pi-mega-compact 0.20.42 → 0.20.44

package/src/vector-cortex/_acceptance-vc2b-heads.ts

@@ -0,0 +1,234 @@
+/** VC2B acceptance — multi-head invariant + independence + forced triad through
+ *  the REAL encode-or-fallback router (no mocks, including unique failure
+ *  injection where the learned asset is deleted after A selection but before
+ *  inference). Extracted from vc2b-acceptance.test.ts so the aggregator stays
+ *  under the soft line limit. Shared context (withFlagsOn, stageVerifyingAssetDir,
+ *  token/order constants) is injected by the aggregator to avoid a circular import.
+ */
+import { test, describe } from "node:test";
+import assert from "node:assert/strict";
+import { rmSync } from "node:fs";
+import { join } from "node:path";
+
+import { ENC_FAIL, ENCODER_HEAD_DIMS } from "./encoder/types.js";
+import { encodeVectorSet, l2Norm } from "./encoder/heads.js";
+import { ENCODER_TRIGRAM_WIDTH } from "./encoder/trigram.js";
+import { embedLexical, ENCODER_LEXICAL_WIDTH } from "./encoder/lexical.js";
+import { encodeOrFallback } from "./encoder/router.js";
+import { createEncoderRuntime } from "./encoder/runtime.js";
+import { createEncoderHeadsReporter } from "./encoder/emit-vc2b.js";
+
+export interface HeadsCtx {
+  withFlagsOn: (fn: () => void) => void;
+  stageVerifyingAssetDir: (maxTokens?: number) => string;
+  SET_TOKENS: number[];
+  EMPTY_TOKENS: number[];
+  ORDERED_DIMS: number[];
+}
+
+export function registerHeads(ctx: HeadsCtx): void {
+  const { withFlagsOn, stageVerifyingAssetDir, SET_TOKENS, EMPTY_TOKENS, ORDERED_DIMS } = ctx;
+
+  // Suite 4 — invariant + unique failure injection + forced triad
+  describe("multi-head invariant + independence + triad", () => {
+    test("invariant: every emitted norm is 0 or within 1e-6 of 1", () => {
+      for (const tokens of [SET_TOKENS, EMPTY_TOKENS, [7], Array.from({ length: 300 }, (_, i) => i)]) {
+        const set = encodeVectorSet(tokens);
+        for (const hv of set.heads) {
+          const n = l2Norm(hv.values);
+          assert.equal(n === 0 || Math.abs(n - 1) <= 1e-6, true, `${hv.head} norm ${n}`);
+        }
+      }
+    });
+
+    test("repeat drift <= 1e-6 across repeated seeded exports (all five heads)", () => {
+      for (let rep = 0; rep < 3; rep++) {
+        const a = encodeVectorSet(SET_TOKENS);
+        const b = encodeVectorSet(SET_TOKENS);
+        for (let i = 0; i < a.heads.length; i++) {
+          for (let j = 0; j < a.heads[i]!.values.length; j++) {
+            assert.equal(Math.abs(a.heads[i]!.values[j]! - b.heads[i]!.values[j]!) <= 1e-6, true);
+          }
+        }
+      }
+    });
+
+    test("unique failure injection: delete model after A selection but before inference; router catches the real load() failure and selects independently initialized B", () => {
+      // This scenario is ON-dependent: it asserts a fallback emission, which is
+      // VC2B-flag-gated, and drives the VC2A runtime into an A load — so it self-pins
+      // both flags ON and is thus valid under either the default-ON run or the
+      // MEGACOMPACT_VC2B=0 parity run.
+      withFlagsOn(() => {
+        // Stage a learned asset the VC2A runtime would VERIFY into mode A, then
+        // REMOVE model.onnx before encoding. The router's load() returns the real
+        // ENC_ASSET_UNREADABLE failure code and must hand off to the independently
+        // initialized asset-free trigram B — emitting vector_cortex_encoder_fallback_selected
+        // from the production seam (S2: a true end-to-end router test, not a
+        // simulated direct call to selectTrigramBFallback).
+        const dir = stageVerifyingAssetDir();
+        const emitted: string[] = [];
+        const reporter = createEncoderHeadsReporter((e) => emitted.push(e));
+        try {
+          // A is selectable at this point (a staging runtime verifies it).
+          const probe = createEncoderRuntime();
+          assert.equal(probe.load(dir).ok, true, "staged asset verifies into A");
+          // "After A selection but before inference": the on-disk model is gone.
+          rmSync(join(dir, "model.onnx"), { force: true });
+          const verdict = encodeOrFallback({ tokens: SET_TOKENS }, dir, { reporter });
+          assert.equal(verdict.ok, true);
+          assert.equal(verdict.mode, "B", "router handoff selects independently initialized B");
+          assert.equal(verdict.width, 512);
+          if (verdict.ok) {
+            assert.equal(verdict.vector.length, 512);
+            assert.equal(verdict.code, ENC_FAIL.ASSET_UNREADABLE, "load() reported the real failure code");
+          }
+          assert.ok(
+            emitted.includes("vector_cortex_encoder_fallback_selected"),
+            "fallback-selected fired from the real router seam: " + emitted.join(","),
+          );
+          // Distinct vectors for distinct inputs — the fallback is not a constant.
+          const again = encodeOrFallback({ tokens: [9, 8, 7, 6] }, dir, { reporter });
+          assert.equal(again.ok, true);
+          if (verdict.ok && again.ok && verdict.mode === "B" && again.mode === "B") {
+            let diff = 0;
+            for (let i = 0; i < verdict.vector.length; i++) diff += Math.abs(verdict.vector[i]! - again.vector[i]!);
+            assert.ok(diff > 1e-3, "independent trigram B is input-sensitive");
+          }
+        } finally {
+          rmSync(dir, { recursive: true, force: true });
+        }
+      });
+    });
+
+    test("forced triad A / B / C through the encode-or-fallback router", () => {
+      // ON-dependent: asserts heads_emitted / fallback_selected emissions, so it
+      // self-pins both flags ON (valid under either the default-ON run or the
+      // MEGACOMPACT_VC2B=0 parity run).
+      withFlagsOn(() => {
+        // A = learned projections: a verifying asset dir routes to a VectorSetV1 with
+        // the five heads in ordered dims (emitting heads_emitted).
+        const dirA = stageVerifyingAssetDir();
+        const emittedA: string[] = [];
+        try {
+          const reporterA = createEncoderHeadsReporter((e) => emittedA.push(e));
+          const a = encodeOrFallback({ tokens: SET_TOKENS }, dirA, { reporter: reporterA });
+          assert.equal(a.ok, true);
+          assert.equal(a.mode, "A");
+          if (a.ok) {
+            assert.equal(a.vectorSet.heads.length, 5);
+            assert.deepEqual(a.vectorSet.heads.map((h) => h.dim), ORDERED_DIMS);
+          }
+          assert.ok(emittedA.includes("vector_cortex_encoder_heads_emitted"));
+        } finally {
+          rmSync(dirA, { recursive: true, force: true });
+        }
+        // B = 512d trigram selected when the learned asset directory is REMOVED: the
+        // router's load() fails (no manual fetch) and hands off to B. Remove a staged
+        // asset dir so the directory is genuinely absent, proving B needs no asset.
+        const dirB = stageVerifyingAssetDir();
+        rmSync(dirB, { recursive: true, force: true }); // asset directory REMOVED
+        const emittedB: string[] = [];
+        const reporterB = createEncoderHeadsReporter((e) => emittedB.push(e));
+        const b = encodeOrFallback({ tokens: SET_TOKENS }, dirB, { reporter: reporterB });
+        assert.equal(b.ok, true);
+        assert.equal(b.mode, "B", "B works without an asset dir");
+        assert.equal(b.width, 512);
+        if (b.ok) {
+          assert.equal(b.vector.length, 512);
+          assert.equal(b.limitation, null);
+        }
+        assert.ok(emittedB.includes("vector_cortex_encoder_fallback_selected"), "B selection emits fallback-selected");
+        // C = token/phrase lexical forced when both A and B runtimes are disabled.
+        const emittedC: string[] = [];
+        const reporterC = createEncoderHeadsReporter((e) => emittedC.push(e));
+        const c = encodeOrFallback({ tokens: SET_TOKENS }, dirB, { reporter: reporterC, forceFallback: "C" });
+        assert.equal(c.ok, true);
+        assert.equal(c.mode, "C");
+        assert.equal(c.width, ENCODER_LEXICAL_WIDTH);
+        if (c.ok && c.mode === "C") {
+          assert.equal(c.vector.length, ENCODER_LEXICAL_WIDTH);
+          assert.ok((c.limitation ?? "").length > 0, "C reports its semantic-context limitation");
+          // Q04: a FORCED C is an intentional selection, not a demotion or rollback,
+          // so it must NOT be stamped with ENC_FAIL.ROLLBACK.
+          assert.equal(c.code, null, "forced C is not a rollback/demotion (code = null)");
+        }
+        assert.ok(emittedC.includes("vector_cortex_encoder_fallback_selected"), "C selection emits fallback-selected");
+        // Widths are disjoint across the triad (no shared feature space).
+        const aWidths = Object.values(ENCODER_HEAD_DIMS);
+        for (const w of [...aWidths, ENCODER_LEXICAL_WIDTH]) assert.notEqual(w, ENCODER_TRIGRAM_WIDTH);
+      });
+    });
+
+    test("forced fallback (B/C) wins over the empty-input degenerate case (Q02)", () => {
+      // Q02: a caller that explicitly forces a fallback mode must get that mode
+      // even for EMPTY input — empty tokens must NOT silently short-circuit to B.
+      // Asserts fallback_selected emissions → ON-dependent, so it self-pins via
+      // withFlagsOn (valid under either external env, same as the forced-triad test).
+      withFlagsOn(() => {
+        const emittedC: string[] = [];
+        const reporterC = createEncoderHeadsReporter((e) => emittedC.push(e));
+        const c = encodeOrFallback({ tokens: EMPTY_TOKENS }, "", { reporter: reporterC, forceFallback: "C" });
+        assert.equal(c.ok, true);
+        assert.equal(c.mode, "C", "forced C must win over empty-input B selection");
+        if (c.ok && c.mode === "C") {
+          assert.equal(c.vector.length, ENCODER_LEXICAL_WIDTH);
+          assert.equal(c.code, null, "forced C carries no rollback/demotion code (Q04)");
+        }
+        assert.ok(emittedC.includes("vector_cortex_encoder_fallback_selected"));
+        // Forced B on empty input also stays B (force mode is honored, no failure
+        // code — a forced mode is not a rollback/demotion, so code is null).
+        const emittedB: string[] = [];
+        const reporterB = createEncoderHeadsReporter((e) => emittedB.push(e));
+        const b = encodeOrFallback({ tokens: EMPTY_TOKENS }, "", { reporter: reporterB, forceFallback: "B" });
+        assert.equal(b.ok, true);
+        assert.equal(b.mode, "B");
+        if (b.ok && b.mode === "B") {
+          assert.equal(b.vector.length, ENCODER_TRIGRAM_WIDTH);
+          assert.equal(b.code, null, "forced B carries no rollback/demotion code (Q04)");
+        }
+      });
+    });
+
+    test("A/B/C use disjoint widths and independent algorithms", () => {
+      // A head widths are 384/128/128/64/32; B is 512; C is 256 — no shared space.
+      const aWidths = Object.values(ENCODER_HEAD_DIMS);
+      const bWidth = ENCODER_TRIGRAM_WIDTH;
+      const cWidth = ENCODER_LEXICAL_WIDTH;
+      for (const w of [...aWidths, cWidth]) assert.notEqual(w, bWidth);
+      // B works with the asset absent; C works with both vector runtimes disabled
+      // (C does not depend on B or A — it embeds tokens directly).
+      const cTokens = embedLexical("independently computed lexical with vector runtimes disabled");
+      assert.equal(cTokens.length, ENCODER_LEXICAL_WIDTH);
+    });
+
+    test("router seam enforces the verified per-manifest token capacity: over-cap input routes to B with ENC_SHAPE_INVALID, never an over-cap A VectorSet", () => {
+      // Q01: the router's mode-A path must enforce the VC2A contract
+      // "only batch1/<=maxTokens verified assets reach inference" at its own seam.
+      // A verified asset declaring maxTokens=64 with an input of 100 tokens must
+      // NOT produce an ok:true mode-A VectorSetV1 whose inputTokens breach the
+      // model's declared capacity — instead the router rejects it and falls back
+      // to the asset-free trigram B, reporting the real shape failure code.
+      withFlagsOn(() => {
+        const dir = stageVerifyingAssetDir(64); // verified low-cap manifest
+        try {
+          const over = encodeOrFallback({ tokens: Array.from({ length: 100 }, (_, i) => i) }, dir);
+          assert.equal(over.ok, true, "over-cap input still yields a usable (fallback) verdict");
+          assert.equal(over.mode, "B", "over-cap input must route to the B fallback, not an A VectorSet");
+          if (over.ok) {
+            assert.equal(over.code, ENC_FAIL.SHAPE_INVALID, "reported the real shape failure code");
+            assert.equal(over.vector.length, ENCODER_TRIGRAM_WIDTH);
+          }
+          // A within-cap input against the SAME verified manifest still reaches a
+          // qualified mode-A VectorSet — the capacity rejection is input-scoped,
+          // not a blanket demotion of the verified asset.
+          const within = encodeOrFallback({ tokens: SET_TOKENS }, dir);
+          assert.equal(within.ok, true);
+          assert.equal(within.mode, "A", "within-cap input still reaches mode A");
+          if (within.ok) assert.equal(within.vectorSet.heads.length, 5);
+        } finally {
+          rmSync(dir, { recursive: true, force: true });
+        }
+      });
+    });
+  });
+}

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

@@ -2,7 +2,8 @@
  * vector-cortex/encoder/asset.ts — VC2A asset verification (task 2).
  *
  * Verifies a ModelManifestV1 before any allocation: SHA-256 the ONNX and
- * tokenizer against the manifest digests, require opset 17, batch exactly 1 and
+ * tokenizer against the manifest digests, require opset 21 (ENC-0a re-baseline),
+ * batch exactly 1 and
  * maximum 512 tokens, and confirm the current platform is in the supported
  * matrix. On ANY of these the caller demotes to mode B (asset-free trigram) —
  * never a remote fetch (PREVENT-PI-004). A truncated/unreadable asset during
@@ -88,7 +89,7 @@ function isManifest(m: unknown): m is ModelManifestV1 {
  *
  *   - manifest parse/shape failure      -> ENC_MANIFEST_INVALID -> mode B
  *   - unsupported platform              -> ENC_PLATFORM_UNSUPPORTED -> mode B
- *   - opset != 17                       -> ENC_OPSET_INVALID -> mode B
+ *   - opset != 21                       -> ENC_OPSET_INVALID -> mode B
  *   - batch != 1                        -> ENC_BATCH_INVALID -> mode B
  *   - maxTokens > 512                   -> ENC_TOKENS_EXCEEDED -> mode B
  *   - on-disk digest != manifest digest -> ENC_DIGEST_MISMATCH (one-byte mutation)

package/src/vector-cortex/encoder/bench-export.ts

@@ -17,7 +17,7 @@ export interface BenchGatesV1 {
   readonly latency: boolean;
   /** steady-state marginal RSS over the process baseline <= 150 MiB. */
   readonly rss: boolean;
-  /** the loaded model's declared opset_import equals 17. */
+  /** the loaded model's declared opset_import equals ENCODER_OPSET (21, ENC-0a re-baseline). */
   readonly opset: boolean;
   /** SHA-256 of the embedding output identical across 3 runs (maxAbsDelta=0). */
   readonly determinism: boolean;
@@ -53,7 +53,7 @@ export interface BenchResultV1 {
   readonly rssBaselineMib: number | null;
   /** rssMib - rssBaselineMib: the encoder's marginal footprint. */
   readonly rssMarginalMib: number | null;
-  /** declared opset_import (17); null when no asset manifest is readable. */
+  /** declared opset_import (21, ENC-0a re-baseline); null when no asset manifest is readable. */
   readonly opset: number | null;
   /** true when the output SHA-256 is identical across 3 runs. */
   readonly deterministic: boolean;

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

@@ -0,0 +1,125 @@
+/**
+ * vector-cortex/encoder/decision.ts — ENC-0a backend-decision contract.
+ *
+ * The durable EncoderBackendDecisionV1 record: which ONNX runtime backend the
+ * real learned encoder ships on (transformers.js/WASM vs onnxruntime-node
+ * native), whether the 80 MiB install budget holds, the per-platform install
+ * matrix, the opset baseline (locked 21 by ENC-0a), the license verdict, and
+ * the pinned model/tokenizer sha256 digests.
+ *
+ * ENC-0a is the DECISION + MEASUREMENT sprint: this contract is what the
+ * deterministic resolver (`scripts/encoder/resolve-backend-decision.mjs`) and
+ * the durable record (`docs/vector-cortex/encoder-backend-decision.md`) both
+ * consume. ENC-0a also owns the opset flip: ENCODER_OPSET is re-baselined to 21
+ * in `types.ts` alongside the placeholder manifest (the 2026-08-05 trunk
+ * research dropped the earlier Xenova opset-17 requirement). ENC-0b asserts the
+ * staged real asset is opset 21 — no further constant change.
+ *
+ * Contract-first (ENGINEERING_PRACTICES §3). Pi-agnostic, dependency-free
+ * (PREVENT-PI-004 — local computation only, never a fetch). No `any`
+ * (PREVENT-011).
+ */
+
+import type { EncoderPlatform } from "./types.js";
+
+/** Backend demotion disposition for a platform row. */
+export type BackendDemotion = "none" | "wasm" | "modeB";
+
+/** One platform row: the concrete runtime + install size + demotion disposition. */
+export interface EncoderPlatformRow {
+  readonly runtime: string;
+  readonly installMiB: number;
+  readonly demotion: BackendDemotion;
+}
+
+/**
+ * EncoderBackendDecisionV1 — the locked learned-encoder runtime-backend decision.
+ *
+ * `opset` is pinned to the literal 21 (the ENC-0a re-baseline); `artifacts`
+ * carries only aggregate digests/sizes — never message content (EVAL-REDACT-002);
+ * `blockedBy` names the open hard-gate items that DEFER part of the decision
+ * (e.g. HG-4 — the darwin-x64 demotion is ENC-0e's job).
+ */
+export interface EncoderBackendDecisionV1 {
+  readonly schema: "encoder-backend-decision-v1";
+  readonly backend: "wasm" | "native";
+  readonly budgetOk: boolean;
+  readonly opset: 21;
+  readonly platformMatrix: Readonly<Record<EncoderPlatform, EncoderPlatformRow>>;
+  readonly license: { readonly spdx: "MIT"; readonly redistribution: true };
+  readonly artifacts: {
+    readonly model: { readonly path: string; readonly bytes: number; readonly sha256: string };
+    readonly tokenizer: { readonly path: string; readonly bytes: number; readonly sha256: string };
+  };
+  readonly p95Ms: number | null;
+  readonly blockedBy: readonly string[];
+}
+
+/**
+ * The normative ENC-0a budget: the 80 MiB install/asset cap (MODEL_ASSET.md
+ * §Qualification). Backend qualifies (budgetOk) iff the shipped byte-count fits.
+ */
+export const ENCODER_INSTALL_BUDGET_MIB = 80;
+
+/** The p95 latency gate at 512 tokens / 4 threads on linux-x64 (ms). */
+export const ENCODER_DECISION_P95_MS = 40;
+
+/**
+ * buildDecision — assemble a valid, platform-complete EncoderBackendDecisionV1.
+ *
+ * Pure helper consumed by the acceptance aggregator (and any TS consumer of the
+ * decision). Every EncoderPlatform must resolve to a row, and the passed
+ * per-platform rows must be complete (no row omitted) — a partial matrix is a
+ * contract violation, not a valid decision. All fields are passed in; this is a
+ * structural constructor, not a rules engine (the resolver owns the decision
+ * rule).
+ */
+export function buildDecision(input: {
+  readonly backend: "wasm" | "native";
+  readonly budgetOk: boolean;
+  readonly p95Ms: number | null;
+  readonly platformMatrix: Readonly<Record<EncoderPlatform, EncoderPlatformRow>>;
+  readonly modelPath: string;
+  readonly modelBytes: number;
+  readonly modelSha256: string;
+  readonly tokenizerPath: string;
+  readonly tokenizerBytes: number;
+  readonly tokenizerSha256: string;
+  readonly blockedBy: readonly string[];
+}): EncoderBackendDecisionV1 {
+  const { platformMatrix } = input;
+  const platforms: readonly EncoderPlatform[] = [
+    "linux-x64",
+    "linux-arm64",
+    "darwin-x64",
+    "darwin-arm64",
+    "win32-x64",
+  ];
+  for (const p of platforms) {
+    if (!Object.prototype.hasOwnProperty.call(platformMatrix, p)) {
+      throw new Error(`platform matrix is incomplete: missing row for ${p}`);
+    }
+  }
+  return {
+    schema: "encoder-backend-decision-v1",
+    backend: input.backend,
+    budgetOk: input.budgetOk,
+    opset: 21,
+    platformMatrix,
+    license: { spdx: "MIT", redistribution: true },
+    artifacts: {
+      model: {
+        path: input.modelPath,
+        bytes: input.modelBytes,
+        sha256: input.modelSha256,
+      },
+      tokenizer: {
+        path: input.tokenizerPath,
+        bytes: input.tokenizerBytes,
+        sha256: input.tokenizerSha256,
+      },
+    },
+    p95Ms: input.p95Ms,
+    blockedBy: input.blockedBy,
+  };
+}

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

@@ -17,16 +17,19 @@ import { VC2A_ENABLED } from "../../config/vector-cortex.js";
 
 export type EncoderEmit = (event: string, fields: Record<string, unknown>) => void;
 
-/** The two-event reporter surface consumed by the runtime seams. */
+/** The reporter surface consumed by the runtime + ONNX seams. */
 export interface EncoderReporter {
   readonly assetVerified: (fields: Record<string, unknown>) => void;
   readonly runtimeDemoted: (fields: Record<string, unknown>) => void;
+  /** ENC-0b: fires when a real ONNX InferenceSession is successfully created. */
+  readonly onnxSessionLoaded: (fields: Record<string, unknown>) => void;
 }
 
 /** A flag-gated no-op reporter (zero emissions, default when none injected). */
 export const NOOP_ENCODER_REPORTER: EncoderReporter = {
   assetVerified: () => {},
   runtimeDemoted: () => {},
+  onnxSessionLoaded: () => {},
 };
 
 /**
@@ -47,5 +50,6 @@ export function createEncoderReporter(emit?: EncoderEmit): EncoderReporter {
   return {
     assetVerified: (fields) => fire("vector_cortex_encoder_asset_verified", fields),
     runtimeDemoted: (fields) => fire("vector_cortex_encoder_runtime_demoted", fields),
+    onnxSessionLoaded: (fields) => fire("vector_cortex_encoder_onnx_loaded", fields),
   };
 }

package/src/vector-cortex/encoder/encoder-onnx-dispatch.ts

@@ -0,0 +1,76 @@
+/**
+ * encoder-onnx-dispatch.ts — ENC-0b ONNX dispatch: session lifecycle glue.
+ *
+ * Extracted from runtime.ts so runtime.ts stays under the 300-line soft limit.
+ * Encapsulates the ENC-0b real ONNX session build during load().
+ *
+ * IMPORTANT (interface constraint): EncoderRuntime.load() and infer() are
+ * SYNCHRONOUS — the router (router.ts:121-132) and all acceptance tests call
+ * them synchronously. ONNX session creation is ASYNC (WASM init). ENC-0b
+ * therefore expose the ONNX session as a SEPARATE async verification: load()
+ * fires-and-forgets the session build, verifySession() awaits it for tests,
+ * and infer() continues serving the LCG placeholder until the router is wired
+ * for async inference (ENC-0c scope). This preserves the sync contract while
+ * proving the real ONNX pipeline work end-to-end.
+ *
+ * ENC-0b gate: when MEGACOMPACT_ENC_0B is OFF, none of this code runs —
+ * the LCG stub serves mode A byte-identical to the predecessor.
+ *
+ * Pi-agnostic (PREVENT-PI-004 / PREVENT-011).
+ */
+
+import { ENC_0B_ENABLED } from "../../config/vector-cortex.js";
+import { buildOnnxSession, type OnnxInferenceSession } from "./onnx.js";
+import type { EncoderReporter } from "./emit.js";
+import type { ModelManifestV1 } from "./types.js";
+
+/** State held by the runtime for the ENC-0b ONNX dispatch. */
+export interface OnnxDispatchState {
+  /** The built ONNX session (null when not built, flag off, or build failed). */
+  readonly session: OnnxInferenceSession | null;
+  /** Resolved when the session build attempt settles (ok or fail). */
+  readonly ready: Promise<void>;
+}
+
+/** No-op state when ENC-0b is off or session build not attempted. */
+export const NO_ONNX: OnnxDispatchState = { session: null, ready: Promise.resolve() };
+
+/**
+ * Attempt a real ONNX session build during load(). Fire-and-forget: returns
+ * immediately with a state whose `ready` Promise resolves once the async
+ * build settles. On failure the state's `session` stays null.
+ */
+export function tryBuildOnnx(
+  assetDir: string,
+  manifest: ModelManifestV1,
+  reporter: EncoderReporter,
+  allocatedBytes: number,
+): OnnxDispatchState {
+  if (!ENC_0B_ENABLED()) return NO_ONNX;
+  let session: OnnxInferenceSession | null = null;
+  const ready = buildOnnxSession(assetDir, manifest, reporter, allocatedBytes)
+    .then((result) => { if (result.ok) session = result.session; })
+    .catch(() => {});
+  // Return a state whose `session` getter reads the mutable binding after
+  // ready resolves (the getter runs at test assertion time, not at load()).
+  return { get session() { return session; }, ready };
+}
+
+/**
+ * Await the ONNX session build and verify it with a real inference.
+ * Returns the embedding on success, null on any failure.
+ * Called by tests and the ENC-0b acceptance aggregator, not by production
+ * infer().
+ */
+export async function verifyOnnxSession(
+  state: OnnxDispatchState,
+  tokens: readonly number[],
+): Promise<Float32Array | null> {
+  await state.ready;
+  if (!state.session) return null;
+  try {
+    return await state.session.infer(tokens);
+  } catch {
+    return null;
+  }
+}