pi-mega-compact 0.20.40 → 0.20.42

package/dist/config/vector-cortex-dedup-attr.js

@@ -0,0 +1,14 @@
+/**
+ * config/vector-cortex-dedup-attr.ts — DEDUP-ATTR tier-attribution flag.
+ *
+ * The single positive sprint flag for the dedup tier-attribution rollup
+ * (external-audit item #6). Default ON; `MEGACOMPACT_DEDUP_ATTR=0` / `=false` /
+ * `_DISABLED=true` disables and is byte-identical to the predecessor: the
+ * `/api/dedup-tier-attribution` endpoint 404s and no rollup cache file is
+ * written.
+ *
+ * Pi-agnostic, dependency-free (PREVENT-PI-004 / PREVENT-011).
+ */
+import { sprintFlag } from "./vector-cortex-flag.js";
+/** Positive sprint flag: `=0` or `_DISABLED=true` disables (default ON). */
+export const DEDUP_ATTR_ENABLED = () => sprintFlag("MEGACOMPACT_DEDUP_ATTR");

package/dist/config/vector-cortex-ml5e.js

@@ -0,0 +1,30 @@
+/**
+ * config/vector-cortex-ml5e.ts — ML5-E nightly retraining cron + feedback loop.
+ *
+ * Sibling extract mirroring vector-cortex-ml5a.ts / ml5b.ts / ml5c.ts / ml5d.ts,
+ * so vector-cortex.ts stays under its 300-line soft limit (soft-as-hard gate).
+ * vector-cortex.ts re-exports the ENUM below and root src/config.ts re-exports
+ * it, so no consumer import path changes.
+ *
+ * ML5-E turns one-shot training (ML5-D's Improve button) into a living loop:
+ * the user's own post-redaction conversation turns become fresh training signal,
+ * the five heads are re-fit nightly, calibration is re-validated, and mode-A
+ * promotion is re-checked without human intervention. The cron is system-
+ * configured (`crontab -e` on the operator device); the extension never installs
+ * or writes a crontab. The flag gates **invocation** only — the scripts
+ * (retrain-nightly.mjs, promotion-gate.mjs, promotion.ts) exist on disk
+ * regardless; `=0` means the runtime never invokes them and the dashboard's
+ * Improve Cortex flow is byte-identical to the ML5-D survivor.
+ *
+ * Pi-agnostic, dependency-free (PREVENT-PI-004 / PREVENT-011).
+ */
+import { sprintFlag } from "./vector-cortex-flag.js";
+/**
+ * ML5-E — nightly retraining cron + feedback loop. Default ON.
+ * `MEGACOMPACT_ML5_E=0` disables and is byte-identical to the ML5-D survivor:
+ * the nightly retraining scripts are never invoked by the runtime, the
+ * promotion gate is inert, and the dashboard's Improve Cortex flow is
+ * byte-identical to the ML5-D tab. The flag gates invocation only — the scripts
+ * exist on disk regardless.
+ */
+export const ML5E_ENABLED = () => sprintFlag("MEGACOMPACT_ML5_E");

package/dist/config/vector-cortex.js

@@ -9,9 +9,8 @@
  * them without re-declaring the ownership boundary. Pi-agnostic, dependency-free.
  */
 import { sprintFlag } from "./vector-cortex-flag.js";
-// VC0/VC1/VC2 foundation-phase flags extracted to vector-cortex-early.ts to keep
-// this file under the 300-line soft limit. Re-exported here so every existing
-// `from "./config/vector-cortex.js"` import keeps resolving unchanged.
+// VC0/VC1/VC2 foundation-phase flags in vector-cortex-early.ts; re-exported
+// so every existing `from "./config/vector-cortex.js"` import resolves unchanged.
 export { VC0A_ENABLED, VC0B_ENABLED, VC0C_ENABLED, VC1A_ENABLED, VC1B_ENABLED, VC1C_ENABLED, VC2A_ENABLED, VC2B_ENABLED, VC2C_ENABLED, } from "./vector-cortex-early.js";
 /**
  * VC3A — capability-gated cortex store (CortexReader/Writer/Admin, CortexRecordV1).
@@ -239,9 +238,8 @@ 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 imports keep resolving.
+// VC8C extracted to vector-cortex-vc8c.ts; re-exported so existing imports keep resolving.
 export { VC8C_ENABLED } from "./vector-cortex-vc8c.js";
-// 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";
@@ -251,5 +249,7 @@ 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";
+export { ML5E_ENABLED } from "./vector-cortex-ml5e.js";
+export { DEDUP_ATTR_ENABLED } from "./vector-cortex-dedup-attr.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, 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";
+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, ML5E_ENABLED, DEDUP_ATTR_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/dedup-attribution.js

@@ -0,0 +1,13 @@
+/**
+ * api-contracts/dedup-attribution.ts — Dedup tier-attribution rollup types (DEDUP-ATTR).
+ *
+ * Types for GET /api/dedup-tier-attribution — the reader-only aggregate that
+ * answers "L0/L1/L2/new percent of dedup decisions in window W". Inputs are the
+ * produced `dedup_audit` events in the local events.log (shape DedupAuditEvent
+ * in src/vectorStore/dedup-audit.ts). Outputs are counts + shares only — never
+ * matched checkpoint paths/text, never raw user query (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

@@ -67,4 +67,11 @@ export const EXTRA_ENDPOINTS = {
         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.",
     },
+    // ─── DEDUP-ATTR tier-attribution rollup ───────────────────────────
+    /** GET /api/dedup-tier-attribution — reader-only per-tier dedup shares. */
+    dedupTierAttribution: {
+        method: "GET",
+        path: "/api/dedup-tier-attribution",
+        description: "Reader-only dedup tier-attribution rollup (DEDUP-ATTR): L0/L1/L2/new percent of dedup decisions in a query window (default 24h, ?windowMs=, capped 30d). Counts + shares only, never matched checkpoint contents or raw query text. Flag-off returns 404.",
+    },
 };

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, handleImproveCortex, handleImproveCortexStatus, } 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, handleDedupTierAttribution, } 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";
@@ -153,5 +153,7 @@ export function dispatchRoutes(req, res, ctx) {
         return true;
     if (handleImproveCortexStatus(req, res, ctx))
         return true;
+    if (handleDedupTierAttribution(req, res, ctx))
+        return true;
     return false;
 }

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

@@ -0,0 +1,150 @@
+/**
+ * dashboard-server/routes-dedup-attribution.ts — DEDUP-ATTR tier attribution route.
+ *
+ * GET /api/dedup-tier-attribution?windowMs=<n> — reader-only aggregate answering
+ * "L0/L1/L2/new percent of dedup decisions in window W". Reads the repo's local
+ * events.log (per-repo stateDir, bounded 8 MiB tail), parses `dedup_audit` lines
+ * (PREVENT-001 guarded), and sends the pure rollup with a shared-derived status.
+ *
+ * Flag-off (MEGACOMPACT_DEDUP_ATTR=0) returns 404 and writes NO durable cache
+ * file — byte-identical predecessor. Reader-only: emits tier counts + shares,
+ * never matched checkpoint paths/text, never raw user query (EVAL-REDACT-002).
+ *
+ * Memoization mirrors routes-vector-cortex-health.ts: an in-memory facts cache
+ * keyed on {events.log mtime+size, windowMs} is served when ≤5s old, so repeated
+ * dashboard polls stay cheap while any on-disk log mutation still invalidates.
+ *
+ * Guardrails: PREVENT-PI-004 (local filesystem read + loopback response only),
+ * PREVENT-001 (JSON.parse guarded), PREVENT-011 (no `any`).
+ */
+import { readFileSync, statSync, writeFileSync, mkdirSync } from "node:fs";
+import { join, dirname } from "node:path";
+import { sendJson } from "./routes-vector-cortex-shared.js";
+import { DEDUP_ATTR_ENABLED } from "../../src/config.js";
+import { computeDedupTierRollup } from "../../src/vector-cortex/dedup-attr/rollup.js";
+import { deriveVcStatus } from "./vc-status.js";
+const DEFAULT_WINDOW_MS = 24 * 60 * 60 * 1000; // 24h
+const MAX_WINDOW_MS = 30 * 24 * 60 * 60 * 1000; // 30 days
+const MAX_TAIL_BYTES = 8 * 1024 * 1024; // 8 MiB
+const MEMO_TTL_MS = 5000; // ≤5s per stateDir
+let memo = null;
+/** Parse one events.log line into a DedupAuditEvent, or null (PREVENT-001). */
+function parseAuditLine(line) {
+    if (!line.includes('"dedup_audit"'))
+        return null;
+    let parsed;
+    try {
+        parsed = JSON.parse(line);
+    }
+    catch {
+        return null;
+    }
+    if (!parsed || typeof parsed !== "object" || Array.isArray(parsed))
+        return null;
+    const obj = parsed;
+    if (obj.type !== "dedup_audit")
+        return null;
+    if (typeof obj.ts !== "string")
+        return null;
+    const tier = obj.tier;
+    const status = obj.status;
+    if (tier !== "L0" && tier !== "L1" && tier !== "L2" && tier !== "new")
+        return null;
+    if (status !== "deduped" && status !== "passed" && status !== "stored")
+        return null;
+    // sessionId is never read by the rollup; a parsed line may omit richer fields.
+    return { type: "dedup_audit", ts: obj.ts, tier, status, sessionId: "" };
+}
+/** Read the bounded tail of events.log into parsed dedup_audit events. */
+function readAuditEvents(eventsPath) {
+    let buf;
+    try {
+        buf = readFileSync(eventsPath); // guardrails-allow PREVENT-PI-004: local events.log read (loopback dashboard)
+    }
+    catch {
+        return [];
+    }
+    const tail = buf.length > MAX_TAIL_BYTES ? buf.subarray(buf.length - MAX_TAIL_BYTES) : buf;
+    const text = tail.toString("utf-8");
+    const lines = text.split("\n");
+    // Drop a leading fragment when we truncated mid-line (first line may be partial).
+    const start = buf.length > MAX_TAIL_BYTES ? 1 : 0;
+    const out = [];
+    for (let i = start; i < lines.length; i++) {
+        const line = lines[i];
+        if (!line)
+            continue;
+        const ev = parseAuditLine(line);
+        if (ev)
+            out.push(ev);
+    }
+    return out;
+}
+/** Resolve windowMs from the ?windowMs= query param (default 24h, capped 30d). */
+function resolveWindowMs(url) {
+    if (!url.includes("?"))
+        return DEFAULT_WINDOW_MS;
+    const params = new URLSearchParams(url.slice(url.indexOf("?")));
+    const raw = params.get("windowMs");
+    if (raw == null)
+        return DEFAULT_WINDOW_MS;
+    const n = Number(raw);
+    if (!Number.isFinite(n) || n <= 0)
+        return DEFAULT_WINDOW_MS;
+    return Math.min(n, MAX_WINDOW_MS);
+}
+/** Best-effort durable snapshot write (never breaks the response). */
+function writeDurableSnapshot(stateDir, rollup) {
+    try {
+        mkdirSync(dirname(join(stateDir, "dedup-tier-attribution.json")), { recursive: true });
+        writeFileSync(join(stateDir, "dedup-tier-attribution.json"), `${JSON.stringify(rollup)}\n`, "utf-8");
+    }
+    catch {
+        /* best-effort — instrumentation must never break the read path */
+    }
+}
+export function handleDedupTierAttribution(req, res, ctx) {
+    const url = req.url ?? "";
+    const path = url.split("?", 1)[0] ?? url;
+    if (path !== "/api/dedup-tier-attribution")
+        return false;
+    if (req.method !== "GET") {
+        sendJson(res, 405, { error: "method_not_allowed" });
+        return true;
+    }
+    if (!DEDUP_ATTR_ENABLED()) {
+        // Flag-off: 404 + no durable cache write — byte-identical predecessor.
+        sendJson(res, 404, { error: "not_found" });
+        return true;
+    }
+    const eventsPath = ctx.eventsPath;
+    const windowMs = resolveWindowMs(url);
+    const now = new Date();
+    // Memoized-facts pattern (mirrors routes-vector-cortex-health.ts): serve a
+    // still-fresh entry keyed on {events.log mtime+size, windowMs}.
+    let keyBase = null;
+    try {
+        const s = statSync(eventsPath); // guardrails-allow PREVENT-PI-004: local events.log stat (loopback)
+        keyBase = { mtimeMs: s.mtimeMs, size: s.size };
+    }
+    catch {
+        keyBase = { mtimeMs: 0, size: 0 };
+    }
+    const key = `${keyBase.mtimeMs}:${keyBase.size}:${windowMs}`;
+    if (memo !== null && memo.key === key && now.getTime() - memo.at <= MEMO_TTL_MS) {
+        sendJson(res, 200, memo.body);
+        return true;
+    }
+    const auditEvents = readAuditEvents(eventsPath);
+    const rollup = computeDedupTierRollup(auditEvents, windowMs, now);
+    const status = deriveVcStatus({
+        enabled: true,
+        hasData: auditEvents.length > 0,
+        structuralOnly: false,
+    });
+    const body = { ...rollup, status };
+    memo = { key, at: now.getTime(), body };
+    writeDurableSnapshot(ctx.stateDir, body);
+    sendJson(res, 200, body);
+    return true;
+}

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

@@ -56,5 +56,7 @@ export const VECTOR_CORTEX_SETTINGS = {
         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),
+        boolDirect("MEGACOMPACT_ML5_E", "ML5-E Nightly Retraining Cron", "ML5-E nightly retraining + feedback loop: the user's own post-redaction conversation turns become fresh training signal, the five heads are re-fit nightly via system cron, calibration is re-validated, and mode-A promotion is re-checked without human intervention. The cron is system-configured (crontab -e); the extension never installs or writes a crontab. Candidates are written to ~/.pi/mega-compact-encoder/candidates/ and promoted via the ML5-D Improve Cortex flow. ON (default) = the scripts may be invoked by the system cron. OFF = scripts are never invoked; byte-identical to ML5-D.", true),
+        boolDirect("MEGACOMPACT_DEDUP_ATTR", "Dedup Tier Attribution Rollup", "Dedup tier-attribution rollup: per-tier dedup catch shares (L0/L1/L2/new percent of dedup decisions) read from the local events.log dedup_audit stream (GET /api/dedup-tier-attribution). OFF = 404 + no cache file, byte-identical predecessor.", true),
     ],
 };

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

@@ -28,3 +28,4 @@ export { handleVectorCortexEvaluation, handleVectorCortexHealth, handleVectorCor
 export { handleSetupCortexStatus } from "./routes-setup-cortex.js";
 export { handleSetupCortexAction, handleSetupCortexActionLog, } from "./routes-setup-cortex-actions.js";
 export { handleImproveCortex, handleImproveCortexStatus, } from "./routes-cortex-improve.js";
+export { handleDedupTierAttribution } from "./routes-dedup-attribution.js";

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

@@ -0,0 +1,14 @@
+/**
+ * config/vector-cortex-dedup-attr.ts — DEDUP-ATTR tier-attribution flag.
+ *
+ * The single positive sprint flag for the dedup tier-attribution rollup
+ * (external-audit item #6). Default ON; `MEGACOMPACT_DEDUP_ATTR=0` / `=false` /
+ * `_DISABLED=true` disables and is byte-identical to the predecessor: the
+ * `/api/dedup-tier-attribution` endpoint 404s and no rollup cache file is
+ * written.
+ *
+ * Pi-agnostic, dependency-free (PREVENT-PI-004 / PREVENT-011).
+ */
+import { sprintFlag } from "./vector-cortex-flag.js";
+/** Positive sprint flag: `=0` or `_DISABLED=true` disables (default ON). */
+export const DEDUP_ATTR_ENABLED = () => sprintFlag("MEGACOMPACT_DEDUP_ATTR");

package/dist/src/config/vector-cortex-ml5e.js

@@ -0,0 +1,30 @@
+/**
+ * config/vector-cortex-ml5e.ts — ML5-E nightly retraining cron + feedback loop.
+ *
+ * Sibling extract mirroring vector-cortex-ml5a.ts / ml5b.ts / ml5c.ts / ml5d.ts,
+ * so vector-cortex.ts stays under its 300-line soft limit (soft-as-hard gate).
+ * vector-cortex.ts re-exports the ENUM below and root src/config.ts re-exports
+ * it, so no consumer import path changes.
+ *
+ * ML5-E turns one-shot training (ML5-D's Improve button) into a living loop:
+ * the user's own post-redaction conversation turns become fresh training signal,
+ * the five heads are re-fit nightly, calibration is re-validated, and mode-A
+ * promotion is re-checked without human intervention. The cron is system-
+ * configured (`crontab -e` on the operator device); the extension never installs
+ * or writes a crontab. The flag gates **invocation** only — the scripts
+ * (retrain-nightly.mjs, promotion-gate.mjs, promotion.ts) exist on disk
+ * regardless; `=0` means the runtime never invokes them and the dashboard's
+ * Improve Cortex flow is byte-identical to the ML5-D survivor.
+ *
+ * Pi-agnostic, dependency-free (PREVENT-PI-004 / PREVENT-011).
+ */
+import { sprintFlag } from "./vector-cortex-flag.js";
+/**
+ * ML5-E — nightly retraining cron + feedback loop. Default ON.
+ * `MEGACOMPACT_ML5_E=0` disables and is byte-identical to the ML5-D survivor:
+ * the nightly retraining scripts are never invoked by the runtime, the
+ * promotion gate is inert, and the dashboard's Improve Cortex flow is
+ * byte-identical to the ML5-D tab. The flag gates invocation only — the scripts
+ * exist on disk regardless.
+ */
+export const ML5E_ENABLED = () => sprintFlag("MEGACOMPACT_ML5_E");

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

@@ -9,9 +9,8 @@
  * them without re-declaring the ownership boundary. Pi-agnostic, dependency-free.
  */
 import { sprintFlag } from "./vector-cortex-flag.js";
-// VC0/VC1/VC2 foundation-phase flags extracted to vector-cortex-early.ts to keep
-// this file under the 300-line soft limit. Re-exported here so every existing
-// `from "./config/vector-cortex.js"` import keeps resolving unchanged.
+// VC0/VC1/VC2 foundation-phase flags in vector-cortex-early.ts; re-exported
+// so every existing `from "./config/vector-cortex.js"` import resolves unchanged.
 export { VC0A_ENABLED, VC0B_ENABLED, VC0C_ENABLED, VC1A_ENABLED, VC1B_ENABLED, VC1C_ENABLED, VC2A_ENABLED, VC2B_ENABLED, VC2C_ENABLED, } from "./vector-cortex-early.js";
 /**
  * VC3A — capability-gated cortex store (CortexReader/Writer/Admin, CortexRecordV1).
@@ -239,9 +238,8 @@ 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 imports keep resolving.
+// VC8C extracted to vector-cortex-vc8c.ts; re-exported so existing imports keep resolving.
 export { VC8C_ENABLED } from "./vector-cortex-vc8c.js";
-// 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";
@@ -251,5 +249,7 @@ 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";
+export { ML5E_ENABLED } from "./vector-cortex-ml5e.js";
+export { DEDUP_ATTR_ENABLED } from "./vector-cortex-dedup-attr.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, 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";
+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, ML5E_ENABLED, DEDUP_ATTR_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/vector-cortex/dedup-attr/rollup.js

@@ -0,0 +1,97 @@
+/**
+ * vector-cortex/dedup-attr/rollup.ts — DEDUP-ATTR pure tier-attribution rollup.
+ *
+ * Pure function: reads an array of parsed dedup-audit events, buckets them by
+ * tier over a query window, and returns the per-tier catch shares. No I/O, no
+ * clock — the route passes `now: Date` (new Date()) at request time, so this
+ * module is deterministically reproducible and trivially testable (DEDUP-ATTR-004).
+ *
+ * Missing `similarity` is fine here: L0/L1 are hash/verify tiers and emit no
+ * score, and the rollup never consults it. Malformed lines (non-object, wrong
+ * type) are skipped silently — the route's events.log reader already guards its
+ * JSON.parse (PREVENT-001), and this layer defends against a hostile/mixed
+ * `dedup_audit` array as well.
+ *
+ * The returned shape is structurally identical to the authoritative contract in
+ * `extensions/dashboard-server/api-contracts/dedup-attribution.ts` (the existing
+ * src-mirrors-dashboard-contract pattern); src/ stays pi-agnostic and never
+ * imports dashboard types. Status here is computed from the pure signal the
+ * rollup owns (a non-empty window) — the route additionally derives the sent
+ * `status` via `deriveVcStatus` from the shared vc-status module.
+ *
+ * Guardrails: PREVENT-PI-004 (zero I/O/network), PREVENT-011 (no `any`).
+ */
+/**
+ * Roll up dedup-audit events over [now - windowMs, now] into per-tier shares.
+ *
+ * Only events whose ISO-8601 `ts` falls inside the window are counted; events of
+ * any other `type` (or a missing/unparseable `ts`) are skipped silently. Shares
+ * are fractions of the window's total decisions, each in [0, 1]; an empty window
+ * yields all-zero shares with `status:"awaiting_data"` — never a fabricated
+ * zero-share table presented as real.
+ */
+export function computeDedupTierRollup(events, windowMs, now) {
+    const winMs = windowMs > 0 ? windowMs : 24 * 60 * 60 * 1000;
+    const windowStart = now.getTime() - winMs;
+    const windowEnd = now.getTime();
+    const acc = {
+        l0: { deduped: 0, passed: 0 },
+        l1: { deduped: 0, passed: 0 },
+        l2: { deduped: 0, passed: 0 },
+        storedNew: 0,
+    };
+    let total = 0;
+    for (const ev of events) {
+        if (!ev || typeof ev !== "object")
+            continue;
+        if (ev.type !== "dedup_audit")
+            continue;
+        const ts = Date.parse(ev.ts);
+        if (Number.isNaN(ts) || ts < windowStart || ts > windowEnd)
+            continue;
+        total += 1;
+        switch (ev.tier) {
+            case "L0":
+                if (ev.status === "deduped")
+                    acc.l0.deduped += 1;
+                else if (ev.status === "passed")
+                    acc.l0.passed += 1;
+                break;
+            case "L1":
+                if (ev.status === "deduped")
+                    acc.l1.deduped += 1;
+                else if (ev.status === "passed")
+                    acc.l1.passed += 1;
+                break;
+            case "L2":
+                if (ev.status === "deduped")
+                    acc.l2.deduped += 1;
+                else if (ev.status === "passed")
+                    acc.l2.passed += 1;
+                break;
+            case "new":
+                acc.storedNew += 1;
+                break;
+            default:
+                break;
+        }
+    }
+    const byTier = {
+        l0: { deduped: acc.l0.deduped, passed: acc.l0.passed },
+        l1: { deduped: acc.l1.deduped, passed: acc.l1.passed },
+        l2: { deduped: acc.l2.deduped, passed: acc.l2.passed },
+        new: acc.storedNew,
+    };
+    const shareOf = (n) => (total > 0 ? n / total : 0);
+    return {
+        schema: "dedup-tier-rollup-v1",
+        windowStart: new Date(windowStart).toISOString(),
+        windowEnd: now.toISOString(),
+        totalDecisions: total,
+        byTier,
+        l0Share: shareOf(acc.l0.deduped + acc.l0.passed),
+        l1Share: shareOf(acc.l1.deduped + acc.l1.passed),
+        l2Share: shareOf(acc.l2.deduped + acc.l2.passed),
+        status: total > 0 ? "live" : "awaiting_data",
+    };
+}

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

@@ -0,0 +1,58 @@
+/**
+ * vector-cortex/encoder/promotion.ts — ML5-E PromotionV1 ledger row type +
+ * append-only manifest helper + atomic digest-swap rollback.
+ *
+ * Pure-function discipline: zero I/O, no clock, no storage, no network
+ * (PREVENT-PI-004 / PREVENT-011). Every function is total over its inputs.
+ * The cron orchestrator (scripts/ml5/retrain-nightly.mjs) and the promotion
+ * gate (scripts/ml5/promotion-gate.mjs) consume these helpers; the acceptance
+ * aggregator (ml5e-acceptance.test.ts) validates them synthetically without a
+ * real training run.
+ *
+ * The promotion manifest is an APPEND-ONLY ledger: every new asset is a new
+ * entry, never an overwrite. Rollback restores a prior entry by SHA-256 via
+ * an atomic digest swap (flip the committed pointer to the prior entry's
+ * digest in one step) — no partial state is possible because the swap is a
+ * single value assignment.
+ */
+// ---------------------------------------------------------------------------
+// PromotionV1 — the ledger row recording one nightly training + promotion run.
+// ---------------------------------------------------------------------------
+/** Schema tag for the promotion ledger row. */
+export const PROMOTION_SCHEMA = "promotion-v1";
+/** Append a new asset entry to the manifest. Never mutates `entries`;
+ *  returns a new manifest with the entry appended and `committed` updated. */
+export function appendAsset(manifest, entry) {
+    return {
+        entries: [...manifest.entries, entry],
+        committed: entry.assetDigest,
+    };
+}
+/** Restore a prior asset by SHA-256: atomically swap the committed pointer
+ *  to the prior entry. The prior entry MUST exist in the manifest (append-only
+ *  guarantees it was never overwritten). Returns null when the digest is not
+ *  found (defensive: append-only means it should always be findable). */
+export function rollbackTo(manifest, digest) {
+    const found = manifest.entries.find((e) => e.assetDigest === digest);
+    if (!found)
+        return null;
+    return {
+        entries: manifest.entries,
+        committed: found.assetDigest,
+    };
+}
+// ---------------------------------------------------------------------------
+// Pure decision rules.
+// ---------------------------------------------------------------------------
+/** Evaluate the promotion gate: promote only when all five heads pass AND the
+ *  new asset beats the committed asset on the held-out dev set. Otherwise
+ *  demote. No training data is ever the held-out set. */
+export function promoteDecision(fiveHeadsOk, heldOutBeat) {
+    return fiveHeadsOk && heldOutBeat ? "promoted" : "demoted";
+}
+/** Whether a rollback is needed: the current asset at position N regressed
+ *  relative to the asset at position N-1. True only when regression is
+ *  confirmed and a prior asset digest exists to restore. */
+export function rollbackNeeded(regressionConfirmed, priorAssetDigest) {
+    return regressionConfirmed && priorAssetDigest !== null;
+}

package/dist/vector-cortex/dedup-attr/rollup.js

@@ -0,0 +1,97 @@
+/**
+ * vector-cortex/dedup-attr/rollup.ts — DEDUP-ATTR pure tier-attribution rollup.
+ *
+ * Pure function: reads an array of parsed dedup-audit events, buckets them by
+ * tier over a query window, and returns the per-tier catch shares. No I/O, no
+ * clock — the route passes `now: Date` (new Date()) at request time, so this
+ * module is deterministically reproducible and trivially testable (DEDUP-ATTR-004).
+ *
+ * Missing `similarity` is fine here: L0/L1 are hash/verify tiers and emit no
+ * score, and the rollup never consults it. Malformed lines (non-object, wrong
+ * type) are skipped silently — the route's events.log reader already guards its
+ * JSON.parse (PREVENT-001), and this layer defends against a hostile/mixed
+ * `dedup_audit` array as well.
+ *
+ * The returned shape is structurally identical to the authoritative contract in
+ * `extensions/dashboard-server/api-contracts/dedup-attribution.ts` (the existing
+ * src-mirrors-dashboard-contract pattern); src/ stays pi-agnostic and never
+ * imports dashboard types. Status here is computed from the pure signal the
+ * rollup owns (a non-empty window) — the route additionally derives the sent
+ * `status` via `deriveVcStatus` from the shared vc-status module.
+ *
+ * Guardrails: PREVENT-PI-004 (zero I/O/network), PREVENT-011 (no `any`).
+ */
+/**
+ * Roll up dedup-audit events over [now - windowMs, now] into per-tier shares.
+ *
+ * Only events whose ISO-8601 `ts` falls inside the window are counted; events of
+ * any other `type` (or a missing/unparseable `ts`) are skipped silently. Shares
+ * are fractions of the window's total decisions, each in [0, 1]; an empty window
+ * yields all-zero shares with `status:"awaiting_data"` — never a fabricated
+ * zero-share table presented as real.
+ */
+export function computeDedupTierRollup(events, windowMs, now) {
+    const winMs = windowMs > 0 ? windowMs : 24 * 60 * 60 * 1000;
+    const windowStart = now.getTime() - winMs;
+    const windowEnd = now.getTime();
+    const acc = {
+        l0: { deduped: 0, passed: 0 },
+        l1: { deduped: 0, passed: 0 },
+        l2: { deduped: 0, passed: 0 },
+        storedNew: 0,
+    };
+    let total = 0;
+    for (const ev of events) {
+        if (!ev || typeof ev !== "object")
+            continue;
+        if (ev.type !== "dedup_audit")
+            continue;
+        const ts = Date.parse(ev.ts);
+        if (Number.isNaN(ts) || ts < windowStart || ts > windowEnd)
+            continue;
+        total += 1;
+        switch (ev.tier) {
+            case "L0":
+                if (ev.status === "deduped")
+                    acc.l0.deduped += 1;
+                else if (ev.status === "passed")
+                    acc.l0.passed += 1;
+                break;
+            case "L1":
+                if (ev.status === "deduped")
+                    acc.l1.deduped += 1;
+                else if (ev.status === "passed")
+                    acc.l1.passed += 1;
+                break;
+            case "L2":
+                if (ev.status === "deduped")
+                    acc.l2.deduped += 1;
+                else if (ev.status === "passed")
+                    acc.l2.passed += 1;
+                break;
+            case "new":
+                acc.storedNew += 1;
+                break;
+            default:
+                break;
+        }
+    }
+    const byTier = {
+        l0: { deduped: acc.l0.deduped, passed: acc.l0.passed },
+        l1: { deduped: acc.l1.deduped, passed: acc.l1.passed },
+        l2: { deduped: acc.l2.deduped, passed: acc.l2.passed },
+        new: acc.storedNew,
+    };
+    const shareOf = (n) => (total > 0 ? n / total : 0);
+    return {
+        schema: "dedup-tier-rollup-v1",
+        windowStart: new Date(windowStart).toISOString(),
+        windowEnd: now.toISOString(),
+        totalDecisions: total,
+        byTier,
+        l0Share: shareOf(acc.l0.deduped + acc.l0.passed),
+        l1Share: shareOf(acc.l1.deduped + acc.l1.passed),
+        l2Share: shareOf(acc.l2.deduped + acc.l2.passed),
+        status: total > 0 ? "live" : "awaiting_data",
+    };
+}