@warmdrift/kgauto-compiler 2.0.0-alpha.72 → 2.0.0-alpha.74

package/dist/{chunk-AI4NNR5T.mjs → chunk-EB2YR7HL.mjs}

@@ -1,5 +1,5 @@
 // src/version.ts
-var LIBRARY_VERSION = "2.0.0-alpha.72";
+var LIBRARY_VERSION = "2.0.0-alpha.74";
 
 // src/key-health.ts
 var JSON_HEADERS = { "Content-Type": "application/json" };

package/dist/glassbox/index.d.mts

@@ -1,6 +1,6 @@
-import { G as GlassboxEvent } from '../types-BDFrJkma.mjs';
-export { A as AdvisoryFiredData, C as CompileDoneData, a as CompileStartData, E as ExecuteAttemptData, b as ExecuteSuccessData, F as FallbackWalkedData, c as GLASSBOX_STREAM_TTL_MS, d as GlassboxEventKind, e as GlassboxPubSub } from '../types-BDFrJkma.mjs';
-import '../ir-CnnJST_N.mjs';
+import { G as GlassboxEvent } from '../types-BgJ5iQB_.mjs';
+export { A as AdvisoryFiredData, C as CompileDoneData, a as CompileStartData, E as ExecuteAttemptData, b as ExecuteSuccessData, F as FallbackWalkedData, c as GLASSBOX_STREAM_TTL_MS, d as GlassboxEventKind, e as GlassboxPubSub } from '../types-BgJ5iQB_.mjs';
+import '../ir-CZukZvDn.mjs';
 import '../dialect.mjs';
 
 /**

package/dist/glassbox/index.d.ts

@@ -1,6 +1,6 @@
-import { G as GlassboxEvent } from '../types-B--CYzMo.js';
-export { A as AdvisoryFiredData, C as CompileDoneData, a as CompileStartData, E as ExecuteAttemptData, b as ExecuteSuccessData, F as FallbackWalkedData, c as GLASSBOX_STREAM_TTL_MS, d as GlassboxEventKind, e as GlassboxPubSub } from '../types-B--CYzMo.js';
-import '../ir-BEQ28muo.js';
+import { G as GlassboxEvent } from '../types-D-PuKCWk.js';
+export { A as AdvisoryFiredData, C as CompileDoneData, a as CompileStartData, E as ExecuteAttemptData, b as ExecuteSuccessData, F as FallbackWalkedData, c as GLASSBOX_STREAM_TTL_MS, d as GlassboxEventKind, e as GlassboxPubSub } from '../types-D-PuKCWk.js';
+import '../ir-B0v2f9NY.js';
 import '../dialect.js';
 
 /**

package/dist/glassbox-routes/format.d.mts

@@ -1,5 +1,5 @@
-import { T as TraceHealth } from '../types-DeGRCTlJ.mjs';
-import '../ir-CnnJST_N.mjs';
+import { T as TraceHealth } from '../types-BOTLfJOp.mjs';
+import '../ir-CZukZvDn.mjs';
 import '../dialect.mjs';
 
 /**

package/dist/glassbox-routes/format.d.ts

@@ -1,5 +1,5 @@
-import { T as TraceHealth } from '../types-DSeJJ6tt.js';
-import '../ir-BEQ28muo.js';
+import { T as TraceHealth } from '../types-Ba68lIs1.js';
+import '../ir-B0v2f9NY.js';
 import '../dialect.js';
 
 /**

package/dist/glassbox-routes/index.d.mts

@@ -1,7 +1,7 @@
-import { G as GlassboxEvent } from '../types-BDFrJkma.mjs';
-import { a as TraceDetail, b as TraceSummary, c as TraceCounterfactual } from '../types-DeGRCTlJ.mjs';
-export { A as AdvisoryRecord, T as TraceHealth, d as TraceSectionRewrite } from '../types-DeGRCTlJ.mjs';
-import '../ir-CnnJST_N.mjs';
+import { G as GlassboxEvent } from '../types-BgJ5iQB_.mjs';
+import { a as TraceDetail, b as TraceSummary, c as TraceCounterfactual } from '../types-BOTLfJOp.mjs';
+export { A as AdvisoryRecord, T as TraceHealth, d as TraceSectionRewrite } from '../types-BOTLfJOp.mjs';
+import '../ir-CZukZvDn.mjs';
 import '../dialect.mjs';
 
 /**

package/dist/glassbox-routes/index.d.ts

@@ -1,7 +1,7 @@
-import { G as GlassboxEvent } from '../types-B--CYzMo.js';
-import { a as TraceDetail, b as TraceSummary, c as TraceCounterfactual } from '../types-DSeJJ6tt.js';
-export { A as AdvisoryRecord, T as TraceHealth, d as TraceSectionRewrite } from '../types-DSeJJ6tt.js';
-import '../ir-BEQ28muo.js';
+import { G as GlassboxEvent } from '../types-D-PuKCWk.js';
+import { a as TraceDetail, b as TraceSummary, c as TraceCounterfactual } from '../types-Ba68lIs1.js';
+export { A as AdvisoryRecord, T as TraceHealth, d as TraceSectionRewrite } from '../types-Ba68lIs1.js';
+import '../ir-B0v2f9NY.js';
 import '../dialect.js';
 
 /**

package/dist/glassbox-routes/react/index.d.mts

@@ -1,6 +1,6 @@
 import * as react_jsx_runtime from 'react/jsx-runtime';
-import { a as TraceDetail } from '../../types-DeGRCTlJ.mjs';
-import '../../ir-CnnJST_N.mjs';
+import { a as TraceDetail } from '../../types-BOTLfJOp.mjs';
+import '../../ir-CZukZvDn.mjs';
 import '../../dialect.mjs';
 
 /**

package/dist/glassbox-routes/react/index.d.ts

@@ -1,6 +1,6 @@
 import * as react_jsx_runtime from 'react/jsx-runtime';
-import { a as TraceDetail } from '../../types-DSeJJ6tt.js';
-import '../../ir-BEQ28muo.js';
+import { a as TraceDetail } from '../../types-Ba68lIs1.js';
+import '../../ir-B0v2f9NY.js';
 import '../../dialect.js';
 
 /**

package/dist/index.d.mts

@@ -1,5 +1,5 @@
-import { C as CompilePolicy, N as NormalizedResponse, A as ApiKeys, P as ProviderOverrides, a as CompiledRequest, b as PromptIR, c as CallOptions, d as CallResult, S as SystemModelMessage, e as CompileResult, B as BestPracticeAdvisory, f as SectionRewrite, R as RecordInput, g as RecordOutcomeInput, O as OutcomeResult, h as OracleScore, i as Adapter, j as PerAxisMetrics, k as Provider, l as ChainEntry, G as Grounding } from './ir-CnnJST_N.mjs';
-export { m as CallAttempt, n as CallError, o as ChainModelEntry, p as ChainWithGrounding, q as Constraints, E as EffortLevel, F as FallbackReason, r as GoldenCaptureOptions, H as HistoryCachePolicy, I as IntentDeclaration, M as Message, s as MutationApplied, t as NormalizedTokens, u as OutcomeKind, v as PerAxisMetricsByModel, w as PromptSection, x as SectionKind, y as ShadowProbeConfig, T as ToolCall, z as ToolDefinition, D as captureGoldenIr, J as parseGoldenCaptureRate, K as resolveGoldenCaptureRate, L as shouldCaptureGolden } from './ir-CnnJST_N.mjs';
+import { C as CompilePolicy, N as NormalizedResponse, A as ApiKeys, P as ProviderOverrides, a as CompiledRequest, b as PromptIR, c as CallOptions, d as CallResult, S as SystemModelMessage, e as CompileResult, B as BestPracticeAdvisory, f as SectionRewrite, R as RecordInput, g as RecordOutcomeInput, O as OutcomeResult, h as OracleScore, i as Adapter, j as PerAxisMetrics, k as Provider, l as ChainEntry, G as Grounding } from './ir-CZukZvDn.mjs';
+export { m as CallAttempt, n as CallError, o as ChainModelEntry, p as ChainWithGrounding, q as Constraints, E as EffortLevel, F as FallbackReason, r as GoldenCaptureOptions, H as HistoryCachePolicy, I as IntentDeclaration, M as Message, s as MutationApplied, t as NormalizedTokens, u as OutcomeKind, v as PerAxisMetricsByModel, w as PromptSection, x as SectionKind, y as ShadowProbeConfig, T as ToolCall, z as ToolDefinition, D as captureGoldenIr, J as parseGoldenCaptureRate, K as resolveGoldenCaptureRate, L as shouldCaptureGolden } from './ir-CZukZvDn.mjs';
 import { ModelProfile, ArchetypeConvention } from './profiles.mjs';
 export { ALIASES, CacheStrategy, CliffRule, LATENCY_TIER_MS, LatencyTier, LoweringSpec, RecoveryRule, StructuredOutputCapability, SystemPromptMode, allProfiles, getProfile, latencyTierOf, profilesByProvider, tryGetProfile } from './profiles.mjs';
 export { BrainForwardConfig, BrainForwardRoutes, createBrainForwardRoutes } from './brain-proxy.mjs';
@@ -120,11 +120,6 @@ declare function execute(request: CompiledRequest, opts?: ExecuteOptions): Promi
  * Hard cap at fallbackChain.length + 1 attempts. Terminal errors short-circuit.
  */
 
-/**
- * Compile, execute, normalize, record. Returns a CallResult once a provider
- * actually serves the request. Throws CallError if the fallback chain is
- * exhausted without success.
- */
 declare function call(ir: PromptIR, opts?: CallOptions): Promise<CallResult>;
 /** Served-side result the AI-SDK consumer already holds from streamText. */
 interface ProbeShadowServed {
@@ -440,6 +435,19 @@ interface BrainQueryConfig {
 interface BrainConfig {
     /** Brain HTTP endpoint base URL (e.g., https://kgauto-brain.vercel.app/api). */
     endpoint: string;
+    /**
+     * alpha.73 — the consumer's app id, declared once so brain-driven levers can
+     * start their first read at configuration time instead of at the first
+     * compile.
+     *
+     * Optional and purely an optimisation: every IR already carries `appId`, and
+     * omitting this changes no verdict — it only means the measured-failure
+     * gate's warm-up starts when `call()` runs rather than when the module
+     * initialises. Declaring it lets a short-lived isolate spend its startup
+     * time on the fetch, so the bounded wait in `call()` usually costs nothing.
+     * Consumers running one app per process should set it.
+     */
+    appId?: string;
     /** Bearer token for auth. */
     apiKey?: string;
     /** Optional error hook for debugging. Defaults to console.warn. */
@@ -1009,7 +1017,7 @@ declare function runGoldenEval(opts: GoldenEvalOptions): Promise<GoldenEvalRunRe
  * guard in `tests/version.test.ts` fails the suite (and therefore
  * `prepublishOnly`) when they diverge — a stale constant cannot reach npm.
  */
-declare const LIBRARY_VERSION = "2.0.0-alpha.72";
+declare const LIBRARY_VERSION = "2.0.0-alpha.74";
 
 /**
  * Oracle contract — how an app tells the brain whether a response was good.
@@ -2997,12 +3005,52 @@ interface GetMeasuredFailureOpts {
     archetype: IntentArchetypeName | string;
     model: string;
 }
+/**
+ * alpha.73 — start the warm-up WITHOUT needing a compile to trigger it.
+ *
+ * The SWR posture below ("cold start returns undefined and warms in the
+ * background") is right for a long-lived process and wrong for a
+ * short-lived isolate, where the first compile IS most compiles. Measured on
+ * playbacksam: on a cold isolate the gate was absent, `claude-haiku-4-5` led,
+ * cliffed, was same-model retried and fell over to gemini — two billed
+ * inferences and ~27–34s, every time an isolate is new. PB has ~1 user, so
+ * its isolates are cold most of the time, which inverts the usual severity
+ * ordering: this hits the LOWEST-traffic consumers hardest, and they are the
+ * ones least able to absorb it.
+ *
+ * Kicking the fetch off at `configureBrain()` moves the refresh into the time
+ * the isolate spends on everything else it does before its first compile.
+ * Fire-and-forget and non-blocking: on its own this narrows the window rather
+ * than closing it, which is why `awaitMeasuredFailureReady()` exists too.
+ *
+ * Returns the in-flight promise (or undefined when not configured / already
+ * fresh) so a caller can await it. NEVER throws.
+ */
+declare function prefetchMeasuredFailure(appId: string): Promise<void> | undefined;
+/**
+ * alpha.73 — bounded await on the warm-up, for async callers only.
+ *
+ * `compile()` is synchronous and must stay so, so it cannot block on a fetch;
+ * `call()` is async and is about to spend seconds on an inference, which makes
+ * a few hundred milliseconds here a trivially good trade against the ~27–34s
+ * and second billed inference a cold-blind gate costs.
+ *
+ * Deliberately bounded and deliberately silent on timeout: the gate is
+ * protective, never required. If the brain is slow or down, the caller
+ * proceeds un-gated exactly as it does today — this can delay a call by at
+ * most `timeoutMs`, and can never fail one.
+ *
+ * NEVER throws.
+ */
+declare function awaitMeasuredFailureReady(appId: string, timeoutMs: number): Promise<void>;
 /**
  * Sync reader. Returns the verdict for `(appId, archetype, model)` or
  * undefined (not configured / cold / below minSample / brain down).
  *
  * NEVER throws. Cold start returns undefined and warms in the background —
- * same posture as every other brain-driven lever in compile().
+ * same posture as every other brain-driven lever in compile(). Async callers
+ * that can afford a bounded wait should call `awaitMeasuredFailureReady()`
+ * first; `call()` does.
  */
 declare function getMeasuredFailureVerdict(opts: GetMeasuredFailureOpts): MeasuredFailureVerdict | undefined;
 declare function _testResetMeasuredFailure(): void;
@@ -3054,4 +3102,4 @@ declare function _testWaitForMeasuredFailureRefresh(): Promise<void>;
  */
 declare function compile(ir: PromptIR, opts?: CompileOptions): CompileResult;
 
-export { ABSOLUTE_FLOOR, type AISDKConvertedMessage, ARCHETYPE_FAMILY_FITS, ARCHETYPE_FLOOR_DEFAULT, type ActionableAdvisory, Adapter, type AdvisoryResolutionSource, type AdvisorySeverity, type AdvisoryStatus, type AdvisorySuggestedFix, ApiKeys, type AppOracle, type ApplySectionRewritesArgs, type ApplySectionRewritesResult, ArchetypeConvention, type ArchetypeFamilyFit, type ArchetypePerfMap, type ArchetypePerfNMap, type ArchetypePerfScoreResult, type AttachCacheControlResult, BestPracticeAdvisory, type BrainConfig, type BrainDeadLetterEntry, type BrainHealthSnapshot, type BrainQueryConfig, type BrainReadEnv, CallOptions, CallResult, ChainEntry, type CompatibilityIntent, type CompileForAISDKv6Result, type CompileOptions, CompilePolicy, CompileResult, CompiledRequest, DEFAULT_FINDINGS_ENDPOINT, DEFAULT_MEASURED_FAILURE_ENDPOINT, DEFAULT_PROMOTIONS_ENDPOINT, type ExclusionFindingRow, type ExclusionResolutionSource, type ExecuteErr, type ExecuteOk, type ExecuteOptions, type ExecuteResult, type FallbackPosture, FamilyResolutionError, type GetActionableAdvisoriesOptions, type GetApplicablePromotionOpts, type GetDefaultFallbackChainOpts, type GetMeasuredFailureOpts, type GetPerAxisMetricsOpts, type GetRecommendedPrimaryOptions, type GoldenEvalCase, type GoldenEvalOptions, type GoldenEvalRunResult, type GoldenIrRecordInput, Grounding, IntentArchetypeName, JUDGE_RUBRICS, LIBRARY_VERSION, type LLMJudgeOptions, MEASURED_FAILURE_CFG, MEASURED_GROUNDING_MIN_N, type MarkAdvisoryResolvedOptions, type MarkExclusionFindingHandledOptions, type MarkPromoteReadyHandledOptions, type MeasuredFailureRuntime, type MeasuredFailureVerdict, type ModelBrainRow, type ModelCompatibility, ModelProfile, NormalizedResponse, type OracleContext, OracleScore, type OutcomePayload, OutcomeResult, OutputMode, PRODUCER_OWNED_RULE_CODES, PROVIDER_ENV_KEYS, PerAxisMetrics, type PricingRow, type ProbeShadowOptions, type ProbeShadowServed, type ProfileToRowOptions, type PromoteReadyFindingRow, type PromoteReadyResolution, type PromotionRow, type PromotionsRuntime, PromptIR, Provider, ProviderOverrides, type ProviderReachability, RULE_SEQUENTIAL_TOOL_CLIFF, type ReachabilityOpts, RecordInput, RecordOutcomeInput, type RunAdvisorPhase2Context, SectionRewrite, type ShadowProbeRecordInput, type SupportedProvider, type SurfaceFailureRow, SystemModelMessage, TRANSLATOR_FLOOR, _testResetMeasuredFailure, _testResetPromotions, _testWaitForMeasuredFailureRefresh, _testWaitForPromotionsRefresh, applyArchetypeConvention, applySectionRewrites, attachCacheControlToStreamTextInput, brainHealth, buildGoldenIrRow, buildLLMJudge, buildPairwiseJudgePrompt, buildShadowProbeRow, call, clearBrain, combineOrderSwappedVerdicts, compile, compileForAISDKv6, configureBrain, configureMeasuredFailureBrain, configurePromotionsBrain, countTokens, deriveFamilyFromModelId, deriveOwnership, execute, findBetterFit, flushBrainDeadLetter, getActionableAdvisories, getAllStarterChains, getAllStarterChainsWithGrounding, getApplicablePromotion, getArchetypePerfScore, getDefaultFallbackChain, getDefaultFallbackChainWithGrounding, getMeasuredFailureVerdict, getModelCompatibility, getPerAxisMetrics, getReachabilityDiagnostic, getRecommendedPrimary, getSequentialStarterChain, getSequentialStarterChainWithGrounding, getStaleExclusionFindings, getStarterChain, getStarterChainWithGrounding, isAutoPromoteEnabledFromEnv, isBrainQueryActiveFor, isBrainSync, isExclusionFindingsBrainActive, isMeasuredFailureBrainActive, isMeasuredFailureGateEnabledFromEnv, isModelReachable, isPromotionsBrainActive, isProviderReachable, judgeMeasuredFailure, loadAliasesFromBrain, loadArchetypePerfFromBrain, loadArchetypePerfNFromBrain, loadChainsFromBrain, loadModelsFromBrain, loadPricingFromBrain, markAdvisoryResolved, markExclusionFindingHandled, markPromoteReadyHandled, parseJudgeVerdict, peekBrainDeadLetter, probeShadow, profileToRow, readBrainReadEnv, record, recordGoldenIr, recordOutcome, recordShadowProbe, renderIrForJudge, resetTokenizer, resolveConventionsForProfile, resolvePricingAt, resolveProviderKey, rubricFor, runAdvisor, runGoldenEval, setTokenizer, wilsonLowerBound };
+export { ABSOLUTE_FLOOR, type AISDKConvertedMessage, ARCHETYPE_FAMILY_FITS, ARCHETYPE_FLOOR_DEFAULT, type ActionableAdvisory, Adapter, type AdvisoryResolutionSource, type AdvisorySeverity, type AdvisoryStatus, type AdvisorySuggestedFix, ApiKeys, type AppOracle, type ApplySectionRewritesArgs, type ApplySectionRewritesResult, ArchetypeConvention, type ArchetypeFamilyFit, type ArchetypePerfMap, type ArchetypePerfNMap, type ArchetypePerfScoreResult, type AttachCacheControlResult, BestPracticeAdvisory, type BrainConfig, type BrainDeadLetterEntry, type BrainHealthSnapshot, type BrainQueryConfig, type BrainReadEnv, CallOptions, CallResult, ChainEntry, type CompatibilityIntent, type CompileForAISDKv6Result, type CompileOptions, CompilePolicy, CompileResult, CompiledRequest, DEFAULT_FINDINGS_ENDPOINT, DEFAULT_MEASURED_FAILURE_ENDPOINT, DEFAULT_PROMOTIONS_ENDPOINT, type ExclusionFindingRow, type ExclusionResolutionSource, type ExecuteErr, type ExecuteOk, type ExecuteOptions, type ExecuteResult, type FallbackPosture, FamilyResolutionError, type GetActionableAdvisoriesOptions, type GetApplicablePromotionOpts, type GetDefaultFallbackChainOpts, type GetMeasuredFailureOpts, type GetPerAxisMetricsOpts, type GetRecommendedPrimaryOptions, type GoldenEvalCase, type GoldenEvalOptions, type GoldenEvalRunResult, type GoldenIrRecordInput, Grounding, IntentArchetypeName, JUDGE_RUBRICS, LIBRARY_VERSION, type LLMJudgeOptions, MEASURED_FAILURE_CFG, MEASURED_GROUNDING_MIN_N, type MarkAdvisoryResolvedOptions, type MarkExclusionFindingHandledOptions, type MarkPromoteReadyHandledOptions, type MeasuredFailureRuntime, type MeasuredFailureVerdict, type ModelBrainRow, type ModelCompatibility, ModelProfile, NormalizedResponse, type OracleContext, OracleScore, type OutcomePayload, OutcomeResult, OutputMode, PRODUCER_OWNED_RULE_CODES, PROVIDER_ENV_KEYS, PerAxisMetrics, type PricingRow, type ProbeShadowOptions, type ProbeShadowServed, type ProfileToRowOptions, type PromoteReadyFindingRow, type PromoteReadyResolution, type PromotionRow, type PromotionsRuntime, PromptIR, Provider, ProviderOverrides, type ProviderReachability, RULE_SEQUENTIAL_TOOL_CLIFF, type ReachabilityOpts, RecordInput, RecordOutcomeInput, type RunAdvisorPhase2Context, SectionRewrite, type ShadowProbeRecordInput, type SupportedProvider, type SurfaceFailureRow, SystemModelMessage, TRANSLATOR_FLOOR, _testResetMeasuredFailure, _testResetPromotions, _testWaitForMeasuredFailureRefresh, _testWaitForPromotionsRefresh, applyArchetypeConvention, applySectionRewrites, attachCacheControlToStreamTextInput, awaitMeasuredFailureReady, brainHealth, buildGoldenIrRow, buildLLMJudge, buildPairwiseJudgePrompt, buildShadowProbeRow, call, clearBrain, combineOrderSwappedVerdicts, compile, compileForAISDKv6, configureBrain, configureMeasuredFailureBrain, configurePromotionsBrain, countTokens, deriveFamilyFromModelId, deriveOwnership, execute, findBetterFit, flushBrainDeadLetter, getActionableAdvisories, getAllStarterChains, getAllStarterChainsWithGrounding, getApplicablePromotion, getArchetypePerfScore, getDefaultFallbackChain, getDefaultFallbackChainWithGrounding, getMeasuredFailureVerdict, getModelCompatibility, getPerAxisMetrics, getReachabilityDiagnostic, getRecommendedPrimary, getSequentialStarterChain, getSequentialStarterChainWithGrounding, getStaleExclusionFindings, getStarterChain, getStarterChainWithGrounding, isAutoPromoteEnabledFromEnv, isBrainQueryActiveFor, isBrainSync, isExclusionFindingsBrainActive, isMeasuredFailureBrainActive, isMeasuredFailureGateEnabledFromEnv, isModelReachable, isPromotionsBrainActive, isProviderReachable, judgeMeasuredFailure, loadAliasesFromBrain, loadArchetypePerfFromBrain, loadArchetypePerfNFromBrain, loadChainsFromBrain, loadModelsFromBrain, loadPricingFromBrain, markAdvisoryResolved, markExclusionFindingHandled, markPromoteReadyHandled, parseJudgeVerdict, peekBrainDeadLetter, prefetchMeasuredFailure, probeShadow, profileToRow, readBrainReadEnv, record, recordGoldenIr, recordOutcome, recordShadowProbe, renderIrForJudge, resetTokenizer, resolveConventionsForProfile, resolvePricingAt, resolveProviderKey, rubricFor, runAdvisor, runGoldenEval, setTokenizer, wilsonLowerBound };

package/dist/index.d.ts

@@ -1,5 +1,5 @@
-import { C as CompilePolicy, N as NormalizedResponse, A as ApiKeys, P as ProviderOverrides, a as CompiledRequest, b as PromptIR, c as CallOptions, d as CallResult, S as SystemModelMessage, e as CompileResult, B as BestPracticeAdvisory, f as SectionRewrite, R as RecordInput, g as RecordOutcomeInput, O as OutcomeResult, h as OracleScore, i as Adapter, j as PerAxisMetrics, k as Provider, l as ChainEntry, G as Grounding } from './ir-BEQ28muo.js';
-export { m as CallAttempt, n as CallError, o as ChainModelEntry, p as ChainWithGrounding, q as Constraints, E as EffortLevel, F as FallbackReason, r as GoldenCaptureOptions, H as HistoryCachePolicy, I as IntentDeclaration, M as Message, s as MutationApplied, t as NormalizedTokens, u as OutcomeKind, v as PerAxisMetricsByModel, w as PromptSection, x as SectionKind, y as ShadowProbeConfig, T as ToolCall, z as ToolDefinition, D as captureGoldenIr, J as parseGoldenCaptureRate, K as resolveGoldenCaptureRate, L as shouldCaptureGolden } from './ir-BEQ28muo.js';
+import { C as CompilePolicy, N as NormalizedResponse, A as ApiKeys, P as ProviderOverrides, a as CompiledRequest, b as PromptIR, c as CallOptions, d as CallResult, S as SystemModelMessage, e as CompileResult, B as BestPracticeAdvisory, f as SectionRewrite, R as RecordInput, g as RecordOutcomeInput, O as OutcomeResult, h as OracleScore, i as Adapter, j as PerAxisMetrics, k as Provider, l as ChainEntry, G as Grounding } from './ir-B0v2f9NY.js';
+export { m as CallAttempt, n as CallError, o as ChainModelEntry, p as ChainWithGrounding, q as Constraints, E as EffortLevel, F as FallbackReason, r as GoldenCaptureOptions, H as HistoryCachePolicy, I as IntentDeclaration, M as Message, s as MutationApplied, t as NormalizedTokens, u as OutcomeKind, v as PerAxisMetricsByModel, w as PromptSection, x as SectionKind, y as ShadowProbeConfig, T as ToolCall, z as ToolDefinition, D as captureGoldenIr, J as parseGoldenCaptureRate, K as resolveGoldenCaptureRate, L as shouldCaptureGolden } from './ir-B0v2f9NY.js';
 import { ModelProfile, ArchetypeConvention } from './profiles.js';
 export { ALIASES, CacheStrategy, CliffRule, LATENCY_TIER_MS, LatencyTier, LoweringSpec, RecoveryRule, StructuredOutputCapability, SystemPromptMode, allProfiles, getProfile, latencyTierOf, profilesByProvider, tryGetProfile } from './profiles.js';
 export { BrainForwardConfig, BrainForwardRoutes, createBrainForwardRoutes } from './brain-proxy.js';
@@ -120,11 +120,6 @@ declare function execute(request: CompiledRequest, opts?: ExecuteOptions): Promi
  * Hard cap at fallbackChain.length + 1 attempts. Terminal errors short-circuit.
  */
 
-/**
- * Compile, execute, normalize, record. Returns a CallResult once a provider
- * actually serves the request. Throws CallError if the fallback chain is
- * exhausted without success.
- */
 declare function call(ir: PromptIR, opts?: CallOptions): Promise<CallResult>;
 /** Served-side result the AI-SDK consumer already holds from streamText. */
 interface ProbeShadowServed {
@@ -440,6 +435,19 @@ interface BrainQueryConfig {
 interface BrainConfig {
     /** Brain HTTP endpoint base URL (e.g., https://kgauto-brain.vercel.app/api). */
     endpoint: string;
+    /**
+     * alpha.73 — the consumer's app id, declared once so brain-driven levers can
+     * start their first read at configuration time instead of at the first
+     * compile.
+     *
+     * Optional and purely an optimisation: every IR already carries `appId`, and
+     * omitting this changes no verdict — it only means the measured-failure
+     * gate's warm-up starts when `call()` runs rather than when the module
+     * initialises. Declaring it lets a short-lived isolate spend its startup
+     * time on the fetch, so the bounded wait in `call()` usually costs nothing.
+     * Consumers running one app per process should set it.
+     */
+    appId?: string;
     /** Bearer token for auth. */
     apiKey?: string;
     /** Optional error hook for debugging. Defaults to console.warn. */
@@ -1009,7 +1017,7 @@ declare function runGoldenEval(opts: GoldenEvalOptions): Promise<GoldenEvalRunRe
  * guard in `tests/version.test.ts` fails the suite (and therefore
  * `prepublishOnly`) when they diverge — a stale constant cannot reach npm.
  */
-declare const LIBRARY_VERSION = "2.0.0-alpha.72";
+declare const LIBRARY_VERSION = "2.0.0-alpha.74";
 
 /**
  * Oracle contract — how an app tells the brain whether a response was good.
@@ -2997,12 +3005,52 @@ interface GetMeasuredFailureOpts {
     archetype: IntentArchetypeName | string;
     model: string;
 }
+/**
+ * alpha.73 — start the warm-up WITHOUT needing a compile to trigger it.
+ *
+ * The SWR posture below ("cold start returns undefined and warms in the
+ * background") is right for a long-lived process and wrong for a
+ * short-lived isolate, where the first compile IS most compiles. Measured on
+ * playbacksam: on a cold isolate the gate was absent, `claude-haiku-4-5` led,
+ * cliffed, was same-model retried and fell over to gemini — two billed
+ * inferences and ~27–34s, every time an isolate is new. PB has ~1 user, so
+ * its isolates are cold most of the time, which inverts the usual severity
+ * ordering: this hits the LOWEST-traffic consumers hardest, and they are the
+ * ones least able to absorb it.
+ *
+ * Kicking the fetch off at `configureBrain()` moves the refresh into the time
+ * the isolate spends on everything else it does before its first compile.
+ * Fire-and-forget and non-blocking: on its own this narrows the window rather
+ * than closing it, which is why `awaitMeasuredFailureReady()` exists too.
+ *
+ * Returns the in-flight promise (or undefined when not configured / already
+ * fresh) so a caller can await it. NEVER throws.
+ */
+declare function prefetchMeasuredFailure(appId: string): Promise<void> | undefined;
+/**
+ * alpha.73 — bounded await on the warm-up, for async callers only.
+ *
+ * `compile()` is synchronous and must stay so, so it cannot block on a fetch;
+ * `call()` is async and is about to spend seconds on an inference, which makes
+ * a few hundred milliseconds here a trivially good trade against the ~27–34s
+ * and second billed inference a cold-blind gate costs.
+ *
+ * Deliberately bounded and deliberately silent on timeout: the gate is
+ * protective, never required. If the brain is slow or down, the caller
+ * proceeds un-gated exactly as it does today — this can delay a call by at
+ * most `timeoutMs`, and can never fail one.
+ *
+ * NEVER throws.
+ */
+declare function awaitMeasuredFailureReady(appId: string, timeoutMs: number): Promise<void>;
 /**
  * Sync reader. Returns the verdict for `(appId, archetype, model)` or
  * undefined (not configured / cold / below minSample / brain down).
  *
  * NEVER throws. Cold start returns undefined and warms in the background —
- * same posture as every other brain-driven lever in compile().
+ * same posture as every other brain-driven lever in compile(). Async callers
+ * that can afford a bounded wait should call `awaitMeasuredFailureReady()`
+ * first; `call()` does.
  */
 declare function getMeasuredFailureVerdict(opts: GetMeasuredFailureOpts): MeasuredFailureVerdict | undefined;
 declare function _testResetMeasuredFailure(): void;
@@ -3054,4 +3102,4 @@ declare function _testWaitForMeasuredFailureRefresh(): Promise<void>;
  */
 declare function compile(ir: PromptIR, opts?: CompileOptions): CompileResult;
 
-export { ABSOLUTE_FLOOR, type AISDKConvertedMessage, ARCHETYPE_FAMILY_FITS, ARCHETYPE_FLOOR_DEFAULT, type ActionableAdvisory, Adapter, type AdvisoryResolutionSource, type AdvisorySeverity, type AdvisoryStatus, type AdvisorySuggestedFix, ApiKeys, type AppOracle, type ApplySectionRewritesArgs, type ApplySectionRewritesResult, ArchetypeConvention, type ArchetypeFamilyFit, type ArchetypePerfMap, type ArchetypePerfNMap, type ArchetypePerfScoreResult, type AttachCacheControlResult, BestPracticeAdvisory, type BrainConfig, type BrainDeadLetterEntry, type BrainHealthSnapshot, type BrainQueryConfig, type BrainReadEnv, CallOptions, CallResult, ChainEntry, type CompatibilityIntent, type CompileForAISDKv6Result, type CompileOptions, CompilePolicy, CompileResult, CompiledRequest, DEFAULT_FINDINGS_ENDPOINT, DEFAULT_MEASURED_FAILURE_ENDPOINT, DEFAULT_PROMOTIONS_ENDPOINT, type ExclusionFindingRow, type ExclusionResolutionSource, type ExecuteErr, type ExecuteOk, type ExecuteOptions, type ExecuteResult, type FallbackPosture, FamilyResolutionError, type GetActionableAdvisoriesOptions, type GetApplicablePromotionOpts, type GetDefaultFallbackChainOpts, type GetMeasuredFailureOpts, type GetPerAxisMetricsOpts, type GetRecommendedPrimaryOptions, type GoldenEvalCase, type GoldenEvalOptions, type GoldenEvalRunResult, type GoldenIrRecordInput, Grounding, IntentArchetypeName, JUDGE_RUBRICS, LIBRARY_VERSION, type LLMJudgeOptions, MEASURED_FAILURE_CFG, MEASURED_GROUNDING_MIN_N, type MarkAdvisoryResolvedOptions, type MarkExclusionFindingHandledOptions, type MarkPromoteReadyHandledOptions, type MeasuredFailureRuntime, type MeasuredFailureVerdict, type ModelBrainRow, type ModelCompatibility, ModelProfile, NormalizedResponse, type OracleContext, OracleScore, type OutcomePayload, OutcomeResult, OutputMode, PRODUCER_OWNED_RULE_CODES, PROVIDER_ENV_KEYS, PerAxisMetrics, type PricingRow, type ProbeShadowOptions, type ProbeShadowServed, type ProfileToRowOptions, type PromoteReadyFindingRow, type PromoteReadyResolution, type PromotionRow, type PromotionsRuntime, PromptIR, Provider, ProviderOverrides, type ProviderReachability, RULE_SEQUENTIAL_TOOL_CLIFF, type ReachabilityOpts, RecordInput, RecordOutcomeInput, type RunAdvisorPhase2Context, SectionRewrite, type ShadowProbeRecordInput, type SupportedProvider, type SurfaceFailureRow, SystemModelMessage, TRANSLATOR_FLOOR, _testResetMeasuredFailure, _testResetPromotions, _testWaitForMeasuredFailureRefresh, _testWaitForPromotionsRefresh, applyArchetypeConvention, applySectionRewrites, attachCacheControlToStreamTextInput, brainHealth, buildGoldenIrRow, buildLLMJudge, buildPairwiseJudgePrompt, buildShadowProbeRow, call, clearBrain, combineOrderSwappedVerdicts, compile, compileForAISDKv6, configureBrain, configureMeasuredFailureBrain, configurePromotionsBrain, countTokens, deriveFamilyFromModelId, deriveOwnership, execute, findBetterFit, flushBrainDeadLetter, getActionableAdvisories, getAllStarterChains, getAllStarterChainsWithGrounding, getApplicablePromotion, getArchetypePerfScore, getDefaultFallbackChain, getDefaultFallbackChainWithGrounding, getMeasuredFailureVerdict, getModelCompatibility, getPerAxisMetrics, getReachabilityDiagnostic, getRecommendedPrimary, getSequentialStarterChain, getSequentialStarterChainWithGrounding, getStaleExclusionFindings, getStarterChain, getStarterChainWithGrounding, isAutoPromoteEnabledFromEnv, isBrainQueryActiveFor, isBrainSync, isExclusionFindingsBrainActive, isMeasuredFailureBrainActive, isMeasuredFailureGateEnabledFromEnv, isModelReachable, isPromotionsBrainActive, isProviderReachable, judgeMeasuredFailure, loadAliasesFromBrain, loadArchetypePerfFromBrain, loadArchetypePerfNFromBrain, loadChainsFromBrain, loadModelsFromBrain, loadPricingFromBrain, markAdvisoryResolved, markExclusionFindingHandled, markPromoteReadyHandled, parseJudgeVerdict, peekBrainDeadLetter, probeShadow, profileToRow, readBrainReadEnv, record, recordGoldenIr, recordOutcome, recordShadowProbe, renderIrForJudge, resetTokenizer, resolveConventionsForProfile, resolvePricingAt, resolveProviderKey, rubricFor, runAdvisor, runGoldenEval, setTokenizer, wilsonLowerBound };
+export { ABSOLUTE_FLOOR, type AISDKConvertedMessage, ARCHETYPE_FAMILY_FITS, ARCHETYPE_FLOOR_DEFAULT, type ActionableAdvisory, Adapter, type AdvisoryResolutionSource, type AdvisorySeverity, type AdvisoryStatus, type AdvisorySuggestedFix, ApiKeys, type AppOracle, type ApplySectionRewritesArgs, type ApplySectionRewritesResult, ArchetypeConvention, type ArchetypeFamilyFit, type ArchetypePerfMap, type ArchetypePerfNMap, type ArchetypePerfScoreResult, type AttachCacheControlResult, BestPracticeAdvisory, type BrainConfig, type BrainDeadLetterEntry, type BrainHealthSnapshot, type BrainQueryConfig, type BrainReadEnv, CallOptions, CallResult, ChainEntry, type CompatibilityIntent, type CompileForAISDKv6Result, type CompileOptions, CompilePolicy, CompileResult, CompiledRequest, DEFAULT_FINDINGS_ENDPOINT, DEFAULT_MEASURED_FAILURE_ENDPOINT, DEFAULT_PROMOTIONS_ENDPOINT, type ExclusionFindingRow, type ExclusionResolutionSource, type ExecuteErr, type ExecuteOk, type ExecuteOptions, type ExecuteResult, type FallbackPosture, FamilyResolutionError, type GetActionableAdvisoriesOptions, type GetApplicablePromotionOpts, type GetDefaultFallbackChainOpts, type GetMeasuredFailureOpts, type GetPerAxisMetricsOpts, type GetRecommendedPrimaryOptions, type GoldenEvalCase, type GoldenEvalOptions, type GoldenEvalRunResult, type GoldenIrRecordInput, Grounding, IntentArchetypeName, JUDGE_RUBRICS, LIBRARY_VERSION, type LLMJudgeOptions, MEASURED_FAILURE_CFG, MEASURED_GROUNDING_MIN_N, type MarkAdvisoryResolvedOptions, type MarkExclusionFindingHandledOptions, type MarkPromoteReadyHandledOptions, type MeasuredFailureRuntime, type MeasuredFailureVerdict, type ModelBrainRow, type ModelCompatibility, ModelProfile, NormalizedResponse, type OracleContext, OracleScore, type OutcomePayload, OutcomeResult, OutputMode, PRODUCER_OWNED_RULE_CODES, PROVIDER_ENV_KEYS, PerAxisMetrics, type PricingRow, type ProbeShadowOptions, type ProbeShadowServed, type ProfileToRowOptions, type PromoteReadyFindingRow, type PromoteReadyResolution, type PromotionRow, type PromotionsRuntime, PromptIR, Provider, ProviderOverrides, type ProviderReachability, RULE_SEQUENTIAL_TOOL_CLIFF, type ReachabilityOpts, RecordInput, RecordOutcomeInput, type RunAdvisorPhase2Context, SectionRewrite, type ShadowProbeRecordInput, type SupportedProvider, type SurfaceFailureRow, SystemModelMessage, TRANSLATOR_FLOOR, _testResetMeasuredFailure, _testResetPromotions, _testWaitForMeasuredFailureRefresh, _testWaitForPromotionsRefresh, applyArchetypeConvention, applySectionRewrites, attachCacheControlToStreamTextInput, awaitMeasuredFailureReady, brainHealth, buildGoldenIrRow, buildLLMJudge, buildPairwiseJudgePrompt, buildShadowProbeRow, call, clearBrain, combineOrderSwappedVerdicts, compile, compileForAISDKv6, configureBrain, configureMeasuredFailureBrain, configurePromotionsBrain, countTokens, deriveFamilyFromModelId, deriveOwnership, execute, findBetterFit, flushBrainDeadLetter, getActionableAdvisories, getAllStarterChains, getAllStarterChainsWithGrounding, getApplicablePromotion, getArchetypePerfScore, getDefaultFallbackChain, getDefaultFallbackChainWithGrounding, getMeasuredFailureVerdict, getModelCompatibility, getPerAxisMetrics, getReachabilityDiagnostic, getRecommendedPrimary, getSequentialStarterChain, getSequentialStarterChainWithGrounding, getStaleExclusionFindings, getStarterChain, getStarterChainWithGrounding, isAutoPromoteEnabledFromEnv, isBrainQueryActiveFor, isBrainSync, isExclusionFindingsBrainActive, isMeasuredFailureBrainActive, isMeasuredFailureGateEnabledFromEnv, isModelReachable, isPromotionsBrainActive, isProviderReachable, judgeMeasuredFailure, loadAliasesFromBrain, loadArchetypePerfFromBrain, loadArchetypePerfNFromBrain, loadChainsFromBrain, loadModelsFromBrain, loadPricingFromBrain, markAdvisoryResolved, markExclusionFindingHandled, markPromoteReadyHandled, parseJudgeVerdict, peekBrainDeadLetter, prefetchMeasuredFailure, probeShadow, profileToRow, readBrainReadEnv, record, recordGoldenIr, recordOutcome, recordShadowProbe, renderIrForJudge, resetTokenizer, resolveConventionsForProfile, resolvePricingAt, resolveProviderKey, rubricFor, runAdvisor, runGoldenEval, setTokenizer, wilsonLowerBound };

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,
@@ -2981,20 +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.`
         });
       }
     }
@@ -4734,6 +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;
@@ -5239,6 +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);
   }
@@ -6295,6 +6344,24 @@ function tryParseJson(s) {
     return void 0;
   }
 }
+function isAuthSignatureBody(body, message) {
+  if (body && typeof body === "object") {
+    const err = body.error;
+    if (err && typeof err === "object") {
+      const e = err;
+      if (Array.isArray(e.details)) {
+        for (const d of e.details) {
+          if (d && typeof d === "object" && d.reason === "API_KEY_INVALID") {
+            return true;
+          }
+        }
+      }
+      if (e.code === "invalid_api_key") return true;
+    }
+  }
+  const m = message.toLowerCase();
+  return m.includes("api key not valid") || m.includes("invalid api key") || m.includes("invalid x-api-key") || m.includes("incorrect api key");
+}
 function classifyHttpError(status, body) {
   const message = extractErrorMessage(body) ?? `HTTP ${status}`;
   if (status === 429)
@@ -6307,8 +6374,11 @@ function classifyHttpError(status, body) {
     return { ok: false, status, errorType: "retryable", errorCode: "model_not_found", message, raw: body };
   if (status === 401 || status === 403)
     return { ok: false, status, errorType: "terminal", errorCode: "auth", message, raw: body };
-  if (status === 400)
+  if (status === 400) {
+    if (isAuthSignatureBody(body, message))
+      return { ok: false, status, errorType: "terminal", errorCode: "auth", message, raw: body };
     return { ok: false, status, errorType: "terminal", errorCode: "invalid_request", message, raw: body };
+  }
   return { ok: false, status, errorType: "terminal", errorCode: "unknown", message, raw: body };
 }
 function extractErrorMessage(body) {
@@ -6573,6 +6643,9 @@ function classifyHttpError2(status, body) {
     return { ok: false, status, errorType: "terminal", errorCode: "auth", message, raw: body };
   }
   if (status === 400) {
+    if (isAuthSignatureBody(body, message)) {
+      return { ok: false, status, errorType: "terminal", errorCode: "auth", message, raw: body };
+    }
     return { ok: false, status, errorType: "terminal", errorCode: "invalid_request", message, raw: body };
   }
   return { ok: false, status, errorType: "terminal", errorCode: "unknown", message, raw: body };
@@ -7258,6 +7331,14 @@ 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(
@@ -7273,6 +7354,7 @@ async function call(ir, opts = {}) {
       )
     })
   );
+  await awaitMeasuredFailureReady(ir.appId, resolveGateWarmupMs(opts));
   const initial = compileAndRegister(ir, opts);
   safeEmit(
     () => emitCompileDone(traceId, ir.appId, {
@@ -8741,7 +8823,7 @@ function createBrainForwardRoutes(config) {
 }
 
 // src/version.ts
-var LIBRARY_VERSION = "2.0.0-alpha.72";
+var LIBRARY_VERSION = "2.0.0-alpha.74";
 
 // src/key-health.ts
 var JSON_HEADERS2 = { "Content-Type": "application/json" };
@@ -9396,6 +9478,7 @@ function compile2(ir, opts) {
   applyArchetypeConvention,
   applySectionRewrites,
   attachCacheControlToStreamTextInput,
+  awaitMeasuredFailureReady,
   brainHealth,
   bucketContext,
   bucketHistory,
@@ -9465,6 +9548,7 @@ 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-AI4NNR5T.mjs";
+} from "./chunk-EB2YR7HL.mjs";
 import {
   ABSOLUTE_FLOOR,
   ARCHETYPE_FLOOR_DEFAULT,
@@ -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.`
         });
       }
     }
@@ -2630,6 +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;
@@ -3135,6 +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);
   }
@@ -4106,6 +4153,24 @@ function tryParseJson(s) {
     return void 0;
   }
 }
+function isAuthSignatureBody(body, message) {
+  if (body && typeof body === "object") {
+    const err = body.error;
+    if (err && typeof err === "object") {
+      const e = err;
+      if (Array.isArray(e.details)) {
+        for (const d of e.details) {
+          if (d && typeof d === "object" && d.reason === "API_KEY_INVALID") {
+            return true;
+          }
+        }
+      }
+      if (e.code === "invalid_api_key") return true;
+    }
+  }
+  const m = message.toLowerCase();
+  return m.includes("api key not valid") || m.includes("invalid api key") || m.includes("invalid x-api-key") || m.includes("incorrect api key");
+}
 function classifyHttpError(status, body) {
   const message = extractErrorMessage(body) ?? `HTTP ${status}`;
   if (status === 429)
@@ -4118,8 +4183,11 @@ function classifyHttpError(status, body) {
     return { ok: false, status, errorType: "retryable", errorCode: "model_not_found", message, raw: body };
   if (status === 401 || status === 403)
     return { ok: false, status, errorType: "terminal", errorCode: "auth", message, raw: body };
-  if (status === 400)
+  if (status === 400) {
+    if (isAuthSignatureBody(body, message))
+      return { ok: false, status, errorType: "terminal", errorCode: "auth", message, raw: body };
     return { ok: false, status, errorType: "terminal", errorCode: "invalid_request", message, raw: body };
+  }
   return { ok: false, status, errorType: "terminal", errorCode: "unknown", message, raw: body };
 }
 function extractErrorMessage(body) {
@@ -4384,6 +4452,9 @@ function classifyHttpError2(status, body) {
     return { ok: false, status, errorType: "terminal", errorCode: "auth", message, raw: body };
   }
   if (status === 400) {
+    if (isAuthSignatureBody(body, message)) {
+      return { ok: false, status, errorType: "terminal", errorCode: "auth", message, raw: body };
+    }
     return { ok: false, status, errorType: "terminal", errorCode: "invalid_request", message, raw: body };
   }
   return { ok: false, status, errorType: "terminal", errorCode: "unknown", message, raw: body };
@@ -4415,6 +4486,14 @@ 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(
@@ -4430,6 +4509,7 @@ async function call(ir, opts = {}) {
       )
     })
   );
+  await awaitMeasuredFailureReady(ir.appId, resolveGateWarmupMs(opts));
   const initial = compileAndRegister(ir, opts);
   safeEmit(
     () => emitCompileDone(traceId, ir.appId, {
@@ -6187,6 +6267,7 @@ export {
   applyArchetypeConvention,
   applySectionRewrites,
   attachCacheControlToStreamTextInput,
+  awaitMeasuredFailureReady,
   brainHealth,
   bucketContext,
   bucketHistory,
@@ -6256,6 +6337,7 @@ 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.72";
+var LIBRARY_VERSION = "2.0.0-alpha.74";
 
 // src/key-health.ts
 var JSON_HEADERS = { "Content-Type": "application/json" };

package/dist/key-health.mjs

@@ -1,6 +1,6 @@
 import {
   createKeyHealthRoute
-} from "./chunk-AI4NNR5T.mjs";
+} from "./chunk-EB2YR7HL.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';
 
 /**

package/dist/profiles.d.ts

@@ -1,4 +1,4 @@
-import { k as Provider } from './ir-BEQ28muo.js';
+import { k as Provider } from './ir-B0v2f9NY.js';
 import { IntentArchetypeName } from './dialect.js';
 
 /**

package/dist/{types-DeGRCTlJ.d.mts → types-BOTLfJOp.d.mts}

@@ -1,4 +1,4 @@
-import { i as Adapter, x as SectionKind } from './ir-CnnJST_N.mjs';
+import { i as Adapter, x as SectionKind } from './ir-CZukZvDn.mjs';
 
 /**
  * Internal config + hook types for createGlassboxRoutes().

package/dist/{types-DSeJJ6tt.d.ts → types-Ba68lIs1.d.ts}

@@ -1,4 +1,4 @@
-import { i as Adapter, x as SectionKind } from './ir-BEQ28muo.js';
+import { i as Adapter, x as SectionKind } from './ir-B0v2f9NY.js';
 
 /**
  * Internal config + hook types for createGlassboxRoutes().

package/dist/{types-BDFrJkma.d.mts → types-BgJ5iQB_.d.mts}

@@ -1,4 +1,4 @@
-import { s as MutationApplied, B as BestPracticeAdvisory, F as FallbackReason, m as CallAttempt } from './ir-CnnJST_N.mjs';
+import { s as MutationApplied, B as BestPracticeAdvisory, F as FallbackReason, m as CallAttempt } from './ir-CZukZvDn.mjs';
 
 /**
  * Glass-Box observability types (alpha.17).

package/dist/{types-B--CYzMo.d.ts → types-D-PuKCWk.d.ts}

@@ -1,4 +1,4 @@
-import { s as MutationApplied, B as BestPracticeAdvisory, F as FallbackReason, m as CallAttempt } from './ir-BEQ28muo.js';
+import { s as MutationApplied, B as BestPracticeAdvisory, F as FallbackReason, m as CallAttempt } from './ir-B0v2f9NY.js';
 
 /**
  * Glass-Box observability types (alpha.17).

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@warmdrift/kgauto-compiler",
-  "version": "2.0.0-alpha.72",
+  "version": "2.0.0-alpha.74",
   "description": "Prompt compiler with executable provider knowledge for multi-model AI apps: normalized multi-provider transport with fallback chains, compile-time cliff guards, a curated model registry, and a telemetry flight recorder. Swap models without rewriting prompts.",
   "main": "./dist/index.js",
   "module": "./dist/index.mjs",