@warmdrift/kgauto-compiler 2.0.0-alpha.97 → 2.0.0-alpha.99

package/dist/probe.js

@@ -912,7 +912,12 @@ var PROFILES_RAW = [
     // "7-8 calls per step" — which was never a claim that only one tool may be
     // offered. See the note in lower.ts:lowerDeepSeek.
     parallelToolCalls: true,
-    structuredOutput: "native",
+    structuredOutput: "grammar",
+    // alpha.99 — MEASURED 2026-08-29: DeepSeek rejects
+    // `response_format: {type:'json_schema'}` outright ("This response_format type is
+    // unavailable now"), so there is no schema-enforced mode for `'native'` to name.
+    // It supports `json_object` only, and only when the prompt contains the literal
+    // word "json" — the lowering emits the instruction that guarantees it.
     systemPromptMode: "inline",
     streaming: true,
     // alpha.96 — the `tool_count@1 -> drop_to_top_relevant` cliff was DELETED.
@@ -1006,7 +1011,12 @@ var PROFILES_RAW = [
     // "7-8 calls per step" — which was never a claim that only one tool may be
     // offered. See the note in lower.ts:lowerDeepSeek.
     parallelToolCalls: true,
-    structuredOutput: "native",
+    structuredOutput: "grammar",
+    // alpha.99 — MEASURED 2026-08-29: DeepSeek rejects
+    // `response_format: {type:'json_schema'}` outright ("This response_format type is
+    // unavailable now"), so there is no schema-enforced mode for `'native'` to name.
+    // It supports `json_object` only, and only when the prompt contains the literal
+    // word "json" — the lowering emits the instruction that guarantees it.
     systemPromptMode: "inline",
     streaming: true,
     // alpha.96 — the `tool_count@1 -> drop_to_top_relevant` cliff was DELETED.
@@ -3655,7 +3665,7 @@ function lowerGoogle(ir, profile, hints) {
   if (hints.forceTerseOutput) {
     generationConfig.maxOutputTokens = 200;
   }
-  if (ir.constraints?.structuredOutput && profile.structuredOutput === "native") {
+  if (ir.constraints?.structuredOutput && profile.structuredOutput !== "none") {
     generationConfig.responseMimeType = "application/json";
   }
   const contents = buildGoogleContents(ir.history ?? [], ir.currentTurn);
@@ -3714,7 +3724,9 @@ function toGoogleTools(tools) {
 }
 function lowerOpenAI(ir, profile, hints) {
   const ordered = sortSections(ir.sections);
-  const systemText = ordered.map((s) => s.text).join("\n\n");
+  const sectionText = ordered.map((s) => s.text).join("\n\n");
+  const jsonInstruction = ir.constraints?.structuredOutput ? "Respond with a single valid JSON value and nothing else. Do not wrap the JSON in a markdown code fence, and do not add any prose before or after it." : "";
+  const systemText = [sectionText, jsonInstruction].filter(Boolean).join("\n\n");
   const systemRole = profile.systemPromptMode === "as_developer" ? "developer" : "system";
   const messages = systemText ? [{ role: systemRole, content: systemText }] : [];
   for (const m of ir.history ?? []) {
@@ -3760,7 +3772,9 @@ function toOpenAITools(tools) {
 }
 function lowerDeepSeek(ir, profile, hints) {
   const ordered = sortSections(ir.sections);
-  const systemText = ordered.map((s) => s.text).join("\n\n");
+  const sectionText = ordered.map((s) => s.text).join("\n\n");
+  const jsonInstruction = ir.constraints?.structuredOutput ? "Respond with a single valid JSON value and nothing else. Do not wrap the JSON in a markdown code fence, and do not add any prose before or after it." : "";
+  const systemText = [sectionText, jsonInstruction].filter(Boolean).join("\n\n");
   const messages = systemText ? [{ role: "system", content: systemText }] : [];
   for (const m of ir.history ?? []) {
     if (m.role === "system") continue;
@@ -3780,6 +3794,9 @@ function lowerDeepSeek(ir, profile, hints) {
       provider: "deepseek",
       model: profile.id,
       messages,
+      // Safe only because `jsonInstruction` above guarantees the literal word
+      // "json" reaches the prompt — DeepSeek 400s otherwise (measurement C).
+      response_format: ir.constraints?.structuredOutput ? { type: "json_object" } : void 0,
       // Every tool the IR carries. Never fewer — see the note above.
       tools: ir.tools && ir.tools.length > 0 ? ir.tools.map((t) => ({
         type: "function",
@@ -3799,7 +3816,9 @@ function lowerDeepSeek(ir, profile, hints) {
 }
 function buildOpenAICompatibleParts(ir) {
   const ordered = sortSections(ir.sections);
-  const systemText = ordered.map((s) => s.text).join("\n\n");
+  const sectionText = ordered.map((s) => s.text).join("\n\n");
+  const jsonInstruction = ir.constraints?.structuredOutput ? "Respond with a single valid JSON value and nothing else. Do not wrap the JSON in a markdown code fence, and do not add any prose before or after it." : "";
+  const systemText = [sectionText, jsonInstruction].filter(Boolean).join("\n\n");
   const messages = systemText ? [{ role: "system", content: systemText }] : [];
   for (const m of ir.history ?? []) {
     if (m.role === "system") continue;
@@ -4284,12 +4303,19 @@ function mapRows(rows) {
     if (!isRawFailureRow(row)) continue;
     const n = coerceCount(row.n);
     const nFail = coerceCount(row.n_fail) ?? 0;
-    if (n === null || n <= 0) continue;
+    const nAttempted = coerceCount(row.n_attempted);
+    const nUnavailable = coerceCount(row.n_unavailable) ?? 0;
+    const nUnclassified = coerceCount(row.n_unclassified) ?? 0;
+    if (n === null) continue;
+    if (n <= 0 && nUnavailable <= 0) continue;
     out.push({
       archetype: row.intent_archetype,
       model: row.model,
       n,
-      nFail
+      nFail,
+      nAttempted: nAttempted ?? n,
+      nUnavailable,
+      nUnclassified
     });
   }
   return out;
@@ -4313,7 +4339,24 @@ var MEASURED_FAILURE_CFG = {
   /** 95% one-sided-ish confidence (standard two-sided z at α=0.05). */
   z: 1.96,
   /** Must match the view's window. Documented here for the advisory text. */
-  windowDays: 28
+  windowDays: 28,
+  /**
+   * alpha.98 — the AVAILABILITY axis. Same statistic, separate vocabulary,
+   * separate verdict, because a 429 and a parse-failure send an operator in
+   * opposite directions (fund or throttle the account vs. distrust the
+   * model's declared structuredOutput). Merging them into one "failure rate"
+   * is the s75 normalization defect.
+   *
+   * Thresholds match the quality axis deliberately: "we are 95% confident
+   * this model is unreachable more often than not" is the same unarguable
+   * bar, and a de-rank is the same remedy. Kept as a distinct config block
+   * so the two can be tuned apart when evidence says they should be.
+   */
+  availability: {
+    minSample: 5,
+    lowerBoundThreshold: 0.5,
+    z: 1.96
+  }
 };
 function wilsonLowerBound(failures, n, z = MEASURED_FAILURE_CFG.z) {
   if (n <= 0) return 0;
@@ -4327,15 +4370,53 @@ function wilsonLowerBound(failures, n, z = MEASURED_FAILURE_CFG.z) {
 }
 function judgeMeasuredFailure(row, cfg = MEASURED_FAILURE_CFG) {
   if (!row) return void 0;
-  const normalized = "nFail" in row && typeof row.n === "number" ? row : mapRows([row])[0];
-  if (!normalized || normalized.n < cfg.minSample) return void 0;
-  const lowerBound = wilsonLowerBound(normalized.nFail, normalized.n, cfg.z);
+  const raw = "nFail" in row && typeof row.n === "number" ? row : mapRows([row])[0];
+  if (!raw) return void 0;
+  const normalized = {
+    ...raw,
+    nAttempted: typeof raw.nAttempted === "number" ? raw.nAttempted : raw.n,
+    nUnavailable: typeof raw.nUnavailable === "number" ? raw.nUnavailable : 0,
+    nUnclassified: typeof raw.nUnclassified === "number" ? raw.nUnclassified : 0
+  };
+  const quality = judgeAxis(
+    normalized.nFail,
+    normalized.n,
+    cfg.minSample,
+    cfg.lowerBoundThreshold,
+    cfg.z
+  );
+  const availability = judgeAxis(
+    normalized.nUnavailable,
+    normalized.nAttempted,
+    cfg.availability.minSample,
+    cfg.availability.lowerBoundThreshold,
+    cfg.availability.z
+  );
+  if (normalized.n < cfg.minSample && normalized.nAttempted < cfg.availability.minSample) {
+    return void 0;
+  }
+  const axis = quality.gated ? "quality" : availability.gated ? "availability" : void 0;
+  const lead = axis === "availability" ? availability : quality;
   return {
-    gated: lowerBound > cfg.lowerBoundThreshold,
-    rate: normalized.nFail / normalized.n,
+    gated: quality.gated || availability.gated,
+    axis,
+    rate: lead.rate,
+    lowerBound: lead.lowerBound,
+    n: lead.n,
+    nFail: lead.nFail,
+    quality,
+    availability,
+    nUnclassified: normalized.nUnclassified
+  };
+}
+function judgeAxis(failures, n, minSample, threshold, z) {
+  const lowerBound = wilsonLowerBound(failures, n, z);
+  return {
+    rate: n > 0 ? failures / n : 0,
     lowerBound,
-    n: normalized.n,
-    nFail: normalized.nFail
+    n,
+    nFail: failures,
+    gated: n >= minSample && lowerBound > threshold
   };
 }
 var snapshots2 = /* @__PURE__ */ new Map();
@@ -5199,9 +5280,13 @@ function detectSingleModelArray(ir, policy) {
   ];
 }
 function suppressedRecommendationReason(ir, archetype, altProfile) {
-  if (getMeasuredFailureVerdict({ appId: ir.appId, archetype, model: altProfile.id })?.gated === true) {
-    return "measured-failure-gate";
-  }
+  const measured = getMeasuredFailureVerdict({
+    appId: ir.appId,
+    archetype,
+    model: altProfile.id
+  });
+  if (measured?.quality.gated === true) return "measured-failure-gate";
+  if (measured?.availability.gated === true) return "measured-unavailable-gate";
   if (ir.constraints?.structuredOutput && effectiveConventions(altProfile).some(
     (c) => c.archetype === archetype && c.structuredOutputHint === "avoid"
   )) {
@@ -5586,7 +5671,8 @@ function compile(ir, opts = {}) {
         rate: verdict.rate,
         lowerBound: verdict.lowerBound,
         n: verdict.n,
-        nFail: verdict.nFail
+        nFail: verdict.nFail,
+        axis: verdict.axis ?? "quality"
       });
     }
   }

package/dist/probe.mjs

@@ -1,12 +1,12 @@
 import {
   compile
-} from "./chunk-AFYBZRQV.mjs";
-import "./chunk-77AHSX6Y.mjs";
+} from "./chunk-G33BUB43.mjs";
+import "./chunk-2YCTWV4S.mjs";
 import "./chunk-FEMPY6EC.mjs";
 import {
   deriveCliffEvidenceClass,
   getProfile
-} from "./chunk-HHIXHL4O.mjs";
+} from "./chunk-PY4CBZFF.mjs";
 
 // src/probe.ts
 function compileSuppressing(ir, opts) {

package/dist/profiles.d.mts

@@ -1,4 +1,4 @@
-import { l as Provider } from './ir-DyMJ84je.mjs';
+import { l as Provider } from './ir-BZ2qOPMp.mjs';
 import { IntentArchetypeName } from './dialect.mjs';
 
 /**

package/dist/profiles.d.ts

@@ -1,4 +1,4 @@
-import { l as Provider } from './ir-DbvOFKF-.js';
+import { l as Provider } from './ir-CpkidtAd.js';
 import { IntentArchetypeName } from './dialect.js';
 
 /**

package/dist/profiles.js

@@ -843,7 +843,12 @@ var PROFILES_RAW = [
     // "7-8 calls per step" — which was never a claim that only one tool may be
     // offered. See the note in lower.ts:lowerDeepSeek.
     parallelToolCalls: true,
-    structuredOutput: "native",
+    structuredOutput: "grammar",
+    // alpha.99 — MEASURED 2026-08-29: DeepSeek rejects
+    // `response_format: {type:'json_schema'}` outright ("This response_format type is
+    // unavailable now"), so there is no schema-enforced mode for `'native'` to name.
+    // It supports `json_object` only, and only when the prompt contains the literal
+    // word "json" — the lowering emits the instruction that guarantees it.
     systemPromptMode: "inline",
     streaming: true,
     // alpha.96 — the `tool_count@1 -> drop_to_top_relevant` cliff was DELETED.
@@ -937,7 +942,12 @@ var PROFILES_RAW = [
     // "7-8 calls per step" — which was never a claim that only one tool may be
     // offered. See the note in lower.ts:lowerDeepSeek.
     parallelToolCalls: true,
-    structuredOutput: "native",
+    structuredOutput: "grammar",
+    // alpha.99 — MEASURED 2026-08-29: DeepSeek rejects
+    // `response_format: {type:'json_schema'}` outright ("This response_format type is
+    // unavailable now"), so there is no schema-enforced mode for `'native'` to name.
+    // It supports `json_object` only, and only when the prompt contains the literal
+    // word "json" — the lowering emits the instruction that guarantees it.
     systemPromptMode: "inline",
     streaming: true,
     // alpha.96 — the `tool_count@1 -> drop_to_top_relevant` cliff was DELETED.

package/dist/profiles.mjs

@@ -14,7 +14,7 @@ import {
   registerProfiles,
   resolveModelAlias,
   tryGetProfile
-} from "./chunk-HHIXHL4O.mjs";
+} from "./chunk-PY4CBZFF.mjs";
 export {
   ALIASES,
   LATENCY_TIER_MS,

package/dist/{types-D441T-KC.d.mts → types-BkdfoONZ.d.mts}

@@ -1,4 +1,4 @@
-import { j as Adapter, F as FallbackReason, y as SectionKind } from './ir-DyMJ84je.mjs';
+import { j as Adapter, F as FallbackReason, y as SectionKind } from './ir-BZ2qOPMp.mjs';
 
 /**
  * Wire contract for the Glass-Box Chrome extension's brain-poll endpoint.

package/dist/{types-5TqjBeZD.d.ts → types-Bn-zF0S4.d.ts}

@@ -1,4 +1,4 @@
-import { j as Adapter, F as FallbackReason, y as SectionKind } from './ir-DbvOFKF-.js';
+import { j as Adapter, F as FallbackReason, y as SectionKind } from './ir-CpkidtAd.js';
 
 /**
  * Wire contract for the Glass-Box Chrome extension's brain-poll endpoint.

package/dist/{types-_myk4bxn.d.mts → types-DJFlO7ux.d.mts}

@@ -1,4 +1,4 @@
-import { t as MutationApplied, B as BestPracticeAdvisory, F as FallbackReason, n as CallAttempt } from './ir-DyMJ84je.mjs';
+import { t as MutationApplied, B as BestPracticeAdvisory, F as FallbackReason, n as CallAttempt } from './ir-BZ2qOPMp.mjs';
 
 /**
  * Glass-Box observability types (alpha.17).

package/dist/{types-BlrbNQfj.d.ts → types-uWypgaAU.d.ts}

@@ -1,4 +1,4 @@
-import { t as MutationApplied, B as BestPracticeAdvisory, F as FallbackReason, n as CallAttempt } from './ir-DbvOFKF-.js';
+import { t as MutationApplied, B as BestPracticeAdvisory, F as FallbackReason, n as CallAttempt } from './ir-CpkidtAd.js';
 
 /**
  * Glass-Box observability types (alpha.17).

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@warmdrift/kgauto-compiler",
-  "version": "2.0.0-alpha.97",
+  "version": "2.0.0-alpha.99",
   "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",