pi-mega-compact 0.20.36 → 0.20.39

package/dist/config/vector-cortex-ml5b.js

@@ -0,0 +1,26 @@
+/**
+ * config/vector-cortex-ml5b.ts — ML5-B production bench harness flag.
+ *
+ * Sibling extract mirroring vector-cortex-ml5a.ts, so vector-cortex.ts stays
+ * under its 300-line soft limit (soft-as-hard gate). This is the ONNX Runtime
+ * evaluation/benchmark 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-B introduces NO runtime code path: the bench harness and corpus export
+ * are developer/evidence tooling (scripts/) plus a consumer-facing TypeScript
+ * shell (src/vector-cortex/encoder/bench.ts) that only writes monitoring
+ * events. The flag records intent and scopes the sprint's evidence assets; it
+ * gates nothing at runtime today. There is no HTTP endpoint and no dashboard
+ * change, so there is no SETTINGS toggle and no EXCLUDED_SETTINGS interaction.
+ *
+ * Pi-agnostic, dependency-free (PREVENT-PI-004 / PREVENT-011).
+ */
+import { sprintFlag } from "./vector-cortex-flag.js";
+/**
+ * ML5-B — production bench harness (ONNX Runtime eval). Default ON.
+ * `MEGACOMPACT_ML5_B=0` disables and is byte-identical to the ML5-A survivor:
+ * no bench endpoint exists and mode B continues to serve all clients exactly as
+ * before. The flag does not gate the harness itself — the harness is an on-demand
+ * developer tool with no runtime path.
+ */
+export const ML5B_ENABLED = () => sprintFlag("MEGACOMPACT_ML5_B");

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.js

@@ -242,12 +242,14 @@ 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.
 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";
 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 { 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, 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, 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/routes-rag-settings-vector-cortex.js

@@ -54,5 +54,6 @@ 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),
     ],
 };

package/dist/monitoring.js

@@ -0,0 +1,172 @@
+/**
+ * monitoring.ts — local dedup monitoring + alerting (Sprint 14, Phase 7).
+ *
+ * Per-decision structured events go to `events.log` (append-only JSON).
+ * Aggregate metrics (hit rate, FP rate, per-tier p95 latency, storage) go to
+ * `dashboard.json` — the SAME local-only file the /dashboard UI reads. There is
+ * NO Prometheus port and NO network listener (PREVENT-PI-004). Alerting is local
+ * only: an FP-rate breach flips the tier to MARK_ONLY and writes a warning.
+ *
+ * Best-effort: logging/metrics never throw into the add()/search() path.
+ */
+import { readFileSync, writeFileSync, existsSync, mkdirSync, appendFileSync } from "node:fs";
+import { dirname, join } from "node:path";
+import { STATE_DIR_DEFAULT } from "./config.js";
+const TIERS = ["L0", "L1", "L2", "RAPTOR"];
+function emptyMetrics() {
+    const dec = {};
+    const dp = {};
+    const fp = {};
+    const lat = {};
+    for (const t of TIERS) {
+        dec[t] = 0;
+        dp[t] = 0;
+        fp[t] = 0;
+        lat[t] = [];
+    }
+    return { decisions: dec, deduped: dp, falsePositives: fp, latency: lat, storageBytes: 0 };
+}
+/** Append a structured decision event to events.log (best-effort). */
+export function logDecision(path, ev) {
+    try {
+        mkdirSync(dirname(path), { recursive: true });
+        appendFileSync(path, `${JSON.stringify(ev)}\n`);
+    }
+    catch {
+        /* never break the extension on a log failure */
+    }
+}
+/**
+ * Load metrics from dashboard.json, or return a fresh empty snapshot.
+ * Kept simple + synchronous (no network).
+ */
+export function loadMetrics(path) {
+    try {
+        if (existsSync(path)) {
+            const raw = readFileSync(path, "utf-8");
+            const parsed = JSON.parse(raw);
+            const base = emptyMetrics();
+            return {
+                decisions: { ...base.decisions, ...(parsed.decisions ?? {}) },
+                deduped: { ...base.deduped, ...(parsed.deduped ?? {}) },
+                falsePositives: { ...base.falsePositives, ...(parsed.falsePositives ?? {}) },
+                latency: { ...base.latency, ...(parsed.latency ?? {}) },
+                storageBytes: parsed.storageBytes ?? 0,
+            };
+        }
+    }
+    catch {
+        /* corrupt metrics → fresh */
+    }
+    return emptyMetrics();
+}
+/** Persist metrics to dashboard.json (best-effort). */
+export function saveMetrics(path, m) {
+    try {
+        mkdirSync(dirname(path), { recursive: true });
+        writeFileSync(path, JSON.stringify(m));
+    }
+    catch {
+        /* never break the extension */
+    }
+}
+/** Compute the p95 latency (ms) for a tier from its samples. */
+export function p95(samples) {
+    if (samples.length === 0)
+        return 0;
+    const sorted = [...samples].sort((a, b) => a - b);
+    const idx = Math.min(sorted.length - 1, Math.floor(sorted.length * 0.95));
+    return sorted[idx];
+}
+/** FP rate for a tier over the current window (0..1). */
+export function fpRate(m, tier) {
+    const decisions = m.decisions[tier] ?? 0;
+    if (decisions === 0)
+        return 0;
+    return (m.falsePositives[tier] ?? 0) / decisions;
+}
+/**
+ * Evaluate FP-rate breaches against the config thresholds. A breached fuzzy tier
+ * (L0 vs L1/L2 have different thresholds) is auto-downgraded to MARK_ONLY — the
+ * local re-map of "alertmanager" (QA #18/#19): record but don't collapse, no
+ * remote alert. Returns the tiers flipped so the caller can mutate its config.
+ */
+export function evaluateAlerts(m, cfg) {
+    const breached = [];
+    const warnings = [];
+    for (const tier of TIERS) {
+        const rate = fpRate(m, tier);
+        const limit = tier === "L0" ? cfg.FP_RATE_L0 : cfg.FP_RATE_L1L2;
+        if (rate > limit) {
+            breached.push(tier);
+            warnings.push(`DEDUP FP BREACH tier=${tier} rate=${rate.toFixed(4)} > ${limit}`);
+        }
+    }
+    return { breached, warnings };
+}
+/**
+ * Record one decision into the metrics snapshot (mutates `m` in place) and
+ * returns the updated snapshot. Caps stored latency samples to keep memory
+ * bounded (last 1000 per tier).
+ */
+export function recordDecision(m, tier, result, latencyMs, falsePositive = false) {
+    m.decisions[tier] = (m.decisions[tier] ?? 0) + 1;
+    if (result === "deduped")
+        m.deduped[tier] = (m.deduped[tier] ?? 0) + 1;
+    if (falsePositive)
+        m.falsePositives[tier] = (m.falsePositives[tier] ?? 0) + 1;
+    const arr = m.latency[tier] ?? (m.latency[tier] = []);
+    arr.push(latencyMs);
+    if (arr.length > 1000)
+        arr.shift();
+    return m;
+}
+/** Default metrics path alongside the state dir. */
+export function defaultMetricsPath(stateDir = STATE_DIR_DEFAULT) {
+    return join(stateDir, "dashboard.json");
+}
+/** Default events-log path alongside the state dir. */
+export function defaultEventsPath(stateDir = STATE_DIR_DEFAULT) {
+    return join(stateDir, "events.log");
+}
+/**
+ * Append a RecallQualityEvent to events.log (best-effort, non-fatal).
+ * Follows the same append-one-JSON-line pattern as logDecision.
+ */
+export function logRecallQuality(path, ev) {
+    try {
+        mkdirSync(dirname(path), { recursive: true });
+        appendFileSync(path, JSON.stringify(ev) + "\n", "utf-8");
+    }
+    catch {
+        /* best-effort — never break the extension */
+    }
+}
+// ---------------------------------------------------------------------------
+// ML5-B encoder bench events (consumer-facing for the dashboard + evidence)
+// ---------------------------------------------------------------------------
+/**
+ * Append a structured ML5-B bench event to events.log (best-effort, non-fatal).
+ * Mirrors the extension's appendEvent schema ({ ts, event, ...fields }) so the
+ * dashboard live-stream tail and evidence tooling parse the four
+ * `vector_cortex_encoder_bench_*` events identically. The bench is developer/
+ * evidence tooling with no runtime gating; this only records its results.
+ */
+export function logBenchEvent(path, event, fields) {
+    try {
+        mkdirSync(dirname(path), { recursive: true });
+        appendFileSync(path, JSON.stringify({ ts: Date.now(), event, ...fields }) + "\n", "utf8");
+    }
+    catch {
+        /* best-effort — never break the caller */
+    }
+}
+// ---------------------------------------------------------------------------
+// Dedup audit trail (external-audit item #2)
+// ---------------------------------------------------------------------------
+// The event shape and its append helper are DEFINED in vectorStore/dedup-audit.ts,
+// co-located with the only recorder that emits them (this file already carries
+// decision events + the metrics snapshot + FP alerting and must stay under its
+// 300-line soft limit). Re-exported here so callers that treat monitoring.ts as
+// the events.log barrel — including the dashboard SSE tail — keep one import.
+export { logDedupAudit } from "./vectorStore/dedup-audit.js";

package/dist/src/config/vector-cortex-ml5b.js

@@ -0,0 +1,26 @@
+/**
+ * config/vector-cortex-ml5b.ts — ML5-B production bench harness flag.
+ *
+ * Sibling extract mirroring vector-cortex-ml5a.ts, so vector-cortex.ts stays
+ * under its 300-line soft limit (soft-as-hard gate). This is the ONNX Runtime
+ * evaluation/benchmark 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-B introduces NO runtime code path: the bench harness and corpus export
+ * are developer/evidence tooling (scripts/) plus a consumer-facing TypeScript
+ * shell (src/vector-cortex/encoder/bench.ts) that only writes monitoring
+ * events. The flag records intent and scopes the sprint's evidence assets; it
+ * gates nothing at runtime today. There is no HTTP endpoint and no dashboard
+ * change, so there is no SETTINGS toggle and no EXCLUDED_SETTINGS interaction.
+ *
+ * Pi-agnostic, dependency-free (PREVENT-PI-004 / PREVENT-011).
+ */
+import { sprintFlag } from "./vector-cortex-flag.js";
+/**
+ * ML5-B — production bench harness (ONNX Runtime eval). Default ON.
+ * `MEGACOMPACT_ML5_B=0` disables and is byte-identical to the ML5-A survivor:
+ * no bench endpoint exists and mode B continues to serve all clients exactly as
+ * before. The flag does not gate the harness itself — the harness is an on-demand
+ * developer tool with no runtime path.
+ */
+export const ML5B_ENABLED = () => sprintFlag("MEGACOMPACT_ML5_B");

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.js

@@ -242,12 +242,14 @@ 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.
 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";
 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 { 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, 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, 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/monitoring.js

@@ -143,6 +143,25 @@ export function logRecallQuality(path, ev) {
     }
 }
 // ---------------------------------------------------------------------------
+// ML5-B encoder bench events (consumer-facing for the dashboard + evidence)
+// ---------------------------------------------------------------------------
+/**
+ * Append a structured ML5-B bench event to events.log (best-effort, non-fatal).
+ * Mirrors the extension's appendEvent schema ({ ts, event, ...fields }) so the
+ * dashboard live-stream tail and evidence tooling parse the four
+ * `vector_cortex_encoder_bench_*` events identically. The bench is developer/
+ * evidence tooling with no runtime gating; this only records its results.
+ */
+export function logBenchEvent(path, event, fields) {
+    try {
+        mkdirSync(dirname(path), { recursive: true });
+        appendFileSync(path, JSON.stringify({ ts: Date.now(), event, ...fields }) + "\n", "utf8");
+    }
+    catch {
+        /* best-effort — never break the caller */
+    }
+}
+// ---------------------------------------------------------------------------
 // Dedup audit trail (external-audit item #2)
 // ---------------------------------------------------------------------------
 // The event shape and its append helper are DEFINED in vectorStore/dedup-audit.ts,

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 (
@@ -92,9 +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.
-    }
     return { processed, updated, duplicatesResolved };
 }
 /** True when no rows remain pending (backfill complete for this state dir). */
@@ -150,10 +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/bench-export.js

@@ -0,0 +1,13 @@
+/**
+ * vector-cortex/encoder/bench-export.ts — ML5-B bench result contract.
+ *
+ * BenchResultV1 is the typed surface `bench.ts` parses from the qualification
+ * harness (`scripts/ml5/bench-onnx-prod.mjs`) and the dashboard / evidence
+ * tooling consume. It carries AGGREGATE measurements + a digest only — never
+ * chunk/message content (EVAL-REDACT-002).
+ *
+ * Contract-first (ENGINEERING_PRACTICES §3): this types file is the reviewed
+ * gate; implementations import from it. Pi-agnostic, dependency-free
+ * (PREVENT-PI-004 / PREVENT-011).
+ */
+export {};

package/dist/src/vector-cortex/encoder/bench.js

@@ -0,0 +1,100 @@
+/**
+ * vector-cortex/encoder/bench.ts — ML5-B bench runner (consumer-facing shell).
+ *
+ * Calls `scripts/ml5/bench-onnx-prod.mjs` via child_process, parses the
+ * BenchResultV1 it emits, and writes the four `vector_cortex_encoder_bench_*`
+ * events to the monitoring events.log (the dashboard / ML5-D surface consume
+ * them later). This is NOT a runtime path — it is developer/evidence tooling.
+ *
+ * Events written (all best-effort / non-fatal):
+ *   - vector_cortex_encoder_bench_p95_ms
+ *   - vector_cortex_encoder_bench_rss_mib
+ *   - vector_cortex_encoder_bench_opset_ok
+ *   - vector_cortex_encoder_bench_deterministic
+ *
+ * Pi-agnostic, dependency-free (PREVENT-PI-004 — the child bench is pure local
+ * computation). No `any` (PREVENT-011).
+ */
+import { spawnSync } from "node:child_process";
+import { dirname, join } from "node:path";
+import { fileURLToPath } from "node:url";
+import { getStateDir } from "../../store.js";
+import { defaultEventsPath, logBenchEvent } from "../../monitoring.js";
+const HERE = dirname(fileURLToPath(import.meta.url));
+const REPO_ROOT = join(HERE, "..", "..", "..");
+const BENCH_SCRIPT = join(REPO_ROOT, "scripts", "ml5", "bench-onnx-prod.mjs");
+/** Default events.log beside the state dir (mirrors defaultEventsPath). */
+function benchEventsPath(stateDir) {
+    return defaultEventsPath(stateDir);
+}
+/**
+ * Run the ONNX bench once and record its four events. Returns the parsed
+ * BenchResultV1. On any failure (script missing, non-zero exit, unparsable
+ * output) it returns a degraded result with gates.all:false — never throws, so
+ * the caller's agent loop is never broken (non-fatal store/write contract).
+ */
+export function runBench(stateDir = getStateDir()) {
+    const noop = (error) => ({
+        timestamp: Date.now(),
+        platform: `${process.platform}-${process.arch}`,
+        encoderNative: false,
+        threads: 4,
+        tokens: 512,
+        corpusTokens: 0,
+        p95Ms: null,
+        rssMib: null,
+        rssBaselineMib: null,
+        rssMarginalMib: null,
+        opset: null,
+        deterministic: false,
+        digest: null,
+        gates: { latency: false, rss: false, opset: false, determinism: false, all: false },
+        error,
+    });
+    const fallback = (error) => {
+        const r = noop(error);
+        emitEvents(stateDir, r);
+        return r;
+    };
+    try {
+        const res = spawnSync(process.execPath, ["--expose-gc", BENCH_SCRIPT], {
+            cwd: REPO_ROOT,
+            encoding: "utf8",
+            timeout: 600_000,
+        });
+        const stdout = (res.stdout ?? "").trim();
+        if (res.status === null) {
+            return fallback("bench timed out or failed to spawn");
+        }
+        const parsed = JSON.parse(stdout || "");
+        if (!isBenchResultV1(parsed)) {
+            return fallback("bench output was not a BenchResultV1");
+        }
+        emitEvents(stateDir, parsed);
+        return parsed;
+    }
+    catch (e) {
+        return fallback(`bench failed: ${e?.message ?? String(e)}`);
+    }
+}
+function isBenchResultV1(v) {
+    if (typeof v !== "object" || v === null)
+        return false;
+    const o = v;
+    return (typeof o.timestamp === "number" &&
+        typeof o.platform === "string" &&
+        typeof o.encoderNative === "boolean" &&
+        typeof o.threads === "number" &&
+        typeof o.tokens === "number" &&
+        typeof o.corpusTokens === "number" &&
+        typeof o.gates === "object" && o.gates !== null &&
+        typeof o.gates.all === "boolean");
+}
+function emitEvents(stateDir, r) {
+    const path = benchEventsPath(stateDir);
+    const run = { platform: r.platform, encoderNative: r.encoderNative, threads: r.threads, tokens: r.tokens, digest: r.digest, corpusTokens: r.corpusTokens };
+    logBenchEvent(path, "vector_cortex_encoder_bench_p95_ms", { ...run, p95Ms: r.p95Ms, pass: r.gates.latency });
+    logBenchEvent(path, "vector_cortex_encoder_bench_rss_mib", { ...run, rssMib: r.rssMib, rssBaselineMib: r.rssBaselineMib, rssMarginalMib: r.rssMarginalMib, pass: r.gates.rss });
+    logBenchEvent(path, "vector_cortex_encoder_bench_opset_ok", { ...run, opset: r.opset, pass: r.gates.opset });
+    logBenchEvent(path, "vector_cortex_encoder_bench_deterministic", { ...run, deterministic: r.deterministic, pass: r.gates.determinism });
+}

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;
+    }
+}