pi-mega-compact 0.20.38 → 0.20.40

package/dist/config/vector-cortex-ml5c.js

@@ -0,0 +1,28 @@
+/**
+ * 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 = () => sprintFlag("MEGACOMPACT_ML5_C");

package/dist/config/vector-cortex-ml5d.js

@@ -0,0 +1,28 @@
+/**
+ * config/vector-cortex-ml5d.ts — ML5-D dashboard "Improve Cortex" flag.
+ *
+ * Sibling extract mirroring vector-cortex-ml5a.ts / vector-cortex-ml5b.ts /
+ * vector-cortex-ml5c.ts, so vector-cortex.ts stays under its 300-line soft
+ * limit (soft-as-hard gate). This is the dashboard "Improve Cortex" surface +
+ * promote workflow 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-D adds the dashboard ModelImprovementCard + the POST /api/cortex/improve
+ * and GET /api/cortex/improve/status/:jobId endpoints. The flag gates that
+ * surface only; when OFF the endpoints return 404/disabled and VectorCortexTab
+ * renders exactly as before (byte-identical to the ML5-C-era tab — no
+ * ModelImprovementCard, no improve job is ever spawned).
+ *
+ * Pi-agnostic, dependency-free (PREVENT-PI-004 / PREVENT-011).
+ */
+import { sprintFlag } from "./vector-cortex-flag.js";
+/**
+ * ML5-D — dashboard "Improve Cortex" surface + promote workflow. Default ON.
+ * `MEGACOMPACT_ML5_D=0` disables and is byte-identical to the ML5-C survivor:
+ * both `/api/cortex/improve*` endpoints return 404 and VectorCortexTab omits the
+ * ModelImprovementCard. The flag gates the dashboard surface only; it does not
+ * gate the underlying ML5-A training pipeline or the encoder's mode-A/B/C
+ * selection (those are governed by ML5-A and the encoder independently).
+ */
+export const ML5D_ENABLED = () => sprintFlag("MEGACOMPACT_ML5_D");

package/dist/config/vector-cortex.js

@@ -239,10 +239,9 @@ export const VC8A_ENABLED = () => sprintFlag("MEGACOMPACT_VC8A");
  * mirroring VC4A..VC8A.
  */
 export const VC8B_ENABLED = () => 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.
+// VC8C (canary selection + Rust parity) extracted to vector-cortex-vc8c.ts; re-exported so existing 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";
@@ -250,5 +249,7 @@ 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";
+export { ML5D_ENABLED } from "./vector-cortex-ml5d.js";
 // Breaker constants (TRIAD_RESILIENCE.md §breaker) extracted to vector-cortex-breakers.ts.
 export { BREAKER_WINDOW_MS, BREAKER_MIN_ATTEMPTS, BREAKER_PERF_FAILURES, BREAKER_PERF_FAILURE_RATE, BREAKER_CORRECTNESS_FAILURES, BREAKER_COOLDOWN_MS, BREAKER_PROBE_COUNT, BREAKER_RETRY_BASE_MS, BREAKER_RETRY_CAP_MS, BREAKER_RETRY_JITTER, BREAKER_HYSTERESIS_FAILURE_RATE, BREAKER_HYSTERESIS_BUDGET_P95_MS, BREAKER_MIN_HEALTHY_RESIDENCE_MS, } from "./vector-cortex-breakers.js";

package/dist/config.js

@@ -114,4 +114,4 @@ export const NEW_UI = () => ragEnabled("MEGACOMPACT_NEW_UI");
 // default ON, `=0`/`_DISABLED` off. Re-exported from src/config/vector-cortex.ts
 // so root consumers share one source of truth.
 // ---------------------------------------------------------------------------
-export { VC0A_ENABLED, VC0B_ENABLED, VC1A_ENABLED, VC0C_ENABLED, VC1B_ENABLED, VC1C_ENABLED, VC2A_ENABLED, VC2B_ENABLED, VC2C_ENABLED, VC3A_ENABLED, VC3B_ENABLED, VC3C_ENABLED, VC4A_ENABLED, VC4B_ENABLED, VC4C_ENABLED, VC5A_ENABLED, VC5B_ENABLED, VC5C_ENABLED, VC6A_ENABLED, VC6B_ENABLED, VC6C_ENABLED, VC7A_ENABLED, VC7B_ENABLED, VC7C_ENABLED, VC8A_ENABLED, VC8B_ENABLED, VC8C_ENABLED, VC9A_ENABLED, VC9B_ENABLED, VC9C_ENABLED, VC9D_ENABLED, PCC_ENABLED, ML5A_ENABLED, ML5B_ENABLED, BREAKER_WINDOW_MS, BREAKER_MIN_ATTEMPTS, BREAKER_PERF_FAILURES, BREAKER_PERF_FAILURE_RATE, BREAKER_CORRECTNESS_FAILURES, BREAKER_COOLDOWN_MS, BREAKER_PROBE_COUNT, BREAKER_RETRY_BASE_MS, BREAKER_RETRY_CAP_MS, BREAKER_RETRY_JITTER, BREAKER_HYSTERESIS_FAILURE_RATE, BREAKER_HYSTERESIS_BUDGET_P95_MS, BREAKER_MIN_HEALTHY_RESIDENCE_MS, } from "./config/vector-cortex.js";
+export { VC0A_ENABLED, VC0B_ENABLED, VC1A_ENABLED, VC0C_ENABLED, VC1B_ENABLED, VC1C_ENABLED, VC2A_ENABLED, VC2B_ENABLED, VC2C_ENABLED, VC3A_ENABLED, VC3B_ENABLED, VC3C_ENABLED, VC4A_ENABLED, VC4B_ENABLED, VC4C_ENABLED, VC5A_ENABLED, VC5B_ENABLED, VC5C_ENABLED, VC6A_ENABLED, VC6B_ENABLED, VC6C_ENABLED, VC7A_ENABLED, VC7B_ENABLED, VC7C_ENABLED, VC8A_ENABLED, VC8B_ENABLED, VC8C_ENABLED, VC9A_ENABLED, VC9B_ENABLED, VC9C_ENABLED, VC9D_ENABLED, PCC_ENABLED, ML5A_ENABLED, ML5B_ENABLED, ML5C_ENABLED, ML5D_ENABLED, BREAKER_WINDOW_MS, BREAKER_MIN_ATTEMPTS, BREAKER_PERF_FAILURES, BREAKER_PERF_FAILURE_RATE, BREAKER_CORRECTNESS_FAILURES, BREAKER_COOLDOWN_MS, BREAKER_PROBE_COUNT, BREAKER_RETRY_BASE_MS, BREAKER_RETRY_CAP_MS, BREAKER_RETRY_JITTER, BREAKER_HYSTERESIS_FAILURE_RATE, BREAKER_HYSTERESIS_BUDGET_P95_MS, BREAKER_MIN_HEALTHY_RESIDENCE_MS, } from "./config/vector-cortex.js";

package/dist/extensions/dashboard-server/api-contracts/cortex-improve.js

@@ -0,0 +1,14 @@
+/**
+ * api-contracts/cortex-improve.ts — ML5-D "Improve Cortex" job API types.
+ *
+ * Types for POST /api/cortex/improve (launch a local ML5-A training job that
+ * re-qualifies the five heads against the latest local corpus) and GET
+ * /api/cortex/improve/status/:jobId (poll the in-process job state to a
+ * terminal qualified / demoted_to_B verdict). Local-only, aggregate surfaces:
+ * mode / verdict / digest / reason / progress — never message content or corpus
+ * rows (EVAL-REDACT-002).
+ *
+ * PREVENT-PI-004: type definitions only, no network code.
+ * PREVENT-011: no `any` type — all types are explicit.
+ */
+export {};

package/dist/extensions/dashboard-server/api-contracts/endpoints/registry-ext.js

@@ -54,4 +54,17 @@ export const EXTRA_ENDPOINTS = {
         path: "/api/prefix-stability",
         description: "Per-turn prompt-cache stable-prefix ratio trend (GET /api/prefix-stability?limit=N) read from prefix_stability rows in the local events log. Flag-off (MEGACOMPACT_PC_C=0) returns 404.",
     },
+    // ─── ML5-D Improve Cortex (dashboard Vector Cortex tab) ───────────
+    /** POST /api/cortex/improve — launch a local ML5-A training job. */
+    improveCortex: {
+        method: "POST",
+        path: "/api/cortex/improve",
+        description: "Launch a local ML5-A training job that re-qualifies the five heads against the latest local corpus and returns an opaque jobId. Requires confirm:true; flag-off returns 404. Local-only — never fetches anything.",
+    },
+    /** GET /api/cortex/improve/status/:jobId — poll an improve job. */
+    improveCortexStatus: {
+        method: "GET",
+        path: "/api/cortex/improve/status/:jobId",
+        description: "Poll an in-process Cortex improve job to terminal qualified / demoted_to_B. Unknown jobId or flag-off returns 404. Read-only, in-memory job state.",
+    },
 };

package/dist/extensions/dashboard-server/route-dispatch.js

@@ -8,7 +8,7 @@
  * PREVENT-PI-004: local filesystem read only, no network.
  * PREVENT-011: no `any` type.
  */
-import { handleIndex, handleRepoIndex, handleEvents, handleGameState, handleGameScores, handlePerfSamples, handlePerf, handleAchievements, handleSessions, handleTopics, handleTurns, handleMaintenance, handleProviderCache, handleMemoryStatus, handleCacheStripes, handlePrefixStability, handleSetupStatus, handleSetupDetect, handleSetupConfigure, handleMemoryMap, handleRaptorTree, handleRaptorBuildHistory, handleContextHealth, handleCachePoison, handleHealthSettings, handleEmbedderHealth, handleRagSettings, handleRagMetrics, handleModelThresholds, handleWiki, handleVectorCortexEvaluation, handleVectorCortexHealth, handleVectorCortexBreakersReset, handleVectorCortexLedger, handleVectorCortexTopology, handleVectorCortexQuery, handleVectorCortexShards, handleVectorCortexResidual, handleVectorCortexReconstruct, handleVectorCortexPlans, handleVectorCortexRender, handleVectorCortexRollout, handleVectorCortexClosureProof, handleVectorCortexRestore, handleSetupCortexStatus, handleSetupCortexAction, handleSetupCortexActionLog, } from "./routes.js";
+import { handleIndex, handleRepoIndex, handleEvents, handleGameState, handleGameScores, handlePerfSamples, handlePerf, handleAchievements, handleSessions, handleTopics, handleTurns, handleMaintenance, handleProviderCache, handleMemoryStatus, handleCacheStripes, handlePrefixStability, handleSetupStatus, handleSetupDetect, handleSetupConfigure, handleMemoryMap, handleRaptorTree, handleRaptorBuildHistory, handleContextHealth, handleCachePoison, handleHealthSettings, handleEmbedderHealth, handleRagSettings, handleRagMetrics, handleModelThresholds, handleWiki, handleVectorCortexEvaluation, handleVectorCortexHealth, handleVectorCortexBreakersReset, handleVectorCortexLedger, handleVectorCortexTopology, handleVectorCortexQuery, handleVectorCortexShards, handleVectorCortexResidual, handleVectorCortexReconstruct, handleVectorCortexPlans, handleVectorCortexRender, handleVectorCortexRollout, handleVectorCortexClosureProof, handleVectorCortexRestore, handleSetupCortexStatus, handleSetupCortexAction, handleSetupCortexActionLog, handleImproveCortex, handleImproveCortexStatus, } from "./routes.js";
 // VC6C repair lives in its own module (routes-vector-cortex-repair.ts) so the
 // heal route file stays well under the 400-line extension soft limit.
 import { handleVectorCortexRepair } from "./routes-vector-cortex-repair.js";
@@ -149,5 +149,9 @@ export function dispatchRoutes(req, res, ctx) {
         return true;
     if (handleSetupCortexActionLog(req, res, ctx))
         return true;
+    if (handleImproveCortex(req, res, ctx))
+        return true;
+    if (handleImproveCortexStatus(req, res, ctx))
+        return true;
     return false;
 }

package/dist/extensions/dashboard-server/routes-cortex-improve.js

@@ -0,0 +1,220 @@
+/**
+ * dashboard-server/routes-cortex-improve.ts — ML5-D "Improve Cortex" job routes.
+ *
+ *   POST /api/cortex/improve          — confirmation-gated local ML5-A training
+ *   GET  /api/cortex/improve/status/:jobId — poll an in-process job to terminal
+ *
+ * The Improve action launches the committed ML5-A training pipeline
+ * (training/vector-cortex/train.py) as a background child_process against the
+ * latest local corpus, then re-qualifies the five heads. Job state is kept
+ * in-process in a Map<jobId, JobState> on this module; nothing is persisted
+ * across restarts (the status endpoint is read-only, in-memory).
+ *
+ * LOCAL ONLY (PREVENT-PI-004): the job spawns a local python process and reads
+ * local files — never a fetch. The `onnxruntime-node` native path is attempted
+ * only when MEGACOMPACT_ENCODER_NATIVE=1 (default OFF); otherwise the harness
+ * uses the WASM/trigram fallback. The status payload surfaces mode/verdict/
+ * digest/reason/progress only — never message or corpus content (EVAL-REDACT-002).
+ *
+ * Guardrails: PREVENT-011 (no `any`), PREVENT-001 (guarded JSON.parse via
+ * readJsonBody), flag-off byte-identical (no card + endpoints 404).
+ */
+import { spawn } from "node:child_process";
+import { createHash, randomBytes } from "node:crypto";
+import { readFileSync, existsSync } from "node:fs";
+import { join, dirname } from "node:path";
+import { fileURLToPath } from "node:url";
+import { sendJson, readJsonBody } from "./routes-vector-cortex-shared.js";
+import { ML5D_ENABLED } from "../../src/config.js";
+import { qualifyDecision } from "../../src/vector-cortex/improve.js";
+/** Resolve the committed training entry by walking up to the repo root. */
+function trainingScript() {
+    let dir = dirname(fileURLToPath(import.meta.url));
+    const rel = join("training", "vector-cortex", "train.py");
+    for (let i = 0; i < 8; i++) {
+        // guardrails-allow PREVENT-PI-004: local script read (loopback)
+        const candidate = join(dir, rel);
+        if (existsSync(candidate))
+            return candidate;
+        const next = dirname(dir);
+        if (next === dir)
+            break;
+        dir = next;
+    }
+    return null;
+}
+/** In-process job registry; nothing survives a server restart. */
+const JOBS = new Map();
+/** Opaque job token: sha256(ts+random) hex-sliced. */
+function newJobId() {
+    return createHash("sha256")
+        .update(`${Date.now()}:${randomBytes(16).toString("hex")}`)
+        .digest("hex")
+        .slice(0, 24);
+}
+/**
+ * Spawn the ML5-A training pipeline in the background and drive the job state
+ * to a terminal qualified / demoted_to_B verdict. Honest: an empty corpus (the
+ * host state) makes train.py no-op and the job ends demoted_to_B.
+ */
+function startJob(jobId) {
+    const script = trainingScript();
+    const boot = Date.now();
+    JOBS.set(jobId, { status: "improving", progress: 0, updatedAt: new Date().toISOString() });
+    if (script === null) {
+        settle(jobId, { status: "demoted_to_B", reason: "ENC_TRAIN_PIPELINE_ABSENT", progress: 1 });
+        return;
+    }
+    const env = { ...process.env };
+    const proc = spawn("python3", [script], {
+        env,
+        cwd: join(dirname(script), "..", ".."),
+        stdio: ["ignore", "pipe", "pipe"],
+    });
+    let output = "";
+    const step = () => {
+        const pct = Math.min(0.9, 0.1 + ((Date.now() - boot) / 60000) * 0.6);
+        const cur = JOBS.get(jobId);
+        if (cur && cur.status === "improving" && pct > cur.progress) {
+            JOBS.set(jobId, { ...cur, progress: pct, updatedAt: new Date().toISOString() });
+        }
+    };
+    const timer = setInterval(step, 2000);
+    const stop = () => clearInterval(timer);
+    proc.stdout?.on("data", (d) => {
+        output += d.toString();
+    });
+    proc.stderr?.on("data", (d) => {
+        output += d.toString();
+    });
+    proc.on("error", (err) => {
+        stop();
+        settle(jobId, {
+            status: "demoted_to_B",
+            reason: `ENC_TRAIN_SPAWN_FAILED:${err.message}`,
+            progress: 1,
+        });
+    });
+    proc.on("close", (code) => {
+        stop();
+        // Read the produced trained-heads artifact digest. The report line includes
+        // `trainedHeadsPath`; verify the file exists after a successful run.
+        const match = /trainedHeadsPath":\s*"([^"]+)"/.exec(output);
+        const assetPath = match?.[1];
+        let digest = null;
+        if (assetPath) {
+            // guardrails-allow PREVENT-PI-004: local artifact read (loopback)
+            try {
+                digest = createHash("sha256")
+                    .update(readFileSync(assetPath))
+                    .digest("hex")
+                    .slice(0, 12);
+            }
+            catch {
+                digest = null;
+            }
+        }
+        const decision = qualifyDecision(code ?? -1, digest);
+        if (decision === "qualified") {
+            settle(jobId, {
+                status: "qualified",
+                progress: 1,
+                assetDigest: digest ?? undefined,
+                verdict: { mode: "A", assetDigestPrefix: digest ?? null, verdict: "qualified" },
+            });
+        }
+        else {
+            settle(jobId, {
+                status: "demoted_to_B",
+                reason: assetPath
+                    ? "trained asset did not verify — demoted to mode B"
+                    : "empty corpus (no groups) — no qualified asset emitted",
+                progress: 1,
+            });
+        }
+    });
+}
+/** Write a terminal job state (idempotent — only advances from improving). */
+function settle(jobId, terminal) {
+    const cur = JOBS.get(jobId);
+    if (cur && cur.status !== "improving")
+        return;
+    JOBS.set(jobId, { ...terminal, updatedAt: new Date().toISOString() });
+}
+/** Flag-off response, byte-identical regardless of request (ML5-D absent). */
+function sendDisabled(res) {
+    sendJson(res, 404, { error: "disabled" });
+}
+/**
+ * POST /api/cortex/improve (ML5-D). Returns true when it claims the request.
+ * Requires `confirm:true` server-side (mirrors the client window.confirm).
+ */
+export function handleImproveCortex(req, res, _ctx) {
+    const url = req.url ?? "";
+    if (url !== "/api/cortex/improve")
+        return false;
+    if (req.method !== "POST") {
+        sendJson(res, 405, { error: "method_not_allowed" });
+        return true;
+    }
+    if (!ML5D_ENABLED()) {
+        sendDisabled(res);
+        return true;
+    }
+    readJsonBody(req, (parsed) => {
+        if (!parsed.ok) {
+            sendJson(res, 400, { error: "invalid_body" });
+            return;
+        }
+        if (parsed.value.confirm !== true) {
+            sendJson(res, 400, { error: "confirmation_required" });
+            return;
+        }
+        const jobId = newJobId();
+        startJob(jobId);
+        const body = { status: "improving", jobId };
+        sendJson(res, 200, body);
+    });
+    return true;
+}
+/**
+ * GET /api/cortex/improve/status/:jobId (ML5-D). Returns true when it claims the
+ * request. Flag-off or unknown jobId → 404. Read-only in-memory.
+ */
+export function handleImproveCortexStatus(req, res, _ctx) {
+    const url = req.url ?? "";
+    const match = /^\/api\/cortex\/improve\/status\/([A-Za-z0-9]+)$/.exec(url);
+    if (!match)
+        return false;
+    if (req.method !== "GET") {
+        sendJson(res, 405, { error: "method_not_allowed" });
+        return true;
+    }
+    if (!ML5D_ENABLED()) {
+        sendDisabled(res);
+        return true;
+    }
+    const job = JOBS.get(match[1]);
+    if (!job) {
+        sendJson(res, 404, { error: "job_not_found" });
+        return true;
+    }
+    // Narrow job.status into the discriminated CortexImproveStatus variant. The
+    // terminal shapes require their payload fields, so construct per-state.
+    const body = job.status === "improving"
+        ? { status: "improving", progress: job.progress }
+        : job.status === "qualified"
+            ? {
+                status: "qualified",
+                progress: job.progress,
+                verdict: job.verdict ?? { mode: "A", assetDigestPrefix: null, verdict: "qualified" },
+                assetDigest: job.assetDigest ?? "",
+            }
+            : {
+                status: "demoted_to_B",
+                progress: job.progress,
+                reason: job.reason ?? "UNKNOWN_TERMINAL",
+            };
+    sendJson(res, 200, body);
+    return true;
+}

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

@@ -54,5 +54,7 @@ export const VECTOR_CORTEX_SETTINGS = {
         boolDirect("MEGACOMPACT_VC9D", "VC9D Embedder Detect Consolidation", "Embedder-detect consolidation + VC9 workstream roll-up: memoizes /api/setup-detect against the mutable input (resolved binary path + mtime) so consecutive requests reuse the result without re-spawning, and unifies the embedder + cortex sub-tabs' 5s poll contract. OFF = byte-identical predecessor (VC9C-era): detect spawns fresh per request and the embedder poll keeps its previous cadence.", true),
         boolDirect("MEGACOMPACT_PC_C", "PC-C Dashboard Cache Visibility", "Dashboard per-turn prompt-cache visibility: surfaces the per-turn stable-prefix ratio trend (GET /api/prefix-stability) in the CacheTab PrefixStabilityCard. Reads aggregate ratios/counts from the local monitoring events log only — no payload bytes. OFF = byte-identical predecessor (PC-B-era): /api/prefix-stability returns 404 and the CacheTab omits the PrefixStabilityCard.", true),
         boolDirect("MEGACOMPACT_ML5_A", "ML5-A Five-Head Training Load", "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),
+        boolDirect("MEGACOMPACT_ML5_D", "ML5-D Dashboard Improve Cortex", "ML5-D dashboard 'Improve Cortex' surface: the ModelImprovementCard + POST /api/cortex/improve + GET /api/cortex/improve/status/:jobId. ON (default) = the card renders and Improve launches a local ML5-A training job. OFF = both improve endpoints return 404 and VectorCortexTab omits the card, byte-identical to ML5-C.", true),
     ],
 };

package/dist/extensions/dashboard-server/routes.js

@@ -27,3 +27,4 @@ export { handleWiki } from "./routes-wiki.js";
 export { handleVectorCortexEvaluation, handleVectorCortexHealth, handleVectorCortexBreakersReset, handleVectorCortexLedger, handleVectorCortexTopology, handleVectorCortexQuery, handleVectorCortexShards, handleVectorCortexResidual, handleVectorCortexReconstruct, handleVectorCortexPlans, handleVectorCortexRender, handleVectorCortexRollout, handleVectorCortexClosureProof, handleVectorCortexRestore, } from "./routes-vector-cortex.js";
 export { handleSetupCortexStatus } from "./routes-setup-cortex.js";
 export { handleSetupCortexAction, handleSetupCortexActionLog, } from "./routes-setup-cortex-actions.js";
+export { handleImproveCortex, handleImproveCortexStatus, } from "./routes-cortex-improve.js";

package/dist/extensions/mega-runtime/dashboard-snapshot.js

@@ -97,8 +97,8 @@ export function buildDashboardSnapshot(ctx) {
             // per-session in rt and there's no cumulative counter yet.
             total: ctx.repo.dedupCollapsed,
             sessionTokensSaved: ctx.rt.cacheHitTokens,
-            // Cumulative tokens saved across all sessions in this repo —
-            // was a placeholder (dedupCollapsed * 100).
+            // Cumulative tokens saved across all sessions in this repo = the
+            // real repo counter (audit Table 1 stub 8 closed — no rolled-up math).
             totalTokensSaved: ctx.repo.tokensSaved,
         },
         compacts: {

package/dist/src/config/vector-cortex-ml5c.js

@@ -0,0 +1,28 @@
+/**
+ * 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 = () => sprintFlag("MEGACOMPACT_ML5_C");

package/dist/src/config/vector-cortex-ml5d.js

@@ -0,0 +1,28 @@
+/**
+ * config/vector-cortex-ml5d.ts — ML5-D dashboard "Improve Cortex" flag.
+ *
+ * Sibling extract mirroring vector-cortex-ml5a.ts / vector-cortex-ml5b.ts /
+ * vector-cortex-ml5c.ts, so vector-cortex.ts stays under its 300-line soft
+ * limit (soft-as-hard gate). This is the dashboard "Improve Cortex" surface +
+ * promote workflow 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-D adds the dashboard ModelImprovementCard + the POST /api/cortex/improve
+ * and GET /api/cortex/improve/status/:jobId endpoints. The flag gates that
+ * surface only; when OFF the endpoints return 404/disabled and VectorCortexTab
+ * renders exactly as before (byte-identical to the ML5-C-era tab — no
+ * ModelImprovementCard, no improve job is ever spawned).
+ *
+ * Pi-agnostic, dependency-free (PREVENT-PI-004 / PREVENT-011).
+ */
+import { sprintFlag } from "./vector-cortex-flag.js";
+/**
+ * ML5-D — dashboard "Improve Cortex" surface + promote workflow. Default ON.
+ * `MEGACOMPACT_ML5_D=0` disables and is byte-identical to the ML5-C survivor:
+ * both `/api/cortex/improve*` endpoints return 404 and VectorCortexTab omits the
+ * ModelImprovementCard. The flag gates the dashboard surface only; it does not
+ * gate the underlying ML5-A training pipeline or the encoder's mode-A/B/C
+ * selection (those are governed by ML5-A and the encoder independently).
+ */
+export const ML5D_ENABLED = () => sprintFlag("MEGACOMPACT_ML5_D");

package/dist/src/config/vector-cortex.js

@@ -239,10 +239,9 @@ export const VC8A_ENABLED = () => sprintFlag("MEGACOMPACT_VC8A");
  * mirroring VC4A..VC8A.
  */
 export const VC8B_ENABLED = () => 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.
+// VC8C (canary selection + Rust parity) extracted to vector-cortex-vc8c.ts; re-exported so existing 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";
@@ -250,5 +249,7 @@ 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";
+export { ML5D_ENABLED } from "./vector-cortex-ml5d.js";
 // Breaker constants (TRIAD_RESILIENCE.md §breaker) extracted to vector-cortex-breakers.ts.
 export { BREAKER_WINDOW_MS, BREAKER_MIN_ATTEMPTS, BREAKER_PERF_FAILURES, BREAKER_PERF_FAILURE_RATE, BREAKER_CORRECTNESS_FAILURES, BREAKER_COOLDOWN_MS, BREAKER_PROBE_COUNT, BREAKER_RETRY_BASE_MS, BREAKER_RETRY_CAP_MS, BREAKER_RETRY_JITTER, BREAKER_HYSTERESIS_FAILURE_RATE, BREAKER_HYSTERESIS_BUDGET_P95_MS, BREAKER_MIN_HEALTHY_RESIDENCE_MS, } from "./vector-cortex-breakers.js";

package/dist/src/config.js

@@ -114,4 +114,4 @@ export const NEW_UI = () => ragEnabled("MEGACOMPACT_NEW_UI");
 // default ON, `=0`/`_DISABLED` off. Re-exported from src/config/vector-cortex.ts
 // so root consumers share one source of truth.
 // ---------------------------------------------------------------------------
-export { VC0A_ENABLED, VC0B_ENABLED, VC1A_ENABLED, VC0C_ENABLED, VC1B_ENABLED, VC1C_ENABLED, VC2A_ENABLED, VC2B_ENABLED, VC2C_ENABLED, VC3A_ENABLED, VC3B_ENABLED, VC3C_ENABLED, VC4A_ENABLED, VC4B_ENABLED, VC4C_ENABLED, VC5A_ENABLED, VC5B_ENABLED, VC5C_ENABLED, VC6A_ENABLED, VC6B_ENABLED, VC6C_ENABLED, VC7A_ENABLED, VC7B_ENABLED, VC7C_ENABLED, VC8A_ENABLED, VC8B_ENABLED, VC8C_ENABLED, VC9A_ENABLED, VC9B_ENABLED, VC9C_ENABLED, VC9D_ENABLED, PCC_ENABLED, ML5A_ENABLED, ML5B_ENABLED, BREAKER_WINDOW_MS, BREAKER_MIN_ATTEMPTS, BREAKER_PERF_FAILURES, BREAKER_PERF_FAILURE_RATE, BREAKER_CORRECTNESS_FAILURES, BREAKER_COOLDOWN_MS, BREAKER_PROBE_COUNT, BREAKER_RETRY_BASE_MS, BREAKER_RETRY_CAP_MS, BREAKER_RETRY_JITTER, BREAKER_HYSTERESIS_FAILURE_RATE, BREAKER_HYSTERESIS_BUDGET_P95_MS, BREAKER_MIN_HEALTHY_RESIDENCE_MS, } from "./config/vector-cortex.js";
+export { VC0A_ENABLED, VC0B_ENABLED, VC1A_ENABLED, VC0C_ENABLED, VC1B_ENABLED, VC1C_ENABLED, VC2A_ENABLED, VC2B_ENABLED, VC2C_ENABLED, VC3A_ENABLED, VC3B_ENABLED, VC3C_ENABLED, VC4A_ENABLED, VC4B_ENABLED, VC4C_ENABLED, VC5A_ENABLED, VC5B_ENABLED, VC5C_ENABLED, VC6A_ENABLED, VC6B_ENABLED, VC6C_ENABLED, VC7A_ENABLED, VC7B_ENABLED, VC7C_ENABLED, VC8A_ENABLED, VC8B_ENABLED, VC8C_ENABLED, VC9A_ENABLED, VC9B_ENABLED, VC9C_ENABLED, VC9D_ENABLED, PCC_ENABLED, ML5A_ENABLED, ML5B_ENABLED, ML5C_ENABLED, ML5D_ENABLED, BREAKER_WINDOW_MS, BREAKER_MIN_ATTEMPTS, BREAKER_PERF_FAILURES, BREAKER_PERF_FAILURE_RATE, BREAKER_CORRECTNESS_FAILURES, BREAKER_COOLDOWN_MS, BREAKER_PROBE_COUNT, BREAKER_RETRY_BASE_MS, BREAKER_RETRY_CAP_MS, BREAKER_RETRY_JITTER, BREAKER_HYSTERESIS_FAILURE_RATE, BREAKER_HYSTERESIS_BUDGET_P95_MS, BREAKER_MIN_HEALTHY_RESIDENCE_MS, } from "./config/vector-cortex.js";

package/dist/src/store/backfill.js

@@ -24,7 +24,6 @@ import { buildRaptorTree } from "../dedup/raptor/tree.js";
 import { defaultEmbedder } from "../embedder.js";
 import { getStateDir } from "../store.js";
 const BATCH = 1000;
-const THROTTLE_MS = 0; // synchronous backfill; no cross-process yield needed
 function ensureProgressTable(db) {
     db.exec(`
     CREATE TABLE IF NOT EXISTS backfill_progress (
@@ -92,10 +91,6 @@ export function backfillContentHashes(stateDir = getStateDir()) {
         withTx(db, () => applyRows(pending));
         db.prepare("INSERT INTO backfill_progress(name, last_session_id, last_id, updated, duplicates_resolved) VALUES('content_hashes',?,?,?,?) ON CONFLICT(name) DO UPDATE SET last_session_id=excluded.last_session_id, last_id=excluded.last_id, updated=excluded.updated, duplicates_resolved=excluded.duplicates_resolved").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 };
 }
 /** True when no rows remain pending (backfill complete for this state dir). */
@@ -151,10 +146,6 @@ export function backfillPhase(phase, sessionId, stateDir, opts = {}) {
         });
         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/dist/src/vector-cortex/encoder/runtime-emit.js

@@ -0,0 +1,42 @@
+/**
+ * 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";
+/**
+ * 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, result) {
+    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/dist/src/vector-cortex/encoder/runtime-native.js

@@ -0,0 +1,77 @@
+/**
+ * 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";
+/** True when `MEGACOMPACT_ENCODER_NATIVE=1` (the native opt-in operator flag). */
+export function nativeOptIn() {
+    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() {
+    try {
+        // @ts-expect-error — optional peer; the shadow type above covers the surface
+        const mod = (await import("onnxruntime-node"));
+        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, options = {}) {
+    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) {
+                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;
+    }
+}