@warmdrift/kgauto-compiler 2.0.0-alpha.71 → 2.0.0-alpha.73

package/dist/index.js

@@ -49,6 +49,7 @@ __export(index_exports, {
   applyArchetypeConvention: () => applyArchetypeConvention,
   applySectionRewrites: () => applySectionRewrites,
   attachCacheControlToStreamTextInput: () => attachCacheControlToStreamTextInput,
+  awaitMeasuredFailureReady: () => awaitMeasuredFailureReady,
   brainHealth: () => brainHealth,
   bucketContext: () => bucketContext,
   bucketHistory: () => bucketHistory,
@@ -118,6 +119,7 @@ __export(index_exports, {
   parseGoldenCaptureRate: () => parseGoldenCaptureRate,
   parseJudgeVerdict: () => parseJudgeVerdict,
   peekBrainDeadLetter: () => peekBrainDeadLetter,
+  prefetchMeasuredFailure: () => prefetchMeasuredFailure,
   probeShadow: () => probeShadow,
   profileToRow: () => profileToRow,
   profilesByProvider: () => profilesByProvider,
@@ -362,7 +364,8 @@ var PROFILES_RAW = [
     id: "claude-fable-5",
     verifiedAgainstDocs: "2026-06-21",
     provider: "anthropic",
-    status: "preview",
+    status: "current",
+    // s76 2026-07-25: GA since 2026-06-09 per docs; was preview.
     maxContextTokens: 1e6,
     maxOutputTokens: 128e3,
     maxTools: 64,
@@ -416,7 +419,8 @@ var PROFILES_RAW = [
     id: "claude-opus-4-8",
     verifiedAgainstDocs: "2026-06-21",
     provider: "anthropic",
-    status: "current",
+    status: "legacy",
+    // s76 2026-07-25: docs list as LEGACY (Opus 5 generation is current).
     maxContextTokens: 1e6,
     maxOutputTokens: 128e3,
     maxTools: 64,
@@ -456,11 +460,78 @@ var PROFILES_RAW = [
       classify: 8
     }
   },
+  {
+    // ── Claude Opus 5 — added s76 (2026-07-25, alpha.72) ──────────────────
+    // The model every seat in the portfolio actually runs, absent from the
+    // roster until now. Found by a catalog QA, not by the release watcher:
+    // the watcher detects unknown IDs on a provider's /models endpoint, but
+    // detection was never wired to ingestion, so the gap sat open. (That is
+    // what `scripts/catalog-sync.mjs` exists to close.)
+    //
+    // PROVENANCE — doc-verified 2026-07-25 against
+    // platform.claude.com/docs/en/docs/about-claude/models/overview:
+    //   1M context · 128k max output · $5/$25 per MTok · adaptive thinking YES
+    //   · extended thinking NO · `effort` defaults to 'high' on the Claude API
+    //   · reliable knowledge cutoff May 2026.
+    //
+    // FAMILY-INHERITED (not doc-read): maxTools, parallelToolCalls,
+    // structuredOutput, systemPromptMode, streaming. These are identical
+    // across opus-4-6 / 4-7 / 4-8 / sonnet-5 / fable-5 — five consecutive
+    // models — so they are a structural provider constant rather than a
+    // per-model value. Deliberately distinguished from the gemini-3.1-flash-lite
+    // failure, where a template-clone of PRICE (the field that actually varies,
+    // there by 2.5-3.75x) went unverified for 65 days.
+    id: "claude-opus-5",
+    verifiedAgainstDocs: "2026-07-25",
+    provider: "anthropic",
+    status: "current",
+    maxContextTokens: 1e6,
+    maxOutputTokens: 128e3,
+    maxTools: 64,
+    parallelToolCalls: true,
+    structuredOutput: "grammar",
+    systemPromptMode: "inline",
+    streaming: true,
+    cliffs: [],
+    costInputPer1m: 5,
+    costOutputPer1m: 25,
+    lowering: ANTHROPIC_LOWERING_BASE,
+    recovery: [
+      {
+        signal: "rate_limit",
+        action: "escalate",
+        reason: "429 from Anthropic \u2014 escalate to fallback chain"
+      },
+      {
+        signal: "model_not_found",
+        action: "escalate",
+        reason: "Model deprecated/renamed \u2014 escalate (L-061)"
+      }
+    ],
+    strengths: ["reasoning", "agentic_coding", "long_context", "reliable_tool_use", "structured_output"],
+    weaknesses: ["cost", "latency"],
+    notes: "Anthropic frontier for complex agentic coding + enterprise work; the Opus-tier current default as of 2026-07-25, superseding 4.8/4.7/4.6 (all now legacy). Doc-verified 2026-07-25: 1M ctx, 128k out, $5/$25, adaptive thinking (extended thinking NOT supported \u2014 that is Haiku 4.5 only in this generation), `effort` defaults to high on the Claude API. Price is unchanged from 4.5 through 5 \u2014 the tier moved without a cost change.",
+    // COLD-START PRIOR cloned from claude-opus-4-8. Self-labelling: below
+    // MEASURED_GROUNDING_MIN_N (10) brain rows these resolve as
+    // grounding:'judgment', so nothing reads them as measured until they are.
+    archetypePerf: {
+      critique: 10,
+      plan: 10,
+      generate: 9,
+      ask: 9,
+      extract: 9,
+      transform: 9,
+      hunt: 8,
+      summarize: 8,
+      classify: 8
+    }
+  },
   {
     id: "claude-opus-4-7",
     verifiedAgainstDocs: "2026-05-08",
     provider: "anthropic",
-    status: "current",
+    status: "legacy",
+    // s76 2026-07-25: docs list as LEGACY (Opus 5 is current). Bundle said current while the brain already said legacy — the two disagreed.
     maxContextTokens: 1e6,
     maxOutputTokens: 128e3,
     maxTools: 64,
@@ -554,7 +625,8 @@ var PROFILES_RAW = [
     id: "claude-sonnet-4-6",
     verifiedAgainstDocs: "2026-05-08",
     provider: "anthropic",
-    status: "current",
+    status: "legacy",
+    // s76 2026-07-25: docs list as LEGACY (Sonnet 5 is current).
     maxContextTokens: 1e6,
     maxOutputTokens: 64e3,
     maxTools: 64,
@@ -1693,7 +1765,8 @@ var PROFILES_RAW = [
     id: "claude-sonnet-5",
     verifiedAgainstDocs: "2026-07-02",
     provider: "anthropic",
-    status: "preview",
+    status: "current",
+    // s76 2026-07-25: GA per docs; was preview.
     maxContextTokens: 1e6,
     maxOutputTokens: 128e3,
     maxTools: 64,
@@ -2910,20 +2983,27 @@ function passScoreTargets(ir, opts) {
       });
     }
     if (qualityGatePenalty > 0) {
+      const rankBefore = rank + qualityGatePenalty;
       if (measuredGate) {
         const pct = (x) => `${(x * 100).toFixed(0)}%`;
         policyMutations.push({
           id: `quality-gate-measured-${modelId}`,
           source: "quality_gate",
           passName: "score_targets",
-          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%). 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.`
+          rankDelta: -qualityGatePenalty,
+          rankBefore,
+          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 {
         policyMutations.push({
           id: `quality-gate-structured-${modelId}`,
           source: "quality_gate",
           passName: "score_targets",
-          description: `Model ${modelId} gated below the quality floor for archetype '${ir.intent.archetype}' \u2014 declared structuredOutput + kgauto convention flags it schema-weak for this contract (structuredOutputHint:'avoid'). Down-ranked out of leadership; retained as graceful fallback only.`
+          rankDelta: -qualityGatePenalty,
+          rankBefore,
+          rankAfter: rank,
+          description: `Model ${modelId} gated below the quality floor for archetype '${ir.intent.archetype}' \u2014 declared structuredOutput + kgauto convention flags it schema-weak for this contract (structuredOutputHint:'avoid'). Rank ${rankBefore.toFixed(2)} \u2192 ${rank.toFixed(2)} (\u2212${qualityGatePenalty.toFixed(2)}). Down-ranked out of leadership; retained as graceful fallback only.`
         });
       }
     }
@@ -3171,6 +3251,12 @@ function lower(ir, profile, hints = {}) {
 }
 function lowerAnthropic(ir, profile, hints) {
   const systemBlocks = buildAnthropicSystemBlocks(ir.sections, profile);
+  if (ir.constraints?.structuredOutput) {
+    systemBlocks.push({
+      type: "text",
+      text: "Respond with a single valid JSON value and nothing else. Do not wrap it in a markdown code fence, and do not add any prose before or after it."
+    });
+  }
   const history = (ir.history ?? []).filter((m) => m.role !== "system");
   const policy = ir.historyCachePolicy;
   const markIndex = resolveHistoryMarkIndex(history.length, policy);
@@ -4657,6 +4743,40 @@ function configureMeasuredFailureBrain(rt) {
 function isMeasuredFailureBrainActive() {
   return runtime6 !== void 0;
 }
+function prefetchMeasuredFailure(appId) {
+  const rt = runtime6;
+  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 asyncRefresh6(rt, appId);
+  return pendingRefreshes5.get(appId);
+}
+async function awaitMeasuredFailureReady(appId, timeoutMs) {
+  if (!runtime6 || !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 = runtime6;
   if (!rt) return void 0;
@@ -5162,6 +5282,12 @@ function configureBrain(config) {
       fetchImpl: config.fetchImpl ?? fetch,
       onError: config.onError
     });
+    if (config.appId) {
+      try {
+        void prefetchMeasuredFailure(config.appId);
+      } catch {
+      }
+    }
   } else {
     configureMeasuredFailureBrain(void 0);
   }
@@ -5788,6 +5914,41 @@ async function recordGoldenIr(input) {
   }
 }
 
+// src/json-loose.ts
+function parseJsonLoose(raw) {
+  const attempt = (s) => {
+    try {
+      return JSON.parse(s);
+    } catch {
+      return FAILED;
+    }
+  };
+  const direct = attempt(raw);
+  if (direct !== FAILED) return { ok: true, value: direct, recovered: false };
+  const cleaned = raw.trim().replace(/^```(?:json|JSON)?\s*/, "").replace(/\s*```$/, "").trim();
+  const fenced = attempt(cleaned);
+  if (fenced !== FAILED) return { ok: true, value: fenced, recovered: true };
+  const objStart = cleaned.indexOf("{");
+  const arrStart = cleaned.indexOf("[");
+  const start = objStart === -1 ? arrStart : arrStart === -1 ? objStart : Math.min(objStart, arrStart);
+  if (start !== -1) {
+    const openChar = cleaned[start];
+    const closeChar = openChar === "{" ? "}" : "]";
+    const end = cleaned.lastIndexOf(closeChar);
+    if (end > start) {
+      const span = attempt(cleaned.slice(start, end + 1));
+      if (span !== FAILED) return { ok: true, value: span, recovered: true };
+    }
+  }
+  try {
+    JSON.parse(raw);
+    return { ok: false, error: "unreachable" };
+  } catch (err) {
+    return { ok: false, error: err instanceof Error ? err.message : String(err) };
+  }
+}
+var FAILED = /* @__PURE__ */ Symbol("parse-failed");
+
 // src/ir.ts
 var CallError = class extends Error {
   attempts;
@@ -7146,6 +7307,14 @@ function emitFallbackWalked(traceId, appId, data) {
 }
 
 // src/call.ts
+function resolveGateWarmupMs(opts) {
+  const declared = opts.gateWarmupMs;
+  if (typeof declared === "number" && Number.isFinite(declared) && declared >= 0) {
+    return declared;
+  }
+  return DEFAULT_GATE_WARMUP_MS;
+}
+var DEFAULT_GATE_WARMUP_MS = 400;
 async function call(ir, opts = {}) {
   const traceId = generateTraceId();
   safeEmit(
@@ -7161,6 +7330,7 @@ async function call(ir, opts = {}) {
       )
     })
   );
+  await awaitMeasuredFailureReady(ir.appId, resolveGateWarmupMs(opts));
   const initial = compileAndRegister(ir, opts);
   safeEmit(
     () => emitCompileDone(traceId, ir.appId, {
@@ -7744,10 +7914,12 @@ function validateStructuredContract(exec, ir) {
   if (!exec.response.text) {
     return { ok: true, response: exec.response };
   }
-  try {
-    const parsed = JSON.parse(exec.response.text);
-    return { ok: true, response: { ...exec.response, structuredOutput: parsed } };
-  } catch (err) {
+  const parsedResult = parseJsonLoose(exec.response.text);
+  if (parsedResult.ok) {
+    return { ok: true, response: { ...exec.response, structuredOutput: parsedResult.value } };
+  }
+  {
+    const err = new Error(parsedResult.error);
     return {
       ok: false,
       status: exec.status,
@@ -8627,7 +8799,7 @@ function createBrainForwardRoutes(config) {
 }
 
 // src/version.ts
-var LIBRARY_VERSION = "2.0.0-alpha.71";
+var LIBRARY_VERSION = "2.0.0-alpha.73";
 
 // src/key-health.ts
 var JSON_HEADERS2 = { "Content-Type": "application/json" };
@@ -9282,6 +9454,7 @@ function compile2(ir, opts) {
   applyArchetypeConvention,
   applySectionRewrites,
   attachCacheControlToStreamTextInput,
+  awaitMeasuredFailureReady,
   brainHealth,
   bucketContext,
   bucketHistory,
@@ -9351,6 +9524,7 @@ function compile2(ir, opts) {
   parseGoldenCaptureRate,
   parseJudgeVerdict,
   peekBrainDeadLetter,
+  prefetchMeasuredFailure,
   probeShadow,
   profileToRow,
   profilesByProvider,

package/dist/index.mjs

@@ -16,7 +16,7 @@ import {
 import {
   LIBRARY_VERSION,
   createKeyHealthRoute
-} from "./chunk-URFQR3SB.mjs";
+} from "./chunk-65KZE7AC.mjs";
 import {
   ABSOLUTE_FLOOR,
   ARCHETYPE_FLOOR_DEFAULT,
@@ -42,7 +42,7 @@ import {
   loadChainsFromBrain,
   readBrainReadEnv,
   resolveProviderKey
-} from "./chunk-SBFSYCQG.mjs";
+} from "./chunk-46U2NVOL.mjs";
 import {
   ALIASES,
   LATENCY_TIER_MS,
@@ -53,7 +53,7 @@ import {
   latencyTierOf,
   profilesByProvider,
   tryGetProfile
-} from "./chunk-4UO4CCSP.mjs";
+} from "./chunk-N36LE3MK.mjs";
 import {
   emitAdvisoryFired,
   emitCompileDone,
@@ -877,20 +877,27 @@ function passScoreTargets(ir, opts) {
       });
     }
     if (qualityGatePenalty > 0) {
+      const rankBefore = rank + qualityGatePenalty;
       if (measuredGate) {
         const pct = (x) => `${(x * 100).toFixed(0)}%`;
         policyMutations.push({
           id: `quality-gate-measured-${modelId}`,
           source: "quality_gate",
           passName: "score_targets",
-          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%). 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.`
+          rankDelta: -qualityGatePenalty,
+          rankBefore,
+          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 {
         policyMutations.push({
           id: `quality-gate-structured-${modelId}`,
           source: "quality_gate",
           passName: "score_targets",
-          description: `Model ${modelId} gated below the quality floor for archetype '${ir.intent.archetype}' \u2014 declared structuredOutput + kgauto convention flags it schema-weak for this contract (structuredOutputHint:'avoid'). Down-ranked out of leadership; retained as graceful fallback only.`
+          rankDelta: -qualityGatePenalty,
+          rankBefore,
+          rankAfter: rank,
+          description: `Model ${modelId} gated below the quality floor for archetype '${ir.intent.archetype}' \u2014 declared structuredOutput + kgauto convention flags it schema-weak for this contract (structuredOutputHint:'avoid'). Rank ${rankBefore.toFixed(2)} \u2192 ${rank.toFixed(2)} (\u2212${qualityGatePenalty.toFixed(2)}). Down-ranked out of leadership; retained as graceful fallback only.`
         });
       }
     }
@@ -1138,6 +1145,12 @@ function lower(ir, profile, hints = {}) {
 }
 function lowerAnthropic(ir, profile, hints) {
   const systemBlocks = buildAnthropicSystemBlocks(ir.sections, profile);
+  if (ir.constraints?.structuredOutput) {
+    systemBlocks.push({
+      type: "text",
+      text: "Respond with a single valid JSON value and nothing else. Do not wrap it in a markdown code fence, and do not add any prose before or after it."
+    });
+  }
   const history = (ir.history ?? []).filter((m) => m.role !== "system");
   const policy = ir.historyCachePolicy;
   const markIndex = resolveHistoryMarkIndex(history.length, policy);
@@ -2624,6 +2637,40 @@ function configureMeasuredFailureBrain(rt) {
 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;
@@ -3129,6 +3176,12 @@ function configureBrain(config) {
       fetchImpl: config.fetchImpl ?? fetch,
       onError: config.onError
     });
+    if (config.appId) {
+      try {
+        void prefetchMeasuredFailure(config.appId);
+      } catch {
+      }
+    }
   } else {
     configureMeasuredFailureBrain(void 0);
   }
@@ -3755,6 +3808,41 @@ async function recordGoldenIr(input) {
   }
 }
 
+// src/json-loose.ts
+function parseJsonLoose(raw) {
+  const attempt = (s) => {
+    try {
+      return JSON.parse(s);
+    } catch {
+      return FAILED;
+    }
+  };
+  const direct = attempt(raw);
+  if (direct !== FAILED) return { ok: true, value: direct, recovered: false };
+  const cleaned = raw.trim().replace(/^```(?:json|JSON)?\s*/, "").replace(/\s*```$/, "").trim();
+  const fenced = attempt(cleaned);
+  if (fenced !== FAILED) return { ok: true, value: fenced, recovered: true };
+  const objStart = cleaned.indexOf("{");
+  const arrStart = cleaned.indexOf("[");
+  const start = objStart === -1 ? arrStart : arrStart === -1 ? objStart : Math.min(objStart, arrStart);
+  if (start !== -1) {
+    const openChar = cleaned[start];
+    const closeChar = openChar === "{" ? "}" : "]";
+    const end = cleaned.lastIndexOf(closeChar);
+    if (end > start) {
+      const span = attempt(cleaned.slice(start, end + 1));
+      if (span !== FAILED) return { ok: true, value: span, recovered: true };
+    }
+  }
+  try {
+    JSON.parse(raw);
+    return { ok: false, error: "unreachable" };
+  } catch (err) {
+    return { ok: false, error: err instanceof Error ? err.message : String(err) };
+  }
+}
+var FAILED = /* @__PURE__ */ Symbol("parse-failed");
+
 // src/ir.ts
 var CallError = class extends Error {
   attempts;
@@ -4374,6 +4462,14 @@ function tryParseJson2(s) {
 }
 
 // src/call.ts
+function resolveGateWarmupMs(opts) {
+  const declared = opts.gateWarmupMs;
+  if (typeof declared === "number" && Number.isFinite(declared) && declared >= 0) {
+    return declared;
+  }
+  return DEFAULT_GATE_WARMUP_MS;
+}
+var DEFAULT_GATE_WARMUP_MS = 400;
 async function call(ir, opts = {}) {
   const traceId = generateTraceId();
   safeEmit(
@@ -4389,6 +4485,7 @@ async function call(ir, opts = {}) {
       )
     })
   );
+  await awaitMeasuredFailureReady(ir.appId, resolveGateWarmupMs(opts));
   const initial = compileAndRegister(ir, opts);
   safeEmit(
     () => emitCompileDone(traceId, ir.appId, {
@@ -4972,10 +5069,12 @@ function validateStructuredContract(exec, ir) {
   if (!exec.response.text) {
     return { ok: true, response: exec.response };
   }
-  try {
-    const parsed = JSON.parse(exec.response.text);
-    return { ok: true, response: { ...exec.response, structuredOutput: parsed } };
-  } catch (err) {
+  const parsedResult = parseJsonLoose(exec.response.text);
+  if (parsedResult.ok) {
+    return { ok: true, response: { ...exec.response, structuredOutput: parsedResult.value } };
+  }
+  {
+    const err = new Error(parsedResult.error);
     return {
       ok: false,
       status: exec.status,
@@ -6144,6 +6243,7 @@ export {
   applyArchetypeConvention,
   applySectionRewrites,
   attachCacheControlToStreamTextInput,
+  awaitMeasuredFailureReady,
   brainHealth,
   bucketContext,
   bucketHistory,
@@ -6213,6 +6313,7 @@ export {
   parseGoldenCaptureRate,
   parseJudgeVerdict,
   peekBrainDeadLetter,
+  prefetchMeasuredFailure,
   probeShadow,
   profileToRow,
   profilesByProvider,

package/dist/{ir-BEQ28muo.d.ts → ir-B0v2f9NY.d.ts}

@@ -419,6 +419,28 @@ type MutationApplied = {
     source: string;
     passName: string;
     description: string;
+    /**
+     * alpha.73 — the ranking delta this mutation applied, when it applied one.
+     * Negative de-ranks, positive boosts. Absent for mutations that do not touch
+     * rank (advisory/labelling mutations).
+     *
+     * Exists because a consumer could previously see *that* a gate fired and
+     * *why*, but not *how much* — which cannot distinguish "the gate did not see
+     * this model" from "the gate saw it and the penalty lost to a price
+     * advantage." PB spent two days and three wrong conclusions inside that gap
+     * (2026-07-25/26), and both wrong conclusions were inferences substituted
+     * for a quantity that was known at emit time and thrown away. Same class as
+     * the `rankedOn` request: a consumer cannot reason about ranking behaviour
+     * it cannot observe, so it will guess.
+     */
+    rankDelta?: number;
+    /**
+     * alpha.73 — the model's rank before and after this mutation, when it
+     * changed one. Together with `rankDelta` this answers the question a
+     * penalty magnitude alone cannot: whether the gated model still WON.
+     */
+    rankBefore?: number;
+    rankAfter?: number;
 };
 /**
  * Target-specific wire request. Shape varies by provider — caller passes the
@@ -1051,6 +1073,23 @@ interface CallOptions {
      * fire-and-forget. See {@link ShadowProbeConfig}.
      */
     shadowProbe?: ShadowProbeConfig;
+    /**
+     * alpha.73 — how long `call()` may wait for the measured-failure gate's
+     * first brain read before compiling, in ms. Default 400.
+     *
+     * Exists because `compile()` is synchronous and therefore reads only what
+     * the SWR cache already holds: on a cold isolate that is nothing, so the
+     * gate silently does not fire on the first compile — measured on
+     * playbacksam as two billed inferences and ~27–34s of added latency, every
+     * time an isolate is new. Low-traffic consumers are cold most of the time,
+     * so this hits them hardest, inverting the usual severity ordering.
+     *
+     * The wait is bounded and never required: on timeout, brain-down, or no
+     * brain configured, the call proceeds un-gated exactly as before. Set `0`
+     * to disable the wait entirely (the background prefetch still runs, so
+     * later calls in the same isolate are gated either way).
+     */
+    gateWarmupMs?: number;
     toolRelevanceThreshold?: number;
     compressHistoryAfter?: number;
     /** Override API keys (defaults: process.env). */

package/dist/{ir-CnnJST_N.d.mts → ir-CZukZvDn.d.mts}

@@ -419,6 +419,28 @@ type MutationApplied = {
     source: string;
     passName: string;
     description: string;
+    /**
+     * alpha.73 — the ranking delta this mutation applied, when it applied one.
+     * Negative de-ranks, positive boosts. Absent for mutations that do not touch
+     * rank (advisory/labelling mutations).
+     *
+     * Exists because a consumer could previously see *that* a gate fired and
+     * *why*, but not *how much* — which cannot distinguish "the gate did not see
+     * this model" from "the gate saw it and the penalty lost to a price
+     * advantage." PB spent two days and three wrong conclusions inside that gap
+     * (2026-07-25/26), and both wrong conclusions were inferences substituted
+     * for a quantity that was known at emit time and thrown away. Same class as
+     * the `rankedOn` request: a consumer cannot reason about ranking behaviour
+     * it cannot observe, so it will guess.
+     */
+    rankDelta?: number;
+    /**
+     * alpha.73 — the model's rank before and after this mutation, when it
+     * changed one. Together with `rankDelta` this answers the question a
+     * penalty magnitude alone cannot: whether the gated model still WON.
+     */
+    rankBefore?: number;
+    rankAfter?: number;
 };
 /**
  * Target-specific wire request. Shape varies by provider — caller passes the
@@ -1051,6 +1073,23 @@ interface CallOptions {
      * fire-and-forget. See {@link ShadowProbeConfig}.
      */
     shadowProbe?: ShadowProbeConfig;
+    /**
+     * alpha.73 — how long `call()` may wait for the measured-failure gate's
+     * first brain read before compiling, in ms. Default 400.
+     *
+     * Exists because `compile()` is synchronous and therefore reads only what
+     * the SWR cache already holds: on a cold isolate that is nothing, so the
+     * gate silently does not fire on the first compile — measured on
+     * playbacksam as two billed inferences and ~27–34s of added latency, every
+     * time an isolate is new. Low-traffic consumers are cold most of the time,
+     * so this hits them hardest, inverting the usual severity ordering.
+     *
+     * The wait is bounded and never required: on timeout, brain-down, or no
+     * brain configured, the call proceeds un-gated exactly as before. Set `0`
+     * to disable the wait entirely (the background prefetch still runs, so
+     * later calls in the same isolate are gated either way).
+     */
+    gateWarmupMs?: number;
     toolRelevanceThreshold?: number;
     compressHistoryAfter?: number;
     /** Override API keys (defaults: process.env). */

package/dist/key-health.js

@@ -25,7 +25,7 @@ __export(key_health_exports, {
 module.exports = __toCommonJS(key_health_exports);
 
 // src/version.ts
-var LIBRARY_VERSION = "2.0.0-alpha.71";
+var LIBRARY_VERSION = "2.0.0-alpha.73";
 
 // src/key-health.ts
 var JSON_HEADERS = { "Content-Type": "application/json" };

package/dist/key-health.mjs

@@ -1,6 +1,6 @@
 import {
   createKeyHealthRoute
-} from "./chunk-URFQR3SB.mjs";
+} from "./chunk-65KZE7AC.mjs";
 export {
   createKeyHealthRoute
 };

package/dist/profiles.d.mts

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