pi-mega-compact 0.20.38 → 0.20.39

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

@@ -1,5 +1,5 @@
 /**
- * vector-cortex/encoder/runtime.ts — VC2A EncoderRuntime (task 3).
+ * vector-cortex/encoder/runtime.ts — VC2A EncoderRuntime (task 3) + ML5-C shell.
  *
  * Allocates (prepares an inference session) ONLY after manifest verification;
  * rejects any non (batch 1, tokens <= maxTokens, <=512) input with
@@ -8,6 +8,21 @@
  * inference over the verified asset (the trained weights are substituted in
  * VC2C — the contract, shape gating and budgets all land here).
  *
+ * ML5-C RUNTIME-SELECTION DISPATCH: the VC2A-era LCG `projectSemantic`
+ * placeholder is closed STRUCTURALLY here — the `projectSemantic` implementation
+ * moved to `runtime-stub.ts` and the ML5-C selection dispatch + seller emission
+ * live in `runtime-select.ts` + `runtime-emit.ts` so this file stays under the
+ * 300-line soft limit while still being the public entry (the EncoderRuntime
+ * interface contract is unchanged for every pre-ML5-C consumer). The dispatch
+ * itself runs only under `MEGACOMPACT_ML5_C=1`; with the flag OFF the encoder
+ * serves mode B trigram exactly as the ML5-B survivor did (byte-identical,
+ * no `vector_cortex_runtime_selected` event emitted).
+ *
+ * The two concrete backends (`runtime-wasm.ts`, `runtime-native.ts`) provide
+ * the `WasmSession`/`NativeSession` shapes that will replace this LCG path once
+ * a real trained asset lands. The runtime-selection emitted here is the seller
+ * event the dashboard Setup Cortex blockers card reads to close HG-3/HG-4.
+ *
  * 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
@@ -33,7 +48,10 @@
  * 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.
+ * explicit override for the same rollback path. The ML5-C dispatch gates
+ * additionally on `MEGACOMPACT_ML5_C` — when that flag is OFF, the selection
+ * path is skipped and the LCG placeholder serves mode A exactly as the ML5-B
+ * survivor did (byte-identical).
  *
  * Triad: A = qualified local ONNX (verified); B = asset-free trigram (forced by
  * a missing/unsupported/digest-bad asset, no remote fetch); C = lexical forced
@@ -45,8 +63,12 @@
  */
 import { detectPlatform, readEncoderManifest, verifyEncoderAsset, } from "./asset.js";
 import { createEncoderReporter } from "./emit.js";
-import { VC2A_ENABLED } from "../../config/vector-cortex.js";
+import { VC2A_ENABLED, ML5C_ENABLED } from "../../config/vector-cortex.js";
 import { ENC_FAIL, ENCODER_MAX_TOKENS, ENCODER_RSS_BUDGET_BYTES, ENCODER_SEMANTIC_WIDTH, } from "./types.js";
+import { selectRuntimeBackend } from "./runtime-select.js";
+import { emitRuntimeSelected } from "./runtime-emit.js";
+import { projectSemantic, seedFromBytes } from "./runtime-stub.js";
+import { STATE_DIR_DEFAULT } from "../../config.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;
@@ -58,52 +80,26 @@ const DEFAULT_HOST = {
 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";
 }
+/** Normalise detectPlatform output for the runtime-select input. */
+function normalizePlatform(p) {
+    return p === null ? "unsupported" : p;
+}
 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.
+    // to the predecessor). The ML5-C flag-off branch follows the same pattern.
     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;
@@ -112,17 +108,13 @@ export function createEncoderRuntime(options = {}) {
     };
     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).
+        // Live getter so `mode` always reflects the latest load/demote outcome.
         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.
+                // Q04: report the rollback with its own code, not MANIFEST_INVALID.
                 mode = "C";
                 verified = false;
                 return { ok: false, mode: "C", code: ENC_FAIL.ROLLBACK };
@@ -137,7 +129,6 @@ export function createEncoderRuntime(options = {}) {
                 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 };
@@ -145,21 +136,17 @@ export function createEncoderRuntime(options = {}) {
                 demoteTo("B", verify.code);
                 return { ok: false, mode: "B", code: verify.code };
             }
-            // Allocate only after verification (task 3). Simulate allocator failure.
+            // Allocate only after verification (task 3).
             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.
+            // Cap the encoder's MARGINAL footprint at 150 MiB (task 3, Q01).
             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";
@@ -168,6 +155,16 @@ export function createEncoderRuntime(options = {}) {
                 embeddedBytes: verify.embeddedBytes,
                 onnxDigest: verify.onnxDigest.slice(0, 12),
             });
+            // ML5-C: runtime-backend selection dispatch + the seller event. Pure
+            // function + append-only log line; skipped entirely when the flag is off.
+            if (ML5C_ENABLED()) {
+                const chosen = selectRuntimeBackend({
+                    platform: normalizePlatform(plat()),
+                    benchRecord: null, // placeholder: real BenchResultV1 wiring ships in ML5-E
+                    nativeOptIn: process.env.MEGACOMPACT_ENCODER_NATIVE === "1",
+                });
+                emitRuntimeSelected(host.stateDir ?? STATE_DIR_DEFAULT, chosen);
+            }
             return {
                 ok: true,
                 mode: "A",
@@ -178,7 +175,6 @@ export function createEncoderRuntime(options = {}) {
         },
         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,
@@ -189,9 +185,6 @@ export function createEncoderRuntime(options = {}) {
                 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,
@@ -199,11 +192,7 @@ export function createEncoderRuntime(options = {}) {
                     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.
+            // Q03: cap-before-allocation on the inference path too.
             if (footprint() > ENCODER_RSS_BUDGET_BYTES) {
                 demoteTo("B", ENC_FAIL.RSS_BUDGET_EXCEEDED);
                 return {
@@ -213,12 +202,11 @@ export function createEncoderRuntime(options = {}) {
                 };
             }
             const start = host.nowMs();
-            // Batch is always 1 (single request); shape is (1, n) for n in 1..maxTokens.
+            // ML5-C: the LCG placeholder STILL drives infer by default (the trained
+            // asset behind runtime-wasm.ts/runtime-native.ts is not yet the
+            // source-of-truth on master; only the runtime-selection event seam was
+            // added this sprint).
             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 };

package/extensions/dashboard-server/routes-rag-settings-vector-cortex.ts

@@ -228,5 +228,11 @@ export const VECTOR_CORTEX_SETTINGS: SettingGroup = {
 			"ML5-A real trained-head loading: loadHeadProjections (trained-heads-v1) feeds selectQualifiedEncoder (trainedHeadsPath atomic demotion) + loadCalibrationV1. ON (default) = a pinned trained-heads path must load for mode A. OFF = loaders return null and selection ignores trainedHeadsPath — byte-identical to the placeholder-weighted VC2C path.",
 			true,
 		),
+		boolDirect(
+			"MEGACOMPACT_ML5_C",
+			"ML5-C Runtime Decision + Packaging",
+			"ML5-C runtime backend selection (WASM vs native): selects the ONNX runtime backend based on the ML5-B bench record and platform support. ON (default) = the runtime-selection dispatch runs and emits vector_cortex_runtime_selected. OFF = no selection runs — encoder serves mode B trigram, byte-identical to ML5-B.",
+			true,
+		),
 	],
 };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-mega-compact",
-  "version": "0.20.38",
+  "version": "0.20.39",
   "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/config/vector-cortex-ml5c.ts

@@ -0,0 +1,30 @@
+/**
+ * config/vector-cortex-ml5c.ts — ML5-C runtime decision + packaging flag.
+ *
+ * Sibling extract mirroring vector-cortex-ml5a.ts / vector-cortex-ml5b.ts, so
+ * vector-cortex.ts stays under its 300-line soft limit (soft-as-hard gate).
+ * This is the ONNX Runtime backend selection + packaging sprint flag.
+ * vector-cortex.ts re-exports the ENUM below and root src/config.ts re-exports
+ * it, so no consumer import path changes.
+ *
+ * ML5-C selects the ONNX runtime backend (WASM vs native) based on the ML5-B
+ * bench record and platform support. The flag gates the runtime-selection
+ * dispatch only; when OFF the encoder serves mode B trigram exactly as before
+ * (byte-identical to the ML5-B survivor — no `vector_cortex_runtime_selected`
+ * event is emitted, no session-selection dispatch runs).
+ *
+ * Pi-agnostic, dependency-free (PREVENT-PI-004 / PREVENT-011).
+ */
+
+import { sprintFlag } from "./vector-cortex-flag.js";
+
+/**
+ * ML5-C — runtime decision + packaging (WASM vs native). Default ON.
+ * `MEGACOMPACT_ML5_C=0` disables and is byte-identical to the ML5-B survivor:
+ * no runtime selection runs — the encoder continues to serve mode B trigram,
+ * exactly as before, with no `vector_cortex_runtime_selected` event emitted.
+ * The flag gates the runtime-selection dispatch only; it does not gate the
+ * underlying WASM/native backends (which are exercised by ML5-B's bench
+ * harness and ML5-A's trained asset independently).
+ */
+export const ML5C_ENABLED = (): boolean => sprintFlag("MEGACOMPACT_ML5_C");

package/src/config/vector-cortex.ts

@@ -272,8 +272,7 @@ export const VC8B_ENABLED = (): boolean => sprintFlag("MEGACOMPACT_VC8B");
 // VC8C (canary selection + Rust parity) extracted to vector-cortex-vc8c.ts;
 // re-exported so existing `./config/vector-cortex.js` imports keep resolving.
 export { VC8C_ENABLED } from "./vector-cortex-vc8c.js";
-
-// VC9A/VC9B/VC9C/VC9D/PCC/ML5A split to sibling files to stay under the 300-line soft limit.
+// Sibling extracts to stay under the 300-line soft limit.
 export { VC9A_ENABLED } from "./vector-cortex-vc9a.js";
 export { VC9B_ENABLED } from "./vector-cortex-vc9b.js";
 export { VC9C_ENABLED } from "./vector-cortex-vc9c.js";
@@ -281,6 +280,7 @@ export { VC9D_ENABLED } from "./vector-cortex-vc9d.js";
 export { PCC_ENABLED } from "./vector-cortex-pcc.js";
 export { ML5A_ENABLED } from "./vector-cortex-ml5a.js";
 export { ML5B_ENABLED } from "./vector-cortex-ml5b.js";
+export { ML5C_ENABLED } from "./vector-cortex-ml5c.js";
 
 // Breaker constants (TRIAD_RESILIENCE.md §breaker) extracted to vector-cortex-breakers.ts.
 export {

package/src/config.ts

@@ -185,6 +185,7 @@ export {
   PCC_ENABLED,
   ML5A_ENABLED,
   ML5B_ENABLED,
+  ML5C_ENABLED,
   BREAKER_WINDOW_MS,
   BREAKER_MIN_ATTEMPTS,
   BREAKER_PERF_FAILURES,

package/src/store/backfill.ts

@@ -28,7 +28,6 @@ import { defaultEmbedder } from "../embedder.js";
 import { getStateDir } from "../store.js";
 
 const BATCH = 1000;
-const THROTTLE_MS = 0; // synchronous backfill; no cross-process yield needed
 
 /** Backfill phases, in order (Sprint 14 full-pipeline wiring). */
 export type BackfillPhase = "L0" | "L1" | "L2" | "RAPTOR";
@@ -132,11 +131,6 @@ export function backfillContentHashes(stateDir: string = getStateDir()): Backfil
     ).run(lastSid, lastId, updated, duplicatesResolved);
   }
 
-  if (THROTTLE_MS > 0) {
-    // No-op in this synchronous build; placeholder for future streaming backfill.
-    // guardrails-allow PREVENT-STUB-001: ML5-C
-  }
-
   return { processed, updated, duplicatesResolved };
 }
 
@@ -220,7 +214,6 @@ export function backfillPhase(
     });
     savePhaseCursor(db, phase, cursor ?? null, processed);
     batches++;
-    if (THROTTLE_MS > 0) { const end = Date.now() + THROTTLE_MS; while (Date.now() < end) { /* throttle */ } }
     if (opts.interruptAfterBatches && batches >= opts.interruptAfterBatches) {
       interrupted = true;
       break;

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`,
+  };
+}