@warmdrift/kgauto-compiler 2.0.0-alpha.77 → 2.0.0-alpha.79

package/dist/index.mjs

@@ -16,10 +16,11 @@ import {
 import {
   LIBRARY_VERSION,
   createKeyHealthRoute
-} from "./chunk-GB7VJQ6C.mjs";
+} from "./chunk-DRYCOR6G.mjs";
 import {
   ABSOLUTE_FLOOR,
   ARCHETYPE_FLOOR_DEFAULT,
+  BRAIN_READ_ENV_NAMES,
   PROVIDER_ENV_KEYS,
   configureBrainQuery,
   createBrainQueryCache,
@@ -42,7 +43,7 @@ import {
   loadChainsFromBrain,
   readBrainReadEnv,
   resolveProviderKey
-} from "./chunk-46U2NVOL.mjs";
+} from "./chunk-FT2FN6ZP.mjs";
 import {
   ALIASES,
   LATENCY_TIER_MS,
@@ -53,7 +54,7 @@ import {
   latencyTierOf,
   profilesByProvider,
   tryGetProfile
-} from "./chunk-N36LE3MK.mjs";
+} from "./chunk-VVRDFE6T.mjs";
 import {
   emitAdvisoryFired,
   emitCompileDone,
@@ -820,11 +821,18 @@ function passScoreTargets(ir, opts) {
       }
     }
     let qualityGatePenalty = 0;
+    let structuredCliffGate;
     if (constraints.structuredOutput) {
       const schemaWeak = effectiveConventions(profile).some(
         (c) => c.archetype === ir.intent.archetype && c.structuredOutputHint === "avoid"
       );
       if (schemaWeak) qualityGatePenalty = QUALITY_GATE_PENALTY;
+      if (!schemaWeak) {
+        structuredCliffGate = profile.cliffs.find(
+          (c) => c.action === "quality_gate_structured" && (!c.whenIntent || c.whenIntent === ir.intent.archetype) && c.metric === "input_tokens" && opts.estimatedInputTokens >= c.threshold
+        );
+        if (structuredCliffGate) qualityGatePenalty = QUALITY_GATE_PENALTY;
+      }
     }
     const measuredGate = opts.measuredFailureGates?.get(modelId);
     if (measuredGate) qualityGatePenalty = QUALITY_GATE_PENALTY;
@@ -889,6 +897,16 @@ function passScoreTargets(ir, opts) {
           rankAfter: rank,
           description: `Model ${modelId} gated below the quality floor for archetype '${ir.intent.archetype}' by MEASURED evidence from this app's own outcomes \u2014 ${measuredGate.nFail} of ${measuredGate.n} attempts failed on the quality axis in the trailing window (${pct(measuredGate.rate)}; 95% lower bound ${pct(measuredGate.lowerBound)} > 50%). Rank ${rankBefore.toFixed(2)} \u2192 ${rank.toFixed(2)} (\u2212${qualityGatePenalty.toFixed(2)}). Down-ranked out of leadership; retained as graceful fallback only. The gate is derived, not stored \u2014 it lifts on its own once the failures age out of the window.`
         });
+      } else if (structuredCliffGate) {
+        policyMutations.push({
+          id: `quality-gate-structured-cliff-${modelId}`,
+          source: "quality_gate",
+          passName: "score_targets",
+          rankDelta: -qualityGatePenalty,
+          rankBefore,
+          rankAfter: rank,
+          description: `Model ${modelId} gated below the quality floor for archetype '${ir.intent.archetype}' \u2014 declared structuredOutput + input_tokens \u2265 ${structuredCliffGate.threshold} trips a measured cliff: ${structuredCliffGate.reason} Rank ${rankBefore.toFixed(2)} \u2192 ${rank.toFixed(2)} (\u2212${qualityGatePenalty.toFixed(2)}). Down-ranked out of leadership; retained as graceful fallback only. Bundled knowledge \u2014 active on cold isolates with no brain.`
+        });
       } else {
         policyMutations.push({
           id: `quality-gate-structured-${modelId}`,
@@ -1572,13 +1590,8 @@ function getArchetypePerfScore(modelId, archetype) {
   return { score, n, grounding };
 }
 
-// src/promote-ready-brain.ts
-function isRawPromoteReadyRow(x) {
-  if (!x || typeof x !== "object") return false;
-  const r = x;
-  return typeof r.intent_archetype === "string" && typeof r.family === "string" && typeof r.candidate_model === "string" && typeof r.current_model === "string" && typeof r.detected_at === "string";
-}
-function coerceNumber(v) {
+// src/measured-failure-brain.ts
+function coerceCount(v) {
   if (typeof v === "number") return Number.isFinite(v) ? v : null;
   if (typeof v === "string") {
     const n = Number(v);
@@ -1586,58 +1599,144 @@ function coerceNumber(v) {
   }
   return null;
 }
-function mapRowsToFindings2(rows) {
+function isRawFailureRow(x) {
+  if (!x || typeof x !== "object") return false;
+  const r = x;
+  return typeof r.intent_archetype === "string" && typeof r.model === "string" && (typeof r.n === "number" || typeof r.n === "string");
+}
+function mapRows(rows) {
   const out = [];
   for (const row of rows) {
-    if (!isRawPromoteReadyRow(row)) continue;
-    const sampleN = coerceNumber(row.sample_n);
-    const passRate = coerceNumber(row.judge_pass_rate);
-    const avgScore = coerceNumber(row.judge_avg_score);
-    if (sampleN === null || passRate === null || avgScore === null) continue;
+    if (!isRawFailureRow(row)) continue;
+    const n = coerceCount(row.n);
+    const nFail = coerceCount(row.n_fail) ?? 0;
+    if (n === null || n <= 0) continue;
     out.push({
       archetype: row.intent_archetype,
-      family: row.family,
-      candidateModel: row.candidate_model,
-      currentModel: row.current_model,
-      sampleN,
-      judgePassRate: passRate,
-      judgeAvgScore: avgScore,
-      costDeltaPct: coerceNumber(row.cost_delta_pct),
-      detectedAt: row.detected_at
+      model: row.model,
+      n,
+      nFail
     });
   }
   return out;
 }
+var MEASURED_FAILURE_CFG = {
+  /**
+   * Hard minimum attempts before ANY gate may be created. Guards against
+   * pathological tiny samples that the confidence bound alone would let
+   * through in edge cases. At 5-for-5 the bound clears the threshold; at
+   * 3-for-3 it does not, which is the behaviour we want (three failures is
+   * a bad day, five in a row is a pattern).
+   */
+  minSample: 5,
+  /**
+   * Gate when we are 95% confident the model fails MORE OFTEN THAN IT
+   * SUCCEEDS on this surface. Deliberately unarguable rather than tuned —
+   * a model that probably fails the majority of the time has no business
+   * leading a surface, whatever its declared scores say.
+   */
+  lowerBoundThreshold: 0.5,
+  /** 95% one-sided-ish confidence (standard two-sided z at α=0.05). */
+  z: 1.96,
+  /** Must match the view's window. Documented here for the advisory text. */
+  windowDays: 28
+};
+function wilsonLowerBound(failures, n, z = MEASURED_FAILURE_CFG.z) {
+  if (n <= 0) return 0;
+  const p = failures / n;
+  const z2 = z * z;
+  const denom = 1 + z2 / n;
+  const centre = p + z2 / (2 * n);
+  const margin = z * Math.sqrt(p * (1 - p) / n + z2 / (4 * n * n));
+  const lower2 = (centre - margin) / denom;
+  return lower2 < 0 ? 0 : lower2;
+}
+function mapMeasuredFailureRows(rows) {
+  return mapRows(rows);
+}
+function judgeMeasuredFailure(row, cfg = MEASURED_FAILURE_CFG) {
+  if (!row) return void 0;
+  const normalized = "nFail" in row && typeof row.n === "number" ? row : mapRows([row])[0];
+  if (!normalized || normalized.n < cfg.minSample) return void 0;
+  const lowerBound = wilsonLowerBound(normalized.nFail, normalized.n, cfg.z);
+  return {
+    gated: lowerBound > cfg.lowerBoundThreshold,
+    rate: normalized.nFail / normalized.n,
+    lowerBound,
+    n: normalized.n,
+    nFail: normalized.nFail
+  };
+}
 var snapshots2 = /* @__PURE__ */ new Map();
 var runtime2;
 var warnedOnce2 = false;
-function isPromoteReadyBrainActive() {
+var DEFAULT_MEASURED_FAILURE_ENDPOINT = "https://kgauto-dashboard.vercel.app/api/kgauto-v2/measured-failure";
+function isMeasuredFailureGateEnabledFromEnv(envSource) {
+  const env = envSource ?? (typeof process !== "undefined" && process.env ? process.env : {});
+  const raw = (env.KGAUTO_MEASURED_FAILURE_GATE ?? "").trim().toLowerCase();
+  return !(raw === "0" || raw === "false");
+}
+function configureMeasuredFailureBrain(rt) {
+  runtime2 = rt;
+  snapshots2.clear();
+  warnedOnce2 = false;
+}
+function isMeasuredFailureBrainActive() {
   return runtime2 !== void 0;
 }
-function loadPromoteReadyFindings(opts) {
+function prefetchMeasuredFailure(appId) {
   const rt = runtime2;
-  if (!rt) return [];
-  const appId = opts.appId;
-  if (!appId) return [];
+  if (!rt || !appId) return void 0;
+  let snap = snapshots2.get(appId);
+  if (!snap) {
+    snap = { data: [], expiresAt: 0, refreshing: false };
+    snapshots2.set(appId, snap);
+  }
+  if (snap.expiresAt > Date.now()) return void 0;
+  const inflight = pendingRefreshes2.get(appId);
+  if (inflight) return inflight;
+  if (snap.refreshing) return void 0;
+  snap.refreshing = true;
+  void asyncRefresh2(rt, appId);
+  return pendingRefreshes2.get(appId);
+}
+async function awaitMeasuredFailureReady(appId, timeoutMs) {
+  if (!runtime2 || !appId) return;
+  const pending = prefetchMeasuredFailure(appId) ?? pendingRefreshes2.get(appId);
+  if (!(timeoutMs > 0)) return;
+  if (!pending) return;
+  let timer;
+  try {
+    await Promise.race([
+      pending,
+      new Promise((resolve) => {
+        timer = setTimeout(resolve, timeoutMs);
+      })
+    ]);
+  } catch {
+  } finally {
+    if (timer) clearTimeout(timer);
+  }
+}
+function getMeasuredFailureVerdict(opts) {
+  const rt = runtime2;
+  if (!rt) return void 0;
+  const { appId, archetype, model } = opts;
+  if (!appId || !archetype || !model) return void 0;
   let snap = snapshots2.get(appId);
   if (!snap) {
     snap = { data: [], expiresAt: 0, refreshing: false };
     snapshots2.set(appId, snap);
   }
   const now = Date.now();
-  const stale = snap.expiresAt <= now;
-  if (stale && !snap.refreshing) {
+  if (snap.expiresAt <= now && !snap.refreshing) {
     snap.refreshing = true;
     void asyncRefresh2(rt, appId);
   }
-  let rows = snap.data;
-  if (opts.archetype) {
-    rows = rows.filter((f) => f.archetype === opts.archetype);
-  }
-  if (opts.family) {
-    rows = rows.filter((f) => f.family === opts.family);
-  }
-  return rows;
+  const row = snap.data.find(
+    (r) => r.archetype === archetype && r.model === model
+  );
+  return judgeMeasuredFailure(row);
 }
 var pendingRefreshes2 = /* @__PURE__ */ new Map();
 async function asyncRefresh2(rt, appId) {
@@ -1661,12 +1760,11 @@ async function doRefresh2(rt, appId) {
   try {
     const res = await rt.fetchImpl(url, { method: "GET" });
     if (!res.ok) {
-      throw new Error(`promote-ready ${res.status}: ${res.statusText}`);
+      throw new Error(`measured-failure ${res.status}: ${res.statusText}`);
     }
     const body = await res.json();
     if (runtime2 !== rt) return;
-    const rows = Array.isArray(body) ? mapRowsToFindings2(body) : [];
-    snap.data = rows;
+    snap.data = Array.isArray(body) ? mapRows(body) : [];
     snap.expiresAt = Date.now() + rt.ttlMs;
     snap.refreshing = false;
   } catch (err) {
@@ -1681,193 +1779,84 @@ async function doRefresh2(rt, appId) {
 }
 function defaultOnError2(err) {
   console.warn(
-    "[kgauto] promote-ready fetch failed (using empty fallback):",
+    "[kgauto] measured-failure fetch failed (gate inactive until next refresh):",
     err
   );
 }
-function resolveFetchImpl(injected) {
-  return injected ?? ((...args) => globalThis.fetch(...args));
-}
-function normalizeEndpoint(endpoint) {
-  return endpoint.replace(/\/+$/, "");
+function _testResetMeasuredFailure() {
+  runtime2 = void 0;
+  snapshots2.clear();
+  pendingRefreshes2 = /* @__PURE__ */ new Map();
+  warnedOnce2 = false;
 }
-async function markPromoteReadyHandled(opts) {
-  const {
-    appId,
-    archetype,
-    family,
-    resolution,
-    resolutionNote,
-    brainEndpoint,
-    brainJwt,
-    brainAnonKey,
-    fetch: injectedFetch
-  } = opts;
-  if (!appId) return { ok: false, reason: "app_id_required" };
-  if (!archetype) return { ok: false, reason: "archetype_required" };
-  if (!family) return { ok: false, reason: "family_required" };
-  if (resolution !== "promoted" && resolution !== "declined" && resolution !== "still-evaluating") {
-    return { ok: false, reason: "resolution_invalid" };
-  }
-  const doFetch = resolveFetchImpl(injectedFetch);
-  const base = normalizeEndpoint(brainEndpoint);
-  const url = `${base}/rest/v1/promote_ready_findings?app_id=eq.${encodeURIComponent(appId)}&intent_archetype=eq.${encodeURIComponent(archetype)}&family=eq.${encodeURIComponent(family)}&resolved_at=is.null`;
-  const patchBody = {
-    resolved_at: (/* @__PURE__ */ new Date()).toISOString(),
-    resolution
-  };
-  if (resolutionNote !== void 0) {
-    patchBody.resolution_note = resolutionNote;
-  }
-  let res;
-  try {
-    res = await doFetch(url, {
-      method: "PATCH",
-      headers: {
-        Authorization: `Bearer ${brainJwt}`,
-        apikey: brainAnonKey,
-        "Content-Type": "application/json",
-        Accept: "application/json",
-        Prefer: "return=minimal"
-      },
-      body: JSON.stringify(patchBody)
-    });
-  } catch (err) {
-    const msg = err instanceof Error ? err.message : String(err);
-    return { ok: false, reason: `network_error:${msg}` };
-  }
-  if (res.status === 401 || res.status === 403) {
-    return { ok: false, reason: "brain_auth_misconfig" };
-  }
-  if (res.status >= 500) {
-    return { ok: false, reason: "brain_unavailable" };
-  }
-  if (!res.ok) {
-    return { ok: false, reason: `patch_failed:${res.status}` };
-  }
-  return { ok: true };
+async function _testWaitForMeasuredFailureRefresh() {
+  const pending = Array.from(pendingRefreshes2.values());
+  if (pending.length > 0) await Promise.all(pending);
 }
 
-// src/advisor-rules/promote-ready.ts
-var PROMOTE_READY_THRESHOLDS = {
-  minPassRate: 0.8,
-  minAvgScore: 4
-};
-function shouldFirePromoteReady(finding, resolvedPrimary) {
-  if (finding.currentModel !== resolvedPrimary) return false;
-  if (finding.judgePassRate < PROMOTE_READY_THRESHOLDS.minPassRate) return false;
-  if (finding.judgeAvgScore < PROMOTE_READY_THRESHOLDS.minAvgScore) return false;
-  return true;
+// src/promotions-brain.ts
+function isRawPromotionRow(x) {
+  if (!x || typeof x !== "object") return false;
+  const r = x;
+  return (typeof r.id === "number" || typeof r.id === "string") && typeof r.intent_archetype === "string" && typeof r.promoted_model === "string" && typeof r.incumbent_model === "string";
 }
-function deriveFamilyLocal(modelId) {
-  if (modelId.startsWith("claude-opus-")) return "claude-opus";
-  if (modelId.startsWith("claude-sonnet-")) return "claude-sonnet";
-  if (modelId.startsWith("claude-haiku-")) return "claude-haiku";
-  if (/^gemini-.*-flash-lite/.test(modelId)) return "gemini-flash-lite";
-  if (/^gemini-.*-flash/.test(modelId)) return "gemini-flash";
-  if (/^gemini-.*-pro/.test(modelId)) return "gemini-pro";
-  if (/^deepseek-.*-pro/.test(modelId)) return "deepseek-reasoner";
-  if (modelId.startsWith("deepseek-")) return "deepseek-chat";
-  if (modelId.startsWith("gpt-")) return "openai-gpt";
+function coerceId(v) {
+  if (typeof v === "number") return Number.isFinite(v) ? v : null;
+  if (typeof v === "string") {
+    const n = Number(v);
+    return Number.isFinite(n) ? n : null;
+  }
   return null;
 }
-function advisorRulePromoteReady(ctx) {
-  if (!isPromoteReadyBrainActive()) return [];
-  if (!ctx.appId) return [];
-  if (!ctx.resolvedPrimary) return [];
-  const family = deriveFamilyLocal(ctx.resolvedPrimary);
-  if (!family) return [];
-  const findings = loadPromoteReadyFindings({
-    appId: ctx.appId,
-    archetype: ctx.archetype,
-    family
-  });
-  if (findings.length === 0) return [];
-  const qualifying = findings.filter(
-    (f) => shouldFirePromoteReady(f, ctx.resolvedPrimary)
-  );
-  if (qualifying.length === 0) return [];
-  qualifying.sort((a, b) => {
-    if (a.judgeAvgScore !== b.judgeAvgScore) {
-      return b.judgeAvgScore - a.judgeAvgScore;
-    }
-    return b.judgePassRate - a.judgePassRate;
-  });
-  const top = qualifying[0];
-  const pctPass = Math.round(top.judgePassRate * 100);
-  const score = top.judgeAvgScore.toFixed(2);
-  let costClause = "";
-  if (top.costDeltaPct !== null) {
-    const sign = top.costDeltaPct < 0 ? "cheaper" : "more expensive";
-    const magnitude = Math.abs(top.costDeltaPct * 100).toFixed(1);
-    costClause = `, cost ${magnitude}% ${sign}`;
-  }
-  const message = `Probe found ${top.candidateModel} produces equivalent-or-better outputs vs ${top.currentModel} on ${top.sampleN} recent ${top.archetype} prompts (pass rate ${pctPass}%, avg score ${score}/5${costClause}). Consider promoting via markPromoteReadyHandled.`;
-  return [
-    {
-      level: "info",
-      code: "promote-ready",
-      message,
-      suggestion: `Migrate ${top.archetype} traffic from ${top.currentModel} to ${top.candidateModel}, then call markPromoteReadyHandled({ appId, archetype: '${top.archetype}', family: '${top.family}', resolution: 'promoted' }) to silence this advisory.`,
-      // alpha.36 architectural field — not a no-ai-needed case.
-      recommendedArchitecture: void 0,
-      docsUrl: "https://github.com/stue/command-center/blob/main/interfaces/kgauto.md#best-practice-advisories"
-    }
-  ];
-}
-
-// src/advisor-rules/consumer-on-stale-model.ts
-function isStaleStatus(v) {
-  return v === "legacy" || v === "deprecated";
-}
-function asString(v) {
-  return typeof v === "string" && v.length > 0 ? v : void 0;
-}
-function mapRowsToFindings3(rows) {
+function mapRowsToPromotions(rows) {
   const out = [];
-  for (const raw of rows) {
-    if (!raw || typeof raw !== "object") continue;
-    const r = raw;
-    const archetype = asString(r.intent_archetype) ?? asString(r.applies_to_archetype);
-    const staleModel = asString(r.stale_model) ?? asString(r.applies_to_model);
-    const staleProvider = asString(r.stale_provider);
-    const recommendedModel = asString(r.recommended_model);
-    const family = asString(r.family);
-    const message = asString(r.message);
-    if (!archetype || !staleModel || !recommendedModel || !family || !message) {
-      continue;
-    }
-    if (!isStaleStatus(r.stale_status)) continue;
-    const row = {
-      archetype,
-      staleModel,
-      staleProvider: staleProvider ?? "unknown",
-      staleStatus: r.stale_status,
-      recommendedModel,
-      family,
-      message
-    };
-    const suggestion = asString(r.suggestion);
-    if (suggestion) row.suggestion = suggestion;
-    if (typeof r.observation_count === "number" && Number.isFinite(r.observation_count)) {
-      row.observationCount = r.observation_count;
-    }
-    out.push(row);
+  for (const row of rows) {
+    if (!isRawPromotionRow(row)) continue;
+    const id = coerceId(row.id);
+    if (id === null) continue;
+    const mode = row.mode === "strategy" ? "strategy" : row.mode === "downswap" || row.mode === void 0 ? "downswap" : null;
+    if (mode === null) continue;
+    out.push({
+      id,
+      archetype: row.intent_archetype,
+      mode,
+      strategy: typeof row.strategy === "string" ? row.strategy : null,
+      promotedModel: row.promoted_model,
+      incumbentModel: row.incumbent_model,
+      evalRunId: coerceId(row.eval_run_id ?? null),
+      suppressQualityGate: row.suppress_quality_gate === true,
+      promotedAt: typeof row.promoted_at === "string" ? row.promoted_at : "",
+      // Pre-.78 endpoints serve no status column and only active rows —
+      // defaulting to 'active' is exact, not optimistic.
+      status: row.status === "rolled_back" ? "rolled_back" : "active",
+      ...typeof row.rolled_back_at === "string" ? { rolledBackAt: row.rolled_back_at } : {},
+      ...typeof row.rollback_class === "string" ? { rollbackClass: row.rollback_class } : {}
+    });
   }
   return out;
 }
 var snapshots3 = /* @__PURE__ */ new Map();
 var runtime3;
 var warnedOnce3 = false;
-var pendingRefreshes3 = /* @__PURE__ */ new Map();
-function isStaleModelFindingsBrainActive() {
+var DEFAULT_PROMOTIONS_ENDPOINT = "https://kgauto-dashboard.vercel.app/api/kgauto-v2/promotions";
+function isAutoPromoteEnabledFromEnv(envSource) {
+  const env = envSource ?? (typeof process !== "undefined" && process.env ? process.env : {});
+  const raw = (env.KGAUTO_AUTO_PROMOTE ?? "").trim().toLowerCase();
+  return raw === "1" || raw === "true";
+}
+function configurePromotionsBrain(rt) {
+  runtime3 = rt;
+  snapshots3.clear();
+  warnedOnce3 = false;
+}
+function isPromotionsBrainActive() {
   return runtime3 !== void 0;
 }
-function getStaleModelFindings(opts) {
+function getApplicablePromotion(opts) {
   const rt = runtime3;
-  if (!rt) return [];
+  if (!rt) return void 0;
   const appId = opts.appId;
-  if (!appId) return [];
+  if (!appId || !opts.archetype || !opts.mode) return void 0;
   let snap = snapshots3.get(appId);
   if (!snap) {
     snap = { data: [], expiresAt: 0, refreshing: false };
@@ -1879,11 +1868,31 @@ function getStaleModelFindings(opts) {
     snap.refreshing = true;
     void asyncRefresh3(rt, appId);
   }
-  if (opts.archetype) {
-    return snap.data.filter((f) => f.archetype === opts.archetype);
+  return snap.data.find(
+    (p) => p.status === "active" && p.archetype === opts.archetype && p.mode === opts.mode
+  );
+}
+var ROLLBACK_SUPPRESSION_WINDOW_DAYS = 28;
+function getRecentRollback(opts) {
+  const rt = runtime3;
+  if (!rt) return void 0;
+  if (!opts.appId || !opts.archetype || !opts.model) return void 0;
+  let snap = snapshots3.get(opts.appId);
+  if (!snap) {
+    snap = { data: [], expiresAt: 0, refreshing: false };
+    snapshots3.set(opts.appId, snap);
   }
-  return snap.data;
+  const now = opts.nowMs ?? Date.now();
+  if (snap.expiresAt <= now && !snap.refreshing) {
+    snap.refreshing = true;
+    void asyncRefresh3(rt, opts.appId);
+  }
+  const windowMs = (opts.windowDays ?? ROLLBACK_SUPPRESSION_WINDOW_DAYS) * 864e5;
+  return snap.data.find(
+    (p) => p.status === "rolled_back" && p.archetype === opts.archetype && p.promotedModel === opts.model && typeof p.rolledBackAt === "string" && now - Date.parse(p.rolledBackAt) <= windowMs
+  );
 }
+var pendingRefreshes3 = /* @__PURE__ */ new Map();
 async function asyncRefresh3(rt, appId) {
   const promise = doRefresh3(rt, appId);
   pendingRefreshes3.set(appId, promise);
@@ -1896,7 +1905,7 @@ async function asyncRefresh3(rt, appId) {
   }
 }
 async function doRefresh3(rt, appId) {
-  const url = `${rt.endpoint}?app_id=${encodeURIComponent(appId)}`;
+  const url = `${rt.endpoint}?app_id=${encodeURIComponent(appId)}&with_rollbacks=1`;
   let snap = snapshots3.get(appId);
   if (!snap) {
     snap = { data: [], expiresAt: 0, refreshing: false };
@@ -1905,11 +1914,11 @@ async function doRefresh3(rt, appId) {
   try {
     const res = await rt.fetchImpl(url, { method: "GET" });
     if (!res.ok) {
-      throw new Error(`stale-model findings ${res.status}: ${res.statusText}`);
+      throw new Error(`promotions ${res.status}: ${res.statusText}`);
     }
     const body = await res.json();
     if (runtime3 !== rt) return;
-    const rows = Array.isArray(body) ? mapRowsToFindings3(body) : [];
+    const rows = Array.isArray(body) ? mapRowsToPromotions(body) : [];
     snap.data = rows;
     snap.expiresAt = Date.now() + rt.ttlMs;
     snap.refreshing = false;
@@ -1925,847 +1934,905 @@ async function doRefresh3(rt, appId) {
 }
 function defaultOnError3(err) {
   console.warn(
-    "[kgauto] stale-model findings fetch failed (using empty fallback):",
+    "[kgauto] promotions fetch failed (promotion boost inactive until next refresh):",
     err
   );
 }
-var CONSUMER_ON_STALE_MODEL_RULE_CODE = "consumer-on-stale-model";
-function advisorRuleConsumerOnStaleModel(ir) {
-  if (!isStaleModelFindingsBrainActive()) return [];
-  if (!ir.appId) return [];
-  const findings = getStaleModelFindings({
-    appId: ir.appId,
-    archetype: ir.intent.archetype
-  });
-  if (findings.length === 0) return [];
-  const ranked = [...findings].sort((a, b) => {
-    if (a.staleStatus !== b.staleStatus) {
-      return a.staleStatus === "deprecated" ? -1 : 1;
-    }
-    return a.staleModel.localeCompare(b.staleModel);
-  });
-  const top = ranked[0];
-  const extraCount = findings.length - 1;
-  const extraNote = extraCount > 0 ? ` (+ ${extraCount} more stale model${extraCount === 1 ? "" : "s"} for this archetype)` : "";
-  return [
-    {
-      level: "warn",
-      code: CONSUMER_ON_STALE_MODEL_RULE_CODE,
-      message: `${top.message}${extraNote}`,
-      suggestion: top.suggestion ?? `Migrate ${top.staleModel} \u2192 ${top.recommendedModel} for archetype "${top.archetype}". The newer model is the current latest in the "${top.family}" family; the stale one is ${top.staleStatus}.`,
-      recommendationType: "model-swap",
-      docsUrl: "https://github.com/stue/command-center/blob/main/interfaces/kgauto.md#best-practice-advisories"
-    }
-  ];
+function _testResetPromotions() {
+  runtime3 = void 0;
+  snapshots3.clear();
+  pendingRefreshes3 = /* @__PURE__ */ new Map();
+  warnedOnce3 = false;
+}
+async function _testWaitForPromotionsRefresh() {
+  const pending = Array.from(pendingRefreshes3.values());
+  if (pending.length > 0) await Promise.all(pending);
 }
 
-// src/archetype-fits.ts
-var ARCHETYPE_FAMILY_FITS = Object.freeze([
-  {
-    archetype: "plan",
-    betterFitFamily: "deepseek-reasoner",
-    reason: "Plan archetype is reasoning-shaped (multi-step chains, hypothesis-and-check, sub-goal decomposition) \u2014 exactly where reasoner-family models excel. Sonnet/Opus produce plans but at higher cost; reasoners produce equivalent-or-better plans at 7-17x lower cost at current promo pricing (deepseek-v4-pro $0.435/$0.87 per 1M promo through 2026-05-31 vs sonnet $3/$15).",
-    costGuidance: "substantially cheaper at current pricing (deepseek-v4-pro promo: ~7-17x cheaper than sonnet)"
-  },
-  {
-    archetype: "critique",
-    betterFitFamily: "deepseek-reasoner",
-    reason: "Critique archetype rewards epistemic humility and explicit reasoning \u2014 reasoner-family default behavior. Sonnet/Opus over-confident on critique tasks; reasoners surface uncertainty productively.",
-    costGuidance: "comparable or cheaper at current pricing"
-  }
-]);
-function findBetterFit(archetype, currentFamily) {
-  for (const fit of ARCHETYPE_FAMILY_FITS) {
-    if (fit.archetype !== archetype) continue;
-    if (fit.betterFitFamily === currentFamily) return null;
-    return fit;
+// src/promote-ready-brain.ts
+function isRawPromoteReadyRow(x) {
+  if (!x || typeof x !== "object") return false;
+  const r = x;
+  return typeof r.intent_archetype === "string" && typeof r.family === "string" && typeof r.candidate_model === "string" && typeof r.current_model === "string" && typeof r.detected_at === "string";
+}
+function coerceNumber(v) {
+  if (typeof v === "number") return Number.isFinite(v) ? v : null;
+  if (typeof v === "string") {
+    const n = Number(v);
+    return Number.isFinite(n) ? n : null;
   }
   return null;
 }
-
-// src/advisor-rules/cross-family-fit.ts
-function familyHasCurrentActiveModel(family) {
-  for (const profile of allProfiles()) {
-    const profileFamily = profile.family ?? deriveFamilyFromModelId(profile.id);
-    if (profileFamily !== family) continue;
-    if (profile.status !== "current") continue;
-    if (profile.active === false) continue;
-    return true;
+function mapRowsToFindings2(rows) {
+  const out = [];
+  for (const row of rows) {
+    if (!isRawPromoteReadyRow(row)) continue;
+    const sampleN = coerceNumber(row.sample_n);
+    const passRate = coerceNumber(row.judge_pass_rate);
+    const avgScore = coerceNumber(row.judge_avg_score);
+    if (sampleN === null || passRate === null || avgScore === null) continue;
+    out.push({
+      archetype: row.intent_archetype,
+      family: row.family,
+      candidateModel: row.candidate_model,
+      currentModel: row.current_model,
+      sampleN,
+      judgePassRate: passRate,
+      judgeAvgScore: avgScore,
+      costDeltaPct: coerceNumber(row.cost_delta_pct),
+      detectedAt: row.detected_at
+    });
   }
-  return false;
+  return out;
 }
-function listCandidatesInFamily(family) {
-  const candidates = [];
-  for (const profile of allProfiles()) {
-    const profileFamily = profile.family ?? deriveFamilyFromModelId(profile.id);
-    if (profileFamily !== family) continue;
-    if (profile.status !== "current") continue;
-    if (profile.active === false) continue;
-    candidates.push(profile.id);
-    if (candidates.length >= 3) break;
-  }
-  return candidates;
+var snapshots4 = /* @__PURE__ */ new Map();
+var runtime4;
+var warnedOnce4 = false;
+function isPromoteReadyBrainActive() {
+  return runtime4 !== void 0;
 }
-function advisorRuleCrossFamilyFit(ctx) {
-  if (!ctx.resolvedPrimary) return [];
-  const currentFamily = deriveFamilyFromModelId(ctx.resolvedPrimary);
-  if (!currentFamily) return [];
-  const fit = findBetterFit(ctx.archetype, currentFamily);
-  if (!fit) return [];
-  if (!familyHasCurrentActiveModel(fit.betterFitFamily)) return [];
-  const candidates = listCandidatesInFamily(fit.betterFitFamily);
-  if (candidates.length === 0) return [];
-  const candidateStr = candidates.join(", ");
-  const message = `Your ${currentFamily} call on ${ctx.archetype} could shift to ${fit.betterFitFamily} \u2014 typically better quality + ${fit.costGuidance}. Suggested candidates: ${candidateStr}.`;
-  return [
-    {
-      level: "info",
-      code: "cross-family-fit-candidate",
-      ownership: "consumer-actionable",
-      message,
-      suggestion: `Swap the model literal in \`ir.models\` to one of: ${candidateStr}. Or call \`getRecommendedPrimary({ family: '${fit.betterFitFamily}', archetype: '${ctx.archetype}', fallback: { id: '${candidates[0]}', reason: 'cross-family-fit-recommendation' } })\` to let kgauto resolve to the current+active family member.`,
-      recommendationType: "model-swap",
-      docsUrl: "https://github.com/stue/command-center/blob/main/interfaces/kgauto.md#best-practice-advisories"
+function loadPromoteReadyFindings(opts) {
+  const rt = runtime4;
+  if (!rt) return [];
+  const appId = opts.appId;
+  if (!appId) return [];
+  let snap = snapshots4.get(appId);
+  if (!snap) {
+    snap = { data: [], expiresAt: 0, refreshing: false };
+    snapshots4.set(appId, snap);
+  }
+  const now = Date.now();
+  const stale = snap.expiresAt <= now;
+  if (stale && !snap.refreshing) {
+    snap.refreshing = true;
+    void asyncRefresh4(rt, appId);
+  }
+  let rows = snap.data;
+  if (opts.archetype) {
+    rows = rows.filter((f) => f.archetype === opts.archetype);
+  }
+  if (opts.family) {
+    rows = rows.filter((f) => f.family === opts.family);
+  }
+  return rows;
+}
+var pendingRefreshes4 = /* @__PURE__ */ new Map();
+async function asyncRefresh4(rt, appId) {
+  const promise = doRefresh4(rt, appId);
+  pendingRefreshes4.set(appId, promise);
+  try {
+    await promise;
+  } finally {
+    if (pendingRefreshes4.get(appId) === promise) {
+      pendingRefreshes4.delete(appId);
     }
-  ];
+  }
 }
-
-// src/advisor.ts
-var QUALITY_FLOOR_FOR_RECOMMENDATION = 6;
-var TIER_DOWN_COST_RATIO = 0.5;
-var COST_MISMATCHED_CHOSEN_SCORE_CEILING = 7;
-var PRODUCER_OWNED_RULE_CODES = Object.freeze(
-  /* @__PURE__ */ new Set(["model-stale-evidence", "promote-ready"])
-);
-function deriveOwnership(code, selfDeclared) {
-  if (selfDeclared) return selfDeclared;
-  return PRODUCER_OWNED_RULE_CODES.has(code) ? "producer-owned" : "consumer-actionable";
+async function doRefresh4(rt, appId) {
+  const url = `${rt.endpoint}?app_id=${encodeURIComponent(appId)}`;
+  let snap = snapshots4.get(appId);
+  if (!snap) {
+    snap = { data: [], expiresAt: 0, refreshing: false };
+    snapshots4.set(appId, snap);
+  }
+  try {
+    const res = await rt.fetchImpl(url, { method: "GET" });
+    if (!res.ok) {
+      throw new Error(`promote-ready ${res.status}: ${res.statusText}`);
+    }
+    const body = await res.json();
+    if (runtime4 !== rt) return;
+    const rows = Array.isArray(body) ? mapRowsToFindings2(body) : [];
+    snap.data = rows;
+    snap.expiresAt = Date.now() + rt.ttlMs;
+    snap.refreshing = false;
+  } catch (err) {
+    if (runtime4 !== rt) return;
+    snap.refreshing = false;
+    snap.expiresAt = Date.now() + rt.ttlMs;
+    if (!warnedOnce4) {
+      warnedOnce4 = true;
+      (rt.onError ?? defaultOnError4)(err);
+    }
+  }
 }
-function runAdvisor(ir, result, profile, policy, phase2) {
-  const out = [];
-  out.push(...detectCachingOff(ir, profile));
-  out.push(...detectSingleChunkSystem(ir, profile));
-  out.push(...detectToolBloat(ir, result));
-  out.push(...detectHistoryUncached(ir, profile));
-  out.push(...detectSingleModelArray(ir, policy));
-  if (policy?.posture !== "locked") {
-    out.push(...detectCostMismatchedArchetype(ir, profile, phase2));
-    out.push(...detectModelStaleEvidence(ir, profile));
-    out.push(...detectTierDown(ir, profile, phase2));
+function defaultOnError4(err) {
+  console.warn(
+    "[kgauto] promote-ready fetch failed (using empty fallback):",
+    err
+  );
+}
+function resolveFetchImpl(injected) {
+  return injected ?? ((...args) => globalThis.fetch(...args));
+}
+function normalizeEndpoint(endpoint) {
+  return endpoint.replace(/\/+$/, "");
+}
+async function markPromoteReadyHandled(opts) {
+  const {
+    appId,
+    archetype,
+    family,
+    resolution,
+    resolutionNote,
+    brainEndpoint,
+    brainJwt,
+    brainAnonKey,
+    fetch: injectedFetch
+  } = opts;
+  if (!appId) return { ok: false, reason: "app_id_required" };
+  if (!archetype) return { ok: false, reason: "archetype_required" };
+  if (!family) return { ok: false, reason: "family_required" };
+  if (resolution !== "promoted" && resolution !== "declined" && resolution !== "still-evaluating") {
+    return { ok: false, reason: "resolution_invalid" };
   }
-  if (!translatorClearedToolCallCliff(phase2)) {
-    out.push(...detectArchetypePerfFloorBreach(ir, profile));
+  const doFetch = resolveFetchImpl(injectedFetch);
+  const base = normalizeEndpoint(brainEndpoint);
+  const url = `${base}/rest/v1/promote_ready_findings?app_id=eq.${encodeURIComponent(appId)}&intent_archetype=eq.${encodeURIComponent(archetype)}&family=eq.${encodeURIComponent(family)}&resolved_at=is.null`;
+  const patchBody = {
+    resolved_at: (/* @__PURE__ */ new Date()).toISOString(),
+    resolution
+  };
+  if (resolutionNote !== void 0) {
+    patchBody.resolution_note = resolutionNote;
   }
-  if (policy?.posture !== "locked") {
-    out.push(...detectStaleExclusionCandidate(ir));
+  let res;
+  try {
+    res = await doFetch(url, {
+      method: "PATCH",
+      headers: {
+        Authorization: `Bearer ${brainJwt}`,
+        apikey: brainAnonKey,
+        "Content-Type": "application/json",
+        Accept: "application/json",
+        Prefer: "return=minimal"
+      },
+      body: JSON.stringify(patchBody)
+    });
+  } catch (err) {
+    const msg = err instanceof Error ? err.message : String(err);
+    return { ok: false, reason: `network_error:${msg}` };
   }
-  if (policy?.posture !== "locked" && ir.appId) {
-    out.push(
-      ...advisorRulePromoteReady({
-        appId: ir.appId,
-        archetype: ir.intent.archetype,
-        resolvedPrimary: profile.id
-      })
-    );
-    out.push(...advisorRuleConsumerOnStaleModel(ir));
+  if (res.status === 401 || res.status === 403) {
+    return { ok: false, reason: "brain_auth_misconfig" };
   }
-  if (policy?.posture !== "locked") {
-    out.push(
-      ...advisorRuleCrossFamilyFit({
-        archetype: ir.intent.archetype,
-        resolvedPrimary: profile.id
-      })
-    );
+  if (res.status >= 500) {
+    return { ok: false, reason: "brain_unavailable" };
   }
-  return out;
-}
-function translatorClearedToolCallCliff(phase2) {
-  const rewrites = phase2?.sectionRewritesApplied;
-  if (!rewrites || rewrites.length === 0) return false;
-  for (const rw of rewrites) {
-    if (rw.kind === "tool_call_contract") return true;
+  if (!res.ok) {
+    return { ok: false, reason: `patch_failed:${res.status}` };
   }
-  return false;
+  return { ok: true };
 }
-function detectCachingOff(ir, profile) {
-  if (profile.provider !== "anthropic") return [];
-  const totalChars = ir.sections.reduce((s, sec) => s + sec.text.length, 0);
-  if (totalChars < 2e3) return [];
-  const anyCacheable = ir.sections.some((s) => s.cacheable === true);
-  if (anyCacheable) return [];
+
+// src/advisor-rules/promote-ready.ts
+var PROMOTE_READY_THRESHOLDS = {
+  minPassRate: 0.8,
+  minAvgScore: 4
+};
+function shouldFirePromoteReady(finding, resolvedPrimary) {
+  if (finding.currentModel !== resolvedPrimary) return false;
+  if (finding.judgePassRate < PROMOTE_READY_THRESHOLDS.minPassRate) return false;
+  if (finding.judgeAvgScore < PROMOTE_READY_THRESHOLDS.minAvgScore) return false;
+  return true;
+}
+function deriveFamilyLocal(modelId) {
+  if (modelId.startsWith("claude-opus-")) return "claude-opus";
+  if (modelId.startsWith("claude-sonnet-")) return "claude-sonnet";
+  if (modelId.startsWith("claude-haiku-")) return "claude-haiku";
+  if (/^gemini-.*-flash-lite/.test(modelId)) return "gemini-flash-lite";
+  if (/^gemini-.*-flash/.test(modelId)) return "gemini-flash";
+  if (/^gemini-.*-pro/.test(modelId)) return "gemini-pro";
+  if (/^deepseek-.*-pro/.test(modelId)) return "deepseek-reasoner";
+  if (modelId.startsWith("deepseek-")) return "deepseek-chat";
+  if (modelId.startsWith("gpt-")) return "openai-gpt";
+  return null;
+}
+function advisorRulePromoteReady(ctx) {
+  if (!isPromoteReadyBrainActive()) return [];
+  if (!ctx.appId) return [];
+  if (!ctx.resolvedPrimary) return [];
+  const family = deriveFamilyLocal(ctx.resolvedPrimary);
+  if (!family) return [];
+  const findings = loadPromoteReadyFindings({
+    appId: ctx.appId,
+    archetype: ctx.archetype,
+    family
+  });
+  if (findings.length === 0) return [];
+  const qualifying = findings.filter(
+    (f) => shouldFirePromoteReady(f, ctx.resolvedPrimary)
+  );
+  if (qualifying.length === 0) return [];
+  qualifying.sort((a, b) => {
+    if (a.judgeAvgScore !== b.judgeAvgScore) {
+      return b.judgeAvgScore - a.judgeAvgScore;
+    }
+    return b.judgePassRate - a.judgePassRate;
+  });
+  const top = qualifying[0];
+  const pctPass = Math.round(top.judgePassRate * 100);
+  const score = top.judgeAvgScore.toFixed(2);
+  let costClause = "";
+  if (top.costDeltaPct !== null) {
+    const sign = top.costDeltaPct < 0 ? "cheaper" : "more expensive";
+    const magnitude = Math.abs(top.costDeltaPct * 100).toFixed(1);
+    costClause = `, cost ${magnitude}% ${sign}`;
+  }
+  const message = `Probe found ${top.candidateModel} produces equivalent-or-better outputs vs ${top.currentModel} on ${top.sampleN} recent ${top.archetype} prompts (pass rate ${pctPass}%, avg score ${score}/5${costClause}). Consider promoting via markPromoteReadyHandled.`;
   return [
     {
-      level: "warn",
-      code: "caching-off-on-claude",
-      message: `System prompt is ${totalChars} chars on Anthropic but no PromptSection has cacheable=true. Anthropic prompt caching cuts cached-prefix input cost by ~90% on subsequent calls; without it, every turn re-pays full price for the static system context.`,
-      suggestion: "Mark stable system sections (role, persona, tool policy) with `cacheable: true`. The lowering pass concatenates cacheable sections into a single cache-controlled block before the dynamic ones.",
+      level: "info",
+      code: "promote-ready",
+      message,
+      suggestion: `Migrate ${top.archetype} traffic from ${top.currentModel} to ${top.candidateModel}, then call markPromoteReadyHandled({ appId, archetype: '${top.archetype}', family: '${top.family}', resolution: 'promoted' }) to silence this advisory.`,
+      // alpha.36 architectural field — not a no-ai-needed case.
+      recommendedArchitecture: void 0,
       docsUrl: "https://github.com/stue/command-center/blob/main/interfaces/kgauto.md#best-practice-advisories"
     }
   ];
 }
-function detectSingleChunkSystem(ir, profile) {
-  if (profile.provider !== "anthropic") return [];
-  if (ir.sections.length !== 1) return [];
-  const only = ir.sections[0];
-  if (!only || only.text.length <= 1e3) return [];
-  return [
-    {
-      level: "info",
-      code: "single-chunk-system",
-      message: `System prompt is a single ${only.text.length}-char chunk. Splitting into NamedChunks (static role/persona vs dynamic context) gives the lowering pass a finer cache-marker boundary \u2014 only the static portion needs to be byte-stable for the cache to hit.`,
-      suggestion: "Refactor the system builder to return an array of `PromptSection` shaped { id, text, cacheable?: boolean }. Static chunks (role, persona, tool policy) get `cacheable: true`; dynamic ones (current context, today's date) don't.",
-      docsUrl: "https://github.com/stue/command-center/blob/main/interfaces/kgauto.md#best-practice-advisories"
-    }
-  ];
+
+// src/advisor-rules/consumer-on-stale-model.ts
+function isStaleStatus(v) {
+  return v === "legacy" || v === "deprecated";
 }
-function detectToolBloat(ir, result) {
-  const SHORT_OUTPUT = /* @__PURE__ */ new Set([
-    "classify",
-    "extract",
-    "summarize",
-    "transform",
-    "critique"
-  ]);
-  if (!ir.tools || ir.tools.length === 0) return [];
-  const toolsKept = result.diagnostics.toolsKept;
-  if (toolsKept <= 10) return [];
-  if (!SHORT_OUTPUT.has(ir.intent.archetype)) return [];
-  return [
-    {
-      level: "warn",
-      code: "tool-bloat",
-      message: `${toolsKept} tools kept after the relevance pass for archetype="${ir.intent.archetype}" (consumer declared ${ir.tools.length}). This archetype is short-output and rarely needs more than 3 tools; each tool definition eats ~350 tokens of context budget.`,
-      suggestion: "Tighten `relevanceByIntent: { [archetype]: 0..1 }` per ToolDefinition. Tools below `toolRelevanceThreshold` (default 0.2) get dropped. Without `relevanceByIntent`, every tool defaults to neutral (0.5) and stays.",
-      docsUrl: "https://github.com/stue/kgauto/blob/main/v2/README.md#tools"
+function asString(v) {
+  return typeof v === "string" && v.length > 0 ? v : void 0;
+}
+function mapRowsToFindings3(rows) {
+  const out = [];
+  for (const raw of rows) {
+    if (!raw || typeof raw !== "object") continue;
+    const r = raw;
+    const archetype = asString(r.intent_archetype) ?? asString(r.applies_to_archetype);
+    const staleModel = asString(r.stale_model) ?? asString(r.applies_to_model);
+    const staleProvider = asString(r.stale_provider);
+    const recommendedModel = asString(r.recommended_model);
+    const family = asString(r.family);
+    const message = asString(r.message);
+    if (!archetype || !staleModel || !recommendedModel || !family || !message) {
+      continue;
     }
-  ];
+    if (!isStaleStatus(r.stale_status)) continue;
+    const row = {
+      archetype,
+      staleModel,
+      staleProvider: staleProvider ?? "unknown",
+      staleStatus: r.stale_status,
+      recommendedModel,
+      family,
+      message
+    };
+    const suggestion = asString(r.suggestion);
+    if (suggestion) row.suggestion = suggestion;
+    if (typeof r.observation_count === "number" && Number.isFinite(r.observation_count)) {
+      row.observationCount = r.observation_count;
+    }
+    out.push(row);
+  }
+  return out;
 }
-function detectHistoryUncached(ir, profile) {
-  if (profile.provider !== "anthropic") return [];
-  if (!ir.history || ir.history.length < 2) return [];
-  if (ir.historyCachePolicy && ir.historyCachePolicy.strategy !== "none") {
-    return [];
+var snapshots5 = /* @__PURE__ */ new Map();
+var runtime5;
+var warnedOnce5 = false;
+var pendingRefreshes5 = /* @__PURE__ */ new Map();
+function isStaleModelFindingsBrainActive() {
+  return runtime5 !== void 0;
+}
+function getStaleModelFindings(opts) {
+  const rt = runtime5;
+  if (!rt) return [];
+  const appId = opts.appId;
+  if (!appId) return [];
+  let snap = snapshots5.get(appId);
+  if (!snap) {
+    snap = { data: [], expiresAt: 0, refreshing: false };
+    snapshots5.set(appId, snap);
   }
-  return [
-    {
-      level: "warn",
-      code: "history-uncached-on-claude",
-      message: `${ir.history.length} history messages on Anthropic with no historyCachePolicy. Every turn re-pays for the full conversation context; with caching, subsequent turns hit the cache at ~10% the input cost.`,
-      suggestion: "Set `historyCachePolicy: { strategy: 'all-but-latest' }` on this IR. The lowering pass marks the message immediately preceding currentTurn with cache_control; subsequent turns whose history prefix matches byte-for-byte hit the cache.",
-      docsUrl: "https://github.com/stue/command-center/blob/main/interfaces/kgauto.md#best-practice-advisories"
-    }
-  ];
+  const now = Date.now();
+  const stale = snap.expiresAt <= now;
+  if (stale && !snap.refreshing) {
+    snap.refreshing = true;
+    void asyncRefresh5(rt, appId);
+  }
+  if (opts.archetype) {
+    return snap.data.filter((f) => f.archetype === opts.archetype);
+  }
+  return snap.data;
 }
-function detectSingleModelArray(ir, policy) {
-  if (ir.models.length !== 1) return [];
-  if (policy?.posture === "locked") return [];
-  const only = ir.models[0];
-  return [
-    {
-      level: "warn",
-      code: "single-model-array",
-      message: `\`ir.models\` has length 1 (only "${only}") and posture is not 'locked'. A single-model chain has no safety net \u2014 the first 429 / 5xx / cliff hits the user as a failure. Master plan \xA71.2 closes the reliability gap with a 2-step minimum.`,
-      suggestion: "Use `getDefaultFallbackChain({ archetype: ir.intent.archetype, primary: '" + only + "', posture: 'preferred' })` for a user-anchored chain, or `getDefaultFallbackChain({ archetype, posture: 'open' })` for library-picked. If single-model is intentional (compliance/brand promise), set `policy.posture = 'locked'` to silence this rule.",
-      docsUrl: "https://github.com/stue/command-center/blob/main/interfaces/kgauto.md#single-model-array"
+async function asyncRefresh5(rt, appId) {
+  const promise = doRefresh5(rt, appId);
+  pendingRefreshes5.set(appId, promise);
+  try {
+    await promise;
+  } finally {
+    if (pendingRefreshes5.get(appId) === promise) {
+      pendingRefreshes5.delete(appId);
     }
-  ];
+  }
 }
-function detectCostMismatchedArchetype(ir, profile, phase2) {
-  if (!phase2 || phase2.fallbackChain.length === 0) return [];
-  if (!phase2.profileResolver) return [];
-  const archetype = ir.intent.archetype;
-  const chosenScore = getArchetypePerfScore(profile.id, archetype);
-  const chosenHasRoomToGrow = chosenScore.grounding === "judgment" || chosenScore.score < COST_MISMATCHED_CHOSEN_SCORE_CEILING;
-  if (!chosenHasRoomToGrow) return [];
-  let bestAlt = null;
-  for (const altId of phase2.fallbackChain) {
-    const altProfile = phase2.profileResolver(altId);
-    if (!altProfile) continue;
-    if (altProfile.id === profile.id) continue;
-    const altScore = getArchetypePerfScore(altProfile.id, archetype);
-    if (altScore.score < QUALITY_FLOOR_FOR_RECOMMENDATION) continue;
-    if (altScore.score < chosenScore.score) continue;
-    if (altProfile.costInputPer1m >= profile.costInputPer1m) continue;
-    if (!bestAlt || altScore.score > bestAlt.score.score || altScore.score === bestAlt.score.score && altProfile.costInputPer1m < bestAlt.profile.costInputPer1m) {
-      bestAlt = { id: altId, profile: altProfile, score: altScore };
-    }
+async function doRefresh5(rt, appId) {
+  const url = `${rt.endpoint}?app_id=${encodeURIComponent(appId)}`;
+  let snap = snapshots5.get(appId);
+  if (!snap) {
+    snap = { data: [], expiresAt: 0, refreshing: false };
+    snapshots5.set(appId, snap);
   }
-  if (!bestAlt) return [];
-  const tierDownWouldFire = bestAlt.score.grounding === "measured" && bestAlt.profile.costInputPer1m <= profile.costInputPer1m * TIER_DOWN_COST_RATIO;
-  if (tierDownWouldFire) return [];
-  const chosenGrounding = chosenScore.grounding === "judgment" ? `archetypePerf.${archetype}=judgment` : `archetypePerf.${archetype}=${chosenScore.score}`;
-  const altGrounding = bestAlt.score.grounding === "measured" ? `archetypePerf.${archetype}=${bestAlt.score.score}, measured, n=${bestAlt.score.n}` : `archetypePerf.${archetype}=${bestAlt.score.score}, judgment`;
-  return [
-    {
-      level: "warn",
-      code: "cost-mismatched-archetype",
-      message: `Cost-mismatched-archetype: target=${profile.id} (${chosenGrounding}) selected for ${archetype}. Alternative ${bestAlt.id} (${altGrounding}) is cheaper ($${bestAlt.profile.costInputPer1m}/$${bestAlt.profile.costOutputPer1m} vs $${profile.costInputPer1m}/$${profile.costOutputPer1m} per 1M) at equal-or-better quality.`,
-      suggestion: `Consider declaring \`${bestAlt.id}\` as the primary model for this archetype, or relax to posture='open' to let kgauto select among the chain. If the chosen model is required for compliance/brand reasons, set \`policy.posture = 'locked'\` to silence this rule.`,
-      recommendationType: profile.provider === bestAlt.profile.provider ? "tier-down" : "model-swap",
-      docsUrl: "https://github.com/stue/command-center/blob/main/interfaces/kgauto.md#best-practice-advisories"
+  try {
+    const res = await rt.fetchImpl(url, { method: "GET" });
+    if (!res.ok) {
+      throw new Error(`stale-model findings ${res.status}: ${res.statusText}`);
     }
-  ];
-}
-function detectModelStaleEvidence(ir, profile) {
-  if (!isBrainQueryActiveFor("kgauto_archetype_perf")) return [];
-  const archetype = ir.intent.archetype;
-  const chosen = getArchetypePerfScore(profile.id, archetype);
-  if (chosen.grounding !== "judgment") return [];
-  return [
-    {
-      level: "info",
-      code: "model-stale-evidence",
-      message: `Model-stale-evidence: target=${profile.id} archetype=${archetype} is judgment-grounded (n=${chosen.n}) despite brain-query mode being active. Measurement substrate is wired but the brain hasn't accumulated >=10 outcomes for this (model, archetype) tuple yet \u2014 routing decisions remain pre-measured for this slot.`,
-      suggestion: "Verify that `record()` is being called on every call() outcome with the appropriate `actualModel` and `mutationsApplied` fields. Once the brain accumulates n>=10 rows on this tuple, the score promotes from judgment to measured automatically (5-min SWR cache). No code change required from your side \u2014 this is the substrate signaling the gap.",
-      recommendationType: "prompt-fix",
-      docsUrl: "https://github.com/stue/command-center/blob/main/interfaces/kgauto.md#best-practice-advisories"
+    const body = await res.json();
+    if (runtime5 !== rt) return;
+    const rows = Array.isArray(body) ? mapRowsToFindings3(body) : [];
+    snap.data = rows;
+    snap.expiresAt = Date.now() + rt.ttlMs;
+    snap.refreshing = false;
+  } catch (err) {
+    if (runtime5 !== rt) return;
+    snap.refreshing = false;
+    snap.expiresAt = Date.now() + rt.ttlMs;
+    if (!warnedOnce5) {
+      warnedOnce5 = true;
+      (rt.onError ?? defaultOnError5)(err);
     }
-  ];
+  }
 }
-function detectTierDown(ir, profile, phase2) {
-  if (!phase2 || phase2.fallbackChain.length === 0) return [];
-  if (!phase2.profileResolver) return [];
-  const archetype = ir.intent.archetype;
-  const chosenScore = getArchetypePerfScore(profile.id, archetype);
-  const chosenCost = profile.costInputPer1m;
-  let bestAlt = null;
-  for (const altId of phase2.fallbackChain) {
-    const altProfile = phase2.profileResolver(altId);
-    if (!altProfile) continue;
-    if (altProfile.id === profile.id) continue;
-    const altScore = getArchetypePerfScore(altProfile.id, archetype);
-    if (altScore.grounding !== "measured") continue;
-    if (altScore.score < QUALITY_FLOOR_FOR_RECOMMENDATION) continue;
-    if (altScore.score < chosenScore.score) continue;
-    if (altProfile.costInputPer1m > chosenCost * TIER_DOWN_COST_RATIO) continue;
-    if (!bestAlt || altProfile.costInputPer1m < bestAlt.profile.costInputPer1m || altProfile.costInputPer1m === bestAlt.profile.costInputPer1m && altScore.score > bestAlt.score.score) {
-      bestAlt = { id: altId, profile: altProfile, score: altScore };
+function defaultOnError5(err) {
+  console.warn(
+    "[kgauto] stale-model findings fetch failed (using empty fallback):",
+    err
+  );
+}
+var CONSUMER_ON_STALE_MODEL_RULE_CODE = "consumer-on-stale-model";
+function advisorRuleConsumerOnStaleModel(ir) {
+  if (!isStaleModelFindingsBrainActive()) return [];
+  if (!ir.appId) return [];
+  const findings = getStaleModelFindings({
+    appId: ir.appId,
+    archetype: ir.intent.archetype
+  });
+  if (findings.length === 0) return [];
+  const ranked = [...findings].sort((a, b) => {
+    if (a.staleStatus !== b.staleStatus) {
+      return a.staleStatus === "deprecated" ? -1 : 1;
     }
-  }
-  if (!bestAlt) return [];
-  const chosenDesc = chosenScore.grounding === "measured" ? `archetypePerf.${archetype}=${chosenScore.score} (measured, n=${chosenScore.n})` : `archetypePerf.${archetype}=${chosenScore.score} (${chosenScore.grounding})`;
+    return a.staleModel.localeCompare(b.staleModel);
+  });
+  const top = ranked[0];
+  const extraCount = findings.length - 1;
+  const extraNote = extraCount > 0 ? ` (+ ${extraCount} more stale model${extraCount === 1 ? "" : "s"} for this archetype)` : "";
   return [
     {
       level: "warn",
-      code: "tier-down",
-      message: `Tier-down: target=${profile.id} (${chosenDesc}) selected for ${archetype}. Brain shows ${bestAlt.id} delivers equal-or-better quality (archetypePerf.${archetype}=${bestAlt.score.score}, measured, n=${bestAlt.score.n}) at $${bestAlt.profile.costInputPer1m}/$${bestAlt.profile.costOutputPer1m} per 1M vs $${profile.costInputPer1m}/$${profile.costOutputPer1m} \u2014 a measured tier-down opportunity.`,
-      suggestion: `Move \`${bestAlt.id}\` to primary for this archetype. The brain has n=${bestAlt.score.n} measured outcomes backing the recommendation; this is data, not opinion. If posture='locked' is required (compliance/brand promise), set it explicitly to silence this rule.`,
-      recommendationType: "tier-down",
+      code: CONSUMER_ON_STALE_MODEL_RULE_CODE,
+      message: `${top.message}${extraNote}`,
+      suggestion: top.suggestion ?? `Migrate ${top.staleModel} \u2192 ${top.recommendedModel} for archetype "${top.archetype}". The newer model is the current latest in the "${top.family}" family; the stale one is ${top.staleStatus}.`,
+      recommendationType: "model-swap",
       docsUrl: "https://github.com/stue/command-center/blob/main/interfaces/kgauto.md#best-practice-advisories"
     }
   ];
 }
-function detectArchetypePerfFloorBreach(ir, profile) {
-  const compat = getModelCompatibility(profile.id, {
-    archetype: ir.intent.archetype,
-    toolOrchestration: ir.constraints?.toolOrchestration
-  });
-  if (compat.status === "compatible") return [];
-  if (compat.status === "requires-adapter") {
-    return [
-      {
-        level: "warn",
-        code: "archetype-perf-floor-breach",
-        message: `${profile.id} sits below the archetype floor for ${ir.intent.archetype} (score ${compat.archetypePerf}/10, floor ${6}). A known adapter would lift it: ${compat.adapter.parameter}=${compat.adapter.value}. ${compat.adapter.consequence}`,
-        suggestion: `Pass \`ir.constraints.${compat.adapter.parameter} = '${compat.adapter.value}'\` for this call, OR pick a model whose archetypePerf for ${ir.intent.archetype} already clears the floor (call \`getModelCompatibility(modelId, { archetype: '${ir.intent.archetype}' })\` to check). Estimated post-adapter score: ${compat.archetypePerfWithAdapter}/10.`,
-        recommendationType: "prompt-fix",
-        suggestedAdaptation: compat.adapter,
-        docsUrl: "https://github.com/stue/command-center/blob/main/interfaces/kgauto.md#best-practice-advisories"
-      }
-    ];
+
+// src/archetype-fits.ts
+var ARCHETYPE_FAMILY_FITS = Object.freeze([
+  {
+    archetype: "plan",
+    betterFitFamily: "deepseek-reasoner",
+    reason: "Plan archetype is reasoning-shaped (multi-step chains, hypothesis-and-check, sub-goal decomposition) \u2014 exactly where reasoner-family models excel. Sonnet/Opus produce plans but at higher cost; reasoners produce equivalent-or-better plans at 7-17x lower cost at current promo pricing (deepseek-v4-pro $0.435/$0.87 per 1M promo through 2026-05-31 vs sonnet $3/$15).",
+    costGuidance: "substantially cheaper at current pricing (deepseek-v4-pro promo: ~7-17x cheaper than sonnet)"
+  },
+  {
+    archetype: "critique",
+    betterFitFamily: "deepseek-reasoner",
+    reason: "Critique archetype rewards epistemic humility and explicit reasoning \u2014 reasoner-family default behavior. Sonnet/Opus over-confident on critique tasks; reasoners surface uncertainty productively.",
+    costGuidance: "comparable or cheaper at current pricing"
+  }
+]);
+function findBetterFit(archetype, currentFamily) {
+  for (const fit of ARCHETYPE_FAMILY_FITS) {
+    if (fit.archetype !== archetype) continue;
+    if (fit.betterFitFamily === currentFamily) return null;
+    return fit;
+  }
+  return null;
+}
+
+// src/advisor-rules/cross-family-fit.ts
+function familyHasCurrentActiveModel(family) {
+  for (const profile of allProfiles()) {
+    const profileFamily = profile.family ?? deriveFamilyFromModelId(profile.id);
+    if (profileFamily !== family) continue;
+    if (profile.status !== "current") continue;
+    if (profile.active === false) continue;
+    return true;
+  }
+  return false;
+}
+function listCandidatesInFamily(family) {
+  const candidates = [];
+  for (const profile of allProfiles()) {
+    const profileFamily = profile.family ?? deriveFamilyFromModelId(profile.id);
+    if (profileFamily !== family) continue;
+    if (profile.status !== "current") continue;
+    if (profile.active === false) continue;
+    candidates.push(profile.id);
+    if (candidates.length >= 3) break;
   }
+  return candidates;
+}
+function advisorRuleCrossFamilyFit(ctx) {
+  if (!ctx.resolvedPrimary) return [];
+  const currentFamily = deriveFamilyFromModelId(ctx.resolvedPrimary);
+  if (!currentFamily) return [];
+  const fit = findBetterFit(ctx.archetype, currentFamily);
+  if (!fit) return [];
+  if (!familyHasCurrentActiveModel(fit.betterFitFamily)) return [];
+  const candidates = listCandidatesInFamily(fit.betterFitFamily);
+  if (candidates.length === 0) return [];
+  const candidateStr = candidates.join(", ");
+  const message = `Your ${currentFamily} call on ${ctx.archetype} could shift to ${fit.betterFitFamily} \u2014 typically better quality + ${fit.costGuidance}. Suggested candidates: ${candidateStr}.`;
   return [
     {
-      level: "critical",
-      code: "archetype-perf-floor-breach",
-      message: `${profile.id} sits below the archetype floor for ${ir.intent.archetype} (score ${compat.archetypePerf}/10, floor ${6}) and no known adapter would lift it. ${compat.reason}`,
-      suggestion: `Swap to a model whose archetypePerf for ${ir.intent.archetype} clears the floor. Use \`getModelCompatibility(candidateId, { archetype: '${ir.intent.archetype}' })\` to vet candidates, or \`getDefaultFallbackChain({ archetype: '${ir.intent.archetype}', posture: 'open' })\` for a library-picked chain that respects the floor by construction.`,
+      level: "info",
+      code: "cross-family-fit-candidate",
+      ownership: "consumer-actionable",
+      message,
+      suggestion: `Swap the model literal in \`ir.models\` to one of: ${candidateStr}. Or call \`getRecommendedPrimary({ family: '${fit.betterFitFamily}', archetype: '${ctx.archetype}', fallback: { id: '${candidates[0]}', reason: 'cross-family-fit-recommendation' } })\` to let kgauto resolve to the current+active family member.`,
       recommendationType: "model-swap",
       docsUrl: "https://github.com/stue/command-center/blob/main/interfaces/kgauto.md#best-practice-advisories"
     }
   ];
 }
-function detectStaleExclusionCandidate(ir) {
-  if (!isExclusionFindingsBrainActive()) return [];
-  if (!ir.appId) return [];
-  const findings = getStaleExclusionFindings({
-    appId: ir.appId,
-    archetype: ir.intent.archetype
-  });
-  if (findings.length === 0) return [];
-  const ranked = [...findings].sort((a, b) => {
-    const sa = a.estimatedSavingsUsd30d ?? -Infinity;
-    const sb = b.estimatedSavingsUsd30d ?? -Infinity;
-    if (sa !== sb) return sb - sa;
-    return confidenceRank(b.confidence) - confidenceRank(a.confidence);
-  });
-  const top = ranked[0];
-  const extraCount = findings.length - 1;
-  const extraNote = extraCount > 0 ? ` (+ ${extraCount} more excluded model${extraCount === 1 ? "" : "s"} for this archetype)` : "";
+
+// src/advisor.ts
+var QUALITY_FLOOR_FOR_RECOMMENDATION = 6;
+var TIER_DOWN_COST_RATIO = 0.5;
+var COST_MISMATCHED_CHOSEN_SCORE_CEILING = 7;
+var PRODUCER_OWNED_RULE_CODES = Object.freeze(
+  /* @__PURE__ */ new Set(["model-stale-evidence", "promote-ready"])
+);
+function deriveOwnership(code, selfDeclared) {
+  if (selfDeclared) return selfDeclared;
+  return PRODUCER_OWNED_RULE_CODES.has(code) ? "producer-owned" : "consumer-actionable";
+}
+function runAdvisor(ir, result, profile, policy, phase2) {
+  const out = [];
+  out.push(...detectCachingOff(ir, profile));
+  out.push(...detectSingleChunkSystem(ir, profile));
+  out.push(...detectToolBloat(ir, result));
+  out.push(...detectHistoryUncached(ir, profile));
+  out.push(...detectSingleModelArray(ir, policy));
+  if (policy?.posture !== "locked") {
+    out.push(...detectCostMismatchedArchetype(ir, profile, phase2));
+    out.push(...detectModelStaleEvidence(ir, profile));
+    out.push(...detectTierDown(ir, profile, phase2));
+  }
+  if (!translatorClearedToolCallCliff(phase2)) {
+    out.push(...detectArchetypePerfFloorBreach(ir, profile));
+  }
+  if (policy?.posture !== "locked") {
+    out.push(...detectStaleExclusionCandidate(ir));
+  }
+  if (policy?.posture !== "locked" && ir.appId) {
+    out.push(
+      ...advisorRulePromoteReady({
+        appId: ir.appId,
+        archetype: ir.intent.archetype,
+        resolvedPrimary: profile.id
+      })
+    );
+    out.push(...advisorRuleConsumerOnStaleModel(ir));
+  }
+  if (policy?.posture !== "locked") {
+    out.push(
+      ...advisorRuleCrossFamilyFit({
+        archetype: ir.intent.archetype,
+        resolvedPrimary: profile.id
+      })
+    );
+  }
+  return out;
+}
+function translatorClearedToolCallCliff(phase2) {
+  const rewrites = phase2?.sectionRewritesApplied;
+  if (!rewrites || rewrites.length === 0) return false;
+  for (const rw of rewrites) {
+    if (rw.kind === "tool_call_contract") return true;
+  }
+  return false;
+}
+function detectCachingOff(ir, profile) {
+  if (profile.provider !== "anthropic") return [];
+  const totalChars = ir.sections.reduce((s, sec) => s + sec.text.length, 0);
+  if (totalChars < 2e3) return [];
+  const anyCacheable = ir.sections.some((s) => s.cacheable === true);
+  if (anyCacheable) return [];
   return [
     {
-      level: "info",
-      code: "stale-exclusion-candidate",
-      message: `${top.message}${extraNote}`,
-      suggestion: top.suggestion,
-      recommendationType: "tier-down",
+      level: "warn",
+      code: "caching-off-on-claude",
+      message: `System prompt is ${totalChars} chars on Anthropic but no PromptSection has cacheable=true. Anthropic prompt caching cuts cached-prefix input cost by ~90% on subsequent calls; without it, every turn re-pays full price for the static system context.`,
+      suggestion: "Mark stable system sections (role, persona, tool policy) with `cacheable: true`. The lowering pass concatenates cacheable sections into a single cache-controlled block before the dynamic ones.",
       docsUrl: "https://github.com/stue/command-center/blob/main/interfaces/kgauto.md#best-practice-advisories"
     }
   ];
 }
-function confidenceRank(c) {
-  if (c === "high") return 3;
-  if (c === "medium") return 2;
-  return 1;
+function detectSingleChunkSystem(ir, profile) {
+  if (profile.provider !== "anthropic") return [];
+  if (ir.sections.length !== 1) return [];
+  const only = ir.sections[0];
+  if (!only || only.text.length <= 1e3) return [];
+  return [
+    {
+      level: "info",
+      code: "single-chunk-system",
+      message: `System prompt is a single ${only.text.length}-char chunk. Splitting into NamedChunks (static role/persona vs dynamic context) gives the lowering pass a finer cache-marker boundary \u2014 only the static portion needs to be byte-stable for the cache to hit.`,
+      suggestion: "Refactor the system builder to return an array of `PromptSection` shaped { id, text, cacheable?: boolean }. Static chunks (role, persona, tool policy) get `cacheable: true`; dynamic ones (current context, today's date) don't. NOTE: the lowering pass HOISTS cacheable sections ahead of dynamic ones on the Anthropic wire (prefix caching requires it) \u2014 if your prompt has a protected ordering (e.g. a voice/persona block that must precede boilerplate), splitting will reorder the compiled output; declining this advisory is then correct. Also: a cacheable block under ~1024 tokens gets NO cache_control marker (provider minimum), so marking small sections is inert, not harmful.",
+      docsUrl: "https://github.com/stue/command-center/blob/main/interfaces/kgauto.md#best-practice-advisories"
+    }
+  ];
 }
-
-// src/translator.ts
-var TRANSLATOR_FLOOR = ARCHETYPE_FLOOR_DEFAULT;
-var RULE_SEQUENTIAL_TOOL_CLIFF = "sequential-tool-cliff-below-floor";
-var RULE_NARRATION_DRIFT_ANTHROPIC = "narration-drift-anthropic";
-var RULE_NARRATION_THINKING_LEAK_DEEPSEEK = "narration-thinking-leak-deepseek";
-var SEQUENTIAL_TOOL_PREAMBLE = "IMPORTANT: Use one tool call per response. Wait for the tool result before deciding the next tool. Do NOT batch tool calls in parallel.";
-var NARRATION_DRIFT_ANTHROPIC_PREAMBLE = "Output ONLY the requested content. Do not narrate your thought process. Each line \u2264 12 words.";
-var NARRATION_THINKING_LEAK_DEEPSEEK_PREAMBLE = "Reasoning is internal. Output ONLY the requested content; do not emit <thinking> blocks or internal monologue as user-facing text.";
-var RULE_DISCIPLINE_GATES_V1 = "discipline-gates-v1";
-var DISCIPLINE_GATES_V1_WITH_TOOLS = `Work through these gates at every judgment point, explicitly:
-1. Evidence before reasoning: cite what you observed before concluding from it.
-2. One extra signal: when a finding feels conclusive, check one more adjacent signal before stating it.
-3. Expand, don't guess: resolve a compressed or referenced item by looking it up rather than inferring its contents.
-4. Innocent explanation first: state the most plausible benign reading before alleging the alarming one.
-5. Label each claim: mark it observed, inferred, or assumed.
-6. A surfaced gap beats a guessed answer: flag what you cannot determine rather than fabricating past it.`;
-var DISCIPLINE_GATES_V1_NO_TOOLS = `Work through these gates at every judgment point, explicitly:
-1. Evidence before reasoning: cite what you observed before concluding from it.
-2. One extra signal: when a finding feels conclusive, check one more adjacent signal before stating it.
-3. Innocent explanation first: state the most plausible benign reading before alleging the alarming one.
-4. Label each claim: mark it observed, inferred, or assumed.
-5. A surfaced gap beats a guessed answer: flag what you cannot determine rather than fabricating past it.`;
-var RULE_DISCIPLINE_GATES_V1_STRUCTURED = "discipline-gates-v1-structured";
-var DISCIPLINE_GATES_V1_STRUCTURED_WITH_TOOLS = `Work through these gates at every judgment point, explicitly:
-1. Evidence before reasoning: cite what you observed before concluding from it.
-2. One extra signal: when a finding feels conclusive, check one more adjacent signal before stating it.
-3. Expand, don't guess: resolve a compressed or referenced item by looking it up rather than inferring its contents.
-4. Innocent explanation first: state the most plausible benign reading before alleging the alarming one.`;
-var DISCIPLINE_GATES_V1_STRUCTURED_NO_TOOLS = `Work through these gates at every judgment point, explicitly:
-1. Evidence before reasoning: cite what you observed before concluding from it.
-2. One extra signal: when a finding feels conclusive, check one more adjacent signal before stating it.
-3. Innocent explanation first: state the most plausible benign reading before alleging the alarming one.`;
-var DISCIPLINE_ELIGIBLE_ARCHETYPES = /* @__PURE__ */ new Set([
-  "hunt",
-  "summarize",
-  "plan",
-  "critique",
-  "judge"
-]);
-function matchRule(kind, profile, archetype, ctx) {
-  if (kind === "discipline_contract") {
-    if (!DISCIPLINE_ELIGIBLE_ARCHETYPES.has(archetype)) return null;
-    if (ctx.outputMode !== "text") {
-      return {
-        id: RULE_DISCIPLINE_GATES_V1_STRUCTURED,
-        preamble: ctx.hasTools ? DISCIPLINE_GATES_V1_STRUCTURED_WITH_TOOLS : DISCIPLINE_GATES_V1_STRUCTURED_NO_TOOLS
-      };
+function detectToolBloat(ir, result) {
+  const SHORT_OUTPUT = /* @__PURE__ */ new Set([
+    "classify",
+    "extract",
+    "summarize",
+    "transform",
+    "critique"
+  ]);
+  if (!ir.tools || ir.tools.length === 0) return [];
+  const toolsKept = result.diagnostics.toolsKept;
+  if (toolsKept <= 10) return [];
+  if (!SHORT_OUTPUT.has(ir.intent.archetype)) return [];
+  return [
+    {
+      level: "warn",
+      code: "tool-bloat",
+      message: `${toolsKept} tools kept after the relevance pass for archetype="${ir.intent.archetype}" (consumer declared ${ir.tools.length}). This archetype is short-output and rarely needs more than 3 tools; each tool definition eats ~350 tokens of context budget.`,
+      suggestion: "Tighten `relevanceByIntent: { [archetype]: 0..1 }` per ToolDefinition. Tools below `toolRelevanceThreshold` (default 0.2) get dropped. Without `relevanceByIntent`, every tool defaults to neutral (0.5) and stays.",
+      docsUrl: "https://github.com/stue/kgauto/blob/main/v2/README.md#tools"
     }
-    return {
-      id: RULE_DISCIPLINE_GATES_V1,
-      preamble: ctx.hasTools ? DISCIPLINE_GATES_V1_WITH_TOOLS : DISCIPLINE_GATES_V1_NO_TOOLS
-    };
-  }
-  if (kind === "tool_call_contract") {
-    if (!profile.archetypePerf) return null;
-    const archetypeScore = profile.archetypePerf[archetype];
-    if (typeof archetypeScore !== "number" || archetypeScore >= TRANSLATOR_FLOOR) {
-      return null;
-    }
-    return {
-      id: RULE_SEQUENTIAL_TOOL_CLIFF,
-      preamble: SEQUENTIAL_TOOL_PREAMBLE,
-      wireOverrides: { parallelToolCalls: false }
-    };
+  ];
+}
+function detectHistoryUncached(ir, profile) {
+  if (profile.provider !== "anthropic") return [];
+  if (!ir.history || ir.history.length < 2) return [];
+  if (ir.historyCachePolicy && ir.historyCachePolicy.strategy !== "none") {
+    return [];
   }
-  if (kind === "narration_contract") {
-    if (profile.provider === "anthropic") {
-      return {
-        id: RULE_NARRATION_DRIFT_ANTHROPIC,
-        preamble: NARRATION_DRIFT_ANTHROPIC_PREAMBLE
-      };
+  return [
+    {
+      level: "warn",
+      code: "history-uncached-on-claude",
+      message: `${ir.history.length} history messages on Anthropic with no historyCachePolicy. Every turn re-pays for the full conversation context; with caching, subsequent turns hit the cache at ~10% the input cost.`,
+      suggestion: "Set `historyCachePolicy: { strategy: 'all-but-latest' }` on this IR. The lowering pass marks the message immediately preceding currentTurn with cache_control; subsequent turns whose history prefix matches byte-for-byte hit the cache.",
+      docsUrl: "https://github.com/stue/command-center/blob/main/interfaces/kgauto.md#best-practice-advisories"
     }
-    if (profile.provider === "deepseek") {
-      return {
-        id: RULE_NARRATION_THINKING_LEAK_DEEPSEEK,
-        preamble: NARRATION_THINKING_LEAK_DEEPSEEK_PREAMBLE
-      };
+  ];
+}
+function detectSingleModelArray(ir, policy) {
+  if (ir.models.length !== 1) return [];
+  if (policy?.posture === "locked") return [];
+  const only = ir.models[0];
+  return [
+    {
+      level: "warn",
+      code: "single-model-array",
+      message: `\`ir.models\` has length 1 (only "${only}") and posture is not 'locked'. A single-model chain has no safety net \u2014 the first 429 / 5xx / cliff hits the user as a failure. Master plan \xA71.2 closes the reliability gap with a 2-step minimum.`,
+      suggestion: "Use `getDefaultFallbackChain({ archetype: ir.intent.archetype, primary: '" + only + "', posture: 'preferred' })` for a user-anchored chain, or `getDefaultFallbackChain({ archetype, posture: 'open' })` for library-picked. If single-model is intentional (compliance/brand promise), set `policy.posture = 'locked'` to silence this rule.",
+      docsUrl: "https://github.com/stue/command-center/blob/main/interfaces/kgauto.md#single-model-array"
     }
-    return null;
-  }
-  return null;
+  ];
 }
-function applySectionRewrites(args) {
-  const { ir, profile, archetype } = args;
-  if (!Array.isArray(ir.sections) || ir.sections.length === 0) {
-    return { rewrittenIR: ir, rewrites: [] };
-  }
-  const outputMode = args.outputMode ?? resolveOutputMode({
-    declared: ir.constraints?.outputMode,
-    structuredOutput: ir.constraints?.structuredOutput,
-    toolCount: ir.tools?.length ?? 0
-  });
-  const hasTools = (ir.tools?.length ?? 0) > 0;
-  const ctx = { outputMode, hasTools };
-  const rewrites = [];
-  const newSections = ir.sections.map((section) => {
-    if (!section.kind || section.kind === "arbitrary") return section;
-    const rule = matchRule(section.kind, profile, archetype, ctx);
-    if (!rule) return section;
-    const originalText = section.text;
-    const transformedText = `${rule.preamble}
-
-${originalText}`;
-    rewrites.push({
-      sectionId: section.id,
-      kind: section.kind,
-      rule: rule.id,
-      originalText,
-      transformedText,
-      ...rule.wireOverrides ? { wireOverrides: rule.wireOverrides } : {}
-    });
-    return { ...section, text: transformedText };
-  });
-  if (rewrites.length === 0) {
-    return { rewrittenIR: ir, rewrites: [] };
+function detectCostMismatchedArchetype(ir, profile, phase2) {
+  if (!phase2 || phase2.fallbackChain.length === 0) return [];
+  if (!phase2.profileResolver) return [];
+  const archetype = ir.intent.archetype;
+  const chosenScore = getArchetypePerfScore(profile.id, archetype);
+  const chosenHasRoomToGrow = chosenScore.grounding === "judgment" || chosenScore.score < COST_MISMATCHED_CHOSEN_SCORE_CEILING;
+  if (!chosenHasRoomToGrow) return [];
+  let bestAlt = null;
+  for (const altId of phase2.fallbackChain) {
+    const altProfile = phase2.profileResolver(altId);
+    if (!altProfile) continue;
+    if (altProfile.id === profile.id) continue;
+    const altScore = getArchetypePerfScore(altProfile.id, archetype);
+    if (altScore.score < QUALITY_FLOOR_FOR_RECOMMENDATION) continue;
+    if (altScore.score < chosenScore.score) continue;
+    if (altProfile.costInputPer1m >= profile.costInputPer1m) continue;
+    if (getMeasuredFailureVerdict({
+      appId: ir.appId,
+      archetype,
+      model: altProfile.id
+    })?.gated === true) {
+      continue;
+    }
+    if (ir.constraints?.structuredOutput && effectiveConventions(altProfile).some(
+      (c) => c.archetype === archetype && c.structuredOutputHint === "avoid"
+    )) {
+      continue;
+    }
+    if (getRecentRollback({
+      appId: ir.appId,
+      archetype,
+      model: altProfile.id
+    }) !== void 0) {
+      continue;
+    }
+    if (!bestAlt || altScore.score > bestAlt.score.score || altScore.score === bestAlt.score.score && altProfile.costInputPer1m < bestAlt.profile.costInputPer1m) {
+      bestAlt = { id: altId, profile: altProfile, score: altScore };
+    }
   }
-  const rewrittenIR = { ...ir, sections: newSections };
-  return { rewrittenIR, rewrites };
+  if (!bestAlt) return [];
+  const tierDownWouldFire = bestAlt.score.grounding === "measured" && bestAlt.profile.costInputPer1m <= profile.costInputPer1m * TIER_DOWN_COST_RATIO;
+  if (tierDownWouldFire) return [];
+  const chosenGrounding = chosenScore.grounding === "judgment" ? `archetypePerf.${archetype}=judgment` : `archetypePerf.${archetype}=${chosenScore.score}`;
+  const altGrounding = bestAlt.score.grounding === "measured" ? `archetypePerf.${archetype}=${bestAlt.score.score}, measured, n=${bestAlt.score.n}` : `archetypePerf.${archetype}=${bestAlt.score.score}, judgment`;
+  return [
+    {
+      level: "warn",
+      code: "cost-mismatched-archetype",
+      message: `Cost-mismatched-archetype: target=${profile.id} (${chosenGrounding}) selected for ${archetype}. Alternative ${bestAlt.id} (${altGrounding}) is cheaper ($${bestAlt.profile.costInputPer1m}/$${bestAlt.profile.costOutputPer1m} vs $${profile.costInputPer1m}/$${profile.costOutputPer1m} per 1M) at equal-or-better quality.`,
+      suggestion: `Consider declaring \`${bestAlt.id}\` as the primary model for this archetype, or relax to posture='open' to let kgauto select among the chain. If the chosen model is required for compliance/brand reasons, set \`policy.posture = 'locked'\` to silence this rule.`,
+      recommendationType: profile.provider === bestAlt.profile.provider ? "tier-down" : "model-swap",
+      docsUrl: "https://github.com/stue/command-center/blob/main/interfaces/kgauto.md#best-practice-advisories"
+    }
+  ];
 }
-
-// src/promotions-brain.ts
-function isRawPromotionRow(x) {
-  if (!x || typeof x !== "object") return false;
-  const r = x;
-  return (typeof r.id === "number" || typeof r.id === "string") && typeof r.intent_archetype === "string" && typeof r.promoted_model === "string" && typeof r.incumbent_model === "string";
+function detectModelStaleEvidence(ir, profile) {
+  if (!isBrainQueryActiveFor("kgauto_archetype_perf")) return [];
+  const archetype = ir.intent.archetype;
+  const chosen = getArchetypePerfScore(profile.id, archetype);
+  if (chosen.grounding !== "judgment") return [];
+  return [
+    {
+      level: "info",
+      code: "model-stale-evidence",
+      message: `Model-stale-evidence: target=${profile.id} archetype=${archetype} is judgment-grounded (n=${chosen.n}, cross-app 90d window) despite brain-query mode being active. Fewer than 10 outcomes back this (model, archetype) tuple across ALL consumers \u2014 routing decisions remain pre-measured for this slot.`,
+      suggestion: "Verify that `record()` is being called on every call() outcome. Counts are cross-app (migration 050 view): once ANY consumers accumulate n>=10 rows on this tuple, the score promotes from judgment to measured automatically within the 5-min SWR window. (Before alpha.78 this promotion was advertised but had no implementing mechanism \u2014 n was never populated; if this advisory has been firing for weeks at n=0 despite real traffic, bump to >=alpha.78 and it will clear on its own.)",
+      recommendationType: "prompt-fix",
+      docsUrl: "https://github.com/stue/command-center/blob/main/interfaces/kgauto.md#best-practice-advisories"
+    }
+  ];
 }
-function coerceId(v) {
-  if (typeof v === "number") return Number.isFinite(v) ? v : null;
-  if (typeof v === "string") {
-    const n = Number(v);
-    return Number.isFinite(n) ? n : null;
+function detectTierDown(ir, profile, phase2) {
+  if (!phase2 || phase2.fallbackChain.length === 0) return [];
+  if (!phase2.profileResolver) return [];
+  const archetype = ir.intent.archetype;
+  const chosenScore = getArchetypePerfScore(profile.id, archetype);
+  const chosenCost = profile.costInputPer1m;
+  let bestAlt = null;
+  for (const altId of phase2.fallbackChain) {
+    const altProfile = phase2.profileResolver(altId);
+    if (!altProfile) continue;
+    if (altProfile.id === profile.id) continue;
+    const altScore = getArchetypePerfScore(altProfile.id, archetype);
+    if (altScore.grounding !== "measured") continue;
+    if (altScore.score < QUALITY_FLOOR_FOR_RECOMMENDATION) continue;
+    if (altScore.score < chosenScore.score) continue;
+    if (altProfile.costInputPer1m > chosenCost * TIER_DOWN_COST_RATIO) continue;
+    if (!bestAlt || altProfile.costInputPer1m < bestAlt.profile.costInputPer1m || altProfile.costInputPer1m === bestAlt.profile.costInputPer1m && altScore.score > bestAlt.score.score) {
+      bestAlt = { id: altId, profile: altProfile, score: altScore };
+    }
   }
-  return null;
+  if (!bestAlt) return [];
+  const chosenDesc = chosenScore.grounding === "measured" ? `archetypePerf.${archetype}=${chosenScore.score} (measured, n=${chosenScore.n})` : `archetypePerf.${archetype}=${chosenScore.score} (${chosenScore.grounding})`;
+  return [
+    {
+      level: "warn",
+      code: "tier-down",
+      message: `Tier-down: target=${profile.id} (${chosenDesc}) selected for ${archetype}. Brain shows ${bestAlt.id} delivers equal-or-better quality (archetypePerf.${archetype}=${bestAlt.score.score}, measured, n=${bestAlt.score.n}) at $${bestAlt.profile.costInputPer1m}/$${bestAlt.profile.costOutputPer1m} per 1M vs $${profile.costInputPer1m}/$${profile.costOutputPer1m} \u2014 a measured tier-down opportunity.`,
+      suggestion: `Move \`${bestAlt.id}\` to primary for this archetype. The brain has n=${bestAlt.score.n} measured outcomes backing the recommendation; this is data, not opinion. If posture='locked' is required (compliance/brand promise), set it explicitly to silence this rule.`,
+      recommendationType: "tier-down",
+      docsUrl: "https://github.com/stue/command-center/blob/main/interfaces/kgauto.md#best-practice-advisories"
+    }
+  ];
 }
-function mapRowsToPromotions(rows) {
-  const out = [];
-  for (const row of rows) {
-    if (!isRawPromotionRow(row)) continue;
-    const id = coerceId(row.id);
-    if (id === null) continue;
-    const mode = row.mode === "strategy" ? "strategy" : row.mode === "downswap" || row.mode === void 0 ? "downswap" : null;
-    if (mode === null) continue;
-    out.push({
-      id,
-      archetype: row.intent_archetype,
-      mode,
-      strategy: typeof row.strategy === "string" ? row.strategy : null,
-      promotedModel: row.promoted_model,
-      incumbentModel: row.incumbent_model,
-      evalRunId: coerceId(row.eval_run_id ?? null),
-      suppressQualityGate: row.suppress_quality_gate === true,
-      promotedAt: typeof row.promoted_at === "string" ? row.promoted_at : ""
-    });
+function detectArchetypePerfFloorBreach(ir, profile) {
+  const compat = getModelCompatibility(profile.id, {
+    archetype: ir.intent.archetype,
+    toolOrchestration: ir.constraints?.toolOrchestration
+  });
+  if (compat.status === "compatible") return [];
+  if (compat.status === "requires-adapter") {
+    return [
+      {
+        level: "warn",
+        code: "archetype-perf-floor-breach",
+        message: `${profile.id} sits below the archetype floor for ${ir.intent.archetype} (score ${compat.archetypePerf}/10, floor ${6}). A known adapter would lift it: ${compat.adapter.parameter}=${compat.adapter.value}. ${compat.adapter.consequence}`,
+        suggestion: `Pass \`ir.constraints.${compat.adapter.parameter} = '${compat.adapter.value}'\` for this call, OR pick a model whose archetypePerf for ${ir.intent.archetype} already clears the floor (call \`getModelCompatibility(modelId, { archetype: '${ir.intent.archetype}' })\` to check). Estimated post-adapter score: ${compat.archetypePerfWithAdapter}/10.`,
+        recommendationType: "prompt-fix",
+        suggestedAdaptation: compat.adapter,
+        docsUrl: "https://github.com/stue/command-center/blob/main/interfaces/kgauto.md#best-practice-advisories"
+      }
+    ];
   }
-  return out;
-}
-var snapshots4 = /* @__PURE__ */ new Map();
-var runtime4;
-var warnedOnce4 = false;
-var DEFAULT_PROMOTIONS_ENDPOINT = "https://kgauto-dashboard.vercel.app/api/kgauto-v2/promotions";
-function isAutoPromoteEnabledFromEnv(envSource) {
-  const env = envSource ?? (typeof process !== "undefined" && process.env ? process.env : {});
-  const raw = (env.KGAUTO_AUTO_PROMOTE ?? "").trim().toLowerCase();
-  return raw === "1" || raw === "true";
-}
-function configurePromotionsBrain(rt) {
-  runtime4 = rt;
-  snapshots4.clear();
-  warnedOnce4 = false;
-}
-function isPromotionsBrainActive() {
-  return runtime4 !== void 0;
+  return [
+    {
+      level: "critical",
+      code: "archetype-perf-floor-breach",
+      message: `${profile.id} sits below the archetype floor for ${ir.intent.archetype} (score ${compat.archetypePerf}/10, floor ${6}) and no known adapter would lift it. ${compat.reason}`,
+      suggestion: `Swap to a model whose archetypePerf for ${ir.intent.archetype} clears the floor. Use \`getModelCompatibility(candidateId, { archetype: '${ir.intent.archetype}' })\` to vet candidates, or \`getDefaultFallbackChain({ archetype: '${ir.intent.archetype}', posture: 'open' })\` for a library-picked chain that respects the floor by construction.`,
+      recommendationType: "model-swap",
+      docsUrl: "https://github.com/stue/command-center/blob/main/interfaces/kgauto.md#best-practice-advisories"
+    }
+  ];
 }
-function getApplicablePromotion(opts) {
-  const rt = runtime4;
-  if (!rt) return void 0;
-  const appId = opts.appId;
-  if (!appId || !opts.archetype || !opts.mode) return void 0;
-  let snap = snapshots4.get(appId);
-  if (!snap) {
-    snap = { data: [], expiresAt: 0, refreshing: false };
-    snapshots4.set(appId, snap);
-  }
-  const now = Date.now();
-  const stale = snap.expiresAt <= now;
-  if (stale && !snap.refreshing) {
-    snap.refreshing = true;
-    void asyncRefresh4(rt, appId);
-  }
-  return snap.data.find(
-    (p) => p.archetype === opts.archetype && p.mode === opts.mode
-  );
-}
-var pendingRefreshes4 = /* @__PURE__ */ new Map();
-async function asyncRefresh4(rt, appId) {
-  const promise = doRefresh4(rt, appId);
-  pendingRefreshes4.set(appId, promise);
-  try {
-    await promise;
-  } finally {
-    if (pendingRefreshes4.get(appId) === promise) {
-      pendingRefreshes4.delete(appId);
-    }
-  }
-}
-async function doRefresh4(rt, appId) {
-  const url = `${rt.endpoint}?app_id=${encodeURIComponent(appId)}`;
-  let snap = snapshots4.get(appId);
-  if (!snap) {
-    snap = { data: [], expiresAt: 0, refreshing: false };
-    snapshots4.set(appId, snap);
-  }
-  try {
-    const res = await rt.fetchImpl(url, { method: "GET" });
-    if (!res.ok) {
-      throw new Error(`promotions ${res.status}: ${res.statusText}`);
-    }
-    const body = await res.json();
-    if (runtime4 !== rt) return;
-    const rows = Array.isArray(body) ? mapRowsToPromotions(body) : [];
-    snap.data = rows;
-    snap.expiresAt = Date.now() + rt.ttlMs;
-    snap.refreshing = false;
-  } catch (err) {
-    if (runtime4 !== rt) return;
-    snap.refreshing = false;
-    snap.expiresAt = Date.now() + rt.ttlMs;
-    if (!warnedOnce4) {
-      warnedOnce4 = true;
-      (rt.onError ?? defaultOnError4)(err);
+function detectStaleExclusionCandidate(ir) {
+  if (!isExclusionFindingsBrainActive()) return [];
+  if (!ir.appId) return [];
+  const findings = getStaleExclusionFindings({
+    appId: ir.appId,
+    archetype: ir.intent.archetype
+  });
+  if (findings.length === 0) return [];
+  const ranked = [...findings].sort((a, b) => {
+    const sa = a.estimatedSavingsUsd30d ?? -Infinity;
+    const sb = b.estimatedSavingsUsd30d ?? -Infinity;
+    if (sa !== sb) return sb - sa;
+    return confidenceRank(b.confidence) - confidenceRank(a.confidence);
+  });
+  const top = ranked[0];
+  const extraCount = findings.length - 1;
+  const extraNote = extraCount > 0 ? ` (+ ${extraCount} more excluded model${extraCount === 1 ? "" : "s"} for this archetype)` : "";
+  return [
+    {
+      level: "info",
+      code: "stale-exclusion-candidate",
+      message: `${top.message}${extraNote}`,
+      suggestion: top.suggestion,
+      recommendationType: "tier-down",
+      docsUrl: "https://github.com/stue/command-center/blob/main/interfaces/kgauto.md#best-practice-advisories"
     }
-  }
-}
-function defaultOnError4(err) {
-  console.warn(
-    "[kgauto] promotions fetch failed (promotion boost inactive until next refresh):",
-    err
-  );
-}
-function _testResetPromotions() {
-  runtime4 = void 0;
-  snapshots4.clear();
-  pendingRefreshes4 = /* @__PURE__ */ new Map();
-  warnedOnce4 = false;
+  ];
 }
-async function _testWaitForPromotionsRefresh() {
-  const pending = Array.from(pendingRefreshes4.values());
-  if (pending.length > 0) await Promise.all(pending);
+function confidenceRank(c) {
+  if (c === "high") return 3;
+  if (c === "medium") return 2;
+  return 1;
 }
 
-// src/measured-failure-brain.ts
-function coerceCount(v) {
-  if (typeof v === "number") return Number.isFinite(v) ? v : null;
-  if (typeof v === "string") {
-    const n = Number(v);
-    return Number.isFinite(n) ? n : null;
-  }
-  return null;
-}
-function isRawFailureRow(x) {
-  if (!x || typeof x !== "object") return false;
-  const r = x;
-  return typeof r.intent_archetype === "string" && typeof r.model === "string" && (typeof r.n === "number" || typeof r.n === "string");
-}
-function mapRows(rows) {
-  const out = [];
-  for (const row of rows) {
-    if (!isRawFailureRow(row)) continue;
-    const n = coerceCount(row.n);
-    const nFail = coerceCount(row.n_fail) ?? 0;
-    if (n === null || n <= 0) continue;
-    out.push({
-      archetype: row.intent_archetype,
-      model: row.model,
-      n,
-      nFail
-    });
-  }
-  return out;
-}
-var MEASURED_FAILURE_CFG = {
-  /**
-   * Hard minimum attempts before ANY gate may be created. Guards against
-   * pathological tiny samples that the confidence bound alone would let
-   * through in edge cases. At 5-for-5 the bound clears the threshold; at
-   * 3-for-3 it does not, which is the behaviour we want (three failures is
-   * a bad day, five in a row is a pattern).
-   */
-  minSample: 5,
-  /**
-   * Gate when we are 95% confident the model fails MORE OFTEN THAN IT
-   * SUCCEEDS on this surface. Deliberately unarguable rather than tuned —
-   * a model that probably fails the majority of the time has no business
-   * leading a surface, whatever its declared scores say.
-   */
-  lowerBoundThreshold: 0.5,
-  /** 95% one-sided-ish confidence (standard two-sided z at α=0.05). */
-  z: 1.96,
-  /** Must match the view's window. Documented here for the advisory text. */
-  windowDays: 28
-};
-function wilsonLowerBound(failures, n, z = MEASURED_FAILURE_CFG.z) {
-  if (n <= 0) return 0;
-  const p = failures / n;
-  const z2 = z * z;
-  const denom = 1 + z2 / n;
-  const centre = p + z2 / (2 * n);
-  const margin = z * Math.sqrt(p * (1 - p) / n + z2 / (4 * n * n));
-  const lower2 = (centre - margin) / denom;
-  return lower2 < 0 ? 0 : lower2;
-}
-function judgeMeasuredFailure(row, cfg = MEASURED_FAILURE_CFG) {
-  if (!row || row.n < cfg.minSample) return void 0;
-  const lowerBound = wilsonLowerBound(row.nFail, row.n, cfg.z);
-  return {
-    gated: lowerBound > cfg.lowerBoundThreshold,
-    rate: row.nFail / row.n,
-    lowerBound,
-    n: row.n,
-    nFail: row.nFail
-  };
-}
-var snapshots5 = /* @__PURE__ */ new Map();
-var runtime5;
-var warnedOnce5 = false;
-var DEFAULT_MEASURED_FAILURE_ENDPOINT = "https://kgauto-dashboard.vercel.app/api/kgauto-v2/measured-failure";
-function isMeasuredFailureGateEnabledFromEnv(envSource) {
-  const env = envSource ?? (typeof process !== "undefined" && process.env ? process.env : {});
-  const raw = (env.KGAUTO_MEASURED_FAILURE_GATE ?? "").trim().toLowerCase();
-  return !(raw === "0" || raw === "false");
-}
-function configureMeasuredFailureBrain(rt) {
-  runtime5 = rt;
-  snapshots5.clear();
-  warnedOnce5 = false;
-}
-function isMeasuredFailureBrainActive() {
-  return runtime5 !== void 0;
-}
-function prefetchMeasuredFailure(appId) {
-  const rt = runtime5;
-  if (!rt || !appId) return void 0;
-  let snap = snapshots5.get(appId);
-  if (!snap) {
-    snap = { data: [], expiresAt: 0, refreshing: false };
-    snapshots5.set(appId, snap);
-  }
-  if (snap.expiresAt > Date.now()) return void 0;
-  const inflight = pendingRefreshes5.get(appId);
-  if (inflight) return inflight;
-  if (snap.refreshing) return void 0;
-  snap.refreshing = true;
-  void asyncRefresh5(rt, appId);
-  return pendingRefreshes5.get(appId);
-}
-async function awaitMeasuredFailureReady(appId, timeoutMs) {
-  if (!runtime5 || !appId) return;
-  const pending = prefetchMeasuredFailure(appId) ?? pendingRefreshes5.get(appId);
-  if (!(timeoutMs > 0)) return;
-  if (!pending) return;
-  let timer;
-  try {
-    await Promise.race([
-      pending,
-      new Promise((resolve) => {
-        timer = setTimeout(resolve, timeoutMs);
-      })
-    ]);
-  } catch {
-  } finally {
-    if (timer) clearTimeout(timer);
-  }
-}
-function getMeasuredFailureVerdict(opts) {
-  const rt = runtime5;
-  if (!rt) return void 0;
-  const { appId, archetype, model } = opts;
-  if (!appId || !archetype || !model) return void 0;
-  let snap = snapshots5.get(appId);
-  if (!snap) {
-    snap = { data: [], expiresAt: 0, refreshing: false };
-    snapshots5.set(appId, snap);
-  }
-  const now = Date.now();
-  if (snap.expiresAt <= now && !snap.refreshing) {
-    snap.refreshing = true;
-    void asyncRefresh5(rt, appId);
-  }
-  const row = snap.data.find(
-    (r) => r.archetype === archetype && r.model === model
-  );
-  return judgeMeasuredFailure(row);
-}
-var pendingRefreshes5 = /* @__PURE__ */ new Map();
-async function asyncRefresh5(rt, appId) {
-  const promise = doRefresh5(rt, appId);
-  pendingRefreshes5.set(appId, promise);
-  try {
-    await promise;
-  } finally {
-    if (pendingRefreshes5.get(appId) === promise) {
-      pendingRefreshes5.delete(appId);
+// src/translator.ts
+var TRANSLATOR_FLOOR = ARCHETYPE_FLOOR_DEFAULT;
+var RULE_SEQUENTIAL_TOOL_CLIFF = "sequential-tool-cliff-below-floor";
+var RULE_NARRATION_DRIFT_ANTHROPIC = "narration-drift-anthropic";
+var RULE_NARRATION_THINKING_LEAK_DEEPSEEK = "narration-thinking-leak-deepseek";
+var SEQUENTIAL_TOOL_PREAMBLE = "IMPORTANT: Use one tool call per response. Wait for the tool result before deciding the next tool. Do NOT batch tool calls in parallel.";
+var NARRATION_DRIFT_ANTHROPIC_PREAMBLE = "Output ONLY the requested content. Do not narrate your thought process. Each line \u2264 12 words.";
+var NARRATION_THINKING_LEAK_DEEPSEEK_PREAMBLE = "Reasoning is internal. Output ONLY the requested content; do not emit <thinking> blocks or internal monologue as user-facing text.";
+var RULE_DISCIPLINE_GATES_V1 = "discipline-gates-v1";
+var DISCIPLINE_GATES_V1_WITH_TOOLS = `Work through these gates at every judgment point, explicitly:
+1. Evidence before reasoning: cite what you observed before concluding from it.
+2. One extra signal: when a finding feels conclusive, check one more adjacent signal before stating it.
+3. Expand, don't guess: resolve a compressed or referenced item by looking it up rather than inferring its contents.
+4. Innocent explanation first: state the most plausible benign reading before alleging the alarming one.
+5. Label each claim: mark it observed, inferred, or assumed.
+6. A surfaced gap beats a guessed answer: flag what you cannot determine rather than fabricating past it.`;
+var DISCIPLINE_GATES_V1_NO_TOOLS = `Work through these gates at every judgment point, explicitly:
+1. Evidence before reasoning: cite what you observed before concluding from it.
+2. One extra signal: when a finding feels conclusive, check one more adjacent signal before stating it.
+3. Innocent explanation first: state the most plausible benign reading before alleging the alarming one.
+4. Label each claim: mark it observed, inferred, or assumed.
+5. A surfaced gap beats a guessed answer: flag what you cannot determine rather than fabricating past it.`;
+var RULE_DISCIPLINE_GATES_V1_STRUCTURED = "discipline-gates-v1-structured";
+var DISCIPLINE_GATES_V1_STRUCTURED_WITH_TOOLS = `Work through these gates at every judgment point, explicitly:
+1. Evidence before reasoning: cite what you observed before concluding from it.
+2. One extra signal: when a finding feels conclusive, check one more adjacent signal before stating it.
+3. Expand, don't guess: resolve a compressed or referenced item by looking it up rather than inferring its contents.
+4. Innocent explanation first: state the most plausible benign reading before alleging the alarming one.`;
+var DISCIPLINE_GATES_V1_STRUCTURED_NO_TOOLS = `Work through these gates at every judgment point, explicitly:
+1. Evidence before reasoning: cite what you observed before concluding from it.
+2. One extra signal: when a finding feels conclusive, check one more adjacent signal before stating it.
+3. Innocent explanation first: state the most plausible benign reading before alleging the alarming one.`;
+var DISCIPLINE_ELIGIBLE_ARCHETYPES = /* @__PURE__ */ new Set([
+  "hunt",
+  "summarize",
+  "plan",
+  "critique",
+  "judge"
+]);
+function matchRule(kind, profile, archetype, ctx) {
+  if (kind === "discipline_contract") {
+    if (!DISCIPLINE_ELIGIBLE_ARCHETYPES.has(archetype)) return null;
+    if (ctx.outputMode !== "text") {
+      return {
+        id: RULE_DISCIPLINE_GATES_V1_STRUCTURED,
+        preamble: ctx.hasTools ? DISCIPLINE_GATES_V1_STRUCTURED_WITH_TOOLS : DISCIPLINE_GATES_V1_STRUCTURED_NO_TOOLS
+      };
     }
+    return {
+      id: RULE_DISCIPLINE_GATES_V1,
+      preamble: ctx.hasTools ? DISCIPLINE_GATES_V1_WITH_TOOLS : DISCIPLINE_GATES_V1_NO_TOOLS
+    };
   }
-}
-async function doRefresh5(rt, appId) {
-  const url = `${rt.endpoint}?app_id=${encodeURIComponent(appId)}`;
-  let snap = snapshots5.get(appId);
-  if (!snap) {
-    snap = { data: [], expiresAt: 0, refreshing: false };
-    snapshots5.set(appId, snap);
+  if (kind === "tool_call_contract") {
+    if (!profile.archetypePerf) return null;
+    const archetypeScore = profile.archetypePerf[archetype];
+    if (typeof archetypeScore !== "number" || archetypeScore >= TRANSLATOR_FLOOR) {
+      return null;
+    }
+    return {
+      id: RULE_SEQUENTIAL_TOOL_CLIFF,
+      preamble: SEQUENTIAL_TOOL_PREAMBLE,
+      wireOverrides: { parallelToolCalls: false }
+    };
   }
-  try {
-    const res = await rt.fetchImpl(url, { method: "GET" });
-    if (!res.ok) {
-      throw new Error(`measured-failure ${res.status}: ${res.statusText}`);
+  if (kind === "narration_contract") {
+    if (profile.provider === "anthropic") {
+      return {
+        id: RULE_NARRATION_DRIFT_ANTHROPIC,
+        preamble: NARRATION_DRIFT_ANTHROPIC_PREAMBLE
+      };
     }
-    const body = await res.json();
-    if (runtime5 !== rt) return;
-    snap.data = Array.isArray(body) ? mapRows(body) : [];
-    snap.expiresAt = Date.now() + rt.ttlMs;
-    snap.refreshing = false;
-  } catch (err) {
-    if (runtime5 !== rt) return;
-    snap.refreshing = false;
-    snap.expiresAt = Date.now() + rt.ttlMs;
-    if (!warnedOnce5) {
-      warnedOnce5 = true;
-      (rt.onError ?? defaultOnError5)(err);
+    if (profile.provider === "deepseek") {
+      return {
+        id: RULE_NARRATION_THINKING_LEAK_DEEPSEEK,
+        preamble: NARRATION_THINKING_LEAK_DEEPSEEK_PREAMBLE
+      };
     }
+    return null;
   }
+  return null;
 }
-function defaultOnError5(err) {
-  console.warn(
-    "[kgauto] measured-failure fetch failed (gate inactive until next refresh):",
-    err
-  );
-}
-function _testResetMeasuredFailure() {
-  runtime5 = void 0;
-  snapshots5.clear();
-  pendingRefreshes5 = /* @__PURE__ */ new Map();
-  warnedOnce5 = false;
-}
-async function _testWaitForMeasuredFailureRefresh() {
-  const pending = Array.from(pendingRefreshes5.values());
-  if (pending.length > 0) await Promise.all(pending);
+function applySectionRewrites(args) {
+  const { ir, profile, archetype } = args;
+  if (!Array.isArray(ir.sections) || ir.sections.length === 0) {
+    return { rewrittenIR: ir, rewrites: [] };
+  }
+  const outputMode = args.outputMode ?? resolveOutputMode({
+    declared: ir.constraints?.outputMode,
+    structuredOutput: ir.constraints?.structuredOutput,
+    toolCount: ir.tools?.length ?? 0
+  });
+  const hasTools = (ir.tools?.length ?? 0) > 0;
+  const ctx = { outputMode, hasTools };
+  const rewrites = [];
+  const newSections = ir.sections.map((section) => {
+    if (!section.kind || section.kind === "arbitrary") return section;
+    const rule = matchRule(section.kind, profile, archetype, ctx);
+    if (!rule) return section;
+    const originalText = section.text;
+    const transformedText = `${rule.preamble}
+
+${originalText}`;
+    rewrites.push({
+      sectionId: section.id,
+      kind: section.kind,
+      rule: rule.id,
+      originalText,
+      transformedText,
+      ...rule.wireOverrides ? { wireOverrides: rule.wireOverrides } : {}
+    });
+    return { ...section, text: transformedText };
+  });
+  if (rewrites.length === 0) {
+    return { rewrittenIR: ir, rewrites: [] };
+  }
+  const rewrittenIR = { ...ir, sections: newSections };
+  return { rewrittenIR, rewrites };
 }
 
 // src/compile.ts
@@ -3895,6 +3962,17 @@ function parseJsonLoose(raw) {
 var FAILED = /* @__PURE__ */ Symbol("parse-failed");
 
 // src/ir.ts
+function mutationId(m) {
+  return typeof m === "string" ? m : m.id;
+}
+function hasMutation(list, idOrPrefix) {
+  if (!Array.isArray(list)) return false;
+  const prefix = idOrPrefix.endsWith("*") ? idOrPrefix.slice(0, -1) : void 0;
+  return list.some((m) => {
+    const id = mutationId(m);
+    return prefix !== void 0 ? id.startsWith(prefix) : id === idOrPrefix;
+  });
+}
 var CallError = class extends Error {
   attempts;
   lastErrorCode;
@@ -3948,12 +4026,46 @@ async function captureGoldenIr(ctx) {
 }
 
 // src/streaming.ts
+function classifyThrownFetchError(err) {
+  const name = err?.name;
+  if (name === "TimeoutError") {
+    return { errorType: "retryable", errorCode: "timeout", message: String(err) };
+  }
+  if (name === "AbortError") {
+    return { errorType: "terminal", errorCode: "aborted", message: "aborted by caller signal" };
+  }
+  return { errorType: "retryable", errorCode: "network_error", message: String(err) };
+}
+function createStallGuard(external, stallTimeoutMs) {
+  const ctl = new AbortController();
+  if (external) {
+    if (external.aborted) ctl.abort(external.reason);
+    else external.addEventListener("abort", () => ctl.abort(external.reason), { once: true });
+  }
+  let timer;
+  const arm = () => {
+    if (!stallTimeoutMs || stallTimeoutMs <= 0) return;
+    if (timer) clearTimeout(timer);
+    timer = setTimeout(() => {
+      ctl.abort(
+        new DOMException(`stream stalled: no bytes for ${stallTimeoutMs}ms`, "TimeoutError")
+      );
+    }, stallTimeoutMs);
+  };
+  const clear = () => {
+    if (timer) clearTimeout(timer);
+    timer = void 0;
+  };
+  return { signal: ctl.signal, arm, clear };
+}
 var ANTHROPIC_URL = "https://api.anthropic.com/v1/messages";
 async function streamAnthropic(request, apiKey, opts) {
   const { provider: _provider, ...body } = request;
   const fetchFn = opts.fetchImpl ?? fetch;
+  const guard = createStallGuard(opts.signal, opts.stallTimeoutMs);
   let res;
   try {
+    guard.arm();
     res = await fetchFn(ANTHROPIC_URL, {
       method: "POST",
       headers: {
@@ -3961,12 +4073,16 @@ async function streamAnthropic(request, apiKey, opts) {
         "anthropic-version": "2023-06-01",
         "content-type": "application/json"
       },
-      body: JSON.stringify({ ...body, stream: true })
+      body: JSON.stringify({ ...body, stream: true }),
+      signal: guard.signal
     });
   } catch (err) {
-    return retryableError(0, "network_error", String(err), null);
+    guard.clear();
+    const c = classifyThrownFetchError(err);
+    return { ok: false, status: 0, errorType: c.errorType, errorCode: c.errorCode, message: c.message, raw: null };
   }
   if (!res.ok) {
+    guard.clear();
     const errBody = await res.json().catch(() => ({}));
     return classifyHttpError(res.status, errBody);
   }
@@ -4029,9 +4145,13 @@ async function streamAnthropic(request, apiKey, opts) {
         if (typeof p.usage?.output_tokens === "number") outputTokens = p.usage.output_tokens;
         return;
       }
-    });
+    }, guard.arm, guard.signal);
   } catch (err) {
-    return retryableError(0, "stream_interrupted", String(err), null);
+    const c = classifyThrownFetchError(err);
+    const code = c.errorCode === "network_error" ? "stream_interrupted" : c.errorCode;
+    return { ok: false, status: 0, errorType: c.errorType, errorCode: code, message: c.message, raw: null };
+  } finally {
+    guard.clear();
   }
   const toolCalls = Array.from(toolBlocks.values()).map((b) => ({
     id: b.id,
@@ -4058,6 +4178,7 @@ async function streamAnthropic(request, apiKey, opts) {
 async function streamOpenAILike(url, request, apiKey, providerLabel, opts) {
   const { provider: _provider, ...body } = request;
   const fetchFn = opts.fetchImpl ?? fetch;
+  const guard = createStallGuard(opts.signal, opts.stallTimeoutMs);
   const reqBody = {
     ...body,
     stream: true,
@@ -4068,18 +4189,23 @@ async function streamOpenAILike(url, request, apiKey, providerLabel, opts) {
   };
   let res;
   try {
+    guard.arm();
     res = await fetchFn(url, {
       method: "POST",
       headers: {
         authorization: `Bearer ${apiKey}`,
         "content-type": "application/json"
       },
-      body: JSON.stringify(reqBody)
+      body: JSON.stringify(reqBody),
+      signal: guard.signal
     });
   } catch (err) {
-    return retryableError(0, "network_error", String(err), null);
+    guard.clear();
+    const c = classifyThrownFetchError(err);
+    return { ok: false, status: 0, errorType: c.errorType, errorCode: c.errorCode, message: c.message, raw: null };
   }
   if (!res.ok) {
+    guard.clear();
     const errBody = await res.json().catch(() => ({}));
     return classifyHttpError(res.status, errBody);
   }
@@ -4132,9 +4258,13 @@ async function streamOpenAILike(url, request, apiKey, providerLabel, opts) {
         const details = usage.prompt_tokens_details;
         if (typeof details?.cached_tokens === "number") cachedTokens = details.cached_tokens;
       }
-    });
+    }, guard.arm, guard.signal);
   } catch (err) {
-    return retryableError(0, "stream_interrupted", String(err), null);
+    const c = classifyThrownFetchError(err);
+    const code = c.errorCode === "network_error" ? "stream_interrupted" : c.errorCode;
+    return { ok: false, status: 0, errorType: c.errorType, errorCode: code, message: c.message, raw: null };
+  } finally {
+    guard.clear();
   }
   const toolCalls = Array.from(toolBuffers.values()).filter((b) => b.name.length > 0).map((b) => ({
     id: b.id,
@@ -4157,15 +4287,28 @@ async function streamOpenAILike(url, request, apiKey, providerLabel, opts) {
   };
   return { ok: true, status: res.status, response };
 }
-async function parseSSEStream(response, handler) {
+async function parseSSEStream(response, handler, onRead, signal) {
   const body = response.body;
   if (!body) throw new Error("Response has no body for SSE parse");
   const reader = body.getReader();
   const decoder = new TextDecoder("utf-8");
   let buffer = "";
+  const abortRace = signal ? new Promise((_resolve, reject) => {
+    if (signal.aborted) return reject(signal.reason);
+    signal.addEventListener("abort", () => reject(signal.reason), { once: true });
+  }) : void 0;
   for (; ; ) {
-    const { value, done } = await reader.read();
+    let readResult;
+    try {
+      readResult = abortRace ? await Promise.race([reader.read(), abortRace]) : await reader.read();
+    } catch (err) {
+      void reader.cancel().catch(() => {
+      });
+      throw err;
+    }
+    const { value, done } = readResult;
     if (done) break;
+    onRead?.();
     buffer += decoder.decode(value, { stream: true });
     let sep;
     while (sep = buffer.indexOf("\n\n"), sep !== -1) {
@@ -4276,9 +4419,6 @@ function extractErrorMessage(body) {
   if (typeof b.message === "string") return b.message;
   return void 0;
 }
-function retryableError(status, code, message, raw) {
-  return { ok: false, status, errorType: "retryable", errorCode: code, message, raw };
-}
 
 // src/execute.ts
 var ANTHROPIC_URL2 = "https://api.anthropic.com/v1/messages";
@@ -4323,7 +4463,9 @@ async function executeAnthropic(request, opts) {
   if (opts.onChunk) {
     return streamAnthropic(request, apiKey, {
       onChunk: opts.onChunk,
-      fetchImpl: opts.fetchImpl
+      fetchImpl: opts.fetchImpl,
+      signal: opts.signal,
+      stallTimeoutMs: opts.stallTimeoutMs
     });
   }
   const { provider: _provider, ...body } = request;
@@ -4338,11 +4480,13 @@ async function executeAnthropic(request, opts) {
         "anthropic-version": "2023-06-01",
         "content-type": "application/json"
       },
-      body: JSON.stringify(body)
+      body: JSON.stringify(body),
+      signal: opts.signal
     });
     json = await res.json().catch(() => ({}));
   } catch (err) {
-    return retryableError2(0, "network_error", String(err), null);
+    const c = classifyThrownFetchError(err);
+    return { ok: false, status: 0, errorType: c.errorType, errorCode: c.errorCode, message: c.message, raw: null };
   }
   if (!res.ok) return classifyHttpError2(res.status, json);
   return { ok: true, status: res.status, response: normalizeAnthropic(json) };
@@ -4374,11 +4518,13 @@ async function executeGoogle(request, opts) {
     res = await fetchFn(url, {
       method: "POST",
       headers: { "content-type": "application/json" },
-      body: JSON.stringify(body)
+      body: JSON.stringify(body),
+      signal: opts.signal
     });
     json = await res.json().catch(() => ({}));
   } catch (err) {
-    return retryableError2(0, "network_error", String(err), null);
+    const c = classifyThrownFetchError(err);
+    return { ok: false, status: 0, errorType: c.errorType, errorCode: c.errorCode, message: c.message, raw: null };
   }
   if (!res.ok) return classifyHttpError2(res.status, json);
   return { ok: true, status: res.status, response: normalizeGoogle(json) };
@@ -4410,7 +4556,9 @@ async function executeOpenAI(request, opts) {
   if (opts.onChunk) {
     return streamOpenAILike(OPENAI_URL, request, apiKey, "openai", {
       onChunk: opts.onChunk,
-      fetchImpl: opts.fetchImpl
+      fetchImpl: opts.fetchImpl,
+      signal: opts.signal,
+      stallTimeoutMs: opts.stallTimeoutMs
     });
   }
   const { provider: _provider, ...body } = request;
@@ -4421,11 +4569,13 @@ async function executeOpenAI(request, opts) {
     res = await fetchFn(OPENAI_URL, {
       method: "POST",
       headers: { authorization: `Bearer ${apiKey}`, "content-type": "application/json" },
-      body: JSON.stringify(body)
+      body: JSON.stringify(body),
+      signal: opts.signal
     });
     json = await res.json().catch(() => ({}));
   } catch (err) {
-    return retryableError2(0, "network_error", String(err), null);
+    const c = classifyThrownFetchError(err);
+    return { ok: false, status: 0, errorType: c.errorType, errorCode: c.errorCode, message: c.message, raw: null };
   }
   if (!res.ok) return classifyHttpError2(res.status, json);
   return { ok: true, status: res.status, response: normalizeOpenAILike(json) };
@@ -4438,7 +4588,9 @@ async function executeDeepSeek(request, opts) {
   if (opts.onChunk) {
     return streamOpenAILike(DEEPSEEK_URL, request, apiKey, "deepseek", {
       onChunk: opts.onChunk,
-      fetchImpl: opts.fetchImpl
+      fetchImpl: opts.fetchImpl,
+      signal: opts.signal,
+      stallTimeoutMs: opts.stallTimeoutMs
     });
   }
   const { provider: _provider, ...body } = request;
@@ -4449,11 +4601,13 @@ async function executeDeepSeek(request, opts) {
     res = await fetchFn(DEEPSEEK_URL, {
       method: "POST",
       headers: { authorization: `Bearer ${apiKey}`, "content-type": "application/json" },
-      body: JSON.stringify(body)
+      body: JSON.stringify(body),
+      signal: opts.signal
     });
     json = await res.json().catch(() => ({}));
   } catch (err) {
-    return retryableError2(0, "network_error", String(err), null);
+    const c = classifyThrownFetchError(err);
+    return { ok: false, status: 0, errorType: c.errorType, errorCode: c.errorCode, message: c.message, raw: null };
   }
   if (!res.ok) return classifyHttpError2(res.status, json);
   return { ok: true, status: res.status, response: normalizeOpenAILike(json) };
@@ -4466,7 +4620,9 @@ async function executeOpenAICompatible(request, opts, spec) {
   if (opts.onChunk) {
     return streamOpenAILike(spec.url, request, apiKey, spec.provider, {
       onChunk: opts.onChunk,
-      fetchImpl: opts.fetchImpl
+      fetchImpl: opts.fetchImpl,
+      signal: opts.signal,
+      stallTimeoutMs: opts.stallTimeoutMs
     });
   }
   const { provider: _provider, ...body } = request;
@@ -4477,11 +4633,13 @@ async function executeOpenAICompatible(request, opts, spec) {
     res = await fetchFn(spec.url, {
       method: "POST",
       headers: { authorization: `Bearer ${apiKey}`, "content-type": "application/json" },
-      body: JSON.stringify(body)
+      body: JSON.stringify(body),
+      signal: opts.signal
     });
     json = await res.json().catch(() => ({}));
   } catch (err) {
-    return retryableError2(0, "network_error", String(err), null);
+    const c = classifyThrownFetchError(err);
+    return { ok: false, status: 0, errorType: c.errorType, errorCode: c.errorCode, message: c.message, raw: null };
   }
   if (!res.ok) return classifyHttpError2(res.status, json);
   return { ok: true, status: res.status, response: normalizeOpenAILike(json) };
@@ -4551,9 +4709,6 @@ function extractErrorMessage2(body) {
 function terminalError(status, code, message) {
   return { ok: false, status, errorType: "terminal", errorCode: code, message, raw: null };
 }
-function retryableError2(status, code, message, raw) {
-  return { ok: false, status, errorType: "retryable", errorCode: code, message, raw };
-}
 function tryParseJson2(s) {
   if (typeof s !== "string" || s.length === 0) return void 0;
   try {
@@ -4708,6 +4863,15 @@ async function call(ir, opts = {}) {
   let retriedSameModel = false;
   for (let i = 0; i < targetsToTry.length; i++) {
     const targetModel = targetsToTry[i];
+    if (opts.abortSignal?.aborted) {
+      attempts.push({
+        model: targetModel,
+        status: "terminal",
+        errorCode: "aborted",
+        message: "Skipped \u2014 caller abortSignal fired before this attempt started"
+      });
+      break;
+    }
     const targetProfile = tryGetProfile(targetModel);
     const providerFailReason = targetProfile ? failedProviders.get(targetProfile.provider) : void 0;
     if (targetProfile && providerFailReason && !opts.noFallback) {
@@ -4744,16 +4908,29 @@ async function call(ir, opts = {}) {
     );
     const targetSupportsStreaming = targetProfile?.streaming === true;
     const streamingOnChunk = opts.onChunk && !opts.noStream && targetSupportsStreaming ? opts.onChunk : void 0;
-    const execOpts = {
+    const mkExecOpts = () => ({
       apiKeys: opts.apiKeys,
       fetchImpl: opts.fetchImpl,
       providerOverrides: opts.providerOverrides,
-      onChunk: streamingOnChunk
-    };
-    const exec = await execute(activeCompile.request, execOpts);
+      onChunk: streamingOnChunk,
+      signal: composeAttemptSignal(
+        streamingOnChunk ? void 0 : opts.attemptTimeoutMs,
+        opts.abortSignal
+      ),
+      stallTimeoutMs: streamingOnChunk ? opts.attemptTimeoutMs : void 0
+    });
+    const exec = await execute(activeCompile.request, mkExecOpts());
     let validated = exec.ok ? validateStructuredContract(exec, ir) : exec;
     let servedByRetry = false;
-    if (!validated.ok && isStructuredContractViolation(validated.errorCode) && sameModelRetryEnabled && !retriedSameModel) {
+    let retrySuppressionNote;
+    if (!validated.ok && isStructuredContractViolation(validated.errorCode) && sameModelRetryEnabled && !retriedSameModel && getMeasuredFailureVerdict({
+      appId: ir.appId,
+      archetype: ir.intent.archetype,
+      model: targetModel
+    })?.gated === true) {
+      retrySuppressionNote = " [sameModelRetry suppressed: this model carries an active measured-failure gate for this archetype \u2014 retrying it would near-certainly bill a second doomed inference; walking the chain instead]";
+    }
+    if (!validated.ok && isStructuredContractViolation(validated.errorCode) && sameModelRetryEnabled && !retriedSameModel && !retrySuppressionNote) {
       retriedSameModel = true;
       attempts.push({
         model: targetModel,
@@ -4765,7 +4942,7 @@ async function call(ir, opts = {}) {
       safeEmit(
         () => emitExecuteAttempt(traceId, ir.appId, { model: targetModel, attemptIndex: i })
       );
-      const retryExec = await execute(retryRequest, execOpts);
+      const retryExec = await execute(retryRequest, mkExecOpts());
       validated = retryExec.ok ? validateStructuredContract(retryExec, ir) : retryExec;
       servedByRetry = true;
     }
@@ -4892,7 +5069,7 @@ async function call(ir, opts = {}) {
       model: targetModel,
       status: validated.errorType,
       errorCode: validated.errorCode,
-      message: validated.message,
+      message: retrySuppressionNote ? validated.message + retrySuppressionNote : validated.message,
       ...servedByRetry ? { sameModelRetry: true } : {}
     });
     lastErr = validated;
@@ -4931,6 +5108,25 @@ async function call(ir, opts = {}) {
     lastErr?.errorCode
   );
 }
+function composeAttemptSignal(attemptTimeoutMs, callerSignal) {
+  const signals = [];
+  if (callerSignal) signals.push(callerSignal);
+  if (typeof attemptTimeoutMs === "number" && attemptTimeoutMs > 0) {
+    signals.push(AbortSignal.timeout(attemptTimeoutMs));
+  }
+  if (signals.length === 0) return void 0;
+  if (signals.length === 1) return signals[0];
+  if (typeof AbortSignal.any === "function") return AbortSignal.any(signals);
+  const ctl = new AbortController();
+  for (const s of signals) {
+    if (s.aborted) {
+      ctl.abort(s.reason);
+      break;
+    }
+    s.addEventListener("abort", () => ctl.abort(s.reason), { once: true });
+  }
+  return ctl.signal;
+}
 function compileAndRegister(ir, opts) {
   const result = compile(ir, {
     policy: opts.policy,
@@ -5985,6 +6181,12 @@ async function runGoldenEval(opts) {
       );
     }
   }
+  if (wins + ties + losses === 0) {
+    notes.push(
+      "zero cases judged \u2014 run row persisted, evidence advisory deliberately NOT written (a verdict on n=0 is not evidence)"
+    );
+    return result;
+  }
   const latestRes = await fetchFn(
     rest(
       `compile_outcomes?app_id=eq.${encodeURIComponent(opts.appId)}&select=id&order=id.desc&limit=1`
@@ -6141,6 +6343,171 @@ function clamp(n) {
   return Math.max(0, Math.min(1, n));
 }
 
+// src/delegate.ts
+function isDelegateEnabledFromEnv(envSource) {
+  const env = envSource ?? (typeof process !== "undefined" && process.env ? process.env : {});
+  const raw = (env.KGAUTO_DELEGATE ?? "").trim().toLowerCase();
+  return raw === "1" || raw === "true";
+}
+var DELEGATE_TOOL_DEFINITION = {
+  name: "delegate",
+  description: "Delegate a self-contained sub-task to a cheaper executor model chosen by kgauto from measured evidence. YOU decide what to delegate and supply a complete input; kgauto decides which model runs it (never outside the declared pool). The call may be refused (budget exhausted, no qualified executor) \u2014 on refusal, continue and compose with what you have. Verify sub-results before composing them into your answer; do not present unverified delegated content as checked.",
+  inputSchema: {
+    type: "object",
+    properties: {
+      sub_archetype: {
+        type: "string",
+        enum: ALL_ARCHETYPES,
+        description: "What kind of work the sub-task is (its routing archetype)."
+      },
+      input: {
+        type: "string",
+        description: "Complete, self-contained sub-task input. Include everything the executor needs \u2014 it sees nothing else."
+      },
+      quality_floor: {
+        type: "number",
+        minimum: 0,
+        maximum: 10,
+        description: "Optional minimum executor quality score (0-10) for this archetype. Omit to accept the evidence-ranked default."
+      }
+    },
+    required: ["sub_archetype", "input"],
+    additionalProperties: false
+  }
+};
+function createDelegate(opts) {
+  const { parentIr, parentHandle } = opts;
+  const enabled = opts.enabled ?? isDelegateEnabledFromEnv();
+  const budget = opts.callOpts?.policy?.maxCostPerTraceUsd;
+  let spentUsd = 0;
+  async function handler(args) {
+    if (!enabled) {
+      return {
+        ok: false,
+        reason: "delegate_not_enabled",
+        detail: "Delegation is not enabled for this consumer (KGAUTO_DELEGATE / CreateDelegateOpts.enabled). Do the sub-task yourself."
+      };
+    }
+    const archetype = args.sub_archetype;
+    if (!ALL_ARCHETYPES.includes(archetype)) {
+      return {
+        ok: false,
+        reason: "invalid_sub_archetype",
+        detail: `Unknown sub_archetype '${String(args.sub_archetype)}'. Valid: ${ALL_ARCHETYPES.join(", ")}. Re-classify or do the sub-task yourself.`
+      };
+    }
+    const pool = parentIr.models;
+    const concreteIds = pool.filter((m) => typeof m === "string");
+    let blockedByFloor = [];
+    if (typeof args.quality_floor === "number") {
+      blockedByFloor = concreteIds.filter(
+        (m) => getArchetypePerfScore(m, archetype).score < args.quality_floor
+      );
+      if (blockedByFloor.length === pool.length) {
+        const scores = concreteIds.map((m) => `${m}=${getArchetypePerfScore(m, archetype).score}`).join(", ");
+        return {
+          ok: false,
+          reason: "no_qualified_executor",
+          detail: `No model in the declared pool clears quality_floor=${args.quality_floor} for '${archetype}' (${scores}). Lower the floor or do the sub-task yourself.`
+        };
+      }
+    }
+    const subIr = {
+      appId: parentIr.appId,
+      intent: { name: `delegate:${archetype}`, archetype },
+      sections: [{ id: "delegated-task", text: args.input }],
+      currentTurn: { role: "user", content: args.input },
+      models: pool
+    };
+    const mergedPolicy = {
+      ...opts.callOpts?.policy ?? {},
+      blockedModels: [
+        ...opts.callOpts?.policy?.blockedModels ?? [],
+        ...blockedByFloor
+      ]
+    };
+    if (typeof budget === "number" && budget > 0) {
+      let estimate = 0;
+      try {
+        estimate = compile(subIr, { policy: mergedPolicy }).estimatedCostUsd;
+      } catch {
+        estimate = 0;
+      }
+      if (spentUsd + estimate > budget) {
+        return {
+          ok: false,
+          reason: "trace_budget_exhausted",
+          detail: `Trace budget $${budget.toFixed(4)} would be exceeded (spent $${spentUsd.toFixed(4)} + estimated $${estimate.toFixed(4)}). Compose your answer from the sub-results you already have.`
+        };
+      }
+    }
+    let result;
+    try {
+      result = await call(subIr, {
+        ...opts.callOpts ?? {},
+        policy: mergedPolicy,
+        parentHandle
+        // R0 linkage — branch row, trace_id = parent
+      });
+    } catch (err) {
+      const detail = err instanceof CallError ? `Sub-call failed after ${err.attempts.length} attempt(s): ${err.attempts.map((a) => a.errorCode).join(" \u2192 ")}.` : String(err);
+      return {
+        ok: false,
+        reason: "call_failed",
+        detail: `${detail} Compose without this sub-result or retry with different input.`
+      };
+    }
+    const profile = tryGetProfile(result.actualModel);
+    const costUsd2 = profile ? result.response.tokens.input / 1e6 * profile.costInputPer1m + result.response.tokens.output / 1e6 * profile.costOutputPer1m : 0;
+    spentUsd += costUsd2;
+    return {
+      ok: true,
+      output: result.response.structuredOutput !== null ? JSON.stringify(result.response.structuredOutput) : result.response.text,
+      subHandle: result.handle,
+      executorModel: result.actualModel,
+      costUsd: costUsd2,
+      latencyMs: result.latencyMs,
+      verification: "trusted"
+    };
+  }
+  async function reportComposition(report) {
+    const env = readBrainReadEnv();
+    if (!env.endpoint || !env.jwt || !env.anonKey) {
+      return { ok: false, reason: `brain_read_not_configured:${env.missingEnv.join(",")}` };
+    }
+    try {
+      const res = await fetch(
+        `${env.endpoint.replace(/\/$/, "")}/rest/v1/kgauto_composition_reports`,
+        {
+          method: "POST",
+          headers: {
+            Authorization: `Bearer ${env.jwt}`,
+            apikey: env.anonKey,
+            "Content-Type": "application/json",
+            Prefer: "return=minimal"
+          },
+          body: JSON.stringify({
+            app_id: parentIr.appId,
+            sub_handle: report.subHandle,
+            disposition: report.disposition,
+            ...report.note ? { note: report.note } : {}
+          })
+        }
+      );
+      if (!res.ok) return { ok: false, reason: `write_failed:${res.status}` };
+      return { ok: true };
+    } catch (err) {
+      return { ok: false, reason: `network_error:${err instanceof Error ? err.message : String(err)}` };
+    }
+  }
+  return {
+    toolDefinition: DELEGATE_TOOL_DEFINITION,
+    handler,
+    reportComposition,
+    traceSpendUsd: () => spentUsd
+  };
+}
+
 // src/advisories-api.ts
 var SEVERITY_SET = /* @__PURE__ */ new Set(["info", "warn", "critical"]);
 var STATUS_SET = /* @__PURE__ */ new Set(["open", "snoozed", "resolved"]);
@@ -6212,19 +6579,32 @@ function resolveFetch(injected) {
 function normalizeEndpoint2(endpoint) {
   return endpoint.replace(/\/+$/, "");
 }
+function resolveBrainReadTrio(opts) {
+  const env = readBrainReadEnv();
+  const brainEndpoint = opts.brainEndpoint ?? env.endpoint;
+  const brainJwt = opts.brainJwt ?? env.jwt;
+  const brainAnonKey = opts.brainAnonKey ?? env.anonKey;
+  if (brainEndpoint && brainJwt && brainAnonKey) {
+    return { ok: true, brainEndpoint, brainJwt, brainAnonKey };
+  }
+  const missingEnv = [];
+  if (!brainEndpoint) missingEnv.push("KGAUTO_V2_BRAIN_SUPABASE_URL");
+  if (!brainJwt) missingEnv.push("KGAUTO_V2_BRAIN_JWT");
+  if (!brainAnonKey) missingEnv.push("KGAUTO_V2_BRAIN_ANON_KEY");
+  return { ok: false, missingEnv };
+}
 async function getActionableAdvisories(opts) {
-  const {
-    appId,
-    severity,
-    status,
-    brainEndpoint,
-    brainJwt,
-    brainAnonKey,
-    fetch: injectedFetch
-  } = opts;
+  const { appId, severity, status, fetch: injectedFetch } = opts;
   if (!appId) {
     throw new Error("getActionableAdvisories: appId is required");
   }
+  const trio = resolveBrainReadTrio(opts);
+  if (!trio.ok) {
+    throw new Error(
+      `getActionableAdvisories: brain-read-not-configured \u2014 missing ${trio.missingEnv.join(", ")} (pass brainEndpoint/brainJwt/brainAnonKey explicitly or set the canonical env vars)`
+    );
+  }
+  const { brainEndpoint, brainJwt, brainAnonKey } = trio;
   const doFetch = resolveFetch(injectedFetch);
   const base = normalizeEndpoint2(brainEndpoint);
   const qs = new URLSearchParams();
@@ -6277,17 +6657,15 @@ async function getActionableAdvisories(opts) {
   return out;
 }
 async function markAdvisoryResolved(opts) {
-  const {
-    id,
-    resolutionNote,
-    brainEndpoint,
-    brainJwt,
-    brainAnonKey,
-    fetch: injectedFetch
-  } = opts;
+  const { id, resolutionNote, fetch: injectedFetch } = opts;
   if (!id) {
     return { ok: false, reason: "id_required" };
   }
+  const trio = resolveBrainReadTrio(opts);
+  if (!trio.ok) {
+    return { ok: false, reason: `brain_read_not_configured:${trio.missingEnv.join(",")}` };
+  }
+  const { brainEndpoint, brainJwt, brainAnonKey } = trio;
   const doFetch = resolveFetch(injectedFetch);
   const base = normalizeEndpoint2(brainEndpoint);
   const lookupUrl = `${base}/rest/v1/actionable_advisories_v?id=eq.${encodeURIComponent(id)}&select=app_id,rule`;
@@ -6372,7 +6750,7 @@ async function markAdvisoryResolved(opts) {
     }
   }
   if (outcomeIds.length === 0) {
-    return { ok: true };
+    return { ok: true, firingsResolved: 0, status: "unknown" };
   }
   const inList = outcomeIds.join(",");
   const patchUrl = `${base}/rest/v1/compile_outcome_advisories?outcome_id=in.(${inList})&code=eq.${encodeURIComponent(code)}&resolved_at=is.null`;
@@ -6392,8 +6770,10 @@ async function markAdvisoryResolved(opts) {
         apikey: brainAnonKey,
         "Content-Type": "application/json",
         Accept: "application/json",
-        // PostgREST default is no return; we don't need the row back.
-        Prefer: "return=minimal"
+        // alpha.78 — count what actually changed. An RLS-filtered PATCH
+        // returns 2xx with zero rows; return=representation makes that
+        // visible instead of success-shaped.
+        Prefer: "return=representation"
       },
       body: JSON.stringify(patchBody)
     });
@@ -6410,7 +6790,33 @@ async function markAdvisoryResolved(opts) {
   if (!patchRes.ok) {
     return { ok: false, reason: `patch_failed:${patchRes.status}` };
   }
-  return { ok: true };
+  let patchedRows = [];
+  try {
+    patchedRows = await patchRes.json();
+  } catch {
+  }
+  const firingsResolved = Array.isArray(patchedRows) ? patchedRows.length : 0;
+  let status = "unknown";
+  try {
+    const statusRes = await doFetch(
+      `${base}/rest/v1/actionable_advisories_v?id=eq.${encodeURIComponent(id)}&select=status`,
+      {
+        method: "GET",
+        headers: {
+          Authorization: `Bearer ${brainJwt}`,
+          apikey: brainAnonKey,
+          Accept: "application/json"
+        }
+      }
+    );
+    if (statusRes.ok) {
+      const rows = await statusRes.json();
+      const s = Array.isArray(rows) ? rows[0]?.status : void 0;
+      if (s === "open" || s === "resolved") status = s;
+    }
+  } catch {
+  }
+  return { ok: true, firingsResolved, status };
 }
 async function markExclusionFindingHandled(opts) {
   const {
@@ -6419,12 +6825,14 @@ async function markExclusionFindingHandled(opts) {
     excludedModel,
     resolution,
     resolutionNote,
-    brainEndpoint,
-    brainJwt,
-    brainAnonKey,
     fetch: injectedFetch
   } = opts;
   if (!appId) return { ok: false, reason: "app_id_required" };
+  const trio = resolveBrainReadTrio(opts);
+  if (!trio.ok) {
+    return { ok: false, reason: `brain_read_not_configured:${trio.missingEnv.join(",")}` };
+  }
+  const { brainEndpoint, brainJwt, brainAnonKey } = trio;
   if (!archetype) return { ok: false, reason: "archetype_required" };
   if (!excludedModel) {
     return { ok: false, reason: "excluded_model_required" };
@@ -6690,6 +7098,7 @@ export {
   ALL_ARCHETYPES,
   ARCHETYPE_FAMILY_FITS,
   ARCHETYPE_FLOOR_DEFAULT,
+  BRAIN_READ_ENV_NAMES,
   COACH_CFG,
   CallError,
   DECOMPOSITION_TEMPLATES,
@@ -6697,6 +7106,7 @@ export {
   DEFAULT_FINDINGS_ENDPOINT,
   DEFAULT_MEASURED_FAILURE_ENDPOINT,
   DEFAULT_PROMOTIONS_ENDPOINT,
+  DELEGATE_TOOL_DEFINITION,
   DIALECT_VERSION,
   DISCIPLINE_GATES_V1_ALT_HEADER,
   FamilyResolutionError,
@@ -6708,6 +7118,7 @@ export {
   MEASURED_GROUNDING_MIN_N,
   PRODUCER_OWNED_RULE_CODES,
   PROVIDER_ENV_KEYS,
+  ROLLBACK_SUPPRESSION_WINDOW_DAYS,
   RULE_DISCIPLINE_GATES_V1,
   RULE_DISCIPLINE_GATES_V1_STRUCTURED,
   RULE_SEQUENTIAL_TOOL_CLIFF,
@@ -6743,6 +7154,7 @@ export {
   configurePromotionsBrain,
   countTokens,
   createBrainForwardRoutes,
+  createDelegate,
   createKeyHealthRoute,
   deriveFamilyFromModelId,
   deriveOwnership,
@@ -6761,17 +7173,20 @@ export {
   getPerAxisMetrics,
   getProfile,
   getReachabilityDiagnostic,
+  getRecentRollback,
   getRecommendedPrimary,
   getSequentialStarterChain,
   getSequentialStarterChainWithGrounding,
   getStaleExclusionFindings,
   getStarterChain,
   getStarterChainWithGrounding,
+  hasMutation,
   hashShape,
   isArchetype,
   isAutoPromoteEnabledFromEnv,
   isBrainQueryActiveFor,
   isBrainSync,
+  isDelegateEnabledFromEnv,
   isExclusionFindingsBrainActive,
   isMeasuredFailureBrainActive,
   isMeasuredFailureGateEnabledFromEnv,
@@ -6787,9 +7202,11 @@ export {
   loadChainsFromBrain,
   loadModelsFromBrain,
   loadPricingFromBrain,
+  mapMeasuredFailureRows,
   markAdvisoryResolved,
   markExclusionFindingHandled,
   markPromoteReadyHandled,
+  mutationId,
   parseGoldenCaptureRate,
   parseJudgeVerdict,
   peekBrainDeadLetter,