@warmdrift/kgauto-compiler 2.0.0-alpha.85 → 2.0.0-alpha.86

package/dist/index.js

@@ -72,6 +72,7 @@ __export(index_exports, {
   applySectionRewrites: () => applySectionRewrites,
   attachCacheControlToStreamTextInput: () => attachCacheControlToStreamTextInput,
   awaitMeasuredFailureReady: () => awaitMeasuredFailureReady,
+  awaitPromotionsReady: () => awaitPromotionsReady,
   brainHealth: () => brainHealth,
   bucketContext: () => bucketContext,
   bucketHistory: () => bucketHistory,
@@ -158,6 +159,7 @@ __export(index_exports, {
   peekBrainDeadLetter: () => peekBrainDeadLetter,
   planDecomposition: () => planDecomposition,
   prefetchMeasuredFailure: () => prefetchMeasuredFailure,
+  prefetchPromotions: () => prefetchPromotions,
   probeShadow: () => probeShadow,
   profileToRow: () => profileToRow,
   profilesByProvider: () => profilesByProvider,
@@ -972,9 +974,21 @@ var PROFILES_RAW = [
   },
   {
     id: "gemini-2.5-pro",
-    verifiedAgainstDocs: "2026-05-08",
+    verifiedAgainstDocs: "2026-08-08",
     provider: "google",
-    status: "current",
+    // alpha.86 — AVAILABILITY demotion, not a provider-lifecycle mirror.
+    // Google's docs still list 2.5-pro as Stable, but the serving layer
+    // disagrees per account cohort: tt-intel measured 4/4 refusals
+    // ("no longer available to new users", 2026-08-02, their bench), and
+    // kgauto's own key 429s on a FreeTier per-day quota (probed
+    // 2026-08-08). A chain anchor that hard-fails for whole account
+    // cohorts is not an anchor; `legacy` removes it from cost-ordered
+    // chains and new selection while keeping the profile resolvable for
+    // consumers who pass it explicitly. Re-promotion is one status flip
+    // if Google restores availability. The registry cannot yet express
+    // account-cohort availability as data — see interfaces/kgauto.md
+    // ## Requested 2026-08-02.
+    status: "legacy",
     maxContextTokens: 1048576,
     maxOutputTokens: 65535,
     maxTools: 128,
@@ -1004,7 +1018,7 @@ var PROFILES_RAW = [
       }
     ],
     strengths: ["reasoning", "1m_context", "structured_output", "tool_use"],
-    weaknesses: ["pricing_above_200k"],
+    weaknesses: ["pricing_above_200k", "availability_restricted_by_account_cohort"],
     // Master plan §3.3 anchor: tier-2 cross-provider in almost every chain.
     // Sits on the frontier at perf-9 — close to Sonnet but cheaper input.
     archetypePerf: {
@@ -2848,9 +2862,12 @@ function passApplyCliffs(ir, profile, estimatedInputTokens) {
     if (!triggered) continue;
     switch (cliff.action) {
       case "drop_to_top_relevant": {
-        const targetCount = Math.min(
-          Math.floor(cliff.threshold * 0.75),
-          Math.max(1, Math.floor((nextIR.tools?.length ?? 0) / 2))
+        const targetCount = Math.max(
+          1,
+          Math.min(
+            Math.floor(cliff.threshold * 0.75),
+            Math.floor((nextIR.tools?.length ?? 0) / 2)
+          )
         );
         if (nextIR.tools && nextIR.tools.length > targetCount) {
           const intent = nextIR.intent.archetype;
@@ -3928,8 +3945,14 @@ var STARTER_CHAINS_GROUNDED = {
   critique: [
     { id: "claude-opus-4-7", grounding: "judgment", reason: "Highest reasoning bar, no degradation tier \u2014 engineer pick, awaiting measured backing" },
     { id: "claude-sonnet-4-6", grounding: "judgment", reason: "Same-provider walk-down from Opus on 429" },
-    { id: "gemini-2.5-pro", grounding: "judgment", reason: "Cross-provider anchor in similar quality bracket" },
-    { id: "gpt-5.5", grounding: "judgment", reason: "alpha.16: third-provider frontier-tier floor (archetypePerf=9)" }
+    // alpha.86: gemini-2.5-pro dropped — availability-restricted (tt-intel
+    // measured 4/4 cohort refusal 2026-08-02; kgauto key free-tier-quota
+    // 429 2026-08-08). No current Google model clears the critique floor
+    // (3.5-flash critique=5; 3.1-pro-preview scores 10 but is preview —
+    // default chains carry promoted models only). deepseek-v4-pro takes the
+    // third-provider tail instead.
+    { id: "gpt-5.5", grounding: "judgment", reason: "alpha.16: third-provider frontier-tier floor (archetypePerf=9)" },
+    { id: "deepseek-v4-pro", grounding: "judgment", reason: "alpha.86: third-provider tail replacing availability-restricted gemini-2.5-pro \u2014 critique=6 clears the floor exactly; last-resort when Anthropic AND OpenAI are down, not a quality peer of the tiers above" }
   ],
   // alpha.62 (eval spine) — pairwise output comparison. Same reasoning-floor
   // posture as critique (a verdict from a weak judge is worse than no
@@ -3938,14 +3961,16 @@ var STARTER_CHAINS_GROUNDED = {
   judge: [
     { id: "claude-opus-4-8", grounding: "judgment", reason: "Highest reasoning bar for pairwise verdicts \u2014 current opus (s64 roster)" },
     { id: "claude-sonnet-4-6", grounding: "judgment", reason: "Same-provider walk-down on 429" },
-    { id: "gemini-2.5-pro", grounding: "judgment", reason: "Cross-provider anchor in similar quality bracket" },
-    { id: "gpt-5.5", grounding: "judgment", reason: "Third-provider frontier-tier floor" }
+    // alpha.86: gemini-2.5-pro dropped (availability-restricted; see critique
+    // note). gpt-5.5 remains the cross-provider frontier floor for verdicts.
+    { id: "gpt-5.5", grounding: "judgment", reason: "Third-provider frontier-tier floor" },
+    { id: "deepseek-v4-pro", grounding: "judgment", reason: "alpha.86: third-provider tail replacing availability-restricted gemini-2.5-pro \u2014 native structured output per the judge contract; last-resort tier" }
   ],
   // Reasoning matters — Sonnet primary; walk UP to Opus on 429.
   plan: [
     { id: "claude-sonnet-4-6", grounding: "judgment", reason: "Reasoning + cost balance \u2014 engineer pick" },
     { id: "claude-opus-4-7", grounding: "judgment", reason: 'Same-provider walk-UP on 429 (rare exception to "always cheaper")' },
-    { id: "gemini-2.5-pro", grounding: "judgment", reason: "Cross-provider anchor" },
+    { id: "gemini-3.5-flash", grounding: "judgment", reason: "alpha.86: cross-provider anchor \u2014 replaces availability-restricted gemini-2.5-pro (tt-intel 4/4 cohort refusal 2026-08-02; probed serving 200 on kgauto key 2026-08-08)" },
     { id: "deepseek-v4-pro", grounding: "judgment", reason: "Tier 3 cost floor \u2014 no brain evidence yet" },
     { id: "gpt-5.4", grounding: "judgment", reason: "alpha.33: third-provider tail (archetypePerf=7) \u2014 closes openai-in-default-fallback-chains" }
   ],
@@ -3953,7 +3978,7 @@ var STARTER_CHAINS_GROUNDED = {
   generate: [
     { id: "claude-sonnet-4-6", grounding: "judgment", reason: "Quality + cost match \u2014 engineer pick" },
     { id: "claude-haiku-4-5", grounding: "judgment", reason: "Same-provider step-down" },
-    { id: "gemini-2.5-pro", grounding: "judgment", reason: "Cross-provider anchor" },
+    { id: "gemini-3.5-flash", grounding: "judgment", reason: "alpha.86: cross-provider anchor \u2014 replaces availability-restricted gemini-2.5-pro (tt-intel 4/4 cohort refusal 2026-08-02; probed serving 200 on kgauto key 2026-08-08)" },
     { id: "gpt-5.4-mini", grounding: "judgment", reason: "alpha.16: third-provider tail (archetypePerf=7) \u2014 closes mono-Anthropic gap" }
   ],
   // ask::sonnet — STARTER_CHAINS calls this "Quality + cost match" but
@@ -3962,7 +3987,7 @@ var STARTER_CHAINS_GROUNDED = {
   ask: [
     { id: "claude-sonnet-4-6", grounding: "judgment", reason: "Quality + cost match \u2014 engineer pick. NOTE: tt-intel s78 prod showed 27% empty rate; placement awaits measurement validation" },
     { id: "claude-haiku-4-5", grounding: "judgment", reason: "Same-provider step-down" },
-    { id: "gemini-2.5-pro", grounding: "judgment", reason: "Cross-provider anchor" },
+    { id: "gemini-3.5-flash", grounding: "judgment", reason: "alpha.86: cross-provider anchor \u2014 replaces availability-restricted gemini-2.5-pro (tt-intel 4/4 cohort refusal 2026-08-02; probed serving 200 on kgauto key 2026-08-08)" },
     { id: "gpt-5.4-mini", grounding: "judgment", reason: "alpha.16: third-provider tail (archetypePerf=7)" }
   ],
   // Structured-output archetype — Flash skipped (alpha.8 MAX_TOKENS cliff,
@@ -3970,14 +3995,14 @@ var STARTER_CHAINS_GROUNDED = {
   extract: [
     { id: "claude-sonnet-4-6", grounding: "judgment", reason: "Reliable structured-output anchor \u2014 engineer pick" },
     { id: "claude-haiku-4-5", grounding: "judgment", reason: "Same-provider step-down with native structured output" },
-    { id: "gemini-2.5-pro", grounding: "judgment", reason: "Cross-provider anchor with structured-output support" },
+    { id: "gemini-3.5-flash", grounding: "judgment", reason: "alpha.86: cross-provider anchor with native structured output \u2014 replaces availability-restricted gemini-2.5-pro" },
     { id: "gpt-5.4", grounding: "capability-fact", reason: "alpha.16: third-provider floor \u2014 native structured-output capability (archetypePerf=8)" }
   ],
   // Forgiving archetype — Sonnet primary but Flash safely floors it.
   transform: [
     { id: "claude-sonnet-4-6", grounding: "judgment", reason: "Quality anchor \u2014 engineer pick" },
     { id: "claude-haiku-4-5", grounding: "judgment", reason: "Same-provider step-down" },
-    { id: "gemini-2.5-pro", grounding: "judgment", reason: "Cross-provider anchor" },
+    { id: "gemini-3.5-flash", grounding: "judgment", reason: "alpha.86: cross-provider anchor \u2014 replaces availability-restricted gemini-2.5-pro (tt-intel 4/4 cohort refusal 2026-08-02; probed serving 200 on kgauto key 2026-08-08)" },
     { id: "gemini-2.5-flash", grounding: "judgment", reason: "Cost floor \u2014 forgiving archetype tolerates Flash" },
     { id: "gpt-5.4-mini", grounding: "judgment", reason: "alpha.33: third-provider tail (archetypePerf=7) \u2014 closes openai-in-default-fallback-chains" }
   ],
@@ -3985,7 +4010,7 @@ var STARTER_CHAINS_GROUNDED = {
   // (capability-fact: Flash 15-75 parallel calls/step vs DeepSeek 7-8).
   hunt: [
     { id: "gemini-2.5-flash", grounding: "capability-fact", reason: "L-040 parallel-tool throughput champion (15-75 calls/step)" },
-    { id: "gemini-2.5-pro", grounding: "capability-fact", reason: "Cross-provider tier 1 with strong parallel-tool support" },
+    { id: "gemini-3.5-flash", grounding: "capability-fact", reason: "alpha.86: parallel-tool agentic tier (hunt archetypePerf 9, L-040 family) \u2014 replaces availability-restricted gemini-2.5-pro" },
     { id: "claude-sonnet-4-6", grounding: "judgment", reason: "Quality safety net for blocked-Flash case" },
     { id: "claude-haiku-4-5", grounding: "judgment", reason: "Reduced tool budget \u2014 cliff at 16 fires" },
     { id: "gpt-5.4-mini", grounding: "judgment", reason: "alpha.33: third-provider tail \u2014 OpenAI parallel-tool capable archetype" }
@@ -4033,9 +4058,9 @@ var STARTER_CHAINS_BY_MODE_GROUNDED = {
         reason: "Cross-provider safety net \u2014 Sonnet handles sequential agentic loops cleanly"
       },
       {
-        id: "gemini-2.5-pro",
+        id: "gemini-3.5-flash",
         grounding: "judgment",
-        reason: "Third-provider tail when no DeepSeek key reachable"
+        reason: "alpha.86: third-provider tail when no DeepSeek key reachable \u2014 replaces availability-restricted gemini-2.5-pro"
       }
     ]
   }
@@ -4601,6 +4626,40 @@ function defaultOnError4(err) {
     err
   );
 }
+function prefetchPromotions(appId) {
+  const rt = runtime4;
+  if (!rt || !appId) return void 0;
+  let snap = snapshots3.get(appId);
+  if (!snap) {
+    snap = { data: [], expiresAt: 0, refreshing: false };
+    snapshots3.set(appId, snap);
+  }
+  if (snap.expiresAt > Date.now()) return void 0;
+  const inflight = pendingRefreshes3.get(appId);
+  if (inflight) return inflight;
+  if (snap.refreshing) return void 0;
+  snap.refreshing = true;
+  void asyncRefresh4(rt, appId);
+  return pendingRefreshes3.get(appId);
+}
+async function awaitPromotionsReady(appId, timeoutMs) {
+  if (!runtime4 || !appId) return;
+  const pending = prefetchPromotions(appId) ?? pendingRefreshes3.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 _testResetPromotions() {
   runtime4 = void 0;
   snapshots3.clear();
@@ -5131,6 +5190,31 @@ function advisorRuleBlockedModelDrift(ctx) {
   return out;
 }
 
+// src/advisor-rules/preferred-blocked-overlap.ts
+var PREFERRED_MODEL_BLOCKED_CODE = "preferred-model-blocked";
+var DOCS_URL2 = "https://github.com/stue/command-center/blob/main/interfaces/kgauto.md#best-practice-advisories";
+function advisorRulePreferredBlockedOverlap(ctx) {
+  const preferred = ctx.policy?.preferredModels;
+  const blocked = ctx.policy?.blockedModels;
+  if (!preferred?.length || !blocked?.length) return [];
+  const blockedSet = canonicalPolicySet(blocked);
+  const deadPins = [...new Set(preferred)].filter((p) => policySetHas(blockedSet, p)).sort();
+  if (deadPins.length === 0) return [];
+  const list = deadPins.map((p) => `\`${p}\``).join(", ");
+  const one = deadPins.length === 1;
+  const preferredSet = canonicalPolicySet(preferred);
+  const servedIsPinned = policySetHas(preferredSet, ctx.selectedModelId);
+  return [
+    {
+      level: "warn",
+      code: PREFERRED_MODEL_BLOCKED_CODE,
+      message: `CompilePolicy.preferredModels entr${one ? "y" : "ies"} ${list} ${one ? "is" : "are"} also in \`blockedModels\` (matched canonically, aliases included). A blocked model is hard-rejected before preference boosts apply, so ${one ? "this pin" : "these pins"} can never serve. ` + (servedIsPinned ? `This call was served by \`${ctx.selectedModelId}\`, which is itself a live pin \u2014 the dead entr${one ? "y is" : "ies are"} latent, not currently rerouting traffic.` : `Substitution is your steady state: this call landed on \`${ctx.selectedModelId}\`, which you did not pin.`),
+      suggestion: `Two readings, and kgauto cannot tell them apart: (1) intentional \u2014 your spend gate deliberately outranks the pin, in which case nothing needs doing and you can filter this code; (2) misconfiguration \u2014 the pin and the block were written at different times and the overlap is an accident. tt-intel hit reading (2) on 2026-08-01: a summarize site pinned \`claude-sonnet\` while \`KGAUTO_BLOCKED_MODELS\` carried the same family, and every call silently substituted a reasoning model whose reasoning burn exceeded the site's \`maxOutputTokens\` \u2014 100% empty payloads under HTTP 200. If the block should win, remove the pin so the policy says what it does. If the pin should win, remove ${one ? "the blocking entry" : "the blocking entries"} or re-scope the block to the exact ids you mean. Check what the substitute costs at your input shape while you are here \u2014 a spend gate that lands traffic on a pricier model than the one it blocked is a cost inversion, not a saving.`,
+      docsUrl: DOCS_URL2
+    }
+  ];
+}
+
 // src/advisor.ts
 var QUALITY_FLOOR_FOR_RECOMMENDATION = 6;
 var TIER_DOWN_COST_RATIO = 0.5;
@@ -5155,6 +5239,12 @@ function runAdvisor(ir, result, profile, policy, phase2) {
       selectedModelId: profile.id
     })
   );
+  out.push(
+    ...advisorRulePreferredBlockedOverlap({
+      policy,
+      selectedModelId: profile.id
+    })
+  );
   if (policy?.posture !== "locked") {
     out.push(...detectCostMismatchedArchetype(ir, profile, phase2));
     out.push(...detectModelStaleEvidence(ir, profile));
@@ -5967,7 +6057,7 @@ function validateFinalFit(ir, profile, tokens) {
 }
 
 // src/version.ts
-var LIBRARY_VERSION = "2.0.0-alpha.85";
+var LIBRARY_VERSION = "2.0.0-alpha.86";
 
 // src/pricing-brain.ts
 function isPricingRow(x) {
@@ -6078,6 +6168,12 @@ function configureBrain(config) {
       fetchImpl: config.fetchImpl ?? fetch,
       onError: config.onError
     });
+    if (config.appId) {
+      try {
+        void prefetchPromotions(config.appId);
+      } catch {
+      }
+    }
   } else {
     configurePromotionsBrain(void 0);
   }
@@ -6449,6 +6545,7 @@ function buildPayload(input, reg) {
     latency_ms: input.latencyMs,
     success: input.success,
     empty_response: input.emptyResponse ?? input.tokensOut === 0,
+    source: input.source,
     error_type: input.errorType,
     tools_called: input.toolsCalled,
     oracle_score: input.oracleScore?.score,
@@ -7848,7 +7945,11 @@ async function call(ir, opts = {}) {
       )
     })
   );
-  await awaitMeasuredFailureReady(ir.appId, resolveGateWarmupMs(opts));
+  const warmupMs = resolveGateWarmupMs(opts);
+  await Promise.all([
+    awaitMeasuredFailureReady(ir.appId, warmupMs),
+    awaitPromotionsReady(ir.appId, warmupMs)
+  ]);
   const initial = compileAndRegister(ir, opts);
   safeEmit(
     () => emitCompileDone(traceId, ir.appId, {
@@ -7888,6 +7989,7 @@ async function call(ir, opts = {}) {
       const latencyMs2 = Date.now() - start;
       await record({
         handle: initial.handle,
+        source: opts.source,
         tokensIn: 0,
         tokensOut: 0,
         latencyMs: latencyMs2,
@@ -7940,6 +8042,7 @@ async function call(ir, opts = {}) {
       const latencyMs2 = Date.now() - start;
       await record({
         handle: initial.handle,
+        source: opts.source,
         tokensIn: 0,
         tokensOut: 0,
         latencyMs: latencyMs2,
@@ -8081,6 +8184,7 @@ async function call(ir, opts = {}) {
       const goldenShapeKey = goldenRate > 0 ? peekRegisteredShapeKey(initial.handle) : void 0;
       await record({
         handle: initial.handle,
+        source: opts.source,
         tokensIn: validated.response.tokens.input,
         tokensOut: validated.response.tokens.output,
         latencyMs: latencyMs2,
@@ -8203,6 +8307,7 @@ async function call(ir, opts = {}) {
   const latencyMs = Date.now() - start;
   await record({
     handle: initial.handle,
+    source: opts.source,
     tokensIn: lastErr?.tokens?.input ?? 0,
     tokensOut: lastErr?.tokens?.output ?? 0,
     latencyMs,
@@ -10743,6 +10848,7 @@ function compile2(ir, opts) {
   applySectionRewrites,
   attachCacheControlToStreamTextInput,
   awaitMeasuredFailureReady,
+  awaitPromotionsReady,
   brainHealth,
   bucketContext,
   bucketHistory,
@@ -10829,6 +10935,7 @@ function compile2(ir, opts) {
   peekBrainDeadLetter,
   planDecomposition,
   prefetchMeasuredFailure,
+  prefetchPromotions,
   probeShadow,
   profileToRow,
   profilesByProvider,

package/dist/index.mjs

@@ -19,7 +19,7 @@ import {
   LIBRARY_VERSION,
   createKeyHealthRoute,
   keyFingerprint
-} from "./chunk-WXN7FNXP.mjs";
+} from "./chunk-V4T2CBRI.mjs";
 import {
   ABSOLUTE_FLOOR,
   ARCHETYPE_FLOOR_DEFAULT,
@@ -56,7 +56,7 @@ import {
   policySetHas,
   readBrainReadEnv,
   resolveProviderKey
-} from "./chunk-LO2JXTGG.mjs";
+} from "./chunk-QDWOMQYN.mjs";
 import {
   ALIASES,
   LATENCY_TIER_MS,
@@ -68,7 +68,7 @@ import {
   profilesByProvider,
   resolveModelAlias,
   tryGetProfile
-} from "./chunk-FD3NFXDC.mjs";
+} from "./chunk-2MQIVVEU.mjs";
 import {
   emitAdvisoryFired,
   emitCompileDone,
@@ -683,9 +683,12 @@ function passApplyCliffs(ir, profile, estimatedInputTokens) {
     if (!triggered) continue;
     switch (cliff.action) {
       case "drop_to_top_relevant": {
-        const targetCount = Math.min(
-          Math.floor(cliff.threshold * 0.75),
-          Math.max(1, Math.floor((nextIR.tools?.length ?? 0) / 2))
+        const targetCount = Math.max(
+          1,
+          Math.min(
+            Math.floor(cliff.threshold * 0.75),
+            Math.floor((nextIR.tools?.length ?? 0) / 2)
+          )
         );
         if (nextIR.tools && nextIR.tools.length > targetCount) {
           const intent = nextIR.intent.archetype;
@@ -1895,6 +1898,40 @@ function defaultOnError3(err) {
     err
   );
 }
+function prefetchPromotions(appId) {
+  const rt = runtime3;
+  if (!rt || !appId) return void 0;
+  let snap = snapshots3.get(appId);
+  if (!snap) {
+    snap = { data: [], expiresAt: 0, refreshing: false };
+    snapshots3.set(appId, snap);
+  }
+  if (snap.expiresAt > Date.now()) return void 0;
+  const inflight = pendingRefreshes3.get(appId);
+  if (inflight) return inflight;
+  if (snap.refreshing) return void 0;
+  snap.refreshing = true;
+  void asyncRefresh3(rt, appId);
+  return pendingRefreshes3.get(appId);
+}
+async function awaitPromotionsReady(appId, timeoutMs) {
+  if (!runtime3 || !appId) return;
+  const pending = prefetchPromotions(appId) ?? pendingRefreshes3.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 _testResetPromotions() {
   runtime3 = void 0;
   snapshots3.clear();
@@ -2425,6 +2462,31 @@ function advisorRuleBlockedModelDrift(ctx) {
   return out;
 }
 
+// src/advisor-rules/preferred-blocked-overlap.ts
+var PREFERRED_MODEL_BLOCKED_CODE = "preferred-model-blocked";
+var DOCS_URL2 = "https://github.com/stue/command-center/blob/main/interfaces/kgauto.md#best-practice-advisories";
+function advisorRulePreferredBlockedOverlap(ctx) {
+  const preferred = ctx.policy?.preferredModels;
+  const blocked = ctx.policy?.blockedModels;
+  if (!preferred?.length || !blocked?.length) return [];
+  const blockedSet = canonicalPolicySet(blocked);
+  const deadPins = [...new Set(preferred)].filter((p) => policySetHas(blockedSet, p)).sort();
+  if (deadPins.length === 0) return [];
+  const list = deadPins.map((p) => `\`${p}\``).join(", ");
+  const one = deadPins.length === 1;
+  const preferredSet = canonicalPolicySet(preferred);
+  const servedIsPinned = policySetHas(preferredSet, ctx.selectedModelId);
+  return [
+    {
+      level: "warn",
+      code: PREFERRED_MODEL_BLOCKED_CODE,
+      message: `CompilePolicy.preferredModels entr${one ? "y" : "ies"} ${list} ${one ? "is" : "are"} also in \`blockedModels\` (matched canonically, aliases included). A blocked model is hard-rejected before preference boosts apply, so ${one ? "this pin" : "these pins"} can never serve. ` + (servedIsPinned ? `This call was served by \`${ctx.selectedModelId}\`, which is itself a live pin \u2014 the dead entr${one ? "y is" : "ies are"} latent, not currently rerouting traffic.` : `Substitution is your steady state: this call landed on \`${ctx.selectedModelId}\`, which you did not pin.`),
+      suggestion: `Two readings, and kgauto cannot tell them apart: (1) intentional \u2014 your spend gate deliberately outranks the pin, in which case nothing needs doing and you can filter this code; (2) misconfiguration \u2014 the pin and the block were written at different times and the overlap is an accident. tt-intel hit reading (2) on 2026-08-01: a summarize site pinned \`claude-sonnet\` while \`KGAUTO_BLOCKED_MODELS\` carried the same family, and every call silently substituted a reasoning model whose reasoning burn exceeded the site's \`maxOutputTokens\` \u2014 100% empty payloads under HTTP 200. If the block should win, remove the pin so the policy says what it does. If the pin should win, remove ${one ? "the blocking entry" : "the blocking entries"} or re-scope the block to the exact ids you mean. Check what the substitute costs at your input shape while you are here \u2014 a spend gate that lands traffic on a pricier model than the one it blocked is a cost inversion, not a saving.`,
+      docsUrl: DOCS_URL2
+    }
+  ];
+}
+
 // src/advisor.ts
 var QUALITY_FLOOR_FOR_RECOMMENDATION = 6;
 var TIER_DOWN_COST_RATIO = 0.5;
@@ -2449,6 +2511,12 @@ function runAdvisor(ir, result, profile, policy, phase2) {
       selectedModelId: profile.id
     })
   );
+  out.push(
+    ...advisorRulePreferredBlockedOverlap({
+      policy,
+      selectedModelId: profile.id
+    })
+  );
   if (policy?.posture !== "locked") {
     out.push(...detectCostMismatchedArchetype(ir, profile, phase2));
     out.push(...detectModelStaleEvidence(ir, profile));
@@ -3369,6 +3437,12 @@ function configureBrain(config) {
       fetchImpl: config.fetchImpl ?? fetch,
       onError: config.onError
     });
+    if (config.appId) {
+      try {
+        void prefetchPromotions(config.appId);
+      } catch {
+      }
+    }
   } else {
     configurePromotionsBrain(void 0);
   }
@@ -3740,6 +3814,7 @@ function buildPayload(input, reg) {
     latency_ms: input.latencyMs,
     success: input.success,
     empty_response: input.emptyResponse ?? input.tokensOut === 0,
+    source: input.source,
     error_type: input.errorType,
     tools_called: input.toolsCalled,
     oracle_score: input.oracleScore?.score,
@@ -4838,7 +4913,11 @@ async function call(ir, opts = {}) {
       )
     })
   );
-  await awaitMeasuredFailureReady(ir.appId, resolveGateWarmupMs(opts));
+  const warmupMs = resolveGateWarmupMs(opts);
+  await Promise.all([
+    awaitMeasuredFailureReady(ir.appId, warmupMs),
+    awaitPromotionsReady(ir.appId, warmupMs)
+  ]);
   const initial = compileAndRegister(ir, opts);
   safeEmit(
     () => emitCompileDone(traceId, ir.appId, {
@@ -4878,6 +4957,7 @@ async function call(ir, opts = {}) {
       const latencyMs2 = Date.now() - start;
       await record({
         handle: initial.handle,
+        source: opts.source,
         tokensIn: 0,
         tokensOut: 0,
         latencyMs: latencyMs2,
@@ -4930,6 +5010,7 @@ async function call(ir, opts = {}) {
       const latencyMs2 = Date.now() - start;
       await record({
         handle: initial.handle,
+        source: opts.source,
         tokensIn: 0,
         tokensOut: 0,
         latencyMs: latencyMs2,
@@ -5071,6 +5152,7 @@ async function call(ir, opts = {}) {
       const goldenShapeKey = goldenRate > 0 ? peekRegisteredShapeKey(initial.handle) : void 0;
       await record({
         handle: initial.handle,
+        source: opts.source,
         tokensIn: validated.response.tokens.input,
         tokensOut: validated.response.tokens.output,
         latencyMs: latencyMs2,
@@ -5193,6 +5275,7 @@ async function call(ir, opts = {}) {
   const latencyMs = Date.now() - start;
   await record({
     handle: initial.handle,
+    source: opts.source,
     tokensIn: lastErr?.tokens?.input ?? 0,
     tokensOut: lastErr?.tokens?.output ?? 0,
     latencyMs,
@@ -7357,6 +7440,7 @@ export {
   applySectionRewrites,
   attachCacheControlToStreamTextInput,
   awaitMeasuredFailureReady,
+  awaitPromotionsReady,
   brainHealth,
   bucketContext,
   bucketHistory,
@@ -7443,6 +7527,7 @@ export {
   peekBrainDeadLetter,
   planDecomposition,
   prefetchMeasuredFailure,
+  prefetchPromotions,
   probeShadow,
   profileToRow,
   profilesByProvider,

package/dist/{ir-JUoybVM-.d.mts → ir-5TJLAYKR.d.mts}

@@ -1084,6 +1084,13 @@ interface ShadowProbeConfig {
     skipSlowTierInSync?: boolean;
 }
 interface CallOptions {
+    /**
+     * alpha.86 — self-mark for synthetic callers (canaries, smokes, eval
+     * replays, probes). Passed through to every outcome row this call
+     * records, including failure rows. OMIT for real traffic. See
+     * {@link OutcomeSource}.
+     */
+    source?: OutcomeSource;
     /**
      * alpha.84 — fires once per FAILED attempt during the fallback walk, before
      * the walk continues. IC-Cairn's filing, 2026-07-29.
@@ -1458,9 +1465,34 @@ interface OracleScore {
     /** Free-form explanation for debugging. */
     rationale?: string;
 }
+/**
+ * alpha.86 — who originated an outcome row (migration 056).
+ *
+ * Absent/undefined means ORGANIC: a real consumer call on behalf of a real
+ * user. Consumers change nothing. Synthetic writers — canaries, smoke
+ * gates, eval replays, probe harnesses — self-mark, so liveness and volume
+ * rules can compute over consumer-originated rows only.
+ *
+ * The incident this closes (cost-watch 2026-08-06 → 08-08): playbacksam's
+ * daily canary wrote exactly 8 rows/day for 13 straight days while PB's
+ * organic traffic was near-zero, and the liveness rule scored PB the
+ * healthiest consumer in the portfolio. On 08-08 EVERY row the brain
+ * received in 24h was canary, and no rule could say so — distinguishing a
+ * replay from real traffic took a hand-reconstructed time-window +
+ * token-fingerprint argument, three mornings running. A consumer whose
+ * only rows are written by kgauto's own machinery is dark, not healthy;
+ * this column is what lets a rule print that sentence.
+ */
+type OutcomeSource = 'canary' | 'smoke' | 'eval' | 'probe' | 'synthetic';
 interface RecordInput {
     /** Handle from CompileResult. */
     handle: string;
+    /**
+     * alpha.86 — self-mark for synthetic writers (migration 056). OMIT for
+     * real traffic; never write an explicit "organic" value. See
+     * {@link OutcomeSource}.
+     */
+    source?: OutcomeSource;
     /** Actual tokens consumed (post-call). */
     tokensIn: number;
     tokensOut: number;

package/dist/{ir-DfrM71pA.d.ts → ir-C1zmvAh6.d.ts}

@@ -1084,6 +1084,13 @@ interface ShadowProbeConfig {
     skipSlowTierInSync?: boolean;
 }
 interface CallOptions {
+    /**
+     * alpha.86 — self-mark for synthetic callers (canaries, smokes, eval
+     * replays, probes). Passed through to every outcome row this call
+     * records, including failure rows. OMIT for real traffic. See
+     * {@link OutcomeSource}.
+     */
+    source?: OutcomeSource;
     /**
      * alpha.84 — fires once per FAILED attempt during the fallback walk, before
      * the walk continues. IC-Cairn's filing, 2026-07-29.
@@ -1458,9 +1465,34 @@ interface OracleScore {
     /** Free-form explanation for debugging. */
     rationale?: string;
 }
+/**
+ * alpha.86 — who originated an outcome row (migration 056).
+ *
+ * Absent/undefined means ORGANIC: a real consumer call on behalf of a real
+ * user. Consumers change nothing. Synthetic writers — canaries, smoke
+ * gates, eval replays, probe harnesses — self-mark, so liveness and volume
+ * rules can compute over consumer-originated rows only.
+ *
+ * The incident this closes (cost-watch 2026-08-06 → 08-08): playbacksam's
+ * daily canary wrote exactly 8 rows/day for 13 straight days while PB's
+ * organic traffic was near-zero, and the liveness rule scored PB the
+ * healthiest consumer in the portfolio. On 08-08 EVERY row the brain
+ * received in 24h was canary, and no rule could say so — distinguishing a
+ * replay from real traffic took a hand-reconstructed time-window +
+ * token-fingerprint argument, three mornings running. A consumer whose
+ * only rows are written by kgauto's own machinery is dark, not healthy;
+ * this column is what lets a rule print that sentence.
+ */
+type OutcomeSource = 'canary' | 'smoke' | 'eval' | 'probe' | 'synthetic';
 interface RecordInput {
     /** Handle from CompileResult. */
     handle: string;
+    /**
+     * alpha.86 — self-mark for synthetic writers (migration 056). OMIT for
+     * real traffic; never write an explicit "organic" value. See
+     * {@link OutcomeSource}.
+     */
+    source?: OutcomeSource;
     /** Actual tokens consumed (post-call). */
     tokensIn: number;
     tokensOut: number;

package/dist/key-health.js

@@ -28,7 +28,7 @@ __export(key_health_exports, {
 module.exports = __toCommonJS(key_health_exports);
 
 // src/version.ts
-var LIBRARY_VERSION = "2.0.0-alpha.85";
+var LIBRARY_VERSION = "2.0.0-alpha.86";
 
 // src/key-health.ts
 var JSON_HEADERS = { "Content-Type": "application/json" };

package/dist/key-health.mjs

@@ -3,7 +3,7 @@ import {
   KEY_FINGERPRINT_LENGTH,
   createKeyHealthRoute,
   keyFingerprint
-} from "./chunk-WXN7FNXP.mjs";
+} from "./chunk-V4T2CBRI.mjs";
 export {
   KEY_FINGERPRINT_DOMAIN,
   KEY_FINGERPRINT_LENGTH,

package/dist/profiles.d.mts

@@ -1,4 +1,4 @@
-import { k as Provider } from './ir-JUoybVM-.mjs';
+import { k as Provider } from './ir-5TJLAYKR.mjs';
 import { IntentArchetypeName } from './dialect.mjs';
 
 /**

package/dist/profiles.d.ts

@@ -1,4 +1,4 @@
-import { k as Provider } from './ir-DfrM71pA.js';
+import { k as Provider } from './ir-C1zmvAh6.js';
 import { IntentArchetypeName } from './dialect.js';
 
 /**