@warmdrift/kgauto-compiler 2.0.0-alpha.52 → 2.0.0-alpha.53

package/dist/{chunk-HHWBB46W.mjs → chunk-IIMPJZNH.mjs}

@@ -1,6 +1,6 @@
 import {
   tryGetProfile
-} from "./chunk-CXH7KC4D.mjs";
+} from "./chunk-QKXTMVCT.mjs";
 
 // src/brain-query.ts
 var FRESH_SNAPSHOT = {

package/dist/{chunk-CXH7KC4D.mjs → chunk-QKXTMVCT.mjs}

@@ -1356,6 +1356,63 @@ var PROFILES_RAW = [
       critique: 5
       // +1 vs 2.5-flash — but still below Sonnet/Opus reasoning floor
     }
+  },
+  {
+    // Auto-onboarded 2026-07-01 from `claude-sonnet-4-6`; VERIFIED 2026-07-02
+    // against the claude-api reference (cc-portfolio ratification pass). The
+    // clone got context right (1M) and pricing right at sticker ($3/$15 —
+    // NOTE an introductory $2/$10 per MTok runs through 2026-08-31; sticker
+    // encoded here per the time-bounded-pricing convention, intro belongs in
+    // brain kgauto_pricing if worth capturing). The clone got max output WRONG:
+    // Sonnet 5 is 128k, not 4-6's 64k — corrected. New tokenizer (~30% more
+    // tokens for the same text vs 4-6): byte-budget consumers should re-baseline.
+    // API quirks (claude-api ref): (a) NON-DEFAULT temperature/top_p/top_k
+    // return 400 — moot for kgauto's own call() path (ANTHROPIC_LOWERING_BASE
+    // emits no sampling params) but a REAL hazard for compileForAISDKv6
+    // consumers that pass temperature themselves (tt-intel scoring uses temp:0
+    // for determinism — that 400s on this model; noted to consumers via the
+    // contract). (b) Omitting `thinking` runs ADAPTIVE by default (4-6 ran
+    // thinking-off) — output spend shifts. (c) Supports effort xhigh. status:
+    // 'preview' per the Fable precedent — no brain evidence yet; promotion to
+    // 'current' is an explicit call. (L-049/L-081.)
+    id: "claude-sonnet-5",
+    verifiedAgainstDocs: "2026-07-02",
+    provider: "anthropic",
+    status: "preview",
+    maxContextTokens: 1e6,
+    maxOutputTokens: 128e3,
+    maxTools: 64,
+    parallelToolCalls: true,
+    structuredOutput: "grammar",
+    systemPromptMode: "inline",
+    streaming: true,
+    cliffs: [],
+    costInputPer1m: 3,
+    costOutputPer1m: 15,
+    lowering: ANTHROPIC_LOWERING_BASE,
+    recovery: [
+      { signal: "rate_limit", action: "escalate", reason: "429 \u2014 escalate" },
+      { signal: "model_not_found", action: "escalate", reason: "Deprecated \u2014 escalate (L-061)" }
+    ],
+    strengths: ["quality", "tool_use", "long_context", "cache_friendly", "extended_thinking"],
+    weaknesses: [],
+    notes: "Sonnet 5 (2026-06): near-Opus quality on coding/agentic work at Sonnet cost. Verified 2026-07-02 against the claude-api reference: 1M ctx, 128k out (clone's 64k corrected), $3/$15 sticker (intro $2/$10 through 2026-08-31). New tokenizer ~30% more tokens vs sonnet-4-6. Consumer hazards: non-default temperature/top_p/top_k 400 (temp:0 rejected \u2014 deterministic-scoring consumers must omit); thinking defaults to adaptive when omitted. status:preview \u2014 no brain evidence yet; earns placement via the machinery.",
+    // Master plan §6.2 anchor. Tier 0 for plan/generate/ask/extract/transform
+    // in starter chains; tier 1 cross-provider for hunt/summarize/classify.
+    archetypePerf: {
+      ask: 9,
+      generate: 9,
+      plan: 9,
+      critique: 9,
+      extract: 9,
+      transform: 9,
+      hunt: 7,
+      // strong but Flash beats on parallel tool throughput
+      summarize: 8,
+      // overkill for tolerant archetype
+      classify: 8
+      // overkill
+    }
   }
 ];
 var ALIASES = {

package/dist/glassbox/index.d.mts

@@ -1,6 +1,6 @@
-import { G as GlassboxEvent } from '../types-vGo-h0tZ.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-vGo-h0tZ.mjs';
-import '../ir-dDcG8Pvu.mjs';
+import { G as GlassboxEvent } from '../types-C_dkNMA_.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-C_dkNMA_.mjs';
+import '../ir-CAlLBu5d.mjs';
 import '../dialect.mjs';
 
 /**

package/dist/glassbox/index.d.ts

@@ -1,6 +1,6 @@
-import { G as GlassboxEvent } from '../types-UXPxWabQ.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-UXPxWabQ.js';
-import '../ir-rUUojj0s.js';
+import { G as GlassboxEvent } from '../types-hEDGehpz.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-hEDGehpz.js';
+import '../ir-BiXAMyji.js';
 import '../dialect.js';
 
 /**

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

@@ -1,5 +1,5 @@
-import { T as TraceHealth } from '../types-BTeRoSvM.mjs';
-import '../ir-dDcG8Pvu.mjs';
+import { T as TraceHealth } from '../types-Dj5FLhBV.mjs';
+import '../ir-CAlLBu5d.mjs';
 import '../dialect.mjs';
 
 /**

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

@@ -1,5 +1,5 @@
-import { T as TraceHealth } from '../types-B_pdPjxm.js';
-import '../ir-rUUojj0s.js';
+import { T as TraceHealth } from '../types-DsA8JFOp.js';
+import '../ir-BiXAMyji.js';
 import '../dialect.js';
 
 /**

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

@@ -1,7 +1,7 @@
-import { G as GlassboxEvent } from '../types-vGo-h0tZ.mjs';
-import { a as TraceDetail, b as TraceSummary, c as TraceCounterfactual } from '../types-BTeRoSvM.mjs';
-export { A as AdvisoryRecord, T as TraceHealth, d as TraceSectionRewrite } from '../types-BTeRoSvM.mjs';
-import '../ir-dDcG8Pvu.mjs';
+import { G as GlassboxEvent } from '../types-C_dkNMA_.mjs';
+import { a as TraceDetail, b as TraceSummary, c as TraceCounterfactual } from '../types-Dj5FLhBV.mjs';
+export { A as AdvisoryRecord, T as TraceHealth, d as TraceSectionRewrite } from '../types-Dj5FLhBV.mjs';
+import '../ir-CAlLBu5d.mjs';
 import '../dialect.mjs';
 
 /**

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

@@ -1,7 +1,7 @@
-import { G as GlassboxEvent } from '../types-UXPxWabQ.js';
-import { a as TraceDetail, b as TraceSummary, c as TraceCounterfactual } from '../types-B_pdPjxm.js';
-export { A as AdvisoryRecord, T as TraceHealth, d as TraceSectionRewrite } from '../types-B_pdPjxm.js';
-import '../ir-rUUojj0s.js';
+import { G as GlassboxEvent } from '../types-hEDGehpz.js';
+import { a as TraceDetail, b as TraceSummary, c as TraceCounterfactual } from '../types-DsA8JFOp.js';
+export { A as AdvisoryRecord, T as TraceHealth, d as TraceSectionRewrite } from '../types-DsA8JFOp.js';
+import '../ir-BiXAMyji.js';
 import '../dialect.js';
 
 /**

package/dist/glassbox-routes/index.js

@@ -1410,6 +1410,63 @@ var PROFILES_RAW = [
       critique: 5
       // +1 vs 2.5-flash — but still below Sonnet/Opus reasoning floor
     }
+  },
+  {
+    // Auto-onboarded 2026-07-01 from `claude-sonnet-4-6`; VERIFIED 2026-07-02
+    // against the claude-api reference (cc-portfolio ratification pass). The
+    // clone got context right (1M) and pricing right at sticker ($3/$15 —
+    // NOTE an introductory $2/$10 per MTok runs through 2026-08-31; sticker
+    // encoded here per the time-bounded-pricing convention, intro belongs in
+    // brain kgauto_pricing if worth capturing). The clone got max output WRONG:
+    // Sonnet 5 is 128k, not 4-6's 64k — corrected. New tokenizer (~30% more
+    // tokens for the same text vs 4-6): byte-budget consumers should re-baseline.
+    // API quirks (claude-api ref): (a) NON-DEFAULT temperature/top_p/top_k
+    // return 400 — moot for kgauto's own call() path (ANTHROPIC_LOWERING_BASE
+    // emits no sampling params) but a REAL hazard for compileForAISDKv6
+    // consumers that pass temperature themselves (tt-intel scoring uses temp:0
+    // for determinism — that 400s on this model; noted to consumers via the
+    // contract). (b) Omitting `thinking` runs ADAPTIVE by default (4-6 ran
+    // thinking-off) — output spend shifts. (c) Supports effort xhigh. status:
+    // 'preview' per the Fable precedent — no brain evidence yet; promotion to
+    // 'current' is an explicit call. (L-049/L-081.)
+    id: "claude-sonnet-5",
+    verifiedAgainstDocs: "2026-07-02",
+    provider: "anthropic",
+    status: "preview",
+    maxContextTokens: 1e6,
+    maxOutputTokens: 128e3,
+    maxTools: 64,
+    parallelToolCalls: true,
+    structuredOutput: "grammar",
+    systemPromptMode: "inline",
+    streaming: true,
+    cliffs: [],
+    costInputPer1m: 3,
+    costOutputPer1m: 15,
+    lowering: ANTHROPIC_LOWERING_BASE,
+    recovery: [
+      { signal: "rate_limit", action: "escalate", reason: "429 \u2014 escalate" },
+      { signal: "model_not_found", action: "escalate", reason: "Deprecated \u2014 escalate (L-061)" }
+    ],
+    strengths: ["quality", "tool_use", "long_context", "cache_friendly", "extended_thinking"],
+    weaknesses: [],
+    notes: "Sonnet 5 (2026-06): near-Opus quality on coding/agentic work at Sonnet cost. Verified 2026-07-02 against the claude-api reference: 1M ctx, 128k out (clone's 64k corrected), $3/$15 sticker (intro $2/$10 through 2026-08-31). New tokenizer ~30% more tokens vs sonnet-4-6. Consumer hazards: non-default temperature/top_p/top_k 400 (temp:0 rejected \u2014 deterministic-scoring consumers must omit); thinking defaults to adaptive when omitted. status:preview \u2014 no brain evidence yet; earns placement via the machinery.",
+    // Master plan §6.2 anchor. Tier 0 for plan/generate/ask/extract/transform
+    // in starter chains; tier 1 cross-provider for hunt/summarize/classify.
+    archetypePerf: {
+      ask: 9,
+      generate: 9,
+      plan: 9,
+      critique: 9,
+      extract: 9,
+      transform: 9,
+      hunt: 7,
+      // strong but Flash beats on parallel tool throughput
+      summarize: 8,
+      // overkill for tolerant archetype
+      classify: 8
+      // overkill
+    }
   }
 ];
 var ALIASES = {

package/dist/glassbox-routes/index.mjs

@@ -1,10 +1,10 @@
 import {
   ARCHETYPE_FLOOR_DEFAULT,
   getDefaultFallbackChain
-} from "../chunk-HHWBB46W.mjs";
+} from "../chunk-IIMPJZNH.mjs";
 import {
   tryGetProfile
-} from "../chunk-CXH7KC4D.mjs";
+} from "../chunk-QKXTMVCT.mjs";
 import {
   subscribe,
   subscribeApp

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-BTeRoSvM.mjs';
-import '../../ir-dDcG8Pvu.mjs';
+import { a as TraceDetail } from '../../types-Dj5FLhBV.mjs';
+import '../../ir-CAlLBu5d.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-B_pdPjxm.js';
-import '../../ir-rUUojj0s.js';
+import { a as TraceDetail } from '../../types-DsA8JFOp.js';
+import '../../ir-BiXAMyji.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-dDcG8Pvu.mjs';
-export { m as CallAttempt, n as CallError, o as ChainModelEntry, p as ChainWithGrounding, q as Constraints, F as FallbackReason, H as HistoryCachePolicy, I as IntentDeclaration, M as Message, r as MutationApplied, s as NormalizedTokens, t as OutcomeKind, u as PerAxisMetricsByModel, v as PromptSection, w as SectionKind, x as ShadowProbeConfig, T as ToolCall, y as ToolDefinition } from './ir-dDcG8Pvu.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-CAlLBu5d.mjs';
+export { m as CallAttempt, n as CallError, o as ChainModelEntry, p as ChainWithGrounding, q as Constraints, F as FallbackReason, H as HistoryCachePolicy, I as IntentDeclaration, M as Message, r as MutationApplied, s as NormalizedTokens, t as OutcomeKind, u as PerAxisMetricsByModel, v as PromptSection, w as SectionKind, x as ShadowProbeConfig, T as ToolCall, y as ToolDefinition } from './ir-CAlLBu5d.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';
 import { IntentArchetypeName } from './dialect.mjs';
@@ -536,13 +536,26 @@ interface ShadowProbeRecordInput {
     /** Wall-clock latency of the candidate replay call, ms (alpha.46). */
     latencyCandidateMs?: number;
     /**
-     * alpha — probe-row class (migration 029). Defaults to `'completed'` (the
-     * pre-029 row shape: candidate ran, columns populated). `'aborted_latency_budget'`
-     * and `'skipped_slow_tier_sync'` are DIAGNOSTIC rows from sync-mode latency
-     * protection — no candidate response, no verdict, no false quality signal. A
-     * verdict rollup MUST filter to `outcome='completed'`.
+     * alpha — probe-row class (migration 029; `candidate_error` added by 032).
+     * Defaults to `'completed'` (the pre-029 row shape: candidate ran, columns
+     * populated). `'aborted_latency_budget'` and `'skipped_slow_tier_sync'` are
+     * DIAGNOSTIC rows from sync-mode latency protection; `'candidate_error'`
+     * (alpha.53) is the failed-candidate diagnostic — the candidate call
+     * errored, with the error class + latency-at-death recorded instead of
+     * silence (the IC demo-night probe deaths were invisible for ~90 min under
+     * the old record-no-row rule). Diagnostic rows carry no candidate response
+     * and no verdict — no false quality signal. A verdict rollup MUST filter to
+     * `outcome='completed'`.
      */
-    outcome?: 'completed' | 'aborted_latency_budget' | 'skipped_slow_tier_sync';
+    outcome?: 'completed' | 'aborted_latency_budget' | 'skipped_slow_tier_sync' | 'candidate_error';
+    /**
+     * alpha.53 (migration 032) — error class for `outcome='candidate_error'`
+     * rows. Values follow execute()'s taxonomy (`rate_limit`, `timeout`,
+     * `server_error`, `model_not_found`, `auth`, `invalid_request`, `unknown`)
+     * plus `execute_rejected` (the execute promise rejected outright). NULL /
+     * omitted on all other row classes.
+     */
+    errorClass?: string;
 }
 /**
  * Pure builder for the `probe_outcomes` row written by the inline shadow-probe.
@@ -574,7 +587,8 @@ declare function buildShadowProbeRow(input: ShadowProbeRecordInput): {
     latency_candidate_ms: number | null;
     prompt_fidelity: number;
     replay_source: 'inline-full-ir';
-    outcome: 'completed' | 'aborted_latency_budget' | 'skipped_slow_tier_sync';
+    outcome: 'completed' | 'aborted_latency_budget' | 'skipped_slow_tier_sync' | 'candidate_error';
+    error_class: string | null;
 };
 /**
  * Persist one inline shadow-probe row to `probe_outcomes`. Fire-and-forget by

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-rUUojj0s.js';
-export { m as CallAttempt, n as CallError, o as ChainModelEntry, p as ChainWithGrounding, q as Constraints, F as FallbackReason, H as HistoryCachePolicy, I as IntentDeclaration, M as Message, r as MutationApplied, s as NormalizedTokens, t as OutcomeKind, u as PerAxisMetricsByModel, v as PromptSection, w as SectionKind, x as ShadowProbeConfig, T as ToolCall, y as ToolDefinition } from './ir-rUUojj0s.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-BiXAMyji.js';
+export { m as CallAttempt, n as CallError, o as ChainModelEntry, p as ChainWithGrounding, q as Constraints, F as FallbackReason, H as HistoryCachePolicy, I as IntentDeclaration, M as Message, r as MutationApplied, s as NormalizedTokens, t as OutcomeKind, u as PerAxisMetricsByModel, v as PromptSection, w as SectionKind, x as ShadowProbeConfig, T as ToolCall, y as ToolDefinition } from './ir-BiXAMyji.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';
 import { IntentArchetypeName } from './dialect.js';
@@ -536,13 +536,26 @@ interface ShadowProbeRecordInput {
     /** Wall-clock latency of the candidate replay call, ms (alpha.46). */
     latencyCandidateMs?: number;
     /**
-     * alpha — probe-row class (migration 029). Defaults to `'completed'` (the
-     * pre-029 row shape: candidate ran, columns populated). `'aborted_latency_budget'`
-     * and `'skipped_slow_tier_sync'` are DIAGNOSTIC rows from sync-mode latency
-     * protection — no candidate response, no verdict, no false quality signal. A
-     * verdict rollup MUST filter to `outcome='completed'`.
+     * alpha — probe-row class (migration 029; `candidate_error` added by 032).
+     * Defaults to `'completed'` (the pre-029 row shape: candidate ran, columns
+     * populated). `'aborted_latency_budget'` and `'skipped_slow_tier_sync'` are
+     * DIAGNOSTIC rows from sync-mode latency protection; `'candidate_error'`
+     * (alpha.53) is the failed-candidate diagnostic — the candidate call
+     * errored, with the error class + latency-at-death recorded instead of
+     * silence (the IC demo-night probe deaths were invisible for ~90 min under
+     * the old record-no-row rule). Diagnostic rows carry no candidate response
+     * and no verdict — no false quality signal. A verdict rollup MUST filter to
+     * `outcome='completed'`.
      */
-    outcome?: 'completed' | 'aborted_latency_budget' | 'skipped_slow_tier_sync';
+    outcome?: 'completed' | 'aborted_latency_budget' | 'skipped_slow_tier_sync' | 'candidate_error';
+    /**
+     * alpha.53 (migration 032) — error class for `outcome='candidate_error'`
+     * rows. Values follow execute()'s taxonomy (`rate_limit`, `timeout`,
+     * `server_error`, `model_not_found`, `auth`, `invalid_request`, `unknown`)
+     * plus `execute_rejected` (the execute promise rejected outright). NULL /
+     * omitted on all other row classes.
+     */
+    errorClass?: string;
 }
 /**
  * Pure builder for the `probe_outcomes` row written by the inline shadow-probe.
@@ -574,7 +587,8 @@ declare function buildShadowProbeRow(input: ShadowProbeRecordInput): {
     latency_candidate_ms: number | null;
     prompt_fidelity: number;
     replay_source: 'inline-full-ir';
-    outcome: 'completed' | 'aborted_latency_budget' | 'skipped_slow_tier_sync';
+    outcome: 'completed' | 'aborted_latency_budget' | 'skipped_slow_tier_sync' | 'candidate_error';
+    error_class: string | null;
 };
 /**
  * Persist one inline shadow-probe row to `probe_outcomes`. Fire-and-forget by

package/dist/index.js

@@ -1635,6 +1635,63 @@ var PROFILES_RAW = [
       critique: 5
       // +1 vs 2.5-flash — but still below Sonnet/Opus reasoning floor
     }
+  },
+  {
+    // Auto-onboarded 2026-07-01 from `claude-sonnet-4-6`; VERIFIED 2026-07-02
+    // against the claude-api reference (cc-portfolio ratification pass). The
+    // clone got context right (1M) and pricing right at sticker ($3/$15 —
+    // NOTE an introductory $2/$10 per MTok runs through 2026-08-31; sticker
+    // encoded here per the time-bounded-pricing convention, intro belongs in
+    // brain kgauto_pricing if worth capturing). The clone got max output WRONG:
+    // Sonnet 5 is 128k, not 4-6's 64k — corrected. New tokenizer (~30% more
+    // tokens for the same text vs 4-6): byte-budget consumers should re-baseline.
+    // API quirks (claude-api ref): (a) NON-DEFAULT temperature/top_p/top_k
+    // return 400 — moot for kgauto's own call() path (ANTHROPIC_LOWERING_BASE
+    // emits no sampling params) but a REAL hazard for compileForAISDKv6
+    // consumers that pass temperature themselves (tt-intel scoring uses temp:0
+    // for determinism — that 400s on this model; noted to consumers via the
+    // contract). (b) Omitting `thinking` runs ADAPTIVE by default (4-6 ran
+    // thinking-off) — output spend shifts. (c) Supports effort xhigh. status:
+    // 'preview' per the Fable precedent — no brain evidence yet; promotion to
+    // 'current' is an explicit call. (L-049/L-081.)
+    id: "claude-sonnet-5",
+    verifiedAgainstDocs: "2026-07-02",
+    provider: "anthropic",
+    status: "preview",
+    maxContextTokens: 1e6,
+    maxOutputTokens: 128e3,
+    maxTools: 64,
+    parallelToolCalls: true,
+    structuredOutput: "grammar",
+    systemPromptMode: "inline",
+    streaming: true,
+    cliffs: [],
+    costInputPer1m: 3,
+    costOutputPer1m: 15,
+    lowering: ANTHROPIC_LOWERING_BASE,
+    recovery: [
+      { signal: "rate_limit", action: "escalate", reason: "429 \u2014 escalate" },
+      { signal: "model_not_found", action: "escalate", reason: "Deprecated \u2014 escalate (L-061)" }
+    ],
+    strengths: ["quality", "tool_use", "long_context", "cache_friendly", "extended_thinking"],
+    weaknesses: [],
+    notes: "Sonnet 5 (2026-06): near-Opus quality on coding/agentic work at Sonnet cost. Verified 2026-07-02 against the claude-api reference: 1M ctx, 128k out (clone's 64k corrected), $3/$15 sticker (intro $2/$10 through 2026-08-31). New tokenizer ~30% more tokens vs sonnet-4-6. Consumer hazards: non-default temperature/top_p/top_k 400 (temp:0 rejected \u2014 deterministic-scoring consumers must omit); thinking defaults to adaptive when omitted. status:preview \u2014 no brain evidence yet; earns placement via the machinery.",
+    // Master plan §6.2 anchor. Tier 0 for plan/generate/ask/extract/transform
+    // in starter chains; tier 1 cross-provider for hunt/summarize/classify.
+    archetypePerf: {
+      ask: 9,
+      generate: 9,
+      plan: 9,
+      critique: 9,
+      extract: 9,
+      transform: 9,
+      hunt: 7,
+      // strong but Flash beats on parallel tool throughput
+      summarize: 8,
+      // overkill for tolerant archetype
+      classify: 8
+      // overkill
+    }
   }
 ];
 var ALIASES = {
@@ -4525,7 +4582,7 @@ async function record(input) {
       });
       if (!res.ok) {
         const text = await res.text().catch(() => "<no body>");
-        throw new Error(`brain ${res.status}: ${text}`);
+        throw new Error(describeBrainWriteFailure(res.status, "outcomes", text));
       }
       outcomeId = await tryExtractOutcomeId(res);
     } catch (err) {
@@ -4563,6 +4620,11 @@ async function record(input) {
 function defaultOnError5(err) {
   console.warn("[kgauto] brain record failed:", err);
 }
+function describeBrainWriteFailure(status, route, body) {
+  const base = `brain ${route} ${status}: ${body}`;
+  if (status !== 404 && status !== 405) return base;
+  return `${base} \u2014 likely cause: the brain proxy at your KGAUTO_V2_BRAIN_URL is missing the /${route} forward route. Self-proxying consumers must implement ALL THREE sibling routes (/outcomes, /probe_outcomes, /compile_outcome_quality) \u2014 see interfaces/kgauto.md "Consumer adoption checklist" route-completeness (L-142).`;
+}
 function buildPayload(input, reg) {
   const compileTarget = reg?.model;
   const actual = input.actualModel ?? compileTarget;
@@ -4691,7 +4753,9 @@ async function recordOutcome(input) {
       });
       if (!res.ok) {
         const text = await res.text().catch(() => "<no body>");
-        const err = new Error(`brain ${res.status}: ${text}`);
+        const err = new Error(
+          describeBrainWriteFailure(res.status, "compile_outcome_quality", text)
+        );
         (config.onError ?? defaultOnError5)(err);
         return { ok: false, reason: "persistence_failed" };
       }
@@ -4733,8 +4797,12 @@ function buildShadowProbeRow(input) {
     prompt_fidelity: 1,
     replay_source: "inline-full-ir",
     // alpha — migration 029. 'completed' is the default completed-probe shape;
-    // diagnostic rows (aborted/skipped) pass the explicit class.
-    outcome: input.outcome ?? "completed"
+    // diagnostic rows (aborted/skipped/errored) pass the explicit class.
+    outcome: input.outcome ?? "completed",
+    // alpha.53 — migration 032. Populated only on candidate_error rows.
+    // Self-proxying consumers must whitelist `error_class` (and `outcome`,
+    // per the alpha.51 PB finding) or diagnostic rows degrade silently.
+    error_class: input.errorClass ?? null
   };
 }
 async function recordShadowProbe(input) {
@@ -4755,7 +4823,7 @@ async function recordShadowProbe(input) {
       });
       if (!res.ok) {
         const text = await res.text().catch(() => "<no body>");
-        throw new Error(`brain probe_outcomes ${res.status}: ${text}`);
+        throw new Error(describeBrainWriteFailure(res.status, "probe_outcomes", text));
       }
     } catch (err) {
       (config.onError ?? defaultOnError5)(err);
@@ -6360,7 +6428,7 @@ async function runProbeCandidates(args) {
     const budgetMs = args.cfg.maxLatencyMs ?? DEFAULT_PROBE_MAX_LATENCY_MS;
     const deadline = args.sync ? Date.now() + budgetMs : Number.POSITIVE_INFINITY;
     const skipSlow = args.sync && (args.cfg.skipSlowTierInSync ?? true);
-    const recordDiagnostic = async (candidate, outcome) => {
+    const recordDiagnostic = async (candidate, outcome, detail) => {
       await recordShadowProbe({
         appId: args.ir.appId,
         archetype: args.ir.intent.archetype,
@@ -6373,7 +6441,9 @@ async function runProbeCandidates(args) {
         tokensCurrentIn: args.servedResponse.tokens.input,
         tokensCurrentOut: args.servedResponse.tokens.output,
         latencyCurrentMs: args.servedLatencyMs,
-        outcome
+        outcome,
+        ...detail?.errorClass ? { errorClass: detail.errorClass } : {},
+        ...detail?.latencyCandidateMs !== void 0 ? { latencyCandidateMs: detail.latencyCandidateMs } : {}
       });
     };
     for (const candidate of candidates) {
@@ -6428,10 +6498,28 @@ async function runProbeCandidates(args) {
           }
           break;
         }
-        if (raced.kind === "failed") continue;
+        if (raced.kind === "failed") {
+          try {
+            await recordDiagnostic(candidate, "candidate_error", {
+              errorClass: "execute_rejected",
+              latencyCandidateMs: Date.now() - candStart
+            });
+          } catch {
+          }
+          continue;
+        }
         const exec = raced.exec;
         const candidateLatencyMs = Date.now() - candStart;
-        if (!exec.ok) continue;
+        if (!exec.ok) {
+          try {
+            await recordDiagnostic(candidate, "candidate_error", {
+              errorClass: exec.errorCode,
+              latencyCandidateMs: candidateLatencyMs
+            });
+          } catch {
+          }
+          continue;
+        }
         await recordShadowProbe({
           appId: args.ir.appId,
           archetype: args.ir.intent.archetype,

package/dist/index.mjs

@@ -33,7 +33,7 @@ import {
   loadChainsFromBrain,
   readBrainReadEnv,
   resolveProviderKey
-} from "./chunk-HHWBB46W.mjs";
+} from "./chunk-IIMPJZNH.mjs";
 import {
   ALIASES,
   LATENCY_TIER_MS,
@@ -44,7 +44,7 @@ import {
   latencyTierOf,
   profilesByProvider,
   tryGetProfile
-} from "./chunk-CXH7KC4D.mjs";
+} from "./chunk-QKXTMVCT.mjs";
 import {
   emitAdvisoryFired,
   emitCompileDone,
@@ -2736,7 +2736,7 @@ async function record(input) {
       });
       if (!res.ok) {
         const text = await res.text().catch(() => "<no body>");
-        throw new Error(`brain ${res.status}: ${text}`);
+        throw new Error(describeBrainWriteFailure(res.status, "outcomes", text));
       }
       outcomeId = await tryExtractOutcomeId(res);
     } catch (err) {
@@ -2774,6 +2774,11 @@ async function record(input) {
 function defaultOnError4(err) {
   console.warn("[kgauto] brain record failed:", err);
 }
+function describeBrainWriteFailure(status, route, body) {
+  const base = `brain ${route} ${status}: ${body}`;
+  if (status !== 404 && status !== 405) return base;
+  return `${base} \u2014 likely cause: the brain proxy at your KGAUTO_V2_BRAIN_URL is missing the /${route} forward route. Self-proxying consumers must implement ALL THREE sibling routes (/outcomes, /probe_outcomes, /compile_outcome_quality) \u2014 see interfaces/kgauto.md "Consumer adoption checklist" route-completeness (L-142).`;
+}
 function buildPayload(input, reg) {
   const compileTarget = reg?.model;
   const actual = input.actualModel ?? compileTarget;
@@ -2902,7 +2907,9 @@ async function recordOutcome(input) {
       });
       if (!res.ok) {
         const text = await res.text().catch(() => "<no body>");
-        const err = new Error(`brain ${res.status}: ${text}`);
+        const err = new Error(
+          describeBrainWriteFailure(res.status, "compile_outcome_quality", text)
+        );
         (config.onError ?? defaultOnError4)(err);
         return { ok: false, reason: "persistence_failed" };
       }
@@ -2944,8 +2951,12 @@ function buildShadowProbeRow(input) {
     prompt_fidelity: 1,
     replay_source: "inline-full-ir",
     // alpha — migration 029. 'completed' is the default completed-probe shape;
-    // diagnostic rows (aborted/skipped) pass the explicit class.
-    outcome: input.outcome ?? "completed"
+    // diagnostic rows (aborted/skipped/errored) pass the explicit class.
+    outcome: input.outcome ?? "completed",
+    // alpha.53 — migration 032. Populated only on candidate_error rows.
+    // Self-proxying consumers must whitelist `error_class` (and `outcome`,
+    // per the alpha.51 PB finding) or diagnostic rows degrade silently.
+    error_class: input.errorClass ?? null
   };
 }
 async function recordShadowProbe(input) {
@@ -2966,7 +2977,7 @@ async function recordShadowProbe(input) {
       });
       if (!res.ok) {
         const text = await res.text().catch(() => "<no body>");
-        throw new Error(`brain probe_outcomes ${res.status}: ${text}`);
+        throw new Error(describeBrainWriteFailure(res.status, "probe_outcomes", text));
       }
     } catch (err) {
       (config.onError ?? defaultOnError4)(err);
@@ -3856,7 +3867,7 @@ async function runProbeCandidates(args) {
     const budgetMs = args.cfg.maxLatencyMs ?? DEFAULT_PROBE_MAX_LATENCY_MS;
     const deadline = args.sync ? Date.now() + budgetMs : Number.POSITIVE_INFINITY;
     const skipSlow = args.sync && (args.cfg.skipSlowTierInSync ?? true);
-    const recordDiagnostic = async (candidate, outcome) => {
+    const recordDiagnostic = async (candidate, outcome, detail) => {
       await recordShadowProbe({
         appId: args.ir.appId,
         archetype: args.ir.intent.archetype,
@@ -3869,7 +3880,9 @@ async function runProbeCandidates(args) {
         tokensCurrentIn: args.servedResponse.tokens.input,
         tokensCurrentOut: args.servedResponse.tokens.output,
         latencyCurrentMs: args.servedLatencyMs,
-        outcome
+        outcome,
+        ...detail?.errorClass ? { errorClass: detail.errorClass } : {},
+        ...detail?.latencyCandidateMs !== void 0 ? { latencyCandidateMs: detail.latencyCandidateMs } : {}
       });
     };
     for (const candidate of candidates) {
@@ -3924,10 +3937,28 @@ async function runProbeCandidates(args) {
           }
           break;
         }
-        if (raced.kind === "failed") continue;
+        if (raced.kind === "failed") {
+          try {
+            await recordDiagnostic(candidate, "candidate_error", {
+              errorClass: "execute_rejected",
+              latencyCandidateMs: Date.now() - candStart
+            });
+          } catch {
+          }
+          continue;
+        }
         const exec = raced.exec;
         const candidateLatencyMs = Date.now() - candStart;
-        if (!exec.ok) continue;
+        if (!exec.ok) {
+          try {
+            await recordDiagnostic(candidate, "candidate_error", {
+              errorClass: exec.errorCode,
+              latencyCandidateMs: candidateLatencyMs
+            });
+          } catch {
+          }
+          continue;
+        }
         await recordShadowProbe({
           appId: args.ir.appId,
           archetype: args.ir.intent.archetype,

package/dist/{ir-rUUojj0s.d.ts → ir-BiXAMyji.d.ts}

@@ -1216,7 +1216,16 @@ type OutcomeKind = 'approved' | 'rejected' | 'partial' | 'engaged' | 'abandoned'
  * outcomes accumulate.
  */
 interface RecordOutcomeInput {
-    /** Joins to compile_outcomes.id. Returned by compile() via CompileResult.outcomeId. */
+    /**
+     * Joins to compile_outcomes.id. Pass the compile `handle` string
+     * (CompileResult.handle / CallResult.handle) — consumer proxies resolve
+     * handle → compile_outcomes.id via the `compile_outcomes.handle` column
+     * (reference implementations: PB `api/kgauto-v2/compile_outcome_quality.js`,
+     * tt-intel/IC `app/api/kgauto/v2/compile_outcome_quality/route.ts`). A
+     * numeric compile_outcomes.id also works if the caller already has it.
+     * (There is no `CompileResult.outcomeId` field — the brain assigns the row
+     * id at insert; `handle` is the consumer-visible correlator.)
+     */
     outcomeId: number | string;
     /** What did the user / system do with this output? */
     outcome: OutcomeKind;

package/dist/{ir-dDcG8Pvu.d.mts → ir-CAlLBu5d.d.mts}

@@ -1216,7 +1216,16 @@ type OutcomeKind = 'approved' | 'rejected' | 'partial' | 'engaged' | 'abandoned'
  * outcomes accumulate.
  */
 interface RecordOutcomeInput {
-    /** Joins to compile_outcomes.id. Returned by compile() via CompileResult.outcomeId. */
+    /**
+     * Joins to compile_outcomes.id. Pass the compile `handle` string
+     * (CompileResult.handle / CallResult.handle) — consumer proxies resolve
+     * handle → compile_outcomes.id via the `compile_outcomes.handle` column
+     * (reference implementations: PB `api/kgauto-v2/compile_outcome_quality.js`,
+     * tt-intel/IC `app/api/kgauto/v2/compile_outcome_quality/route.ts`). A
+     * numeric compile_outcomes.id also works if the caller already has it.
+     * (There is no `CompileResult.outcomeId` field — the brain assigns the row
+     * id at insert; `handle` is the consumer-visible correlator.)
+     */
     outcomeId: number | string;
     /** What did the user / system do with this output? */
     outcome: OutcomeKind;

package/dist/profiles.d.mts

@@ -1,4 +1,4 @@
-import { k as Provider } from './ir-dDcG8Pvu.mjs';
+import { k as Provider } from './ir-CAlLBu5d.mjs';
 import { IntentArchetypeName } from './dialect.mjs';
 
 /**

package/dist/profiles.d.ts

@@ -1,4 +1,4 @@
-import { k as Provider } from './ir-rUUojj0s.js';
+import { k as Provider } from './ir-BiXAMyji.js';
 import { IntentArchetypeName } from './dialect.js';
 
 /**

package/dist/profiles.js

@@ -1388,6 +1388,63 @@ var PROFILES_RAW = [
       critique: 5
       // +1 vs 2.5-flash — but still below Sonnet/Opus reasoning floor
     }
+  },
+  {
+    // Auto-onboarded 2026-07-01 from `claude-sonnet-4-6`; VERIFIED 2026-07-02
+    // against the claude-api reference (cc-portfolio ratification pass). The
+    // clone got context right (1M) and pricing right at sticker ($3/$15 —
+    // NOTE an introductory $2/$10 per MTok runs through 2026-08-31; sticker
+    // encoded here per the time-bounded-pricing convention, intro belongs in
+    // brain kgauto_pricing if worth capturing). The clone got max output WRONG:
+    // Sonnet 5 is 128k, not 4-6's 64k — corrected. New tokenizer (~30% more
+    // tokens for the same text vs 4-6): byte-budget consumers should re-baseline.
+    // API quirks (claude-api ref): (a) NON-DEFAULT temperature/top_p/top_k
+    // return 400 — moot for kgauto's own call() path (ANTHROPIC_LOWERING_BASE
+    // emits no sampling params) but a REAL hazard for compileForAISDKv6
+    // consumers that pass temperature themselves (tt-intel scoring uses temp:0
+    // for determinism — that 400s on this model; noted to consumers via the
+    // contract). (b) Omitting `thinking` runs ADAPTIVE by default (4-6 ran
+    // thinking-off) — output spend shifts. (c) Supports effort xhigh. status:
+    // 'preview' per the Fable precedent — no brain evidence yet; promotion to
+    // 'current' is an explicit call. (L-049/L-081.)
+    id: "claude-sonnet-5",
+    verifiedAgainstDocs: "2026-07-02",
+    provider: "anthropic",
+    status: "preview",
+    maxContextTokens: 1e6,
+    maxOutputTokens: 128e3,
+    maxTools: 64,
+    parallelToolCalls: true,
+    structuredOutput: "grammar",
+    systemPromptMode: "inline",
+    streaming: true,
+    cliffs: [],
+    costInputPer1m: 3,
+    costOutputPer1m: 15,
+    lowering: ANTHROPIC_LOWERING_BASE,
+    recovery: [
+      { signal: "rate_limit", action: "escalate", reason: "429 \u2014 escalate" },
+      { signal: "model_not_found", action: "escalate", reason: "Deprecated \u2014 escalate (L-061)" }
+    ],
+    strengths: ["quality", "tool_use", "long_context", "cache_friendly", "extended_thinking"],
+    weaknesses: [],
+    notes: "Sonnet 5 (2026-06): near-Opus quality on coding/agentic work at Sonnet cost. Verified 2026-07-02 against the claude-api reference: 1M ctx, 128k out (clone's 64k corrected), $3/$15 sticker (intro $2/$10 through 2026-08-31). New tokenizer ~30% more tokens vs sonnet-4-6. Consumer hazards: non-default temperature/top_p/top_k 400 (temp:0 rejected \u2014 deterministic-scoring consumers must omit); thinking defaults to adaptive when omitted. status:preview \u2014 no brain evidence yet; earns placement via the machinery.",
+    // Master plan §6.2 anchor. Tier 0 for plan/generate/ask/extract/transform
+    // in starter chains; tier 1 cross-provider for hunt/summarize/classify.
+    archetypePerf: {
+      ask: 9,
+      generate: 9,
+      plan: 9,
+      critique: 9,
+      extract: 9,
+      transform: 9,
+      hunt: 7,
+      // strong but Flash beats on parallel tool throughput
+      summarize: 8,
+      // overkill for tolerant archetype
+      classify: 8
+      // overkill
+    }
   }
 ];
 var ALIASES = {

package/dist/profiles.mjs

@@ -8,7 +8,7 @@ import {
   latencyTierOf,
   profilesByProvider,
   tryGetProfile
-} from "./chunk-CXH7KC4D.mjs";
+} from "./chunk-QKXTMVCT.mjs";
 export {
   ALIASES,
   LATENCY_TIER_MS,

package/dist/{types-vGo-h0tZ.d.mts → types-C_dkNMA_.d.mts}

@@ -1,4 +1,4 @@
-import { r as MutationApplied, B as BestPracticeAdvisory, F as FallbackReason, m as CallAttempt } from './ir-dDcG8Pvu.mjs';
+import { r as MutationApplied, B as BestPracticeAdvisory, F as FallbackReason, m as CallAttempt } from './ir-CAlLBu5d.mjs';
 
 /**
  * Glass-Box observability types (alpha.17).

package/dist/{types-BTeRoSvM.d.mts → types-Dj5FLhBV.d.mts}

@@ -1,4 +1,4 @@
-import { i as Adapter, w as SectionKind } from './ir-dDcG8Pvu.mjs';
+import { i as Adapter, w as SectionKind } from './ir-CAlLBu5d.mjs';
 
 /**
  * Internal config + hook types for createGlassboxRoutes().

package/dist/{types-B_pdPjxm.d.ts → types-DsA8JFOp.d.ts}

@@ -1,4 +1,4 @@
-import { i as Adapter, w as SectionKind } from './ir-rUUojj0s.js';
+import { i as Adapter, w as SectionKind } from './ir-BiXAMyji.js';
 
 /**
  * Internal config + hook types for createGlassboxRoutes().

package/dist/{types-UXPxWabQ.d.ts → types-hEDGehpz.d.ts}

@@ -1,4 +1,4 @@
-import { r as MutationApplied, B as BestPracticeAdvisory, F as FallbackReason, m as CallAttempt } from './ir-rUUojj0s.js';
+import { r as MutationApplied, B as BestPracticeAdvisory, F as FallbackReason, m as CallAttempt } from './ir-BiXAMyji.js';
 
 /**
  * Glass-Box observability types (alpha.17).

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@warmdrift/kgauto-compiler",
-  "version": "2.0.0-alpha.52",
+  "version": "2.0.0-alpha.53",
   "description": "Prompt compiler + central learning brain for multi-model AI apps. Swap models without rewriting prompts.",
   "main": "./dist/index.js",
   "module": "./dist/index.mjs",