pi-mega-compact 0.20.64 → 0.20.65

package/README.md

@@ -90,7 +90,7 @@ Set env vars before starting pi. Defaults are in `src/config/dedup.ts`.
 | `MEGACOMPACT_WIKI_INCREMENTAL` | `true` | Incremental wiki updates (no full rebuild) |
 | `MEGACOMPACT_AUTO_WIKI` | `true` | Auto-generate wiki from turns |
 | `MEGACOMPACT_NEW_UI` | `true` | Use the new Tailwind/shadcn dashboard shell |
-| `MEGACOMPACT_COST_API_ENABLED` | `false` | Opt-in: fetch model pricing from an external API (PREVENT-PI-004 network exception). Enriches dashboard cost data for models not in the local pricing table |
+| `MEGACOMPACT_COST_API_ENABLED` | `false` | Opt-in: fetch model pricing from an external API (PREVENT-PI-004 applies to defaults; opt-in features are exempt). Enriches dashboard cost data for models not in the local pricing table |
 | `MEGACOMPACT_COST_API_URL` | _(unset)_ | OpenRouter-compatible model pricing endpoint (e.g. `https://openrouter.ai/api/v1/models`). Only contacted when `MEGACOMPACT_COST_API_ENABLED=true` |
 
 Full config reference: [`docs/CONFIGURATION.md`](docs/CONFIGURATION.md)

package/dist/config/vector-cortex-enc2b.js

@@ -0,0 +1,36 @@
+/**
+ * config/vector-cortex-enc2b.ts — ENC-2b native onnxruntime qualification
+ * retest flag.
+ *
+ * Sibling of vector-cortex-enc2a.ts / vector-cortex-enc2budget.ts so
+ * vector-cortex.ts stays under the 300-line soft limit (soft-as-hard gate).
+ * vector-cortex.ts re-exports the flag below and root src/config.ts re-exports
+ * it, so no consumer import path changes.
+ *
+ * ENC-2b re-probes and re-qualifies the installed native onnxruntime binding
+ * (`onnxruntime-node` in `~/.pi/mega-compact/native-ort/`, operator-installed
+ * via the ENC-2a guide). When the operator installs the binding the read path
+ * does not automatically re-evaluate the qualification verdict; this sprint
+ * delivers a reader-only retest that loads the LOCAL on-disk binding, runs a
+ * bounded warmup + p95 probe against it, measures RSS, computes a fresh
+ * qualification verdict against the ENC-0f p95 budget + the operator
+ * install-budget, and surfaces it on the Setup Cortex sub-tab and the GET
+ * `/api/setup-status` response.
+ *
+ * `MEGACOMPACT_ENC_2B=0` disables: the GET omits both new retest fields, the
+ * POST ignores the retest-request key, the Cortex sub-tab renders no retest
+ * card — byte-identical to the ENC-2a predecessor. The flag MUST also be a
+ * dashboard SETTINGS toggle (visible in config UI, never in EXCLUDED_SETTINGS
+ * — see `feedback_dashboard-flags-toggleable` memory).
+ *
+ * Pi-agnostic, dependency-free (PREVENT-PI-004 / PREVENT-011).
+ */
+import { sprintFlag } from "./vector-cortex-flag.js";
+/**
+ * ENC-2b — native onnxruntime qualification retest (detect installed binding,
+ * re-probe p95/RSS, re-qualify). Default ON. `MEGACOMPACT_ENC_2B=0` disables and
+ * is byte-identical to the ENC-2a predecessor: no retest GET fields, no retest
+ * POST branch, no retest card. This flag MUST also be a dashboard SETTINGS
+ * toggle (visible in config UI, never in EXCLUDED_SETTINGS).
+ */
+export const ENC_2B_ENABLED = () => sprintFlag("MEGACOMPACT_ENC_2B");

package/dist/config/vector-cortex.js

@@ -52,5 +52,7 @@ export { ENC_1B_ENABLED, ENC_1B_MAX_EMBEDDING_DIM, ENC_1B_EMBEDDING_DIM_ENV, ENC
 export { ENC_2BUDGET_ENABLED, ENC_2BUDGET_NATIVE_ORT_BUDGET_ENV, ENC_2BUDGET_MAX_MIB, ENC_2BUDGET_DEFAULT_MIB, } from "./vector-cortex-enc2budget.js";
 // ENC-2a native onnxruntime install-guide flag extracted to vector-cortex-enc2a.ts.
 export { ENC_2A_ENABLED } from "./vector-cortex-enc2a.js";
+// ENC-2b native onnxruntime qualification-retest flag extracted to vector-cortex-enc2b.ts.
+export { ENC_2B_ENABLED } from "./vector-cortex-enc2b.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, ML5E_ENABLED, DEDUP_ATTR_ENABLED, ENC_0A_ENABLED, ENC_0B_ENABLED, ENC_0C_ENABLED, ENC_0D_ENABLED, ENC_0E_ENABLED, ENC_0F_ENABLED, ENC_0G_ENABLED, ENC_1A_ENABLED, ENC_1B_ENABLED, ENC_2BUDGET_ENABLED, ENC_2BUDGET_NATIVE_ORT_BUDGET_ENV, ENC_2BUDGET_MAX_MIB, ENC_2BUDGET_DEFAULT_MIB, ENC_2A_ENABLED, COSINE_FP_BENCH_ENABLED, COSINE_FP_REAL_ENABLED, REPO_CORPUS_ENABLED, DASH_0A_ENABLED, DASH_0B_ENABLED, DASH_0C_ENABLED, DASH_0D_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, ENC_0A_ENABLED, ENC_0B_ENABLED, ENC_0C_ENABLED, ENC_0D_ENABLED, ENC_0E_ENABLED, ENC_0F_ENABLED, ENC_0G_ENABLED, ENC_1A_ENABLED, ENC_1B_ENABLED, ENC_2BUDGET_ENABLED, ENC_2BUDGET_NATIVE_ORT_BUDGET_ENV, ENC_2BUDGET_MAX_MIB, ENC_2BUDGET_DEFAULT_MIB, ENC_2A_ENABLED, ENC_2B_ENABLED, COSINE_FP_BENCH_ENABLED, COSINE_FP_REAL_ENABLED, REPO_CORPUS_ENABLED, DASH_0A_ENABLED, DASH_0B_ENABLED, DASH_0C_ENABLED, DASH_0D_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

@@ -71,6 +71,7 @@ export const VECTOR_CORTEX_SETTINGS = {
         boolDirect("MEGACOMPACT_ENC_1B", "ENC-1b ONNX runtime backend + embedder API settings", "ENC-1b completes the embedder API Settings surface (dim / headers / allow-remote persisted to the per-repo .mega-compact.env) and surfaces the ONNX runtime selection knob (MEGACOMPACT_ENCODER_NATIVE) as a toggle plus the current effective backend + demotion reason. OFF = no new GET fields, no writer branch, no new Settings rows or toggle (byte-identical ENC-1a predecessor).", true),
         boolDirect("MEGACOMPACT_ENC_2BUDGET", "ENC-2a operator-configurable native install-budget knob", "ENC-2a exposes the runtime operator knob MEGACOMPACT_NATIVE_ORT_BUDGET_MIB (default 300 MiB, clamp 8192) as a dashboard Settings field. The knob gates budgetOk in the encoder runtime decision (decision.ts installBudgetMib + runtime-select.ts selectRuntimeBackend). Persisted to the per-repo .mega-compact.env via the setup-configure writer branch; the GET status route surfaces the persisted value AND the effective integer installBudgetMib() resolves to (so the dashboard shows the runtime's actual operand). OFF = no new GET fields, no writer branch, no new Settings rows (byte-identical ENC-1b predecessor).", true),
         boolDirect("MEGACOMPACT_ENC_2A", "ENC-2a native onnxruntime install GUIDE (detect + copy-paste)", "ENC-2a surfaces the platform-matched native onnxruntime install-guide (three copy-paste steps + script path) on the Setup Cortex sub-tab when the operator opted into the native backend but the effective runtime is still wasm (onnxruntime-node absent). Pure guidance — the extension NEVER executes or fetches (PREVENT-PI-004 unchanged); the committed operator script scripts/encoder/install-native-ort.mjs runs in the USER's shell. OFF = no guide GET fields, no guide-request POST branch, no install card (byte-identical ENC-1b predecessor).", true),
+        boolDirect("MEGACOMPACT_ENC_2B", "ENC-2b native ONNX runtime qualification retest (detect + re-qualify)", "ENC-2b re-probes and re-qualifies the installed native onnxruntime binding (onnxruntime-node in ~/.pi/mega-compact/native-ort/, operator-installed via the ENC-2a guide): loads the LOCAL on-disk binding, runs a bounded warmup + p95 probe, measures RSS, and computes a fresh qualification verdict against the ENC-0f p95 budget + the operator install-budget. Never downloads or trains (PREVENT-PI-004 / HG-1 unchanged); surfaces nativeOrtRetestResult + nativeOrtBackendEffective on the GET /api/setup-status. OFF = no retest GET fields, no retest POST branch, no retest card (byte-identical ENC-2a predecessor).", true),
         boolDirect("MEGACOMPACT_COSINE_FP_BENCH", "COS-FP-A Synthetic FP Bench", "Synthetic FP harness + L2 cosine threshold calibration (COS-FP-A): drives the deterministic synthetic-corpus grid sweep (scripts/cosine-fp/bench.mjs) that measures L2 cosine false-positive rates and emits a recommended default, and serves the reader-only GET /api/cosine-fp-report. Synthetic-corpus-only — never reads real session/ledger bytes. OFF = bench inert, endpoint 404s, no report write, L2_COSINE plain MEGACOMPACT_L2_THRESHOLD (byte-identical predecessor).", true),
         boolDirect("MEGACOMPACT_COSINE_FP_REAL", "COS-FP-R Real-Corpus Validation", "Real-corpus L2 cosine FP validation (COS-FP-R): runs the real-corpus grid sweep (scripts/cosine-fp/real-bench.mjs) against donated, consent-approved sessions only (SECURITY_PRIVACY Lifecycle/Consent), scoring FP/FN with Wilson intervals + session-grouped bootstrap, and appends CI-backed per-content-type recommendations to the cosine-threshold report. Runs only when a valid consented corpus exists; absent corpus = no_corpus (normative pre-donation state, nothing written). OFF = script inert, nothing executes, no writes (byte-identical predecessor).", true),
         boolDirect("MEGACOMPACT_REPO_CORPUS", "Cross-Repo Corpus (REPO-A)", "Cross-repo corpus preparation (REPO-A): the governed, consent-gated reader-only path that builds a pseudonymous corpus manifest from per-repo events.log slices and serves GET /api/repo-corpus answering which pseudonymous repos/sessions are in the corpus, what cross-repo overlap exists, and whether consent is active for each — counts + IDs + status only, never payload content. Consent is append-only (consent.mjs, CLI/ops-only); revocation freezes the builder + route. OFF = builder refuses to run, reader route 404s, single-repo recall untouched (byte-identical predecessor).", true),

package/dist/extensions/dashboard-server/routes-setup-enc2b.js

@@ -0,0 +1,77 @@
+/**
+ * dashboard-server/routes-setup-enc2b.ts — ENC-2b native onnxruntime
+ * qualification retest (reader-only probe) + POST retest action.
+ *
+ * Extracted out of routes-setup.ts (which hovers at the 300-line source soft
+ * cap, mirroring the ENC-1b/ENC-2a sibling extracts) so the ENC-2b additive
+ * GET field — the native qualification retest result + effective backend —
+ * lives in a sibling impl file and routes-setup.ts stays a thin host.
+ *
+ * The retest's job is RE-QUALIFICATION: when the operator installed
+ * onnxruntime-node via the ENC-2a guide, the encoder read path does not
+ * automatically re-evaluate the qualification verdict. This sibling surface
+ * runs `runNativeRetest` (a bounded, reader-only, local probe) and:
+ *
+ *   - surfaces `nativeOrtRetestResult` + `nativeOrtBackendEffective` on the GET
+ *     when the flag is on AND a binding is installed — absent otherwise
+ *     (byte-identical to the ENC-2a-era shape on flag-off/none);
+ *   - handles the POST `nativeOrtRetest` key (`false` → 400
+ *     `retest_rejected_false_nothing_to_do`; `true` → run + echo the fresh
+ *     result; flag-off → the key is unrecognized, byte-identical predecessor).
+ *
+ * `nativeOrtBackendEffective` is `"native"` only when the retest verdict is
+ * `qualified`; any `degraded`/`failed` verdict keeps the backend `"wasm"` — the
+ * runtime never silently switches to a backend whose retest failed.
+ *
+ * Guardrails: PREVENT-PI-004 (local filesystem + in-process binding load only,
+ * zero network — the retest NEVER fetches), PREVENT-011 (no `any`).
+ */
+import { ENC_2B_ENABLED } from "../../src/config/vector-cortex.js";
+import { runNativeRetest, } from "../../src/vector-cortex/encoder/native-qualify-retest.js";
+/**
+ * Resolve the ENC-2b additive GET fields. Returns BOTH fields when the flag is
+ * on AND a native binding is installed (the retest ran); returns `{}` on
+ * flag-off AND when no binding is installed — both fields are OMITTED (absent,
+ * not null), so the GET stays byte-identical to the ENC-2a-era shape and the
+ * client hides the retest card. Flag-off omits, never nulls: the ENC-2a-era
+ * host spreads `...{}` so no new key appears. Mirror of the ENC-2a/ENC-1b
+ * sibling idiom; async because the retest loads the binding in-process.
+ */
+export async function readEnc2bRetest(stateDir) {
+    if (!ENC_2B_ENABLED())
+        return {};
+    const result = await runNativeRetest(stateDir);
+    if (result === null)
+        return {};
+    return {
+        nativeOrtRetestResult: result,
+        nativeOrtBackendEffective: result.verdict === "qualified" ? "native" : "wasm",
+    };
+}
+/**
+ * ENC-2b POST retest-action key handling (flag-gated, additive).
+ *
+ * `true`  → { action: "run" } (the host runs the retest and returns the result)
+ * `false` → { action: "reject" } (400 `retest_rejected_false_nothing_to_do`)
+ * other   → null (key absent or flag off — falls through unrecognized,
+ *            byte-identical ENC-2a-era predecessor).
+ */
+export function enc2bRetestRequest(body) {
+    if (!ENC_2B_ENABLED())
+        return null;
+    if (body.nativeOrtRetest === true)
+        return "run";
+    if (body.nativeOrtRetest === false)
+        return "reject";
+    return null;
+}
+/**
+ * Run the ENC-2b retest now (bounded, synchronous on the request) and return
+ * the fresh result. Flag-off or no binding → null. Used by the POST "run"
+ * branch; the result is returned on the response body, never persisted.
+ */
+export async function runEnc2bRetest(stateDir) {
+    if (!ENC_2B_ENABLED())
+        return null;
+    return runNativeRetest(stateDir);
+}

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

@@ -15,6 +15,7 @@ import { enc1aStatusFields, writeEnc1aEnv, tryEnc1aConfigure, wantsEnc1a } from
 import { enc1bStatusFields, tryEnc1bConfigure, tryEnc1bInto, enc1bValidateCombined } from "./routes-setup-enc1b.js";
 import { enc2BudgetStatusFields, tryEnc2BudgetInto, enc2BudgetValidateCombined, } from "./routes-setup-enc2budget.js";
 import { readEnc2aGuide, enc2aGuideRequest } from "./routes-setup-enc2a.js";
+import { readEnc2bRetest, enc2bRetestRequest, runEnc2bRetest } from "./routes-setup-enc2b.js";
 import { writeEmbedderEnv } from "./routes-setup-env-upsert.js";
 import { detectOllama, detectLlamaCpp, detectOnnx, memoizedDetectOllama, memoizedDetectLlamaCpp, memoizedDetectOnnx, } from "./routes-setup-detect-cache.js";
 // ---------------------------------------------------------------------------
@@ -69,7 +70,7 @@ export function handleSetupStatus(req, res, ctx) {
     // ENC-1a (flag-gated, additive): echo the persisted endpoint URL and an
     // embeddingApiKeySet boolean ONLY — the raw API key is never returned. When
     // the flag is off (or neither is set) both fields are simply omitted.
-    const body = {
+    const base = {
         currentEmbedder: active,
         configuredEmbedder: configured.embedder,
         configuredUrl: configured.url,
@@ -82,9 +83,20 @@ export function handleSetupStatus(req, res, ctx) {
         ...(enc2a.guide !== null ? { nativeOrtInstallGuide: enc2a.guide } : {}),
         ...(enc2a.installedVersion !== null ? { nativeOrtInstalledVersion: enc2a.installedVersion } : {}),
     };
+    // ENC-2b (flag-gated, additive): the native qualification retest fields. The
+    // retest loads the on-disk binding in-process (bounded, never a fetch), so
+    // readEnc2bRetest is async; the handler returns true synchronously for the
+    // dispatch and writes the response after the await. Flag-off/none mutates
+    // the body additively and stays byte-identical to the ENC-2a-era shape.
     // guardrails-allow PREVENT-PI-004: loopback dashboard response (local)
-    res.writeHead(200, { "Content-Type": "application/json" });
-    res.end(JSON.stringify(body));
+    readEnc2bRetest(ctx.stateDir).then((enc2b) => {
+        const body = { ...base, ...enc2b };
+        res.writeHead(200, { "Content-Type": "application/json" });
+        res.end(JSON.stringify(body));
+    }).catch(() => {
+        res.writeHead(200, { "Content-Type": "application/json" });
+        res.end(JSON.stringify(base));
+    });
     return true;
 }
 // ---------------------------------------------------------------------------
@@ -204,6 +216,30 @@ export function handleSetupConfigure(req, res, ctx) {
             res.end(JSON.stringify({ nativeOrtInstallGuide: readEnc2aGuide(ctx.stateDir).guide }));
             return;
         }
+        // ENC-2b retest-action key (flag-gated, additive): `false` → 400 (an
+        // operator-driven no-op — nothing to retest), `true` → run the bounded
+        // retest synchronously and return the fresh result. Flag-off: the key
+        // falls through unrecognized (byte-identical ENC-2a-era predecessor).
+        // The retest never fetches — the binding is on disk from the ENC-2a
+        // install (PREVENT-PI-004 clean).
+        const retestReq = enc2bRetestRequest(body);
+        // guardrails-allow PREVENT-PI-004: loopback dashboard response (local)
+        if (retestReq === "reject") {
+            res.writeHead(400, { "Content-Type": "application/json" });
+            res.end(JSON.stringify({ error: "retest_rejected_false_nothing_to_do" }));
+            return;
+        }
+        if (retestReq === "run") {
+            // guardrails-allow PREVENT-PI-004: loopback dashboard response (local)
+            runEnc2bRetest(ctx.stateDir).then((fresh) => {
+                res.writeHead(200, { "Content-Type": "application/json" });
+                res.end(JSON.stringify({ nativeOrtRetestResult: fresh }));
+            }).catch(() => {
+                res.writeHead(500, { "Content-Type": "application/json" });
+                res.end(JSON.stringify({ error: "retest_failed" }));
+            });
+            return;
+        }
         if (tryEnc1aConfigure(body, res, ctx))
             return;
         if (tryEnc1bConfigure(body, res, ctx))

package/dist/src/config/vector-cortex-enc2b.js

@@ -0,0 +1,36 @@
+/**
+ * config/vector-cortex-enc2b.ts — ENC-2b native onnxruntime qualification
+ * retest flag.
+ *
+ * Sibling of vector-cortex-enc2a.ts / vector-cortex-enc2budget.ts so
+ * vector-cortex.ts stays under the 300-line soft limit (soft-as-hard gate).
+ * vector-cortex.ts re-exports the flag below and root src/config.ts re-exports
+ * it, so no consumer import path changes.
+ *
+ * ENC-2b re-probes and re-qualifies the installed native onnxruntime binding
+ * (`onnxruntime-node` in `~/.pi/mega-compact/native-ort/`, operator-installed
+ * via the ENC-2a guide). When the operator installs the binding the read path
+ * does not automatically re-evaluate the qualification verdict; this sprint
+ * delivers a reader-only retest that loads the LOCAL on-disk binding, runs a
+ * bounded warmup + p95 probe against it, measures RSS, computes a fresh
+ * qualification verdict against the ENC-0f p95 budget + the operator
+ * install-budget, and surfaces it on the Setup Cortex sub-tab and the GET
+ * `/api/setup-status` response.
+ *
+ * `MEGACOMPACT_ENC_2B=0` disables: the GET omits both new retest fields, the
+ * POST ignores the retest-request key, the Cortex sub-tab renders no retest
+ * card — byte-identical to the ENC-2a predecessor. The flag MUST also be a
+ * dashboard SETTINGS toggle (visible in config UI, never in EXCLUDED_SETTINGS
+ * — see `feedback_dashboard-flags-toggleable` memory).
+ *
+ * Pi-agnostic, dependency-free (PREVENT-PI-004 / PREVENT-011).
+ */
+import { sprintFlag } from "./vector-cortex-flag.js";
+/**
+ * ENC-2b — native onnxruntime qualification retest (detect installed binding,
+ * re-probe p95/RSS, re-qualify). Default ON. `MEGACOMPACT_ENC_2B=0` disables and
+ * is byte-identical to the ENC-2a predecessor: no retest GET fields, no retest
+ * POST branch, no retest card. This flag MUST also be a dashboard SETTINGS
+ * toggle (visible in config UI, never in EXCLUDED_SETTINGS).
+ */
+export const ENC_2B_ENABLED = () => sprintFlag("MEGACOMPACT_ENC_2B");

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

@@ -52,5 +52,7 @@ export { ENC_1B_ENABLED, ENC_1B_MAX_EMBEDDING_DIM, ENC_1B_EMBEDDING_DIM_ENV, ENC
 export { ENC_2BUDGET_ENABLED, ENC_2BUDGET_NATIVE_ORT_BUDGET_ENV, ENC_2BUDGET_MAX_MIB, ENC_2BUDGET_DEFAULT_MIB, } from "./vector-cortex-enc2budget.js";
 // ENC-2a native onnxruntime install-guide flag extracted to vector-cortex-enc2a.ts.
 export { ENC_2A_ENABLED } from "./vector-cortex-enc2a.js";
+// ENC-2b native onnxruntime qualification-retest flag extracted to vector-cortex-enc2b.ts.
+export { ENC_2B_ENABLED } from "./vector-cortex-enc2b.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, ML5E_ENABLED, DEDUP_ATTR_ENABLED, ENC_0A_ENABLED, ENC_0B_ENABLED, ENC_0C_ENABLED, ENC_0D_ENABLED, ENC_0E_ENABLED, ENC_0F_ENABLED, ENC_0G_ENABLED, ENC_1A_ENABLED, ENC_1B_ENABLED, ENC_2BUDGET_ENABLED, ENC_2BUDGET_NATIVE_ORT_BUDGET_ENV, ENC_2BUDGET_MAX_MIB, ENC_2BUDGET_DEFAULT_MIB, ENC_2A_ENABLED, COSINE_FP_BENCH_ENABLED, COSINE_FP_REAL_ENABLED, REPO_CORPUS_ENABLED, DASH_0A_ENABLED, DASH_0B_ENABLED, DASH_0C_ENABLED, DASH_0D_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, ENC_0A_ENABLED, ENC_0B_ENABLED, ENC_0C_ENABLED, ENC_0D_ENABLED, ENC_0E_ENABLED, ENC_0F_ENABLED, ENC_0G_ENABLED, ENC_1A_ENABLED, ENC_1B_ENABLED, ENC_2BUDGET_ENABLED, ENC_2BUDGET_NATIVE_ORT_BUDGET_ENV, ENC_2BUDGET_MAX_MIB, ENC_2BUDGET_DEFAULT_MIB, ENC_2A_ENABLED, ENC_2B_ENABLED, COSINE_FP_BENCH_ENABLED, COSINE_FP_REAL_ENABLED, REPO_CORPUS_ENABLED, DASH_0A_ENABLED, DASH_0B_ENABLED, DASH_0C_ENABLED, DASH_0D_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/encoder/native-qualify-retest.js

@@ -0,0 +1,225 @@
+/**
+ * vector-cortex/encoder/native-qualify-retest.ts — ENC-2b native onnxruntime
+ * qualification retest (reader-only probe).
+ *
+ * When an operator installs onnxruntime-node via the ENC-2a guide, the encoder
+ * read path does not automatically re-evaluate the qualification verdict — the
+ * encoder still reports encoderBackend "wasm" until a retest runs. This module
+ * is that retest: a pure, bounded, reader-only probe that
+ *
+ *   (a) resolves the installed binding `<native-ort>/node_modules/onnxruntime-node/`;
+ *   (b) reads its `package.json` version;
+ *   (c) dynamic-imports the binding;
+ *   (d) finds a LOCAL onnx model to probe against (under `~/.pi/mega-compact/`);
+ *   (e) runs a bounded warmup (3 passes) + 10 timed inference passes on a fixed
+ *       512-token synthetic input (the bench-onnx.mjs harness shape, in-process);
+ *   (f) measures p95 (sorted index) and RSS (`process.memoryUsage().rss`);
+ *   (g) computes a fresh qualification verdict against the ENC-0f p95 budget
+ *       (`ENCODER_LATENCY_P95_MS`) and the operator install-budget
+ *       (`installBudgetMib()` as the RSS ceiling — the ENC-2budget derivative).
+ *
+ * Results: `{platform, version, verdict: "qualified"|"degraded"|"failed",
+ * p95Ms, rssMiB, testedAt}` or `null` when no binding is installed. It NEVER
+ * throws into the route layer: every binding-load/session/model failure is
+ * caught and reported as a `failed` verdict with the error message surfaced in
+ * the engine, never raised. NO network (PREVENT-PI-004 — the binding + model
+ * are on disk from the operator's ENC-2a install), NO training (HG-1 unchanged).
+ *
+ * Pi-agnostic, dependency-free. No `any` (PREVENT-011).
+ */
+import { readFileSync, existsSync } from "node:fs";
+import { join } from "node:path";
+import { ENCODER_LATENCY_P95_MS, ENCODER_MAX_TOKENS } from "./types.js";
+import { installBudgetMib } from "./decision.js";
+import { detectPlatform } from "./asset.js";
+import { NATIVE_ORT_PACKAGE } from "./native-install-artifacts.js";
+/** Candidate ONNX runtime native-ort prefixes. The ENC-2a guide installs to
+ *  the global `~/.pi/mega-compact/native-ort/`; `stateDir` (per-repo) is a
+ *  secondary probe root so the retest resolves either install location. */
+function nativeOrtRootCandidates(stateDir) {
+    return [
+        join(process.env.HOME ?? "", ".pi", "mega-compact", "native-ort"),
+        join(stateDir, "native-ort"),
+    ];
+}
+/** A single bounded inference pass on the fixed 512-token synthetic input;
+ *  returns latency ms, or null when the session cannot run (caller reports
+ *  failed). The input mirrors bench-onnx.mjs: batch 1, 512 int64 ids. */
+async function timedPass(ort, session, ids, mask, types) {
+    try {
+        const dims = [1, ENCODER_MAX_TOKENS];
+        const o = ort;
+        const s = session;
+        const feeds = {};
+        if (s.inputNames.includes("input_ids")) {
+            feeds.input_ids = new o.Tensor("int64", ids, dims);
+        }
+        if (s.inputNames.includes("attention_mask")) {
+            feeds.attention_mask = new o.Tensor("int64", mask, dims);
+        }
+        if (s.inputNames.includes("token_type_ids")) {
+            feeds.token_type_ids = new o.Tensor("int64", types, dims);
+        }
+        const t = process.hrtime.bigint();
+        await s.run(feeds);
+        return Number(process.hrtime.bigint() - t) / 1e6;
+    }
+    catch {
+        return null;
+    }
+}
+/** Resolve the first existing installed-binding package root, or the first
+ *  candidate's path (for a missing-binding probe) when none exists. */
+function bindingRootDir(stateDir) {
+    for (const root of nativeOrtRootCandidates(stateDir)) {
+        const pkg = join(root, "node_modules", NATIVE_ORT_PACKAGE);
+        if (existsSync(pkg))
+            return pkg;
+    }
+    return join(nativeOrtRootCandidates(stateDir)[0], "node_modules", NATIVE_ORT_PACKAGE);
+}
+/** Read the installed binding version from its package.json; null when absent. */
+function readInstalledVersion(pkgJsonPath) {
+    try {
+        const raw = JSON.parse(readFileSync(pkgJsonPath, "utf8"));
+        if (typeof raw === "object" &&
+            raw !== null &&
+            typeof raw.version === "string") {
+            return raw.version;
+        }
+        return null;
+    }
+    catch {
+        return null;
+    }
+}
+/** Locate a LOCAL onnx model to probe against, under the native-ort roots.
+ *  Null when none exists (the retest cannot qualify without a probe target). */
+function findLocalModel(stateDir) {
+    const candidates = [
+        join(nativeOrtRootCandidates(stateDir)[0], "model.onnx"),
+        join(stateDir, "model.onnx"),
+    ];
+    for (const c of candidates) {
+        if (existsSync(c))
+            return c;
+    }
+    return null;
+}
+/** p95 of sorted latencies (sorted-index percentile, bench-onnx idiom). */
+function p95Of(sorted) {
+    if (sorted.length === 0)
+        return 0;
+    const idx = Math.min(sorted.length - 1, Math.floor(sorted.length * 0.95));
+    return +sorted[idx].toFixed(2);
+}
+/**
+ * Run the native onnxruntime qualification retest. Returns `RetestResult` on a
+ * completed (pass/fail/degrade) probe, or `null` when no binding is installed.
+ * Never throws: binding-load / session / model failures produce a `failed`
+ * verdict with the message on the result, and the `testedAt`/RSS are filled.
+ */
+export async function runNativeRetest(stateDir) {
+    const rootDir = bindingRootDir(stateDir);
+    const pkgJsonPath = join(rootDir, "package.json");
+    if (!existsSync(pkgJsonPath))
+        return null;
+    const version = readInstalledVersion(pkgJsonPath);
+    const platform = detectPlatform();
+    const platformStr = platform ?? `${process.platform}-${process.arch}`;
+    const testedAt = new Date().toISOString();
+    const rssMiB = +((process.memoryUsage().rss / 1048576).toFixed(1));
+    // Binding present but version unreadable — report failed (never throw).
+    if (version === null) {
+        return {
+            platform: platformStr,
+            version: "unknown",
+            verdict: "failed",
+            p95Ms: 0,
+            rssMiB,
+            testedAt,
+        };
+    }
+    // Load the LOCAL binding via dynamic import. Load failure -> failed verdict.
+    let ort;
+    try {
+        ort = await import(join(rootDir, "lib", "index.js"));
+    }
+    catch {
+        try {
+            ort = await import(join(rootDir, "dist", "ort.node.mjs"));
+        }
+        catch {
+            return {
+                platform: platformStr,
+                version,
+                verdict: "failed",
+                p95Ms: 0,
+                rssMiB,
+                testedAt,
+            };
+        }
+    }
+    // Find a local probe model; without one the binding cannot be qualified.
+    const modelPath = findLocalModel(stateDir);
+    if (modelPath === null) {
+        return {
+            platform: platformStr,
+            version,
+            verdict: "failed",
+            p95Ms: 0,
+            rssMiB,
+            testedAt,
+        };
+    }
+    // Create the session (bounded); failure -> failed.
+    try {
+        const factory = ort;
+        const session = await factory.InferenceSession.create(modelPath, {
+            executionProviders: ["cpu"],
+            intraOpNumThreads: 4,
+            graphOptimizationLevel: "all",
+        });
+        const n = ENCODER_MAX_TOKENS;
+        const ids = BigInt64Array.from({ length: n }, (_, i) => BigInt(i === 0 ? 101 : i === n - 1 ? 102 : 2000 + (i % 500)));
+        const mask = BigInt64Array.from({ length: n }, () => 1n);
+        const types = new BigInt64Array(n);
+        // Bounded warmup (3 passes) + 10 timed passes.
+        for (let i = 0; i < 3; i++)
+            await timedPass(ort, session, ids, mask, types);
+        const lat = [];
+        for (let i = 0; i < 10; i++) {
+            const ms = await timedPass(ort, session, ids, mask, types);
+            if (ms === null)
+                break;
+            lat.push(ms);
+        }
+        if (lat.length === 0) {
+            return {
+                platform: platformStr,
+                version,
+                verdict: "failed",
+                p95Ms: 0,
+                rssMiB,
+                testedAt,
+            };
+        }
+        lat.sort((a, b) => a - b);
+        const p95Ms = p95Of(lat);
+        const rssBudgetMib = installBudgetMib();
+        const verdict = p95Ms <= ENCODER_LATENCY_P95_MS && rssMiB <= rssBudgetMib
+            ? "qualified"
+            : "degraded";
+        return { platform: platformStr, version, verdict, p95Ms, rssMiB, testedAt };
+    }
+    catch {
+        return {
+            platform: platformStr,
+            version,
+            verdict: "failed",
+            p95Ms: 0,
+            rssMiB,
+            testedAt,
+        };
+    }
+}