@warmdrift/kgauto-compiler 2.0.0-alpha.86 → 2.0.0-alpha.88

package/dist/index.js

@@ -61,7 +61,9 @@ __export(index_exports, {
   STRATEGY_AUTHORSHIP_INDEPENDENT: () => STRATEGY_AUTHORSHIP_INDEPENDENT,
   STRATEGY_AUTHORSHIP_LIMITATION: () => STRATEGY_AUTHORSHIP_LIMITATION,
   TRANSLATOR_FLOOR: () => TRANSLATOR_FLOOR,
+  _testClearConsumerProfiles: () => _testClearConsumerProfiles,
   _testResetMeasuredFailure: () => _testResetMeasuredFailure,
+  _testResetOutcomeSourceWarning: () => _testResetOutcomeSourceWarning,
   _testResetPromotions: () => _testResetPromotions,
   _testWaitForMeasuredFailureRefresh: () => _testWaitForMeasuredFailureRefresh,
   _testWaitForPromotionsRefresh: () => _testWaitForPromotionsRefresh,
@@ -73,6 +75,7 @@ __export(index_exports, {
   attachCacheControlToStreamTextInput: () => attachCacheControlToStreamTextInput,
   awaitMeasuredFailureReady: () => awaitMeasuredFailureReady,
   awaitPromotionsReady: () => awaitPromotionsReady,
+  bestEffortProfile: () => bestEffortProfile,
   brainHealth: () => brainHealth,
   bucketContext: () => bucketContext,
   bucketHistory: () => bucketHistory,
@@ -127,6 +130,7 @@ __export(index_exports, {
   getStarterChainWithGrounding: () => getStarterChainWithGrounding,
   hasMutation: () => hasMutation,
   hashShape: () => hashShape,
+  inferProviderFromId: () => inferProviderFromId,
   isAltStrategy: () => isAltStrategy,
   isArchetype: () => isArchetype,
   isAutoPromoteEnabledFromEnv: () => isAutoPromoteEnabledFromEnv,
@@ -168,11 +172,13 @@ __export(index_exports, {
   recordGoldenIr: () => recordGoldenIr,
   recordOutcome: () => recordOutcome,
   recordShadowProbe: () => recordShadowProbe,
+  registerProfiles: () => registerProfiles,
   renderIrForJudge: () => renderIrForJudge,
   resetTokenizer: () => resetTokenizer,
   resolveConventionsForProfile: () => resolveConventionsForProfile,
   resolveGoldenCaptureRate: () => resolveGoldenCaptureRate,
   resolveModelAlias: () => resolveModelAlias,
+  resolveOutcomeSource: () => resolveOutcomeSource,
   resolvePricingAt: () => resolvePricingAt,
   resolveProviderKey: () => resolveProviderKey,
   rowToAdvisory: () => rowToAdvisory,
@@ -199,6 +205,9 @@ var FRESH_SNAPSHOT = {
 };
 var snapshot = { ...FRESH_SNAPSHOT };
 var runtime;
+function configFreshness() {
+  return runtime && snapshot.data ? "brain" : "bundled";
+}
 function configureBrainQuery(rt) {
   runtime = rt;
   snapshot = { ...FRESH_SNAPSHOT };
@@ -574,7 +583,7 @@ var PROFILES_RAW = [
   },
   {
     id: "claude-opus-4-7",
-    verifiedAgainstDocs: "2026-05-08",
+    verifiedAgainstDocs: "2026-08-24",
     provider: "anthropic",
     status: "legacy",
     // s76 2026-07-25: docs list as LEGACY (Opus 5 is current). Bundle said current while the brain already said legacy — the two disagreed.
@@ -624,7 +633,7 @@ var PROFILES_RAW = [
   },
   {
     id: "claude-opus-4-6",
-    verifiedAgainstDocs: "2026-05-08",
+    verifiedAgainstDocs: "2026-08-24",
     provider: "anthropic",
     status: "legacy",
     maxContextTokens: 1e6,
@@ -669,7 +678,7 @@ var PROFILES_RAW = [
   },
   {
     id: "claude-sonnet-4-6",
-    verifiedAgainstDocs: "2026-05-08",
+    verifiedAgainstDocs: "2026-08-24",
     provider: "anthropic",
     status: "legacy",
     // s76 2026-07-25: docs list as LEGACY (Sonnet 5 is current).
@@ -710,7 +719,7 @@ var PROFILES_RAW = [
   },
   {
     id: "claude-haiku-4-5",
-    verifiedAgainstDocs: "2026-05-08",
+    verifiedAgainstDocs: "2026-08-24",
     provider: "anthropic",
     status: "current",
     maxContextTokens: 2e5,
@@ -773,7 +782,7 @@ var PROFILES_RAW = [
   // ── Google ──
   {
     id: "gemini-2.5-flash",
-    verifiedAgainstDocs: "2026-05-08",
+    verifiedAgainstDocs: "2026-08-24",
     provider: "google",
     status: "current",
     maxContextTokens: 1048576,
@@ -879,7 +888,7 @@ var PROFILES_RAW = [
     // thresholds. The brain will validate/relax these as evidence accumulates
     // per (archetype, model) tuple. Currently ZERO brain rows for this model.
     id: "gemini-2.5-flash-lite",
-    verifiedAgainstDocs: "2026-05-13",
+    verifiedAgainstDocs: "2026-08-24",
     provider: "google",
     status: "current",
     maxContextTokens: 1048576,
@@ -1036,7 +1045,7 @@ var PROFILES_RAW = [
   },
   {
     id: "gemini-3.1-pro-preview",
-    verifiedAgainstDocs: "2026-05-08",
+    verifiedAgainstDocs: "2026-08-24",
     provider: "google",
     status: "preview",
     maxContextTokens: 1048576,
@@ -1096,7 +1105,7 @@ var PROFILES_RAW = [
   // deepseek-reasoner registered as aliases (see ALIASES below).
   {
     id: "deepseek-v4-flash",
-    verifiedAgainstDocs: "2026-05-08",
+    verifiedAgainstDocs: "2026-08-24",
     provider: "deepseek",
     status: "current",
     maxContextTokens: 1e6,
@@ -1114,8 +1123,8 @@ var PROFILES_RAW = [
         reason: "Sequential tool calls only \u2014 L-040"
       }
     ],
-    costInputPer1m: 0.14,
-    costOutputPer1m: 0.28,
+    costInputPer1m: 0.44,
+    costOutputPer1m: 1.32,
     lowering: {
       system: { mode: "inline" },
       cache: { strategy: "unsupported" },
@@ -1134,7 +1143,7 @@ var PROFILES_RAW = [
     // baseQuality (it carries 'reasoning') with no latency counterweight,
     // leapfrogged gemini-2.5-flash as PB's summarize leader once reachable.
     latencyTier: "slow",
-    notes: "Cheap workhorse. 1M context, 384k max output. Cache-hit input $0.0028/M (1/50\xD7 of miss). Aliased as `deepseek-chat` (non-thinking) and `deepseek-reasoner` (thinking) \u2014 see ALIASES.",
+    notes: "Cheap workhorse. 1M context, 384k max output. Cache-hit input $0.0028/M (1/50\xD7 of miss). Aliased as `deepseek-chat` (non-thinking) and `deepseek-reasoner` (thinking) \u2014 see ALIASES. PEAK/OFF-PEAK since 2026-08-16 (page fetched 2026-08-24): peak = weekdays 01:00-04:00 + 06:00-10:00 UTC, off-peak = all other hours at HALF peak. Encoded prices ARE the peak rate (conservative: never understate cost); off-peak traffic costs half what the estimate says. alpha.87 shipped the DEAD pre-transition prices ($0.14/$0.28 flash, $0.435/$0.87 pro) with a fresh verification stamp \u2014 corrected here (alpha.88). A single price field cannot represent a two-rate schedule; catalog-sync extractions oscillate between the rates day-to-day (peak read 08-21, off-peak read 08-22) \u2014 time-of-day pricing remains unmodeled, roadmap.",
     // Master plan §6.2 anchor. Brain-validated tier 1 cross-provider for
     // classify (169 rows, 0% empty). Tier 0 for summarize-with-no-tools.
     // Falls off on hunt (sequential tools — L-040) and reasoning depth.
@@ -1169,7 +1178,7 @@ var PROFILES_RAW = [
   },
   {
     id: "deepseek-v4-pro",
-    verifiedAgainstDocs: "2026-06-22",
+    verifiedAgainstDocs: "2026-08-24",
     provider: "deepseek",
     status: "current",
     maxContextTokens: 1e6,
@@ -1193,8 +1202,8 @@ var PROFILES_RAW = [
     // cache-hit in; no promo caveat). The old $1.74/$3.48 "regular post-promo"
     // was a projected revert that never happened (L-073/L-081) — it 4×-overstated
     // cost and suppressed V4-Pro in cost-aware ranking.
-    costInputPer1m: 0.435,
-    costOutputPer1m: 0.87,
+    costInputPer1m: 1.32,
+    costOutputPer1m: 3.96,
     lowering: {
       system: { mode: "inline" },
       cache: { strategy: "unsupported" },
@@ -1210,7 +1219,7 @@ var PROFILES_RAW = [
     // of the served set. Tag derivation would say 'medium'; the measurement says
     // otherwise.
     latencyTier: "slow",
-    notes: "Pro tier. 1M context, 384k max output. $0.435/$0.87 per 1M (cache-hit in $0.003625) \u2014 verified live 2026-06-22; the launch promo became the standing price (no revert). Default mode = thinking.",
+    notes: "Pro tier. 1M context, 384k max output. $0.435/$0.87 per 1M (cache-hit in $0.003625) \u2014 verified live 2026-06-22; the launch promo became the standing price (no revert). Default mode = thinking. PEAK/OFF-PEAK since 2026-08-16 (page fetched 2026-08-24): peak = weekdays 01:00-04:00 + 06:00-10:00 UTC, off-peak = all other hours at HALF peak. Encoded prices ARE the peak rate (conservative: never understate cost); off-peak traffic costs half what the estimate says. alpha.87 shipped the DEAD pre-transition prices ($0.14/$0.28 flash, $0.435/$0.87 pro) with a fresh verification stamp \u2014 corrected here (alpha.88). A single price field cannot represent a two-rate schedule; catalog-sync extractions oscillate between the rates day-to-day (peak read 08-21, off-peak read 08-22) \u2014 time-of-day pricing remains unmodeled, roadmap.",
     // Master plan §3.3: tier 3 cross-provider for plan chain. Reasoning
     // bumped one notch over V4-Flash; same parallel-tool ceiling.
     archetypePerf: {
@@ -1818,13 +1827,111 @@ var PROFILES_RAW = [
       // +1 vs 2.5-flash — but still below Sonnet/Opus reasoning floor
     }
   },
+  // ─────────────────────────────────────────────────────────────────────────
+  // gemini-3.6-flash — hand-onboarded s86 (2026-08-14), step-zero against
+  // Google docs (models/gemini-3.6-flash page + pricing via catalog-sync's
+  // two-agreeing-extraction the same day). Docs list it as Stable and call
+  // it "previous-generation Flash" (3.5-flash is now "legacy" in Google's
+  // wording) — but kgauto status is 'preview' per the standing entry rule:
+  // a new model enters on capability data verified only at the docs level,
+  // preview is unreachable by starter chains, and promotion to routing
+  // happens through brain evidence, not through Google's marketing tier.
+  // This also means NO family-primary retarget rides this release.
+  // ─────────────────────────────────────────────────────────────────────────
+  {
+    id: "gemini-3.6-flash",
+    verifiedAgainstDocs: "2026-08-24",
+    versionAdded: "2.0.0-alpha.87",
+    provider: "google",
+    status: "preview",
+    maxContextTokens: 1048576,
+    maxOutputTokens: 65536,
+    maxTools: 128,
+    parallelToolCalls: true,
+    structuredOutput: "native",
+    systemPromptMode: "separate",
+    streaming: true,
+    cliffs: [
+      {
+        metric: "input_tokens",
+        threshold: 8e3,
+        action: "downgrade_quality_warning",
+        reason: "Inherited Flash-family guard (2.5\u21923.5 lineage); no 3.6 brain evidence yet"
+      },
+      {
+        metric: "tool_count",
+        threshold: 20,
+        action: "drop_to_top_relevant",
+        reason: "Inherited from Flash family: tool reliability drops above ~20 (despite 128 hard limit). Validate per (archetype, model) after n\u226520."
+      },
+      {
+        metric: "thinking_with_short_output",
+        threshold: 1,
+        action: "force_thinking_budget_zero",
+        reason: "Thinking supported per Google docs; same drain risk as 2.5/3.5-flash \u2014 thinking tokens consume maxOutputTokens"
+      },
+      {
+        metric: "tool_count",
+        threshold: 1,
+        whenIntent: "summarize",
+        action: "strip_tools",
+        reason: "Inherited from 2.5-flash s11 cliff: summarize+tools \u2192 empty response. Preemptive family guard until brain evidence on 3.6-flash specifically."
+      }
+    ],
+    costInputPer1m: 0.75,
+    // promo through 2026-12-31; $1.50 from 2027-01-01 (TIME_BOUNDED_PRICING_ROWS carries the transition)
+    costOutputPer1m: 3.75,
+    // promo through 2026-12-31; $7.50 from 2027-01-01
+    lowering: {
+      ...GOOGLE_LOWERING_BASE,
+      cache: { ...GOOGLE_LOWERING_BASE.cache, discount: 0.1 },
+      thinking: { field: "generationConfig.thinkingConfig.thinkingBudget", default: "auto" }
+    },
+    recovery: [
+      {
+        signal: "empty_response_after_tool",
+        action: "retry_with_params",
+        retryParams: { "generationConfig.thinkingConfig.thinkingBudget": 0 },
+        maxRetries: 1,
+        reason: "Inherited Flash-family pattern: empty after tool result \u2014 retry with thinking off"
+      },
+      {
+        signal: "empty_response",
+        action: "retry_with_params",
+        retryParams: { "generationConfig.thinkingConfig.thinkingBudget": 0 },
+        maxRetries: 1,
+        reason: "Empty response \u2014 try with thinking off"
+      },
+      {
+        signal: "malformed_function_call",
+        action: "escalate",
+        reason: "MALFORMED_FUNCTION_CALL maps to stop \u2014 escalate to next target"
+      }
+    ],
+    strengths: ["agentic_loops", "coding", "1m_context", "parallel_tools", "thinking_mode"],
+    weaknesses: ["no_brain_evidence_yet"],
+    notes: "Hand-onboarded s86 (2026-08-24; the session ran under a 10-day-slow machine clock, hence 08-14 stamps in kin artifacts), step-zero verified: 1,048,576 ctx / 65,536 out (Google model page). PRICING (page 2026-08-24): $0.75/$3.75 promo through 2026-12-31, $1.50/$7.50 from 2027-01-01 \u2014 the alpha.87 profile carried the post-promo rate as if current; corrected alpha.88. Docs status Stable; enters kgauto as preview per standing rule (availability != lifecycle, promotion via brain evidence). Function calling + structured outputs + thinking per docs; parallel tools assumed per Flash family (docs silent). First flagged by the release watcher s73 (2026-07-24); ~31 days in the onboarding queue.",
+    archetypePerf: {
+      hunt: 9,
+      // Flash-family parallel-tool tier (L-040); inherited from 3.5-flash
+      classify: 7,
+      summarize: 7,
+      transform: 7,
+      ask: 8,
+      generate: 8,
+      plan: 7,
+      extract: 7,
+      critique: 5
+    }
+  },
   {
     // 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:
+    // clone got context right (1M). PRICING RE-BASED 2026-08-14 (s86): the
+    // introductory $2/$10 was made PERMANENT — Anthropic's pricing page now
+    // states the scheduled Sept-1 increase to $3/$15 "will not occur", so
+    // $2/$10 IS the sticker. (The $3/$15 encoded 07-02 was correct under the
+    // time-bounded-pricing convention at the time; the bound was cancelled.) 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
@@ -1837,7 +1944,7 @@ var PROFILES_RAW = [
     // '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",
+    verifiedAgainstDocs: "2026-08-24",
     provider: "anthropic",
     status: "current",
     // s76 2026-07-25: GA per docs; was preview.
@@ -1849,8 +1956,9 @@ var PROFILES_RAW = [
     systemPromptMode: "inline",
     streaming: true,
     cliffs: [],
-    costInputPer1m: 3,
-    costOutputPer1m: 15,
+    costInputPer1m: 2,
+    // 2026-08-24: intro price made PERMANENT (docs: "the previously scheduled increase to $3/$15 on September 1, 2026 will not occur")
+    costOutputPer1m: 10,
     lowering: ANTHROPIC_LOWERING_BASE,
     recovery: [
       { signal: "rate_limit", action: "escalate", reason: "429 \u2014 escalate" },
@@ -2041,6 +2149,55 @@ function _setProfileBrainHook(hook) {
 function canonicalId(id) {
   return brainHook.resolveAlias?.(id) ?? ALIASES[id] ?? id;
 }
+var CONSUMER_INDEX = /* @__PURE__ */ new Map();
+function registerProfiles(profiles) {
+  for (const p of profiles) CONSUMER_INDEX.set(p.id, p);
+}
+function _testClearConsumerProfiles() {
+  CONSUMER_INDEX.clear();
+}
+function inferProviderFromId(id) {
+  const lower2 = id.toLowerCase();
+  if (lower2.startsWith("claude")) return "anthropic";
+  if (lower2.startsWith("gemini") || lower2.startsWith("palm")) return "google";
+  if (lower2.startsWith("gpt") || /^o[0-9]/.test(lower2) || lower2.startsWith("chatgpt")) return "openai";
+  if (lower2.startsWith("deepseek")) return "deepseek";
+  if (lower2.startsWith("glm")) return "zai";
+  if (lower2.startsWith("kimi") || lower2.startsWith("moonshot")) return "moonshot";
+  return void 0;
+}
+function bestEffortProfile(id) {
+  const canonical = canonicalId(id);
+  const existing = tryGetProfile(canonical);
+  if (existing) return existing;
+  const provider = inferProviderFromId(canonical);
+  if (!provider) return void 0;
+  const donor = PROFILES_RAW.find((p) => p.provider === provider && p.status === "current") ?? PROFILES_RAW.find((p) => p.provider === provider);
+  if (!donor) return void 0;
+  const synth = {
+    id: canonical,
+    provider,
+    bestEffort: true,
+    status: "preview",
+    maxContextTokens: 128e3,
+    maxOutputTokens: 4096,
+    maxTools: donor.maxTools,
+    parallelToolCalls: false,
+    structuredOutput: "none",
+    systemPromptMode: donor.systemPromptMode,
+    streaming: donor.streaming,
+    cliffs: [],
+    costInputPer1m: 0,
+    costOutputPer1m: 0,
+    lowering: donor.lowering,
+    recovery: [],
+    strengths: [],
+    weaknesses: ["best-effort synthesized profile \u2014 no verified capability data"],
+    notes: `Synthesized by onUnprofiledModel:'best-effort' (provider ${provider} inferred from id; wire mechanics borrowed from ${donor.id}). No cliffs, no measured knowledge, cost UNKNOWN (recorded as 0). Replace via registerProfiles() for real guards.`
+  };
+  CONSUMER_INDEX.set(canonical, synth);
+  return synth;
+}
 function resolveModelAlias(id) {
   return canonicalId(id);
 }
@@ -2049,18 +2206,22 @@ var PROFILE_INDEX = new Map(
 );
 function getProfile(id) {
   const canonical = canonicalId(id);
+  const fromConsumer = CONSUMER_INDEX.get(canonical);
+  if (fromConsumer) return fromConsumer;
   const fromBrain = brainHook.getProfile?.(canonical);
   if (fromBrain) return fromBrain;
   const p = PROFILE_INDEX.get(canonical);
   if (!p) {
     const known = [...PROFILE_INDEX.keys(), ...Object.keys(ALIASES)].join(", ");
-    throw new Error(`Unknown model id: "${id}". Known: ${known}`);
+    throw new Error(
+      `Unknown model id: "${id}". Known: ${known}. For a model kgauto has not onboarded: registerProfiles([...]) supplies a full profile, or policy.onUnprofiledModel:'best-effort' compiles with guards skipped (alpha.87).`
+    );
   }
   return p;
 }
 function tryGetProfile(id) {
   const canonical = canonicalId(id);
-  return brainHook.getProfile?.(canonical) ?? PROFILE_INDEX.get(canonical);
+  return CONSUMER_INDEX.get(canonical) ?? brainHook.getProfile?.(canonical) ?? PROFILE_INDEX.get(canonical);
 }
 function allProfiles() {
   return PROFILES_RAW;
@@ -2339,6 +2500,9 @@ function mapRowsToFindings(rows) {
   return out;
 }
 var snapshots = /* @__PURE__ */ new Map();
+function exclusionFindingsFreshness(appId) {
+  return snapshots.get(appId)?.loaded ? "warm" : "cold";
+}
 var runtime2;
 var warnedOnce = false;
 var DEFAULT_FINDINGS_ENDPOINT = "https://kgauto-dashboard.vercel.app/api/kgauto-v2/findings/exclusions";
@@ -2399,6 +2563,7 @@ async function doRefresh2(rt, appId) {
     if (runtime2 !== rt) return;
     const rows = Array.isArray(body) ? mapRowsToFindings(body) : [];
     snap.data = rows;
+    snap.loaded = true;
     snap.expiresAt = Date.now() + rt.ttlMs;
     snap.refreshing = false;
   } catch (err) {
@@ -4356,6 +4521,9 @@ function judgeMeasuredFailure(row, cfg = MEASURED_FAILURE_CFG) {
   };
 }
 var snapshots2 = /* @__PURE__ */ new Map();
+function measuredFailureFreshness(appId) {
+  return snapshots2.get(appId)?.loaded ? "warm" : "cold";
+}
 var runtime3;
 var warnedOnce2 = false;
 var DEFAULT_MEASURED_FAILURE_ENDPOINT = "https://kgauto-dashboard.vercel.app/api/kgauto-v2/measured-failure";
@@ -4453,6 +4621,7 @@ async function doRefresh3(rt, appId) {
     const body = await res.json();
     if (runtime3 !== rt) return;
     snap.data = Array.isArray(body) ? mapRows(body) : [];
+    snap.loaded = true;
     snap.expiresAt = Date.now() + rt.ttlMs;
     snap.refreshing = false;
   } catch (err) {
@@ -4524,6 +4693,9 @@ function mapRowsToPromotions(rows) {
   return out;
 }
 var snapshots3 = /* @__PURE__ */ new Map();
+function promotionsFreshness(appId) {
+  return snapshots3.get(appId)?.loaded ? "warm" : "cold";
+}
 var runtime4;
 var warnedOnce3 = false;
 var DEFAULT_PROMOTIONS_ENDPOINT = "https://kgauto-dashboard.vercel.app/api/kgauto-v2/promotions";
@@ -4608,6 +4780,7 @@ async function doRefresh4(rt, appId) {
     if (runtime4 !== rt) return;
     const rows = Array.isArray(body) ? mapRowsToPromotions(body) : [];
     snap.data = rows;
+    snap.loaded = true;
     snap.expiresAt = Date.now() + rt.ttlMs;
     snap.refreshing = false;
   } catch (err) {
@@ -4708,6 +4881,9 @@ function mapRowsToFindings2(rows) {
   return out;
 }
 var snapshots4 = /* @__PURE__ */ new Map();
+function promoteReadyFreshness(appId) {
+  return snapshots4.get(appId)?.loaded ? "warm" : "cold";
+}
 var runtime5;
 var warnedOnce4 = false;
 function isPromoteReadyBrainActive() {
@@ -4766,6 +4942,7 @@ async function doRefresh5(rt, appId) {
     if (runtime5 !== rt) return;
     const rows = Array.isArray(body) ? mapRowsToFindings2(body) : [];
     snap.data = rows;
+    snap.loaded = true;
     snap.expiresAt = Date.now() + rt.ttlMs;
     snap.refreshing = false;
   } catch (err) {
@@ -5715,7 +5892,16 @@ function makeHandle() {
   return `c${Date.now().toString(36)}-${counter.toString(36)}-${Math.random().toString(36).slice(2, 6)}`;
 }
 function compile(ir, opts = {}) {
-  const resolver = opts.profileResolver ?? getProfile;
+  const baseResolver = opts.profileResolver ?? getProfile;
+  const resolver = opts.policy?.onUnprofiledModel === "best-effort" ? (id) => {
+    try {
+      return baseResolver(id);
+    } catch (err) {
+      const synth = bestEffortProfile(id);
+      if (!synth) throw err;
+      return synth;
+    }
+  } : baseResolver;
   validateIR(ir);
   ir = resolveModelEntries(ir);
   const sliced = passSlice(ir);
@@ -5775,11 +5961,26 @@ function compile(ir, opts = {}) {
   accumulatedMutations.push(...scores.mutations);
   const target = pickTarget(workingIR, scores.value);
   if (!target) {
+    const unknownIds = scores.value.filter((s) => s.rejectReasons?.includes("unknown_model_id")).map((s) => s.modelId);
+    const uninferable = unknownIds.filter((id) => !inferProviderFromId(id));
+    const unknownHint = unknownIds.length > 0 ? ` Unprofiled model id(s): ${unknownIds.join(", ")}. registerProfiles([...]) supplies a full profile; policy.onUnprofiledModel:'best-effort' compiles with guards skipped.` + (uninferable.length > 0 ? ` Note: provider is not inferable from ${uninferable.join(", ")}, so best-effort cannot lower it \u2014 registerProfiles is the only path for those.` : "") : "";
     throw new Error(
-      `compile(): no allowed model fits the request. Scores: ${JSON.stringify(scores.value, null, 2)}`
+      `compile(): no allowed model fits the request.${unknownHint} Scores: ${JSON.stringify(scores.value, null, 2)}`
     );
   }
   const profile = resolver(target.modelId);
+  const bestEffortWarnings = [];
+  if (profile.bestEffort) {
+    bestEffortWarnings.push(
+      `best-effort profile in use for "${profile.id}" (provider ${profile.provider} inferred from the id; wire mechanics borrowed). SKIPPED for this model: cliff guards, recovery rules, measured archetype knowledge, structured-output capability (treated as 'none'), parallel tools (treated as sequential). Cost is UNKNOWN and recorded as $0 \u2014 cost gates and cost attribution are meaningless for this model. registerProfiles() restores real guards.`
+    );
+    accumulatedMutations.push({
+      id: `best_effort_profile_${profile.id}`,
+      source: "best_effort",
+      passName: "resolve-profile",
+      description: `Unprofiled model "${profile.id}" compiled under onUnprofiledModel:'best-effort' \u2014 guards skipped, cost unknown. See diagnostics.cliffWarnings for the full list.`
+    });
+  }
   const fallbackChain = scores.value.filter((s) => s.modelId !== target.modelId && s.fits).sort((a, b) => b.rank - a.rank).map((s) => s.modelId);
   const cliffs = passApplyCliffs(workingIR, profile, inputTokens);
   workingIR = cliffs.value.ir;
@@ -5865,6 +6066,23 @@ function compile(ir, opts = {}) {
   const historyCacheMarkIndex = computeHistoryCacheMarkIndex(workingIR);
   const systemMessages = buildSystemMessages(workingIR, profile.provider);
   const systemCacheMarkIndex = lastCacheableSystemIndex(systemMessages);
+  const keptSectionIds = new Set(workingIR.sections.map((s) => s.id));
+  const cachePrefixActive = lowered.diagnostics.cacheableTokens > 0;
+  const cacheReadFraction = profile.lowering.cache.discount ?? 0.25;
+  const sectionAttribution = ir.sections.map((s) => {
+    const tokens = countTokens(s.text);
+    const kept = keptSectionIds.has(s.id);
+    const estCostUsdFirstCall = kept ? tokens / 1e6 * profile.costInputPer1m : 0;
+    const inActiveCachePrefix = kept && s.cacheable === true && cachePrefixActive;
+    return {
+      id: s.id,
+      tokens,
+      kept,
+      cacheable: s.cacheable === true,
+      estCostUsdFirstCall,
+      estCostUsdWarm: inActiveCachePrefix ? estCostUsdFirstCall * cacheReadFraction : estCostUsdFirstCall
+    };
+  });
   const diagnostics = {
     sectionsKept: workingIR.sections.length,
     sectionsDropped: ir.sections.length - workingIR.sections.length,
@@ -5888,11 +6106,25 @@ function compile(ir, opts = {}) {
     // text the consumer can route on without changing behavior).
     cliffWarnings: [
       ...cliffs.value.loweringHints.qualityWarning ?? [],
-      ...conventions.value.cliffWarnings
+      ...conventions.value.cliffWarnings,
+      // alpha.87 — best-effort profile in use (loud on EVERY compile).
+      ...bestEffortWarnings
     ],
     // alpha.68 / Release A — measured discipline gate-token tax (§5.D). 0 when
     // the gate didn't fire.
-    disciplineGateTokens
+    disciplineGateTokens,
+    // alpha.87 — per-source freshness stamped at the moment the compile's
+    // reads have all happened, from state each source already tracks (the
+    // "one-line stamp at the point where the source is already known and
+    // discarded" this filing reduced to once alpha.73/.86 warmed the caches).
+    freshness: {
+      config: configFreshness(),
+      measuredFailure: measuredFailureFreshness(ir.appId),
+      promotions: promotionsFreshness(ir.appId),
+      exclusions: exclusionFindingsFreshness(ir.appId),
+      promoteReady: promoteReadyFreshness(ir.appId)
+    },
+    sections: sectionAttribution
   };
   if (ir.intent.archetype === "hunt" && ir.constraints?.toolOrchestration === "sequential") {
     accumulatedMutations.push({
@@ -6057,7 +6289,7 @@ function validateFinalFit(ir, profile, tokens) {
 }
 
 // src/version.ts
-var LIBRARY_VERSION = "2.0.0-alpha.86";
+var LIBRARY_VERSION = "2.0.0-alpha.88";
 
 // src/pricing-brain.ts
 function isPricingRow(x) {
@@ -6431,6 +6663,26 @@ function maybeWarnDiscardedCompile(reg, handle) {
     `[kgauto] compile-then-discard: compileForAISDKv6() output for handle ${handle} (archetype=${reg.archetype}, model=${reg.model}) was recorded but neither .system, .model, nor .raw was read before record(). The brain row's mutations_applied / estimated_tokens_in / system_prompt_chars describe a compile that did NOT shape the served call. Pass compiled.system + compiled.model into your generateText/streamText (or use call()). Silence via configureBrain({ warnOnDiscardedCompile: false }).`
   );
 }
+var OUTCOME_SOURCES = ["canary", "smoke", "eval", "probe", "synthetic"];
+var warnedInvalidEnvSource = false;
+function resolveOutcomeSource(explicit) {
+  if (explicit) return explicit;
+  const raw = typeof process !== "undefined" ? process.env?.KGAUTO_SOURCE : void 0;
+  if (!raw) return void 0;
+  const trimmed = raw.trim();
+  if (!trimmed) return void 0;
+  if (OUTCOME_SOURCES.includes(trimmed)) return trimmed;
+  if (!warnedInvalidEnvSource) {
+    warnedInvalidEnvSource = true;
+    console.warn(
+      `[kgauto] KGAUTO_SOURCE="${trimmed}" is not one of ${OUTCOME_SOURCES.join("|")} \u2014 recording as 'synthetic' so the row survives the brain's CHECK constraint and stays out of organic-liveness stats.`
+    );
+  }
+  return "synthetic";
+}
+function _testResetOutcomeSourceWarning() {
+  warnedInvalidEnvSource = false;
+}
 async function record(input) {
   const reg = compileRegistry.get(input.handle);
   if (reg) {
@@ -6545,7 +6797,9 @@ function buildPayload(input, reg) {
     latency_ms: input.latencyMs,
     success: input.success,
     empty_response: input.emptyResponse ?? input.tokensOut === 0,
-    source: input.source,
+    // alpha.87 — explicit source wins; KGAUTO_SOURCE env fills the gap (the
+    // zero-code synthetic self-mark for harness/smoke processes).
+    source: resolveOutcomeSource(input.source),
     error_type: input.errorType,
     tools_called: input.toolsCalled,
     oracle_score: input.oracleScore?.score,
@@ -10837,7 +11091,9 @@ function compile2(ir, opts) {
   STRATEGY_AUTHORSHIP_INDEPENDENT,
   STRATEGY_AUTHORSHIP_LIMITATION,
   TRANSLATOR_FLOOR,
+  _testClearConsumerProfiles,
   _testResetMeasuredFailure,
+  _testResetOutcomeSourceWarning,
   _testResetPromotions,
   _testWaitForMeasuredFailureRefresh,
   _testWaitForPromotionsRefresh,
@@ -10849,6 +11105,7 @@ function compile2(ir, opts) {
   attachCacheControlToStreamTextInput,
   awaitMeasuredFailureReady,
   awaitPromotionsReady,
+  bestEffortProfile,
   brainHealth,
   bucketContext,
   bucketHistory,
@@ -10903,6 +11160,7 @@ function compile2(ir, opts) {
   getStarterChainWithGrounding,
   hasMutation,
   hashShape,
+  inferProviderFromId,
   isAltStrategy,
   isArchetype,
   isAutoPromoteEnabledFromEnv,
@@ -10944,11 +11202,13 @@ function compile2(ir, opts) {
   recordGoldenIr,
   recordOutcome,
   recordShadowProbe,
+  registerProfiles,
   renderIrForJudge,
   resetTokenizer,
   resolveConventionsForProfile,
   resolveGoldenCaptureRate,
   resolveModelAlias,
+  resolveOutcomeSource,
   resolvePricingAt,
   resolveProviderKey,
   rowToAdvisory,