/error record
+   * (gates.all:false) forces the native fallback per the decision rule.
+   */
+  readonly benchRecord: BenchResultV1 | null;
+  /** True when the operator opted into onnxruntime-node via MEGACOMPACT_ENCODER_NATIVE=1. */
+  readonly nativeOptIn: boolean;
+}
+
+/** The output of the ML5-C decision-rule dispatch. */
+export interface RuntimeSelectionResult {
+  /** The chosen backend. */
+  readonly backend: RuntimeBackendChoice;
+  /** HG-3 closure state: whether the 80 MiB install budget is satisfied. */
+  readonly budgetOk: boolean;
+  /** The measured p95 (ms) that drove the decision, from BenchResultV1. */
+  readonly p95Ms: number | null;
+  /** The platform the selection is valid for. */
+  readonly platform: string;
+  /** Short human-readable rationale for the choice (record on the event). */
+  readonly rationale: string;
+}
+
+/** The 80 MiB install budget in bytes (the HG-3 ceiling, unamended). */
+export const RUNTIME_NATIVE_INSTALL_BUDGET_MIB = 80;
+
+/**
+ * Per-platform optionalDependency footprint of onnxruntime-node (MiB, approx).
+ * These sum to ~160 MiB SHIPPED in the npm package (every target platform is
+ * included so the resolver lands on a concrete row at install time) — which is
+ * what exceeds the 80 MiB HG-3 budget and forces the amendment the fixtures
+ * record (ML5-RUNTIME-001). The per-host INSTALLED footprint (one row only)
+ * is 28–35 MiB and irrelevant to the 80 MiB ceiling — the budget covers the
+ * shipped tarball, not the single-platform install.
+ */
+export const NATIVE_FOOTPRINT_MIB: Readonly<Record<EncoderPlatform, number>> = {
+  "linux-x64": 33,
+  "darwin-arm64": 28,
+  "darwin-x64": 33,
+  "linux-arm64": 31,
+  "win32-x64": 35,
+};
+
+/**
+ * ML5-C decision-rule dispatch: choose the ONNX runtime backend (pure).
+ *
+ * When the flag is OFF (`MEGACOMPACT_ML5_C=0`), returns mode B trigram — byte-
+ * identical to the ML5-B survivor with no selection event emitted.
+ *
+ * The rule (from the sprint spec):
+ *   - If nativeOptIn && platform is supported → native (Option N)
+ *   - If benchRecord has p95Ms <= 40 ms on linux-x64 → WASM (Option W)
+ *   - Else → native (Option N) with the budget amendment recorded (p95 exceeds
+ *     the WASM gate or is absent — the placeholder has no measured p95)
+ *   - darwin-x64 → WASM or mode B demotion per HG-4 (never native here)
+ */
+export function selectRuntimeBackend(input: RuntimeSelectionInput): RuntimeSelectionResult {
+  if (!ML5C_ENABLED()) {
+    return {
+      backend: "modeB",
+      budgetOk: true,
+      p95Ms: null,
+      platform: input.platform,
+      rationale: "flag-off: byte-identical mode-B trigram (no selection)",
+    };
+  }
+
+  // HG-4: Intel Mac (darwin-x64) is out-of-scope per HG-1's deferral — always demote.
+  if (input.platform === "darwin-x64") {
+    return {
+      backend: "wasm",
+      budgetOk: true,
+      p95Ms: null,
+      platform: input.platform,
+      rationale: "darwin-x64 demoted to WASM per HG-4 (never native on this platform)",
+    };
+  }
+
+  // Native opt-in short-circuits: operator explicitly wants the native path.
+  // The HG-3 budget compares the SHIPPED byte-count (sum across every platform
+  // row in the package's optionalDependencies map) against the 80 MiB ceiling
+  // — not the single-platform install size. Native always exceeds 80 MiB across
+  // 5 platforms (~160 MiB shipped), so budgetOk is false and the evidence
+  // records the amended budget (the ML5-C spec, HG-3 closure).
+  if (input.nativeOptIn) {
+    const shippedMib = Object.values(NATIVE_FOOTPRINT_MIB).reduce((a, b) => a + b, 0);
+    return {
+      backend: "native",
+      budgetOk: shippedMib <= RUNTIME_NATIVE_INSTALL_BUDGET_MIB,
+      p95Ms: input.benchRecord?.p95Ms ?? null,
+      platform: input.platform,
+      rationale: `native opt-in (MEGACOMPACT_ENCODER_NATIVE=1); shipped ${shippedMib} MiB across 5 platforms → budget amended to ${shippedMib} MiB`,
+    };
+  }
+
+  // No bench record or degraded (gates.all:false) → WASM cannot qualify (the
+  // placeholder 42-byte asset has no measured real p95), so native is selected
+  // with the SAME amended-budget disposition as the opt-in path above: the
+  // evidence records the closed HG-3 amendment.
+  if (!input.benchRecord || !input.benchRecord.gates.all || input.benchRecord.p95Ms === null) {
+    const shippedMib = Object.values(NATIVE_FOOTPRINT_MIB).reduce((a, b) => a + b, 0);
+    return {
+      backend: "native",
+      budgetOk: false, // amended: native ships > 80 MiB across the 5-platform matrix
+      p95Ms: input.benchRecord?.p95Ms ?? null,
+      platform: input.platform,
+      rationale: `no qualifying bench record — native fallback with budget amendment (${shippedMib} MiB shipped, HG-3 amendment recorded)`,
+    };
+  }
+
+  // The decision rule: WASM iff p95 <= 40 ms (linux-x64, 512 tokens, 4 threads) —
+  // native required otherwise, with the same budget amendment recorded.
+  if (input.benchRecord.p95Ms <= ENCODER_LATENCY_P95_MS) {
+    return {
+      backend: "wasm",
+      budgetOk: true,
+      p95Ms: input.benchRecord.p95Ms,
+      platform: input.platform,
+      rationale: `WASM qualifies: p95 ${input.benchRecord.p95Ms}ms <= ${ENCODER_LATENCY_P95_MS}ms`,
+    };
+  }
+
+  return {
+    backend: "native",
+    budgetOk: false, // amended: native exceeds the 80 MiB budget per the evidence
+    p95Ms: input.benchRecord.p95Ms,
+    platform: input.platform,
+    rationale: `native required: p95 ${input.benchRecord.p95Ms}ms > ${ENCODER_LATENCY_P95_MS}ms on WASM`,
+  };
+}

package/src/vector-cortex/encoder/runtime-stub.ts

@@ -0,0 +1,38 @@
+/**
+ * vector-cortex/encoder/runtime-stub.ts — the VC2A-era deterministic LCG
+ * placeholder (`projectSemantic`) + token-seed helper, split out of runtime.ts
+ * so `runtime.ts` stays under its 300-line soft limit as the delegate-shell.
+ *
+ * This placeholder is the VC2A-era stand-in for a real ONNX EncoderRuntime
+ * inference result (the actual weights were VC2C and the runtime-selection
+ * dispatch is ML5-C). It remains exported so tests driving end-to-end shape
+ * keep working even when no backend session is active.
+ *
+ * Pi-agnostic, dependency-free (PREVENT-PI-004 / PREVENT-011).
+ */
+
+import { ENCODER_SEMANTIC_WIDTH } from "./types.js";
+
+/** A deterministic seeded projection so the mode-A inference path is testable
+ *  end-to-end without onnxruntime (real weights + execution are VC2C). */
+export function projectSemantic(seed: number, n: number): Float32Array {
+  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. */
+export function seedFromBytes(embeddedBytes: number): number {
+  return (embeddedBytes * 2654435761) >>> 0;
+}
+
+/** The semantic embedding width from the normative types barrel. */
+export { ENCODER_SEMANTIC_WIDTH };

package/src/vector-cortex/encoder/runtime-wasm.ts

@@ -0,0 +1,110 @@
+/**
+ * vector-cortex/encoder/runtime-wasm.ts — ML5-C WASM backend (Option W).
+ *
+ * Loads an `InferenceSession` from the `onnxruntime-web` WASM execution
+ * provider for the committed encoder-v1 ONNX asset. This is the default
+ * backend when the WASM path is selected by `select.ts` — it covers all Node
+ * platforms (no per-platform optionalDependencies), is pure JS + WASM (~9 MiB),
+ * and never fetches from the network (PREVENT-PI-004).
+ *
+ * The package is NOT declared in package.json dependencies — it is a lazily-
+ * resolved peer that the runtime loads ONLY when the WASM backend is actually
+ * selected. Loading uses dynamic `import()` so the module graph compiles
+ * cleanly on hosts without the package; absent installs return null (never
+ * throw), so the ML5-C dispatch demotes to mode B trigram rather than
+ * breaking (ML5-B-bench precedent: the fixtures declare the shape even when
+ * the package is absent).
+ *
+ * Pi-agnostic, dependency-free (PREVENT-PI-004 — the WASM artifact is loaded
+ * from the committed local path; no fetch/HTTP anywhere). No `any`
+ * (PREVENT-011).
+ */
+
+import {
+  ENCODER_OPSET,
+  ENCODER_SEMANTIC_WIDTH,
+  ENCODER_MAX_TOKENS,
+} from "./types.js";
+
+/** The shape of the optionalImport result when onnxruntime-web is present.
+ *  Shadow-types instead of `import("onnxruntime-web")` so the module graph builds
+ *  without the package being declared in package.json (ML5-B precedent). */
+export interface OrtWasmModule {
+  InferenceSession: {
+    create(
+      path: string,
+      opts: { executionProviders: string[]; intraOpNumThreads: number },
+    ): Promise<{
+      run(
+        feeds: Record<string, Float32Array>,
+        outputNames: string[],
+      ): Promise<Record<string, { data: Float32Array }>>;
+    }>;
+  };
+}
+
+/** The backend's inference session — a thin wrapper over the real WASM session. */
+export interface WasmSession {
+  /** The declared ONNX opset in the loaded manifest (normative 17). */
+  readonly opset: number;
+  /** The semantic embedding width (normative 384). */
+  readonly semanticWidth: number;
+  /** The per-asset token capacity cap (normative <= 512). */
+  readonly maxTokens: number;
+  /** Run one inference over already shape-checked input tokens. */
+  infer(inputIds: Float32Array): Promise<Float32Array>;
+}
+
+/** True if `onnxruntime-web` resolves on this host (loading is best-effort).
+ *  Absent installs return null (never throw) so the ML5-C dispatch can demote
+ *  to mode B trigram cleanly. */
+async function loadOrtWasm(): Promise<OrtWasmModule | null> {
+  try {
+    // @ts-expect-error — optional peer; the shadow type above covers the surface
+    const mod = (await import("onnxruntime-web")) as OrtWasmModule;
+    return mod;
+  } catch {
+    return null;
+  }
+}
+
+/**
+ * Create a WASM-backed `WasmSession` over the committed ONNX asset. Returns
+ * null (never throws) on any failure (absent package, unreadable asset, bad
+ * session creation) so the caller demotes to mode B trigram per HG-4 mode-B
+ * disposition when the WASM path is unavailable on a darwin-x64 host.
+ */
+export async function createWasmSession(
+  modelPath: string,
+  options: { threads?: number; maxTokens?: number } = {},
+): Promise<WasmSession | null> {
+  const ort = await loadOrtWasm();
+  if (!ort || !ort.InferenceSession?.create) return null;
+
+  const threads = options.threads ?? 4;
+  const maxTokens = options.maxTokens ?? ENCODER_MAX_TOKENS;
+
+  try {
+    const session = await ort.InferenceSession.create(modelPath, {
+      executionProviders: ["wasm"],
+      intraOpNumThreads: threads,
+    });
+
+    return {
+      opset: ENCODER_OPSET,
+      semanticWidth: ENCODER_SEMANTIC_WIDTH,
+      maxTokens,
+      async infer(inputIds: Float32Array): Promise<Float32Array> {
+        const feeds = { input_ids: inputIds };
+        const results = await session.run(feeds, ["embedding"]);
+        const out = results["embedding"];
+        if (!out || !(out.data instanceof Float32Array)) {
+          return new Float32Array(ENCODER_SEMANTIC_WIDTH);
+        }
+        return out.data;
+      },
+    };
+  } catch {
+    return null;
+  }
+}