@tangle-network/agent-eval 0.50.1 → 0.51.0

package/CHANGELOG.md

@@ -4,6 +4,45 @@ All notable changes to `@tangle-network/agent-eval` and its sibling `agent-eval-
 
 ---
 
+## [0.51.0] — 2026-05-27 — skillOptDriver (SkillOpt methodology as a substrate driver)
+
+### Added
+
+- **`skillOptDriver`** in `/campaign`. A section-aware, bounded-edit `ImprovementDriver` for structured natural-language procedures (SKILL.md files, runbooks, sectioned system prompts, judge rubrics with dimensions). Implements the SkillOpt methodology (Microsoft, 2026): treat the skill document as a trainable optimization target, train the procedure not the weights, constrain each generation to ≤N targeted edits to prevent useful-rule overwrites.
+  - **Edit-budget enforcement** — candidates that exceed `editBudget * 2` sentence-level diffs vs baseline are rejected at parse time. SkillOpt's "edit budget functions as a textual learning rate."
+  - **Section preservation** — H2 headings (or an explicit `preserveSections` allowlist) MUST appear unchanged in every candidate. Candidates that delete or rename sections are rejected.
+  - **Surface-typed** — throws on non-string surfaces; agent-runtime's `improvementDriver` handles code-tier.
+- `extractH2Sections(text)` + `countSentenceEdits(a, b)` exported as named helpers for consumers writing custom drivers with similar invariants.
+
+### Scope (honest)
+
+This is **batch SkillOpt** — one LLM call per generation produces all N candidates with the budget enforced as a prompt instruction + post-parse rejection. **Per-edit iteration** (propose 1 edit → validate → accept-or-reject → propose next) is a future 0.52.0 enhancement that needs a new `IncrementalImprovementDriver` interface; the substrate's current batch `ImprovementDriver` can run SkillOpt-style behavior with `populationSize=1` + `maxGenerations=N`, but a single driver invocation can't iterate per-edit yet. Tracked.
+
+### Notes
+
+Selectable alongside `gepaDriver` and `evolutionaryDriver`. Use when the surface IS a structured doc; use `gepaDriver` when the surface is unstructured prose.
+
+---
+
+## [0.50.2] — 2026-05-27 — actionability fixes from real-data dogfood
+
+### Added
+
+- **`ScalarDistribution.tailRuns?: Array<{runId, score}>`** — populated for the composite distribution. The report now names the 5 worst runs a customer should inspect first, instead of telling them to "investigate the lower tail" anonymously.
+- **`InsightReport.costQuality.degraded?: {cost?, pareto?}`** — explicit per-axis degradation reasons when `costUsd` is all zero (cost axis carries no signal) or only a single candidate appears (Pareto collapses to a single point). Replaces the prior silent emission of meaningless single-point Pareto figures.
+- **Composite-distribution recommendations.** When `composite.mean < 0.3`, the report emits a `critical/investigate` recommendation with the worst-5 runIds enumerated in the detail. Between 0.3 and 0.5, a `high/investigate` recommendation with the worst-3. Closes the gap where `recommendations: []` was being emitted for completely broken corpora.
+- **Missing-judges flag.** When `judges` is empty across the corpus, the report emits a `medium/expand-corpus` recommendation pointing at `outcome.judgeScores.perJudge` enrichment. Before, the customer had no signal that per-dimension / calibration was unavailable because of input shape, not substrate failure.
+
+### Fixed
+
+- `analyzeRuns()` on the legal-agent canonical run (n=36, mean composite = 0.002) now emits actionable recommendations naming specific failing scenarios; previously it returned `recommendations: []` for a fully-broken agent.
+
+### Notes
+
+The four behavior changes are additive — fields are optional, no existing field shape changed. Dogfood-driven: surfaced by running `analyzeRuns()` against three real consumer datasets (legal-agent, agent-builder, gtm-agent golden run) and observing where the report was silent when it should have been loud.
+
+---
+
 ## [0.50.1] — 2026-05-27 — docs + examples
 
 ### Added

package/dist/adapters/otel.d.ts

@@ -1,4 +1,4 @@
-import { T as TraceSpanEvent, H as HostedClient } from '../index-BRxz6qov.js';
+import { T as TraceSpanEvent, H as HostedClient } from '../index-DQHtWQ57.js';
 import '../types-Dbj5gu8n.js';
 import '../summary-report-B7gNRX-r.js';
 import '../run-record-BGY6bHRh.js';

package/dist/campaign/index.d.ts

@@ -1,13 +1,72 @@
 export { C as CampaignStorage, D as DefaultProductionGateOptions, E as EvolutionaryDriverOptions, G as GepaDriverOptions, H as HeldOutGateOptions, O as OpenAutoPrOptions, a as OpenAutoPrResult, R as RunCampaignOptions, b as RunEvalOptions, c as RunImprovementLoopOptions, d as RunImprovementLoopResult, e as RunOptimizationOptions, f as RunOptimizationResult, g as composeGate, h as defaultProductionGate, i as evolutionaryDriver, j as fsCampaignStorage, k as gepaDriver, l as heldOutGate, m as inMemoryCampaignStorage, o as openAutoPr, r as runCampaign, n as runEval, p as runImprovementLoop, q as runOptimization, s as surfaceHash } from '../run-improvement-loop-BPMjNKMJ.js';
-import { L as LabeledScenarioStore, c as LabeledScenarioWrite, d as LabeledScenarioSampleArgs, e as LabeledScenarioRecord, C as CodeSurface } from '../types-Dbj5gu8n.js';
-export { f as CampaignAggregates, g as CampaignArtifactWriter, h as CampaignCellResult, i as CampaignCostMeter, j as CampaignResult, k as CampaignTraceWriter, b as DispatchContext, D as DispatchFn, G as Gate, l as GateContext, m as GateDecision, n as GateResult, o as GenerationCandidate, p as GenerationRecord, I as ImprovementDriver, q as JudgeAggregate, a as JudgeConfig, r as JudgeDimension, J as JudgeScore, s as LabeledScenarioSource, M as MutableSurface, t as Mutator, O as OptimizerConfig, P as ProposeContext, R as RedactionStatus, S as Scenario, u as ScenarioAggregate, v as SessionScript, T as TraceSpan } from '../types-Dbj5gu8n.js';
-import '../llm-client-BXVRUZyX.js';
-import '../errors-mje_cKOs.js';
-import '../raw-provider-sink-C46HDghv.js';
+import { L as LlmClientOptions } from '../llm-client-BXVRUZyX.js';
+import { I as ImprovementDriver, L as LabeledScenarioStore, c as LabeledScenarioWrite, d as LabeledScenarioSampleArgs, e as LabeledScenarioRecord, C as CodeSurface } from '../types-Dbj5gu8n.js';
+export { f as CampaignAggregates, g as CampaignArtifactWriter, h as CampaignCellResult, i as CampaignCostMeter, j as CampaignResult, k as CampaignTraceWriter, b as DispatchContext, D as DispatchFn, G as Gate, l as GateContext, m as GateDecision, n as GateResult, o as GenerationCandidate, p as GenerationRecord, q as JudgeAggregate, a as JudgeConfig, r as JudgeDimension, J as JudgeScore, s as LabeledScenarioSource, M as MutableSurface, t as Mutator, O as OptimizerConfig, P as ProposeContext, R as RedactionStatus, S as Scenario, u as ScenarioAggregate, v as SessionScript, T as TraceSpan } from '../types-Dbj5gu8n.js';
 import '../red-team-30II1T4o.js';
 import '../dataset-BlwAtYYf.js';
+import '../errors-mje_cKOs.js';
 import '../store-Db2Bv8Cf.js';
 import '../run-record-BGY6bHRh.js';
+import '../raw-provider-sink-C46HDghv.js';
+
+/**
+ * @experimental
+ *
+ * `skillOptDriver` — a section-aware, bounded-edit `ImprovementDriver` for
+ * structured natural-language procedures (SKILL.md files, runbooks, sectioned
+ * system prompts, judge rubrics with dimensions). Implements the SkillOpt
+ * methodology (Microsoft, 2026): treat the skill document as a trainable
+ * optimization target, train the procedure not the weights, constrain each
+ * generation to ≤N targeted edits to prevent useful-rule overwrites.
+ *
+ * Differs from `gepaDriver` in two specific ways:
+ *
+ * 1. **Bounded edits.** Each candidate must differ from the baseline by at
+ *    most `editBudget` sentence-level changes. The "edit budget functions
+ *    as a textual learning rate" — without it, an LLM proposal can rewrite
+ *    so much that useful prior rules get overwritten.
+ *
+ * 2. **Section preservation.** When the surface is a structured doc, the
+ *    H2 headers (and an opt-in `preserveSections` allowlist) are
+ *    load-bearing for discoverability. Candidates that delete or rename
+ *    preserved sections are rejected at parse time.
+ *
+ * Selectable alongside `gepaDriver` and `evolutionaryDriver`. Use this when
+ * the surface IS a structured doc; use `gepaDriver` when the surface is
+ * unstructured prose.
+ */
+
+interface SkillOptDriverOptions {
+    /** Router transport (apiKey/baseUrl). */
+    llm: LlmClientOptions;
+    /** Model that performs the reflection. */
+    model: string;
+    /** What is being optimized — appears in the reflection prompt for orientation. */
+    target: string;
+    /** Max edits per generation — SkillOpt's "textual learning rate".
+     *  Default 3. Lower = more conservative, higher = more exploratory. */
+    editBudget?: number;
+    /** Section headings the driver MUST preserve. When the surface is a
+     *  structured skill doc, sections are load-bearing for discoverability.
+     *  Default: auto-detected from H2 headers in the baseline. */
+    preserveSections?: string[];
+    /** Surface-specific mutation levers offered to the model. */
+    mutationPrimitives?: string[];
+    /** Top/bottom scenarios surfaced as evidence each generation. Default 3. */
+    evidenceK?: number;
+    /** Reflection sampling temperature. Default 0.7. */
+    temperature?: number;
+    /** Reflection max tokens. Default 6000. */
+    maxTokens?: number;
+}
+/** Internal — exported for tests. */
+declare function extractH2Sections(text: string): string[];
+/** Sentence-level edit distance. Counts distinct sentence add/remove/replace
+ *  ops between baseline and candidate using a normalised line-by-line diff.
+ *  Imperfect (treats trivial whitespace as identical) but tight enough to
+ *  bound an LLM rewrite. Exported for tests. */
+declare function countSentenceEdits(baseline: string, candidate: string): number;
+declare function skillOptDriver(opts: SkillOptDriverOptions): ImprovementDriver;
 
 /**
  * @experimental
@@ -124,4 +183,4 @@ declare function gitWorktreeAdapter(opts: GitWorktreeAdapterOptions): WorktreeAd
  *  as a ref under the adapter's worktree dir. */
 declare function resolveWorktreePath(surface: CodeSurface, worktreeDir?: string): string;
 
-export { CodeSurface, FsLabeledScenarioStore, type FsLabeledScenarioStoreOptions, type GitWorktreeAdapterOptions, LabeledScenarioRecord, LabeledScenarioSampleArgs, LabeledScenarioStore, LabeledScenarioStoreError, LabeledScenarioWrite, type Worktree, type WorktreeAdapter, WorktreeAdapterError, gitWorktreeAdapter, resolveWorktreePath };
+export { CodeSurface, FsLabeledScenarioStore, type FsLabeledScenarioStoreOptions, type GitWorktreeAdapterOptions, ImprovementDriver, LabeledScenarioRecord, LabeledScenarioSampleArgs, LabeledScenarioStore, LabeledScenarioStoreError, LabeledScenarioWrite, type SkillOptDriverOptions, type Worktree, type WorktreeAdapter, WorktreeAdapterError, countSentenceEdits, extractH2Sections, gitWorktreeAdapter, resolveWorktreePath, skillOptDriver };

package/dist/campaign/index.js

@@ -15,15 +15,133 @@ import {
   inMemoryCampaignStorage,
   runCampaign
 } from "../chunk-J3EIOI3O.js";
-import "../chunk-N4SBKEPJ.js";
+import {
+  buildReflectionPrompt,
+  parseReflectionResponse
+} from "../chunk-N4SBKEPJ.js";
 import "../chunk-YV7J7X5N.js";
 import "../chunk-WP7SY7AI.js";
 import "../chunk-GGE4NNQT.js";
-import "../chunk-VXNVVBZO.js";
+import {
+  callLlm
+} from "../chunk-VXNVVBZO.js";
 import "../chunk-PC4UYEBM.js";
 import "../chunk-QYJT52YW.js";
 import "../chunk-NSBPE2FW.js";
 
+// src/campaign/drivers/skillopt.ts
+var REFLECTION_SYSTEM = 'You are an expert prompt engineer applying the SkillOpt methodology. You will edit a structured natural-language procedure under TWO HARD CONSTRAINTS: (1) preserve every H2 section heading verbatim \u2014 do NOT delete, rename, or merge sections; (2) make at most EDIT_BUDGET targeted sentence-level edits per candidate \u2014 bounded edits prevent overwriting useful prior rules. Output ONLY a JSON object of shape {"proposals":[{"label":string,"rationale":string,"payload":string}]} where each `payload` is the FULL improved skill text. No prose outside the JSON.';
+function extractH2Sections(text) {
+  const out = [];
+  for (const line of text.split("\n")) {
+    const match = /^##\s+(.+?)\s*$/.exec(line);
+    if (match) out.push(match[1]);
+  }
+  return out;
+}
+function countSentenceEdits(baseline, candidate) {
+  const norm = (s) => s.split(/(?<=[.!?])\s+|\n/g).map((p) => p.trim()).filter((p) => p.length > 0);
+  const a = new Set(norm(baseline));
+  const b = new Set(norm(candidate));
+  let edits = 0;
+  for (const s of a) if (!b.has(s)) edits++;
+  for (const s of b) if (!a.has(s)) edits++;
+  return edits;
+}
+function skillOptDriver(opts) {
+  const evidenceK = opts.evidenceK ?? 3;
+  const editBudget = opts.editBudget ?? 3;
+  if (editBudget < 1) {
+    throw new Error(
+      `skillOptDriver: editBudget must be >= 1, got ${editBudget} (use evolutionaryDriver with a noop mutator for measure-only runs)`
+    );
+  }
+  return {
+    kind: "skillopt",
+    async propose(ctx) {
+      if (typeof ctx.currentSurface !== "string") {
+        throw new Error(
+          `skillOptDriver: surface must be a string skill document; got ${typeof ctx.currentSurface}. Use evolutionaryDriver with a CodeSurface mutator for code-tier surfaces.`
+        );
+      }
+      const baseline = ctx.currentSurface;
+      const preserveSections = opts.preserveSections ?? extractH2Sections(baseline);
+      const { top, bottom, target } = buildEvidence(ctx, evidenceK, opts.target);
+      const reflectionUser = buildReflectionPrompt({
+        target,
+        parentPayload: baseline,
+        topTrials: top,
+        bottomTrials: bottom,
+        childCount: ctx.populationSize,
+        mutationPrimitives: opts.mutationPrimitives
+      });
+      const constraintPreamble = [
+        "",
+        "## SkillOpt constraints (hard rules \u2014 violations rejected)",
+        "",
+        `- Edit budget: at most ${editBudget} sentence-level edits per candidate.`,
+        "- Section preservation: every H2 heading below must appear unchanged in your output.",
+        ...preserveSections.map((s) => `  - \`## ${s}\``),
+        "",
+        "Reject any candidate in your own thinking that would delete a section, rename a heading, or exceed the edit budget. Make TARGETED, surgical edits \u2014 not rewrites.",
+        ""
+      ].join("\n");
+      const userPrompt = `${reflectionUser}${constraintPreamble}`;
+      const system = REFLECTION_SYSTEM.replace("EDIT_BUDGET", String(editBudget));
+      const result = await callLlm(
+        {
+          model: opts.model,
+          messages: [
+            { role: "system", content: system },
+            { role: "user", content: userPrompt }
+          ],
+          jsonMode: true,
+          temperature: opts.temperature ?? 0.7,
+          maxTokens: opts.maxTokens ?? 6e3
+        },
+        opts.llm
+      );
+      const proposals = parseReflectionResponse(result.content, ctx.populationSize);
+      const out = [];
+      for (const proposal of proposals) {
+        const text = typeof proposal.payload === "string" ? proposal.payload.trim() : "";
+        if (!text || text === baseline) continue;
+        if (!validateSections(text, preserveSections)) continue;
+        if (countSentenceEdits(baseline, text) > editBudget * 2) continue;
+        if (out.includes(text)) continue;
+        out.push(text);
+      }
+      return out;
+    }
+  };
+}
+function validateSections(candidate, required) {
+  if (required.length === 0) return true;
+  const have = new Set(extractH2Sections(candidate));
+  for (const section of required) {
+    if (!have.has(section)) return false;
+  }
+  return true;
+}
+function buildEvidence(ctx, evidenceK, baseTarget) {
+  const last = ctx.history.at(-1);
+  if (!last || last.candidates.length === 0) {
+    return { top: [], bottom: [], target: baseTarget };
+  }
+  const best = [...last.candidates].sort((a, b) => b.composite - a.composite)[0];
+  if (!best) return { top: [], bottom: [], target: baseTarget };
+  const byScore = [...best.scenarios].sort((a, b) => b.composite - a.composite);
+  const toTrace = (s) => ({
+    id: s.scenarioId,
+    score: s.composite
+  });
+  const top = byScore.slice(0, evidenceK).map(toTrace);
+  const bottom = byScore.slice(-evidenceK).reverse().map(toTrace);
+  const weakest = Object.entries(best.dimensions).sort((a, b) => a[1] - b[1]).slice(0, 3).map(([dim, value]) => `${dim} (${value.toFixed(2)})`);
+  const target = weakest.length > 0 ? `${baseTarget} \u2014 weakest dimensions: ${weakest.join(", ")}` : baseTarget;
+  return { top, bottom, target };
+}
+
 // src/campaign/labeled-store/fs-adapter.ts
 import { createHash } from "crypto";
 import { existsSync, mkdirSync, readFileSync, writeFileSync } from "fs";
@@ -283,8 +401,10 @@ export {
   LabeledScenarioStoreError,
   WorktreeAdapterError,
   composeGate,
+  countSentenceEdits,
   defaultProductionGate,
   evolutionaryDriver,
+  extractH2Sections,
   fsCampaignStorage,
   gepaDriver,
   gitWorktreeAdapter,
@@ -296,6 +416,7 @@ export {
   runEval,
   runImprovementLoop,
   runOptimization,
+  skillOptDriver,
   surfaceHash
 };
 //# sourceMappingURL=index.js.map

package/dist/campaign/index.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../../src/campaign/labeled-store/fs-adapter.ts","../../src/campaign/worktree/index.ts"],"sourcesContent":["/**\n * @experimental\n *\n * Filesystem `LabeledScenarioStore` adapter. The default capture sink for\n * traces + eval artifacts. Production deployments typically swap for a\n * Turso/SQLite adapter (same interface).\n *\n * Records land as one JSONL file per source under `<root>/<source>.jsonl`.\n * Each line is a `LabeledScenarioRecord`. Append-only — no in-place edits.\n *\n * Safety properties enforced at write-time:\n *\n *   - **Provenance required**: writes without `source`, `sourceVersionHash`,\n *     `capturedAt`, `redactionStatus` are rejected. Closes the alignment\n *     reviewer's data-poisoning gap.\n *   - **Per-source rate limits**: optional `rateLimitBucket` + `maxWritesPerMinute`\n *     stops a single tenant/source from flooding the store.\n *\n * Safety properties enforced at sample-time:\n *\n *   - **Required split + capturedBefore**: substrate refuses to sample without\n *     an explicit `split` ('train' | 'test') AND a temporal cutoff. Eliminates\n *     accidental train/test contamination.\n *   - **Default training-source filter**: when the store is sampled with\n *     `split: 'train'`, production-trace records are EXCLUDED unless the\n *     caller passes `filter.source: 'production-trace'` explicitly. Closes\n *     the contamination-by-default gap flagged by the senior eval engineer.\n */\n\nimport { createHash } from 'node:crypto'\nimport { existsSync, mkdirSync, readFileSync, writeFileSync } from 'node:fs'\nimport { join } from 'node:path'\nimport type {\n  LabeledScenarioRecord,\n  LabeledScenarioSampleArgs,\n  LabeledScenarioSource,\n  LabeledScenarioStore,\n  LabeledScenarioWrite,\n} from '../types'\n\nexport interface FsLabeledScenarioStoreOptions {\n  /** Root directory for JSONL files. Created if missing. */\n  root: string\n  /** Per-source rate limit. When set, writes exceeding the cap are rejected\n   *  with a typed error. Default: no limit. */\n  maxWritesPerMinutePerBucket?: number\n  /** Test seam — override `Date.now()` for deterministic tests. */\n  now?: () => number\n}\n\nexport class LabeledScenarioStoreError extends Error {\n  constructor(\n    public readonly code: string,\n    message: string,\n  ) {\n    super(message)\n    this.name = 'LabeledScenarioStoreError'\n  }\n}\n\ninterface RateLimitState {\n  bucket: string\n  windowStartMs: number\n  count: number\n}\n\nexport class FsLabeledScenarioStore implements LabeledScenarioStore {\n  private readonly now: () => number\n  private readonly rateLimits = new Map<string, RateLimitState>()\n\n  constructor(private readonly options: FsLabeledScenarioStoreOptions) {\n    if (!existsSync(options.root)) mkdirSync(options.root, { recursive: true })\n    this.now = options.now ?? Date.now\n  }\n\n  async observe(write: LabeledScenarioWrite): Promise<void> {\n    this.assertProvenance(write)\n    this.assertRateLimit(write)\n    const record = this.toRecord(write)\n    const path = this.pathForSource(write.source)\n    const line = `${JSON.stringify(record)}\\n`\n    // Append atomically. For high-throughput a writev-friendly buffered\n    // implementation lands in the Turso adapter; FS adapter is for tests +\n    // local dev + small workloads.\n    appendLine(path, line)\n  }\n\n  async sample(args: LabeledScenarioSampleArgs): Promise<LabeledScenarioRecord[]> {\n    if (!args.split) {\n      throw new LabeledScenarioStoreError(\n        'split_required',\n        'sample() requires an explicit `split` (train | test) — substrate refuses ambiguous reads',\n      )\n    }\n    if (!args.capturedBefore) {\n      throw new LabeledScenarioStoreError(\n        'capturedBefore_required',\n        'sample() requires an explicit `capturedBefore` timestamp for temporal-split discipline',\n      )\n    }\n\n    const all: LabeledScenarioRecord[] = []\n    for (const source of ALL_SOURCES) {\n      // Default training-source filter: when sampling train, EXCLUDE\n      // production-trace records unless the caller asks for them.\n      if (args.split === 'train' && source === 'production-trace') {\n        const explicit = sourceFilterContains(args.filter?.source, 'production-trace')\n        if (!explicit) continue\n      }\n      const path = this.pathForSource(source)\n      if (!existsSync(path)) continue\n      const lines = readFileSync(path, 'utf8').split('\\n').filter(Boolean)\n      for (const line of lines) {\n        let record: LabeledScenarioRecord\n        try {\n          record = JSON.parse(line) as LabeledScenarioRecord\n        } catch {\n          continue\n        }\n        if (!matchesFilter(record, args, source)) continue\n        all.push(record)\n      }\n    }\n\n    // Deterministic order: by capturedAt ascending, then recordHash.\n    all.sort((a, b) => {\n      if (a.capturedAt !== b.capturedAt) return a.capturedAt.localeCompare(b.capturedAt)\n      return a.recordHash.localeCompare(b.recordHash)\n    })\n\n    return all.slice(0, args.count)\n  }\n\n  async size(): Promise<{ train: number; test: number; bySource: Record<string, number> }> {\n    const bySource: Record<string, number> = {}\n    let total = 0\n    for (const source of ALL_SOURCES) {\n      const path = this.pathForSource(source)\n      if (!existsSync(path)) {\n        bySource[source] = 0\n        continue\n      }\n      const count = readFileSync(path, 'utf8').split('\\n').filter(Boolean).length\n      bySource[source] = count\n      total += count\n    }\n    // FS adapter doesn't track split assignments per-record (split is\n    // computed at sample-time based on `capturedBefore`). For size(), we\n    // report `train`+`test` as the same total — split is a sampling concept.\n    return { train: total, test: total, bySource }\n  }\n\n  private assertProvenance(write: LabeledScenarioWrite): void {\n    if (!write.source) {\n      throw new LabeledScenarioStoreError(\n        'missing_source',\n        'LabeledScenarioWrite requires `source`',\n      )\n    }\n    if (!write.sourceVersionHash || write.sourceVersionHash.length === 0) {\n      throw new LabeledScenarioStoreError(\n        'missing_source_version',\n        'LabeledScenarioWrite requires `sourceVersionHash` (git sha or substrate version)',\n      )\n    }\n    if (!write.capturedAt) {\n      throw new LabeledScenarioStoreError(\n        'missing_captured_at',\n        'LabeledScenarioWrite requires `capturedAt` ISO timestamp',\n      )\n    }\n    if (!write.redactionStatus) {\n      throw new LabeledScenarioStoreError(\n        'missing_redaction_status',\n        'LabeledScenarioWrite requires explicit `redactionStatus` — raw / redacted-pii / redacted-secrets / fully-redacted',\n      )\n    }\n    if (!ALL_SOURCES.includes(write.source)) {\n      throw new LabeledScenarioStoreError(\n        'unknown_source',\n        `LabeledScenarioWrite.source must be one of: ${ALL_SOURCES.join(', ')}`,\n      )\n    }\n  }\n\n  private assertRateLimit(write: LabeledScenarioWrite): void {\n    const cap = this.options.maxWritesPerMinutePerBucket\n    if (!cap || !write.rateLimitBucket) return\n    const now = this.now()\n    const windowMs = 60_000\n    let state = this.rateLimits.get(write.rateLimitBucket)\n    if (!state || now - state.windowStartMs >= windowMs) {\n      state = { bucket: write.rateLimitBucket, windowStartMs: now, count: 0 }\n      this.rateLimits.set(write.rateLimitBucket, state)\n    }\n    if (state.count >= cap) {\n      throw new LabeledScenarioStoreError(\n        'rate_limit_exceeded',\n        `LabeledScenarioStore: bucket ${write.rateLimitBucket} exceeded ${cap} writes/min`,\n      )\n    }\n    state.count += 1\n  }\n\n  private toRecord(write: LabeledScenarioWrite): LabeledScenarioRecord {\n    const recordHash = sha256(\n      JSON.stringify({\n        id: write.scenario.id,\n        src: write.source,\n        at: write.capturedAt,\n        ver: write.sourceVersionHash,\n      }),\n    )\n    // FS adapter assigns split at sample-time, but we cache a hint here\n    // based on capturedAt vs the world's \"now\" — sampler overrides this.\n    return {\n      ...write,\n      recordHash,\n      split: 'train',\n    }\n  }\n\n  private pathForSource(source: string): string {\n    return join(this.options.root, `${source}.jsonl`)\n  }\n}\n\nconst ALL_SOURCES: LabeledScenarioWrite['source'][] = [\n  'production-trace',\n  'eval-run',\n  'manual',\n  'red-team',\n  'synthetic',\n]\n\nfunction sourceFilterContains(\n  filter: LabeledScenarioSource | LabeledScenarioSource[] | undefined,\n  needle: LabeledScenarioSource,\n): boolean {\n  if (!filter) return false\n  if (Array.isArray(filter)) return filter.includes(needle)\n  return filter === needle\n}\n\nfunction matchesFilter(\n  record: LabeledScenarioRecord,\n  args: LabeledScenarioSampleArgs,\n  source: string,\n): boolean {\n  // Temporal cutoff — train must be capturedAt < capturedBefore.\n  if (args.split === 'train' && record.capturedAt >= args.capturedBefore) return false\n  if (args.split === 'test' && record.capturedAt < args.capturedBefore) return false\n\n  const f = args.filter\n  if (!f) return true\n  if (f.kind && record.scenario.kind !== f.kind) return false\n  if (f.source) {\n    const sources = Array.isArray(f.source) ? f.source : [f.source]\n    if (!sources.includes(source as never)) return false\n  }\n  if (f.minComposite !== undefined || f.maxComposite !== undefined) {\n    const composites = Object.values(record.judgeScores).map((s) => s.composite)\n    const max = composites.length === 0 ? 0 : Math.max(...composites)\n    if (f.minComposite !== undefined && max < f.minComposite) return false\n    if (f.maxComposite !== undefined && max > f.maxComposite) return false\n  }\n  return true\n}\n\nfunction sha256(input: string): string {\n  return createHash('sha256').update(input).digest('hex').slice(0, 16)\n}\n\nfunction appendLine(path: string, line: string): void {\n  if (existsSync(path)) {\n    const existing = readFileSync(path, 'utf8')\n    writeFileSync(path, existing + line)\n  } else {\n    writeFileSync(path, line)\n  }\n}\n","/**\n * @experimental\n *\n * VCS-pluggable worktree adapter. One improvement = one worktree, PR-like\n * (multiple commits allowed). A code-tier driver's `propose()` creates a\n * worktree, an agent commits the change into it, and `finalize()` returns a\n * `CodeSurface{ worktreeRef }` the measurement checks out to run the worker\n * against the changed code. On promotion the worktree becomes the PR branch.\n *\n * The interface is VCS-agnostic so a future `jj` ([jj-vcs](https://github.com/jj-vcs/jj))\n * adapter can slot in without touching driver code. Only the git adapter\n * ships today. See `docs/design/self-improvement-engine.md`.\n */\n\nimport { execFileSync } from 'node:child_process'\nimport { existsSync } from 'node:fs'\nimport { basename, isAbsolute, join } from 'node:path'\nimport type { CodeSurface } from '../types'\n\nexport interface Worktree {\n  /** Absolute path to the checked-out worktree directory. */\n  path: string\n  /** The branch the worktree is on (becomes the PR branch on promotion). */\n  branch: string\n  /** The ref the worktree was forked from. */\n  baseRef: string\n}\n\nexport interface WorktreeAdapter {\n  /** Create an isolated worktree on a fresh branch off `baseRef`. */\n  create(opts: { baseRef: string; label: string }): Promise<Worktree>\n  /** Commit any pending changes in the worktree, then return a CodeSurface\n   *  pointing at it. The agent has already written its change into\n   *  `worktree.path` by the time this is called. */\n  finalize(worktree: Worktree, summary: string): Promise<CodeSurface>\n  /** Remove the worktree (and its branch) — called for losing candidates. */\n  discard(worktree: Worktree): Promise<void>\n}\n\nexport class WorktreeAdapterError extends Error {\n  constructor(\n    message: string,\n    readonly cause?: unknown,\n  ) {\n    super(message)\n    this.name = 'WorktreeAdapterError'\n  }\n}\n\nexport interface GitWorktreeAdapterOptions {\n  /** Repo root the worktrees fork from. */\n  repoRoot: string\n  /** Directory worktrees are created under. Default: `<repoRoot>/.worktrees`. */\n  worktreeDir?: string\n  /** Branch-name prefix. Default: `improve`. */\n  branchPrefix?: string\n  /** Test seam — defaults to a real `git` runner. */\n  git?: (args: string[], cwd: string) => string\n}\n\nfunction defaultGit(args: string[], cwd: string): string {\n  try {\n    return execFileSync('git', args, { cwd, encoding: 'utf8' }).trim()\n  } catch (err) {\n    const stderr =\n      err && typeof err === 'object' && 'stderr' in err\n        ? String((err as { stderr: unknown }).stderr)\n        : ''\n    throw new WorktreeAdapterError(`git ${args.join(' ')} failed: ${stderr || String(err)}`, err)\n  }\n}\n\n/** Slugify a label into a branch-safe segment. */\nfunction slug(label: string): string {\n  return (\n    label\n      .toLowerCase()\n      .replace(/[^a-z0-9]+/g, '-')\n      .replace(/^-+|-+$/g, '')\n      .slice(0, 48) || 'candidate'\n  )\n}\n\nexport function gitWorktreeAdapter(opts: GitWorktreeAdapterOptions): WorktreeAdapter {\n  const git = opts.git ?? defaultGit\n  const worktreeDir = opts.worktreeDir ?? join(opts.repoRoot, '.worktrees')\n  const branchPrefix = opts.branchPrefix ?? 'improve'\n\n  return {\n    async create({ baseRef, label }) {\n      const id = `${slug(label)}-${Date.now().toString(36)}-${Math.random().toString(36).slice(2, 6)}`\n      const branch = `${branchPrefix}/${id}`\n      const path = join(worktreeDir, id)\n      git(['worktree', 'add', '-b', branch, path, baseRef], opts.repoRoot)\n      return { path, branch, baseRef }\n    },\n\n    async finalize(worktree, summary) {\n      // Stage + commit any pending changes the agent left in the worktree.\n      // A no-op commit is refused by git, so only commit when the tree is dirty.\n      const status = git(['status', '--porcelain'], worktree.path)\n      if (status.length > 0) {\n        git(['add', '-A'], worktree.path)\n        git(['commit', '-m', summary], worktree.path)\n      }\n      return {\n        kind: 'code',\n        worktreeRef: worktree.path,\n        baseRef: worktree.baseRef,\n        summary,\n      }\n    },\n\n    async discard(worktree) {\n      // Remove the worktree, then delete its branch. Force-remove because the\n      // worktree may hold uncommitted experiment state we're discarding.\n      git(['worktree', 'remove', '--force', worktree.path], opts.repoRoot)\n      git(['branch', '-D', worktree.branch], opts.repoRoot)\n    },\n  }\n}\n\n/** Resolve a `CodeSurface`'s worktreeRef to a directory the measurement can\n *  run the worker in. A path ref is returned as-is; anything else is treated\n *  as a ref under the adapter's worktree dir. */\nexport function resolveWorktreePath(surface: CodeSurface, worktreeDir?: string): string {\n  if (isAbsolute(surface.worktreeRef) && existsSync(surface.worktreeRef)) return surface.worktreeRef\n  if (worktreeDir) return join(worktreeDir, basename(surface.worktreeRef))\n  return surface.worktreeRef\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;AA6BA,SAAS,kBAAkB;AAC3B,SAAS,YAAY,WAAW,cAAc,qBAAqB;AACnE,SAAS,YAAY;AAmBd,IAAM,4BAAN,cAAwC,MAAM;AAAA,EACnD,YACkB,MAChB,SACA;AACA,UAAM,OAAO;AAHG;AAIhB,SAAK,OAAO;AAAA,EACd;AAAA,EALkB;AAMpB;AAQO,IAAM,yBAAN,MAA6D;AAAA,EAIlE,YAA6B,SAAwC;AAAxC;AAC3B,QAAI,CAAC,WAAW,QAAQ,IAAI,EAAG,WAAU,QAAQ,MAAM,EAAE,WAAW,KAAK,CAAC;AAC1E,SAAK,MAAM,QAAQ,OAAO,KAAK;AAAA,EACjC;AAAA,EAH6B;AAAA,EAHZ;AAAA,EACA,aAAa,oBAAI,IAA4B;AAAA,EAO9D,MAAM,QAAQ,OAA4C;AACxD,SAAK,iBAAiB,KAAK;AAC3B,SAAK,gBAAgB,KAAK;AAC1B,UAAM,SAAS,KAAK,SAAS,KAAK;AAClC,UAAM,OAAO,KAAK,cAAc,MAAM,MAAM;AAC5C,UAAM,OAAO,GAAG,KAAK,UAAU,MAAM,CAAC;AAAA;AAItC,eAAW,MAAM,IAAI;AAAA,EACvB;AAAA,EAEA,MAAM,OAAO,MAAmE;AAC9E,QAAI,CAAC,KAAK,OAAO;AACf,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AACA,QAAI,CAAC,KAAK,gBAAgB;AACxB,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AAEA,UAAM,MAA+B,CAAC;AACtC,eAAW,UAAU,aAAa;AAGhC,UAAI,KAAK,UAAU,WAAW,WAAW,oBAAoB;AAC3D,cAAM,WAAW,qBAAqB,KAAK,QAAQ,QAAQ,kBAAkB;AAC7E,YAAI,CAAC,SAAU;AAAA,MACjB;AACA,YAAM,OAAO,KAAK,cAAc,MAAM;AACtC,UAAI,CAAC,WAAW,IAAI,EAAG;AACvB,YAAM,QAAQ,aAAa,MAAM,MAAM,EAAE,MAAM,IAAI,EAAE,OAAO,OAAO;AACnE,iBAAW,QAAQ,OAAO;AACxB,YAAI;AACJ,YAAI;AACF,mBAAS,KAAK,MAAM,IAAI;AAAA,QAC1B,QAAQ;AACN;AAAA,QACF;AACA,YAAI,CAAC,cAAc,QAAQ,MAAM,MAAM,EAAG;AAC1C,YAAI,KAAK,MAAM;AAAA,MACjB;AAAA,IACF;AAGA,QAAI,KAAK,CAAC,GAAG,MAAM;AACjB,UAAI,EAAE,eAAe,EAAE,WAAY,QAAO,EAAE,WAAW,cAAc,EAAE,UAAU;AACjF,aAAO,EAAE,WAAW,cAAc,EAAE,UAAU;AAAA,IAChD,CAAC;AAED,WAAO,IAAI,MAAM,GAAG,KAAK,KAAK;AAAA,EAChC;AAAA,EAEA,MAAM,OAAmF;AACvF,UAAM,WAAmC,CAAC;AAC1C,QAAI,QAAQ;AACZ,eAAW,UAAU,aAAa;AAChC,YAAM,OAAO,KAAK,cAAc,MAAM;AACtC,UAAI,CAAC,WAAW,IAAI,GAAG;AACrB,iBAAS,MAAM,IAAI;AACnB;AAAA,MACF;AACA,YAAM,QAAQ,aAAa,MAAM,MAAM,EAAE,MAAM,IAAI,EAAE,OAAO,OAAO,EAAE;AACrE,eAAS,MAAM,IAAI;AACnB,eAAS;AAAA,IACX;AAIA,WAAO,EAAE,OAAO,OAAO,MAAM,OAAO,SAAS;AAAA,EAC/C;AAAA,EAEQ,iBAAiB,OAAmC;AAC1D,QAAI,CAAC,MAAM,QAAQ;AACjB,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AACA,QAAI,CAAC,MAAM,qBAAqB,MAAM,kBAAkB,WAAW,GAAG;AACpE,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AACA,QAAI,CAAC,MAAM,YAAY;AACrB,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AACA,QAAI,CAAC,MAAM,iBAAiB;AAC1B,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AACA,QAAI,CAAC,YAAY,SAAS,MAAM,MAAM,GAAG;AACvC,YAAM,IAAI;AAAA,QACR;AAAA,QACA,+CAA+C,YAAY,KAAK,IAAI,CAAC;AAAA,MACvE;AAAA,IACF;AAAA,EACF;AAAA,EAEQ,gBAAgB,OAAmC;AACzD,UAAM,MAAM,KAAK,QAAQ;AACzB,QAAI,CAAC,OAAO,CAAC,MAAM,gBAAiB;AACpC,UAAM,MAAM,KAAK,IAAI;AACrB,UAAM,WAAW;AACjB,QAAI,QAAQ,KAAK,WAAW,IAAI,MAAM,eAAe;AACrD,QAAI,CAAC,SAAS,MAAM,MAAM,iBAAiB,UAAU;AACnD,cAAQ,EAAE,QAAQ,MAAM,iBAAiB,eAAe,KAAK,OAAO,EAAE;AACtE,WAAK,WAAW,IAAI,MAAM,iBAAiB,KAAK;AAAA,IAClD;AACA,QAAI,MAAM,SAAS,KAAK;AACtB,YAAM,IAAI;AAAA,QACR;AAAA,QACA,gCAAgC,MAAM,eAAe,aAAa,GAAG;AAAA,MACvE;AAAA,IACF;AACA,UAAM,SAAS;AAAA,EACjB;AAAA,EAEQ,SAAS,OAAoD;AACnE,UAAM,aAAa;AAAA,MACjB,KAAK,UAAU;AAAA,QACb,IAAI,MAAM,SAAS;AAAA,QACnB,KAAK,MAAM;AAAA,QACX,IAAI,MAAM;AAAA,QACV,KAAK,MAAM;AAAA,MACb,CAAC;AAAA,IACH;AAGA,WAAO;AAAA,MACL,GAAG;AAAA,MACH;AAAA,MACA,OAAO;AAAA,IACT;AAAA,EACF;AAAA,EAEQ,cAAc,QAAwB;AAC5C,WAAO,KAAK,KAAK,QAAQ,MAAM,GAAG,MAAM,QAAQ;AAAA,EAClD;AACF;AAEA,IAAM,cAAgD;AAAA,EACpD;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF;AAEA,SAAS,qBACP,QACA,QACS;AACT,MAAI,CAAC,OAAQ,QAAO;AACpB,MAAI,MAAM,QAAQ,MAAM,EAAG,QAAO,OAAO,SAAS,MAAM;AACxD,SAAO,WAAW;AACpB;AAEA,SAAS,cACP,QACA,MACA,QACS;AAET,MAAI,KAAK,UAAU,WAAW,OAAO,cAAc,KAAK,eAAgB,QAAO;AAC/E,MAAI,KAAK,UAAU,UAAU,OAAO,aAAa,KAAK,eAAgB,QAAO;AAE7E,QAAM,IAAI,KAAK;AACf,MAAI,CAAC,EAAG,QAAO;AACf,MAAI,EAAE,QAAQ,OAAO,SAAS,SAAS,EAAE,KAAM,QAAO;AACtD,MAAI,EAAE,QAAQ;AACZ,UAAM,UAAU,MAAM,QAAQ,EAAE,MAAM,IAAI,EAAE,SAAS,CAAC,EAAE,MAAM;AAC9D,QAAI,CAAC,QAAQ,SAAS,MAAe,EAAG,QAAO;AAAA,EACjD;AACA,MAAI,EAAE,iBAAiB,UAAa,EAAE,iBAAiB,QAAW;AAChE,UAAM,aAAa,OAAO,OAAO,OAAO,WAAW,EAAE,IAAI,CAAC,MAAM,EAAE,SAAS;AAC3E,UAAM,MAAM,WAAW,WAAW,IAAI,IAAI,KAAK,IAAI,GAAG,UAAU;AAChE,QAAI,EAAE,iBAAiB,UAAa,MAAM,EAAE,aAAc,QAAO;AACjE,QAAI,EAAE,iBAAiB,UAAa,MAAM,EAAE,aAAc,QAAO;AAAA,EACnE;AACA,SAAO;AACT;AAEA,SAAS,OAAO,OAAuB;AACrC,SAAO,WAAW,QAAQ,EAAE,OAAO,KAAK,EAAE,OAAO,KAAK,EAAE,MAAM,GAAG,EAAE;AACrE;AAEA,SAAS,WAAW,MAAc,MAAoB;AACpD,MAAI,WAAW,IAAI,GAAG;AACpB,UAAM,WAAW,aAAa,MAAM,MAAM;AAC1C,kBAAc,MAAM,WAAW,IAAI;AAAA,EACrC,OAAO;AACL,kBAAc,MAAM,IAAI;AAAA,EAC1B;AACF;;;AC1QA,SAAS,oBAAoB;AAC7B,SAAS,cAAAA,mBAAkB;AAC3B,SAAS,UAAU,YAAY,QAAAC,aAAY;AAuBpC,IAAM,uBAAN,cAAmC,MAAM;AAAA,EAC9C,YACE,SACS,OACT;AACA,UAAM,OAAO;AAFJ;AAGT,SAAK,OAAO;AAAA,EACd;AAAA,EAJW;AAKb;AAaA,SAAS,WAAW,MAAgB,KAAqB;AACvD,MAAI;AACF,WAAO,aAAa,OAAO,MAAM,EAAE,KAAK,UAAU,OAAO,CAAC,EAAE,KAAK;AAAA,EACnE,SAAS,KAAK;AACZ,UAAM,SACJ,OAAO,OAAO,QAAQ,YAAY,YAAY,MAC1C,OAAQ,IAA4B,MAAM,IAC1C;AACN,UAAM,IAAI,qBAAqB,OAAO,KAAK,KAAK,GAAG,CAAC,YAAY,UAAU,OAAO,GAAG,CAAC,IAAI,GAAG;AAAA,EAC9F;AACF;AAGA,SAAS,KAAK,OAAuB;AACnC,SACE,MACG,YAAY,EACZ,QAAQ,eAAe,GAAG,EAC1B,QAAQ,YAAY,EAAE,EACtB,MAAM,GAAG,EAAE,KAAK;AAEvB;AAEO,SAAS,mBAAmB,MAAkD;AACnF,QAAM,MAAM,KAAK,OAAO;AACxB,QAAM,cAAc,KAAK,eAAeA,MAAK,KAAK,UAAU,YAAY;AACxE,QAAM,eAAe,KAAK,gBAAgB;AAE1C,SAAO;AAAA,IACL,MAAM,OAAO,EAAE,SAAS,MAAM,GAAG;AAC/B,YAAM,KAAK,GAAG,KAAK,KAAK,CAAC,IAAI,KAAK,IAAI,EAAE,SAAS,EAAE,CAAC,IAAI,KAAK,OAAO,EAAE,SAAS,EAAE,EAAE,MAAM,GAAG,CAAC,CAAC;AAC9F,YAAM,SAAS,GAAG,YAAY,IAAI,EAAE;AACpC,YAAM,OAAOA,MAAK,aAAa,EAAE;AACjC,UAAI,CAAC,YAAY,OAAO,MAAM,QAAQ,MAAM,OAAO,GAAG,KAAK,QAAQ;AACnE,aAAO,EAAE,MAAM,QAAQ,QAAQ;AAAA,IACjC;AAAA,IAEA,MAAM,SAAS,UAAU,SAAS;AAGhC,YAAM,SAAS,IAAI,CAAC,UAAU,aAAa,GAAG,SAAS,IAAI;AAC3D,UAAI,OAAO,SAAS,GAAG;AACrB,YAAI,CAAC,OAAO,IAAI,GAAG,SAAS,IAAI;AAChC,YAAI,CAAC,UAAU,MAAM,OAAO,GAAG,SAAS,IAAI;AAAA,MAC9C;AACA,aAAO;AAAA,QACL,MAAM;AAAA,QACN,aAAa,SAAS;AAAA,QACtB,SAAS,SAAS;AAAA,QAClB;AAAA,MACF;AAAA,IACF;AAAA,IAEA,MAAM,QAAQ,UAAU;AAGtB,UAAI,CAAC,YAAY,UAAU,WAAW,SAAS,IAAI,GAAG,KAAK,QAAQ;AACnE,UAAI,CAAC,UAAU,MAAM,SAAS,MAAM,GAAG,KAAK,QAAQ;AAAA,IACtD;AAAA,EACF;AACF;AAKO,SAAS,oBAAoB,SAAsB,aAA8B;AACtF,MAAI,WAAW,QAAQ,WAAW,KAAKD,YAAW,QAAQ,WAAW,EAAG,QAAO,QAAQ;AACvF,MAAI,YAAa,QAAOC,MAAK,aAAa,SAAS,QAAQ,WAAW,CAAC;AACvE,SAAO,QAAQ;AACjB;","names":["existsSync","join"]}
+{"version":3,"sources":["../../src/campaign/drivers/skillopt.ts","../../src/campaign/labeled-store/fs-adapter.ts","../../src/campaign/worktree/index.ts"],"sourcesContent":["/**\n * @experimental\n *\n * `skillOptDriver` — a section-aware, bounded-edit `ImprovementDriver` for\n * structured natural-language procedures (SKILL.md files, runbooks, sectioned\n * system prompts, judge rubrics with dimensions). Implements the SkillOpt\n * methodology (Microsoft, 2026): treat the skill document as a trainable\n * optimization target, train the procedure not the weights, constrain each\n * generation to ≤N targeted edits to prevent useful-rule overwrites.\n *\n * Differs from `gepaDriver` in two specific ways:\n *\n * 1. **Bounded edits.** Each candidate must differ from the baseline by at\n *    most `editBudget` sentence-level changes. The \"edit budget functions\n *    as a textual learning rate\" — without it, an LLM proposal can rewrite\n *    so much that useful prior rules get overwritten.\n *\n * 2. **Section preservation.** When the surface is a structured doc, the\n *    H2 headers (and an opt-in `preserveSections` allowlist) are\n *    load-bearing for discoverability. Candidates that delete or rename\n *    preserved sections are rejected at parse time.\n *\n * Selectable alongside `gepaDriver` and `evolutionaryDriver`. Use this when\n * the surface IS a structured doc; use `gepaDriver` when the surface is\n * unstructured prose.\n */\n\nimport { callLlm, type LlmClientOptions } from '../../llm-client'\nimport {\n  buildReflectionPrompt,\n  parseReflectionResponse,\n  type TrialTrace,\n} from '../../reflective-mutation'\nimport type { ImprovementDriver, MutableSurface, ProposeContext } from '../types'\n\nconst REFLECTION_SYSTEM =\n  'You are an expert prompt engineer applying the SkillOpt methodology. ' +\n  'You will edit a structured natural-language procedure under TWO HARD ' +\n  'CONSTRAINTS: (1) preserve every H2 section heading verbatim — do NOT ' +\n  'delete, rename, or merge sections; (2) make at most EDIT_BUDGET targeted ' +\n  'sentence-level edits per candidate — bounded edits prevent overwriting ' +\n  'useful prior rules. Output ONLY a JSON object of shape ' +\n  '{\"proposals\":[{\"label\":string,\"rationale\":string,\"payload\":string}]} ' +\n  'where each `payload` is the FULL improved skill text. No prose outside the JSON.'\n\nexport interface SkillOptDriverOptions {\n  /** Router transport (apiKey/baseUrl). */\n  llm: LlmClientOptions\n  /** Model that performs the reflection. */\n  model: string\n  /** What is being optimized — appears in the reflection prompt for orientation. */\n  target: string\n\n  /** Max edits per generation — SkillOpt's \"textual learning rate\".\n   *  Default 3. Lower = more conservative, higher = more exploratory. */\n  editBudget?: number\n\n  /** Section headings the driver MUST preserve. When the surface is a\n   *  structured skill doc, sections are load-bearing for discoverability.\n   *  Default: auto-detected from H2 headers in the baseline. */\n  preserveSections?: string[]\n\n  /** Surface-specific mutation levers offered to the model. */\n  mutationPrimitives?: string[]\n  /** Top/bottom scenarios surfaced as evidence each generation. Default 3. */\n  evidenceK?: number\n  /** Reflection sampling temperature. Default 0.7. */\n  temperature?: number\n  /** Reflection max tokens. Default 6000. */\n  maxTokens?: number\n}\n\n/** Internal — exported for tests. */\nexport function extractH2Sections(text: string): string[] {\n  const out: string[] = []\n  for (const line of text.split('\\n')) {\n    const match = /^##\\s+(.+?)\\s*$/.exec(line)\n    if (match) out.push(match[1]!)\n  }\n  return out\n}\n\n/** Sentence-level edit distance. Counts distinct sentence add/remove/replace\n *  ops between baseline and candidate using a normalised line-by-line diff.\n *  Imperfect (treats trivial whitespace as identical) but tight enough to\n *  bound an LLM rewrite. Exported for tests. */\nexport function countSentenceEdits(baseline: string, candidate: string): number {\n  const norm = (s: string) =>\n    s\n      .split(/(?<=[.!?])\\s+|\\n/g)\n      .map((p) => p.trim())\n      .filter((p) => p.length > 0)\n  const a = new Set(norm(baseline))\n  const b = new Set(norm(candidate))\n  let edits = 0\n  for (const s of a) if (!b.has(s)) edits++ // deletions\n  for (const s of b) if (!a.has(s)) edits++ // additions\n  return edits\n}\n\nexport function skillOptDriver(opts: SkillOptDriverOptions): ImprovementDriver {\n  const evidenceK = opts.evidenceK ?? 3\n  const editBudget = opts.editBudget ?? 3\n  if (editBudget < 1) {\n    throw new Error(\n      `skillOptDriver: editBudget must be >= 1, got ${editBudget} (use evolutionaryDriver with a noop mutator for measure-only runs)`,\n    )\n  }\n  return {\n    kind: 'skillopt',\n    async propose(ctx: ProposeContext): Promise<MutableSurface[]> {\n      if (typeof ctx.currentSurface !== 'string') {\n        throw new Error(\n          `skillOptDriver: surface must be a string skill document; got ${typeof ctx.currentSurface}. Use evolutionaryDriver with a CodeSurface mutator for code-tier surfaces.`,\n        )\n      }\n      const baseline = ctx.currentSurface\n      const preserveSections = opts.preserveSections ?? extractH2Sections(baseline)\n\n      const { top, bottom, target } = buildEvidence(ctx, evidenceK, opts.target)\n\n      const reflectionUser = buildReflectionPrompt({\n        target,\n        parentPayload: baseline,\n        topTrials: top,\n        bottomTrials: bottom,\n        childCount: ctx.populationSize,\n        mutationPrimitives: opts.mutationPrimitives,\n      })\n      const constraintPreamble = [\n        '',\n        '## SkillOpt constraints (hard rules — violations rejected)',\n        '',\n        `- Edit budget: at most ${editBudget} sentence-level edits per candidate.`,\n        '- Section preservation: every H2 heading below must appear unchanged in your output.',\n        ...preserveSections.map((s) => `  - \\`## ${s}\\``),\n        '',\n        'Reject any candidate in your own thinking that would delete a section, rename a heading, or exceed the edit budget. Make TARGETED, surgical edits — not rewrites.',\n        '',\n      ].join('\\n')\n      const userPrompt = `${reflectionUser}${constraintPreamble}`\n      const system = REFLECTION_SYSTEM.replace('EDIT_BUDGET', String(editBudget))\n\n      const result = await callLlm(\n        {\n          model: opts.model,\n          messages: [\n            { role: 'system', content: system },\n            { role: 'user', content: userPrompt },\n          ],\n          jsonMode: true,\n          temperature: opts.temperature ?? 0.7,\n          maxTokens: opts.maxTokens ?? 6000,\n        },\n        opts.llm,\n      )\n\n      const proposals = parseReflectionResponse(result.content, ctx.populationSize)\n      const out: MutableSurface[] = []\n      for (const proposal of proposals) {\n        const text = typeof proposal.payload === 'string' ? proposal.payload.trim() : ''\n        if (!text || text === baseline) continue\n        if (!validateSections(text, preserveSections)) continue\n        if (countSentenceEdits(baseline, text) > editBudget * 2) continue // x2: add+remove pair per edit\n        if (out.includes(text)) continue\n        out.push(text)\n      }\n      return out\n    },\n  }\n}\n\nfunction validateSections(candidate: string, required: string[]): boolean {\n  if (required.length === 0) return true\n  const have = new Set(extractH2Sections(candidate))\n  for (const section of required) {\n    if (!have.has(section)) return false\n  }\n  return true\n}\n\n/** Reused from gepaDriver pattern — build evidence from prior best candidate. */\nfunction buildEvidence(\n  ctx: ProposeContext,\n  evidenceK: number,\n  baseTarget: string,\n): { top: TrialTrace[]; bottom: TrialTrace[]; target: string } {\n  const last = ctx.history.at(-1)\n  if (!last || last.candidates.length === 0) {\n    return { top: [], bottom: [], target: baseTarget }\n  }\n  const best = [...last.candidates].sort((a, b) => b.composite - a.composite)[0]\n  if (!best) return { top: [], bottom: [], target: baseTarget }\n\n  const byScore = [...best.scenarios].sort((a, b) => b.composite - a.composite)\n  const toTrace = (s: { scenarioId: string; composite: number }): TrialTrace => ({\n    id: s.scenarioId,\n    score: s.composite,\n  })\n  const top = byScore.slice(0, evidenceK).map(toTrace)\n  const bottom = byScore.slice(-evidenceK).reverse().map(toTrace)\n\n  const weakest = Object.entries(best.dimensions)\n    .sort((a, b) => a[1] - b[1])\n    .slice(0, 3)\n    .map(([dim, value]) => `${dim} (${value.toFixed(2)})`)\n  const target =\n    weakest.length > 0 ? `${baseTarget} — weakest dimensions: ${weakest.join(', ')}` : baseTarget\n\n  return { top, bottom, target }\n}\n","/**\n * @experimental\n *\n * Filesystem `LabeledScenarioStore` adapter. The default capture sink for\n * traces + eval artifacts. Production deployments typically swap for a\n * Turso/SQLite adapter (same interface).\n *\n * Records land as one JSONL file per source under `<root>/<source>.jsonl`.\n * Each line is a `LabeledScenarioRecord`. Append-only — no in-place edits.\n *\n * Safety properties enforced at write-time:\n *\n *   - **Provenance required**: writes without `source`, `sourceVersionHash`,\n *     `capturedAt`, `redactionStatus` are rejected. Closes the alignment\n *     reviewer's data-poisoning gap.\n *   - **Per-source rate limits**: optional `rateLimitBucket` + `maxWritesPerMinute`\n *     stops a single tenant/source from flooding the store.\n *\n * Safety properties enforced at sample-time:\n *\n *   - **Required split + capturedBefore**: substrate refuses to sample without\n *     an explicit `split` ('train' | 'test') AND a temporal cutoff. Eliminates\n *     accidental train/test contamination.\n *   - **Default training-source filter**: when the store is sampled with\n *     `split: 'train'`, production-trace records are EXCLUDED unless the\n *     caller passes `filter.source: 'production-trace'` explicitly. Closes\n *     the contamination-by-default gap flagged by the senior eval engineer.\n */\n\nimport { createHash } from 'node:crypto'\nimport { existsSync, mkdirSync, readFileSync, writeFileSync } from 'node:fs'\nimport { join } from 'node:path'\nimport type {\n  LabeledScenarioRecord,\n  LabeledScenarioSampleArgs,\n  LabeledScenarioSource,\n  LabeledScenarioStore,\n  LabeledScenarioWrite,\n} from '../types'\n\nexport interface FsLabeledScenarioStoreOptions {\n  /** Root directory for JSONL files. Created if missing. */\n  root: string\n  /** Per-source rate limit. When set, writes exceeding the cap are rejected\n   *  with a typed error. Default: no limit. */\n  maxWritesPerMinutePerBucket?: number\n  /** Test seam — override `Date.now()` for deterministic tests. */\n  now?: () => number\n}\n\nexport class LabeledScenarioStoreError extends Error {\n  constructor(\n    public readonly code: string,\n    message: string,\n  ) {\n    super(message)\n    this.name = 'LabeledScenarioStoreError'\n  }\n}\n\ninterface RateLimitState {\n  bucket: string\n  windowStartMs: number\n  count: number\n}\n\nexport class FsLabeledScenarioStore implements LabeledScenarioStore {\n  private readonly now: () => number\n  private readonly rateLimits = new Map<string, RateLimitState>()\n\n  constructor(private readonly options: FsLabeledScenarioStoreOptions) {\n    if (!existsSync(options.root)) mkdirSync(options.root, { recursive: true })\n    this.now = options.now ?? Date.now\n  }\n\n  async observe(write: LabeledScenarioWrite): Promise<void> {\n    this.assertProvenance(write)\n    this.assertRateLimit(write)\n    const record = this.toRecord(write)\n    const path = this.pathForSource(write.source)\n    const line = `${JSON.stringify(record)}\\n`\n    // Append atomically. For high-throughput a writev-friendly buffered\n    // implementation lands in the Turso adapter; FS adapter is for tests +\n    // local dev + small workloads.\n    appendLine(path, line)\n  }\n\n  async sample(args: LabeledScenarioSampleArgs): Promise<LabeledScenarioRecord[]> {\n    if (!args.split) {\n      throw new LabeledScenarioStoreError(\n        'split_required',\n        'sample() requires an explicit `split` (train | test) — substrate refuses ambiguous reads',\n      )\n    }\n    if (!args.capturedBefore) {\n      throw new LabeledScenarioStoreError(\n        'capturedBefore_required',\n        'sample() requires an explicit `capturedBefore` timestamp for temporal-split discipline',\n      )\n    }\n\n    const all: LabeledScenarioRecord[] = []\n    for (const source of ALL_SOURCES) {\n      // Default training-source filter: when sampling train, EXCLUDE\n      // production-trace records unless the caller asks for them.\n      if (args.split === 'train' && source === 'production-trace') {\n        const explicit = sourceFilterContains(args.filter?.source, 'production-trace')\n        if (!explicit) continue\n      }\n      const path = this.pathForSource(source)\n      if (!existsSync(path)) continue\n      const lines = readFileSync(path, 'utf8').split('\\n').filter(Boolean)\n      for (const line of lines) {\n        let record: LabeledScenarioRecord\n        try {\n          record = JSON.parse(line) as LabeledScenarioRecord\n        } catch {\n          continue\n        }\n        if (!matchesFilter(record, args, source)) continue\n        all.push(record)\n      }\n    }\n\n    // Deterministic order: by capturedAt ascending, then recordHash.\n    all.sort((a, b) => {\n      if (a.capturedAt !== b.capturedAt) return a.capturedAt.localeCompare(b.capturedAt)\n      return a.recordHash.localeCompare(b.recordHash)\n    })\n\n    return all.slice(0, args.count)\n  }\n\n  async size(): Promise<{ train: number; test: number; bySource: Record<string, number> }> {\n    const bySource: Record<string, number> = {}\n    let total = 0\n    for (const source of ALL_SOURCES) {\n      const path = this.pathForSource(source)\n      if (!existsSync(path)) {\n        bySource[source] = 0\n        continue\n      }\n      const count = readFileSync(path, 'utf8').split('\\n').filter(Boolean).length\n      bySource[source] = count\n      total += count\n    }\n    // FS adapter doesn't track split assignments per-record (split is\n    // computed at sample-time based on `capturedBefore`). For size(), we\n    // report `train`+`test` as the same total — split is a sampling concept.\n    return { train: total, test: total, bySource }\n  }\n\n  private assertProvenance(write: LabeledScenarioWrite): void {\n    if (!write.source) {\n      throw new LabeledScenarioStoreError(\n        'missing_source',\n        'LabeledScenarioWrite requires `source`',\n      )\n    }\n    if (!write.sourceVersionHash || write.sourceVersionHash.length === 0) {\n      throw new LabeledScenarioStoreError(\n        'missing_source_version',\n        'LabeledScenarioWrite requires `sourceVersionHash` (git sha or substrate version)',\n      )\n    }\n    if (!write.capturedAt) {\n      throw new LabeledScenarioStoreError(\n        'missing_captured_at',\n        'LabeledScenarioWrite requires `capturedAt` ISO timestamp',\n      )\n    }\n    if (!write.redactionStatus) {\n      throw new LabeledScenarioStoreError(\n        'missing_redaction_status',\n        'LabeledScenarioWrite requires explicit `redactionStatus` — raw / redacted-pii / redacted-secrets / fully-redacted',\n      )\n    }\n    if (!ALL_SOURCES.includes(write.source)) {\n      throw new LabeledScenarioStoreError(\n        'unknown_source',\n        `LabeledScenarioWrite.source must be one of: ${ALL_SOURCES.join(', ')}`,\n      )\n    }\n  }\n\n  private assertRateLimit(write: LabeledScenarioWrite): void {\n    const cap = this.options.maxWritesPerMinutePerBucket\n    if (!cap || !write.rateLimitBucket) return\n    const now = this.now()\n    const windowMs = 60_000\n    let state = this.rateLimits.get(write.rateLimitBucket)\n    if (!state || now - state.windowStartMs >= windowMs) {\n      state = { bucket: write.rateLimitBucket, windowStartMs: now, count: 0 }\n      this.rateLimits.set(write.rateLimitBucket, state)\n    }\n    if (state.count >= cap) {\n      throw new LabeledScenarioStoreError(\n        'rate_limit_exceeded',\n        `LabeledScenarioStore: bucket ${write.rateLimitBucket} exceeded ${cap} writes/min`,\n      )\n    }\n    state.count += 1\n  }\n\n  private toRecord(write: LabeledScenarioWrite): LabeledScenarioRecord {\n    const recordHash = sha256(\n      JSON.stringify({\n        id: write.scenario.id,\n        src: write.source,\n        at: write.capturedAt,\n        ver: write.sourceVersionHash,\n      }),\n    )\n    // FS adapter assigns split at sample-time, but we cache a hint here\n    // based on capturedAt vs the world's \"now\" — sampler overrides this.\n    return {\n      ...write,\n      recordHash,\n      split: 'train',\n    }\n  }\n\n  private pathForSource(source: string): string {\n    return join(this.options.root, `${source}.jsonl`)\n  }\n}\n\nconst ALL_SOURCES: LabeledScenarioWrite['source'][] = [\n  'production-trace',\n  'eval-run',\n  'manual',\n  'red-team',\n  'synthetic',\n]\n\nfunction sourceFilterContains(\n  filter: LabeledScenarioSource | LabeledScenarioSource[] | undefined,\n  needle: LabeledScenarioSource,\n): boolean {\n  if (!filter) return false\n  if (Array.isArray(filter)) return filter.includes(needle)\n  return filter === needle\n}\n\nfunction matchesFilter(\n  record: LabeledScenarioRecord,\n  args: LabeledScenarioSampleArgs,\n  source: string,\n): boolean {\n  // Temporal cutoff — train must be capturedAt < capturedBefore.\n  if (args.split === 'train' && record.capturedAt >= args.capturedBefore) return false\n  if (args.split === 'test' && record.capturedAt < args.capturedBefore) return false\n\n  const f = args.filter\n  if (!f) return true\n  if (f.kind && record.scenario.kind !== f.kind) return false\n  if (f.source) {\n    const sources = Array.isArray(f.source) ? f.source : [f.source]\n    if (!sources.includes(source as never)) return false\n  }\n  if (f.minComposite !== undefined || f.maxComposite !== undefined) {\n    const composites = Object.values(record.judgeScores).map((s) => s.composite)\n    const max = composites.length === 0 ? 0 : Math.max(...composites)\n    if (f.minComposite !== undefined && max < f.minComposite) return false\n    if (f.maxComposite !== undefined && max > f.maxComposite) return false\n  }\n  return true\n}\n\nfunction sha256(input: string): string {\n  return createHash('sha256').update(input).digest('hex').slice(0, 16)\n}\n\nfunction appendLine(path: string, line: string): void {\n  if (existsSync(path)) {\n    const existing = readFileSync(path, 'utf8')\n    writeFileSync(path, existing + line)\n  } else {\n    writeFileSync(path, line)\n  }\n}\n","/**\n * @experimental\n *\n * VCS-pluggable worktree adapter. One improvement = one worktree, PR-like\n * (multiple commits allowed). A code-tier driver's `propose()` creates a\n * worktree, an agent commits the change into it, and `finalize()` returns a\n * `CodeSurface{ worktreeRef }` the measurement checks out to run the worker\n * against the changed code. On promotion the worktree becomes the PR branch.\n *\n * The interface is VCS-agnostic so a future `jj` ([jj-vcs](https://github.com/jj-vcs/jj))\n * adapter can slot in without touching driver code. Only the git adapter\n * ships today. See `docs/design/self-improvement-engine.md`.\n */\n\nimport { execFileSync } from 'node:child_process'\nimport { existsSync } from 'node:fs'\nimport { basename, isAbsolute, join } from 'node:path'\nimport type { CodeSurface } from '../types'\n\nexport interface Worktree {\n  /** Absolute path to the checked-out worktree directory. */\n  path: string\n  /** The branch the worktree is on (becomes the PR branch on promotion). */\n  branch: string\n  /** The ref the worktree was forked from. */\n  baseRef: string\n}\n\nexport interface WorktreeAdapter {\n  /** Create an isolated worktree on a fresh branch off `baseRef`. */\n  create(opts: { baseRef: string; label: string }): Promise<Worktree>\n  /** Commit any pending changes in the worktree, then return a CodeSurface\n   *  pointing at it. The agent has already written its change into\n   *  `worktree.path` by the time this is called. */\n  finalize(worktree: Worktree, summary: string): Promise<CodeSurface>\n  /** Remove the worktree (and its branch) — called for losing candidates. */\n  discard(worktree: Worktree): Promise<void>\n}\n\nexport class WorktreeAdapterError extends Error {\n  constructor(\n    message: string,\n    readonly cause?: unknown,\n  ) {\n    super(message)\n    this.name = 'WorktreeAdapterError'\n  }\n}\n\nexport interface GitWorktreeAdapterOptions {\n  /** Repo root the worktrees fork from. */\n  repoRoot: string\n  /** Directory worktrees are created under. Default: `<repoRoot>/.worktrees`. */\n  worktreeDir?: string\n  /** Branch-name prefix. Default: `improve`. */\n  branchPrefix?: string\n  /** Test seam — defaults to a real `git` runner. */\n  git?: (args: string[], cwd: string) => string\n}\n\nfunction defaultGit(args: string[], cwd: string): string {\n  try {\n    return execFileSync('git', args, { cwd, encoding: 'utf8' }).trim()\n  } catch (err) {\n    const stderr =\n      err && typeof err === 'object' && 'stderr' in err\n        ? String((err as { stderr: unknown }).stderr)\n        : ''\n    throw new WorktreeAdapterError(`git ${args.join(' ')} failed: ${stderr || String(err)}`, err)\n  }\n}\n\n/** Slugify a label into a branch-safe segment. */\nfunction slug(label: string): string {\n  return (\n    label\n      .toLowerCase()\n      .replace(/[^a-z0-9]+/g, '-')\n      .replace(/^-+|-+$/g, '')\n      .slice(0, 48) || 'candidate'\n  )\n}\n\nexport function gitWorktreeAdapter(opts: GitWorktreeAdapterOptions): WorktreeAdapter {\n  const git = opts.git ?? defaultGit\n  const worktreeDir = opts.worktreeDir ?? join(opts.repoRoot, '.worktrees')\n  const branchPrefix = opts.branchPrefix ?? 'improve'\n\n  return {\n    async create({ baseRef, label }) {\n      const id = `${slug(label)}-${Date.now().toString(36)}-${Math.random().toString(36).slice(2, 6)}`\n      const branch = `${branchPrefix}/${id}`\n      const path = join(worktreeDir, id)\n      git(['worktree', 'add', '-b', branch, path, baseRef], opts.repoRoot)\n      return { path, branch, baseRef }\n    },\n\n    async finalize(worktree, summary) {\n      // Stage + commit any pending changes the agent left in the worktree.\n      // A no-op commit is refused by git, so only commit when the tree is dirty.\n      const status = git(['status', '--porcelain'], worktree.path)\n      if (status.length > 0) {\n        git(['add', '-A'], worktree.path)\n        git(['commit', '-m', summary], worktree.path)\n      }\n      return {\n        kind: 'code',\n        worktreeRef: worktree.path,\n        baseRef: worktree.baseRef,\n        summary,\n      }\n    },\n\n    async discard(worktree) {\n      // Remove the worktree, then delete its branch. Force-remove because the\n      // worktree may hold uncommitted experiment state we're discarding.\n      git(['worktree', 'remove', '--force', worktree.path], opts.repoRoot)\n      git(['branch', '-D', worktree.branch], opts.repoRoot)\n    },\n  }\n}\n\n/** Resolve a `CodeSurface`'s worktreeRef to a directory the measurement can\n *  run the worker in. A path ref is returned as-is; anything else is treated\n *  as a ref under the adapter's worktree dir. */\nexport function resolveWorktreePath(surface: CodeSurface, worktreeDir?: string): string {\n  if (isAbsolute(surface.worktreeRef) && existsSync(surface.worktreeRef)) return surface.worktreeRef\n  if (worktreeDir) return join(worktreeDir, basename(surface.worktreeRef))\n  return surface.worktreeRef\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAmCA,IAAM,oBACJ;AAqCK,SAAS,kBAAkB,MAAwB;AACxD,QAAM,MAAgB,CAAC;AACvB,aAAW,QAAQ,KAAK,MAAM,IAAI,GAAG;AACnC,UAAM,QAAQ,kBAAkB,KAAK,IAAI;AACzC,QAAI,MAAO,KAAI,KAAK,MAAM,CAAC,CAAE;AAAA,EAC/B;AACA,SAAO;AACT;AAMO,SAAS,mBAAmB,UAAkB,WAA2B;AAC9E,QAAM,OAAO,CAAC,MACZ,EACG,MAAM,mBAAmB,EACzB,IAAI,CAAC,MAAM,EAAE,KAAK,CAAC,EACnB,OAAO,CAAC,MAAM,EAAE,SAAS,CAAC;AAC/B,QAAM,IAAI,IAAI,IAAI,KAAK,QAAQ,CAAC;AAChC,QAAM,IAAI,IAAI,IAAI,KAAK,SAAS,CAAC;AACjC,MAAI,QAAQ;AACZ,aAAW,KAAK,EAAG,KAAI,CAAC,EAAE,IAAI,CAAC,EAAG;AAClC,aAAW,KAAK,EAAG,KAAI,CAAC,EAAE,IAAI,CAAC,EAAG;AAClC,SAAO;AACT;AAEO,SAAS,eAAe,MAAgD;AAC7E,QAAM,YAAY,KAAK,aAAa;AACpC,QAAM,aAAa,KAAK,cAAc;AACtC,MAAI,aAAa,GAAG;AAClB,UAAM,IAAI;AAAA,MACR,gDAAgD,UAAU;AAAA,IAC5D;AAAA,EACF;AACA,SAAO;AAAA,IACL,MAAM;AAAA,IACN,MAAM,QAAQ,KAAgD;AAC5D,UAAI,OAAO,IAAI,mBAAmB,UAAU;AAC1C,cAAM,IAAI;AAAA,UACR,gEAAgE,OAAO,IAAI,cAAc;AAAA,QAC3F;AAAA,MACF;AACA,YAAM,WAAW,IAAI;AACrB,YAAM,mBAAmB,KAAK,oBAAoB,kBAAkB,QAAQ;AAE5E,YAAM,EAAE,KAAK,QAAQ,OAAO,IAAI,cAAc,KAAK,WAAW,KAAK,MAAM;AAEzE,YAAM,iBAAiB,sBAAsB;AAAA,QAC3C;AAAA,QACA,eAAe;AAAA,QACf,WAAW;AAAA,QACX,cAAc;AAAA,QACd,YAAY,IAAI;AAAA,QAChB,oBAAoB,KAAK;AAAA,MAC3B,CAAC;AACD,YAAM,qBAAqB;AAAA,QACzB;AAAA,QACA;AAAA,QACA;AAAA,QACA,0BAA0B,UAAU;AAAA,QACpC;AAAA,QACA,GAAG,iBAAiB,IAAI,CAAC,MAAM,YAAY,CAAC,IAAI;AAAA,QAChD;AAAA,QACA;AAAA,QACA;AAAA,MACF,EAAE,KAAK,IAAI;AACX,YAAM,aAAa,GAAG,cAAc,GAAG,kBAAkB;AACzD,YAAM,SAAS,kBAAkB,QAAQ,eAAe,OAAO,UAAU,CAAC;AAE1E,YAAM,SAAS,MAAM;AAAA,QACnB;AAAA,UACE,OAAO,KAAK;AAAA,UACZ,UAAU;AAAA,YACR,EAAE,MAAM,UAAU,SAAS,OAAO;AAAA,YAClC,EAAE,MAAM,QAAQ,SAAS,WAAW;AAAA,UACtC;AAAA,UACA,UAAU;AAAA,UACV,aAAa,KAAK,eAAe;AAAA,UACjC,WAAW,KAAK,aAAa;AAAA,QAC/B;AAAA,QACA,KAAK;AAAA,MACP;AAEA,YAAM,YAAY,wBAAwB,OAAO,SAAS,IAAI,cAAc;AAC5E,YAAM,MAAwB,CAAC;AAC/B,iBAAW,YAAY,WAAW;AAChC,cAAM,OAAO,OAAO,SAAS,YAAY,WAAW,SAAS,QAAQ,KAAK,IAAI;AAC9E,YAAI,CAAC,QAAQ,SAAS,SAAU;AAChC,YAAI,CAAC,iBAAiB,MAAM,gBAAgB,EAAG;AAC/C,YAAI,mBAAmB,UAAU,IAAI,IAAI,aAAa,EAAG;AACzD,YAAI,IAAI,SAAS,IAAI,EAAG;AACxB,YAAI,KAAK,IAAI;AAAA,MACf;AACA,aAAO;AAAA,IACT;AAAA,EACF;AACF;AAEA,SAAS,iBAAiB,WAAmB,UAA6B;AACxE,MAAI,SAAS,WAAW,EAAG,QAAO;AAClC,QAAM,OAAO,IAAI,IAAI,kBAAkB,SAAS,CAAC;AACjD,aAAW,WAAW,UAAU;AAC9B,QAAI,CAAC,KAAK,IAAI,OAAO,EAAG,QAAO;AAAA,EACjC;AACA,SAAO;AACT;AAGA,SAAS,cACP,KACA,WACA,YAC6D;AAC7D,QAAM,OAAO,IAAI,QAAQ,GAAG,EAAE;AAC9B,MAAI,CAAC,QAAQ,KAAK,WAAW,WAAW,GAAG;AACzC,WAAO,EAAE,KAAK,CAAC,GAAG,QAAQ,CAAC,GAAG,QAAQ,WAAW;AAAA,EACnD;AACA,QAAM,OAAO,CAAC,GAAG,KAAK,UAAU,EAAE,KAAK,CAAC,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS,EAAE,CAAC;AAC7E,MAAI,CAAC,KAAM,QAAO,EAAE,KAAK,CAAC,GAAG,QAAQ,CAAC,GAAG,QAAQ,WAAW;AAE5D,QAAM,UAAU,CAAC,GAAG,KAAK,SAAS,EAAE,KAAK,CAAC,GAAG,MAAM,EAAE,YAAY,EAAE,SAAS;AAC5E,QAAM,UAAU,CAAC,OAA8D;AAAA,IAC7E,IAAI,EAAE;AAAA,IACN,OAAO,EAAE;AAAA,EACX;AACA,QAAM,MAAM,QAAQ,MAAM,GAAG,SAAS,EAAE,IAAI,OAAO;AACnD,QAAM,SAAS,QAAQ,MAAM,CAAC,SAAS,EAAE,QAAQ,EAAE,IAAI,OAAO;AAE9D,QAAM,UAAU,OAAO,QAAQ,KAAK,UAAU,EAC3C,KAAK,CAAC,GAAG,MAAM,EAAE,CAAC,IAAI,EAAE,CAAC,CAAC,EAC1B,MAAM,GAAG,CAAC,EACV,IAAI,CAAC,CAAC,KAAK,KAAK,MAAM,GAAG,GAAG,KAAK,MAAM,QAAQ,CAAC,CAAC,GAAG;AACvD,QAAM,SACJ,QAAQ,SAAS,IAAI,GAAG,UAAU,+BAA0B,QAAQ,KAAK,IAAI,CAAC,KAAK;AAErF,SAAO,EAAE,KAAK,QAAQ,OAAO;AAC/B;;;ACrLA,SAAS,kBAAkB;AAC3B,SAAS,YAAY,WAAW,cAAc,qBAAqB;AACnE,SAAS,YAAY;AAmBd,IAAM,4BAAN,cAAwC,MAAM;AAAA,EACnD,YACkB,MAChB,SACA;AACA,UAAM,OAAO;AAHG;AAIhB,SAAK,OAAO;AAAA,EACd;AAAA,EALkB;AAMpB;AAQO,IAAM,yBAAN,MAA6D;AAAA,EAIlE,YAA6B,SAAwC;AAAxC;AAC3B,QAAI,CAAC,WAAW,QAAQ,IAAI,EAAG,WAAU,QAAQ,MAAM,EAAE,WAAW,KAAK,CAAC;AAC1E,SAAK,MAAM,QAAQ,OAAO,KAAK;AAAA,EACjC;AAAA,EAH6B;AAAA,EAHZ;AAAA,EACA,aAAa,oBAAI,IAA4B;AAAA,EAO9D,MAAM,QAAQ,OAA4C;AACxD,SAAK,iBAAiB,KAAK;AAC3B,SAAK,gBAAgB,KAAK;AAC1B,UAAM,SAAS,KAAK,SAAS,KAAK;AAClC,UAAM,OAAO,KAAK,cAAc,MAAM,MAAM;AAC5C,UAAM,OAAO,GAAG,KAAK,UAAU,MAAM,CAAC;AAAA;AAItC,eAAW,MAAM,IAAI;AAAA,EACvB;AAAA,EAEA,MAAM,OAAO,MAAmE;AAC9E,QAAI,CAAC,KAAK,OAAO;AACf,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AACA,QAAI,CAAC,KAAK,gBAAgB;AACxB,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AAEA,UAAM,MAA+B,CAAC;AACtC,eAAW,UAAU,aAAa;AAGhC,UAAI,KAAK,UAAU,WAAW,WAAW,oBAAoB;AAC3D,cAAM,WAAW,qBAAqB,KAAK,QAAQ,QAAQ,kBAAkB;AAC7E,YAAI,CAAC,SAAU;AAAA,MACjB;AACA,YAAM,OAAO,KAAK,cAAc,MAAM;AACtC,UAAI,CAAC,WAAW,IAAI,EAAG;AACvB,YAAM,QAAQ,aAAa,MAAM,MAAM,EAAE,MAAM,IAAI,EAAE,OAAO,OAAO;AACnE,iBAAW,QAAQ,OAAO;AACxB,YAAI;AACJ,YAAI;AACF,mBAAS,KAAK,MAAM,IAAI;AAAA,QAC1B,QAAQ;AACN;AAAA,QACF;AACA,YAAI,CAAC,cAAc,QAAQ,MAAM,MAAM,EAAG;AAC1C,YAAI,KAAK,MAAM;AAAA,MACjB;AAAA,IACF;AAGA,QAAI,KAAK,CAAC,GAAG,MAAM;AACjB,UAAI,EAAE,eAAe,EAAE,WAAY,QAAO,EAAE,WAAW,cAAc,EAAE,UAAU;AACjF,aAAO,EAAE,WAAW,cAAc,EAAE,UAAU;AAAA,IAChD,CAAC;AAED,WAAO,IAAI,MAAM,GAAG,KAAK,KAAK;AAAA,EAChC;AAAA,EAEA,MAAM,OAAmF;AACvF,UAAM,WAAmC,CAAC;AAC1C,QAAI,QAAQ;AACZ,eAAW,UAAU,aAAa;AAChC,YAAM,OAAO,KAAK,cAAc,MAAM;AACtC,UAAI,CAAC,WAAW,IAAI,GAAG;AACrB,iBAAS,MAAM,IAAI;AACnB;AAAA,MACF;AACA,YAAM,QAAQ,aAAa,MAAM,MAAM,EAAE,MAAM,IAAI,EAAE,OAAO,OAAO,EAAE;AACrE,eAAS,MAAM,IAAI;AACnB,eAAS;AAAA,IACX;AAIA,WAAO,EAAE,OAAO,OAAO,MAAM,OAAO,SAAS;AAAA,EAC/C;AAAA,EAEQ,iBAAiB,OAAmC;AAC1D,QAAI,CAAC,MAAM,QAAQ;AACjB,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AACA,QAAI,CAAC,MAAM,qBAAqB,MAAM,kBAAkB,WAAW,GAAG;AACpE,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AACA,QAAI,CAAC,MAAM,YAAY;AACrB,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AACA,QAAI,CAAC,MAAM,iBAAiB;AAC1B,YAAM,IAAI;AAAA,QACR;AAAA,QACA;AAAA,MACF;AAAA,IACF;AACA,QAAI,CAAC,YAAY,SAAS,MAAM,MAAM,GAAG;AACvC,YAAM,IAAI;AAAA,QACR;AAAA,QACA,+CAA+C,YAAY,KAAK,IAAI,CAAC;AAAA,MACvE;AAAA,IACF;AAAA,EACF;AAAA,EAEQ,gBAAgB,OAAmC;AACzD,UAAM,MAAM,KAAK,QAAQ;AACzB,QAAI,CAAC,OAAO,CAAC,MAAM,gBAAiB;AACpC,UAAM,MAAM,KAAK,IAAI;AACrB,UAAM,WAAW;AACjB,QAAI,QAAQ,KAAK,WAAW,IAAI,MAAM,eAAe;AACrD,QAAI,CAAC,SAAS,MAAM,MAAM,iBAAiB,UAAU;AACnD,cAAQ,EAAE,QAAQ,MAAM,iBAAiB,eAAe,KAAK,OAAO,EAAE;AACtE,WAAK,WAAW,IAAI,MAAM,iBAAiB,KAAK;AAAA,IAClD;AACA,QAAI,MAAM,SAAS,KAAK;AACtB,YAAM,IAAI;AAAA,QACR;AAAA,QACA,gCAAgC,MAAM,eAAe,aAAa,GAAG;AAAA,MACvE;AAAA,IACF;AACA,UAAM,SAAS;AAAA,EACjB;AAAA,EAEQ,SAAS,OAAoD;AACnE,UAAM,aAAa;AAAA,MACjB,KAAK,UAAU;AAAA,QACb,IAAI,MAAM,SAAS;AAAA,QACnB,KAAK,MAAM;AAAA,QACX,IAAI,MAAM;AAAA,QACV,KAAK,MAAM;AAAA,MACb,CAAC;AAAA,IACH;AAGA,WAAO;AAAA,MACL,GAAG;AAAA,MACH;AAAA,MACA,OAAO;AAAA,IACT;AAAA,EACF;AAAA,EAEQ,cAAc,QAAwB;AAC5C,WAAO,KAAK,KAAK,QAAQ,MAAM,GAAG,MAAM,QAAQ;AAAA,EAClD;AACF;AAEA,IAAM,cAAgD;AAAA,EACpD;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF;AAEA,SAAS,qBACP,QACA,QACS;AACT,MAAI,CAAC,OAAQ,QAAO;AACpB,MAAI,MAAM,QAAQ,MAAM,EAAG,QAAO,OAAO,SAAS,MAAM;AACxD,SAAO,WAAW;AACpB;AAEA,SAAS,cACP,QACA,MACA,QACS;AAET,MAAI,KAAK,UAAU,WAAW,OAAO,cAAc,KAAK,eAAgB,QAAO;AAC/E,MAAI,KAAK,UAAU,UAAU,OAAO,aAAa,KAAK,eAAgB,QAAO;AAE7E,QAAM,IAAI,KAAK;AACf,MAAI,CAAC,EAAG,QAAO;AACf,MAAI,EAAE,QAAQ,OAAO,SAAS,SAAS,EAAE,KAAM,QAAO;AACtD,MAAI,EAAE,QAAQ;AACZ,UAAM,UAAU,MAAM,QAAQ,EAAE,MAAM,IAAI,EAAE,SAAS,CAAC,EAAE,MAAM;AAC9D,QAAI,CAAC,QAAQ,SAAS,MAAe,EAAG,QAAO;AAAA,EACjD;AACA,MAAI,EAAE,iBAAiB,UAAa,EAAE,iBAAiB,QAAW;AAChE,UAAM,aAAa,OAAO,OAAO,OAAO,WAAW,EAAE,IAAI,CAAC,MAAM,EAAE,SAAS;AAC3E,UAAM,MAAM,WAAW,WAAW,IAAI,IAAI,KAAK,IAAI,GAAG,UAAU;AAChE,QAAI,EAAE,iBAAiB,UAAa,MAAM,EAAE,aAAc,QAAO;AACjE,QAAI,EAAE,iBAAiB,UAAa,MAAM,EAAE,aAAc,QAAO;AAAA,EACnE;AACA,SAAO;AACT;AAEA,SAAS,OAAO,OAAuB;AACrC,SAAO,WAAW,QAAQ,EAAE,OAAO,KAAK,EAAE,OAAO,KAAK,EAAE,MAAM,GAAG,EAAE;AACrE;AAEA,SAAS,WAAW,MAAc,MAAoB;AACpD,MAAI,WAAW,IAAI,GAAG;AACpB,UAAM,WAAW,aAAa,MAAM,MAAM;AAC1C,kBAAc,MAAM,WAAW,IAAI;AAAA,EACrC,OAAO;AACL,kBAAc,MAAM,IAAI;AAAA,EAC1B;AACF;;;AC1QA,SAAS,oBAAoB;AAC7B,SAAS,cAAAA,mBAAkB;AAC3B,SAAS,UAAU,YAAY,QAAAC,aAAY;AAuBpC,IAAM,uBAAN,cAAmC,MAAM;AAAA,EAC9C,YACE,SACS,OACT;AACA,UAAM,OAAO;AAFJ;AAGT,SAAK,OAAO;AAAA,EACd;AAAA,EAJW;AAKb;AAaA,SAAS,WAAW,MAAgB,KAAqB;AACvD,MAAI;AACF,WAAO,aAAa,OAAO,MAAM,EAAE,KAAK,UAAU,OAAO,CAAC,EAAE,KAAK;AAAA,EACnE,SAAS,KAAK;AACZ,UAAM,SACJ,OAAO,OAAO,QAAQ,YAAY,YAAY,MAC1C,OAAQ,IAA4B,MAAM,IAC1C;AACN,UAAM,IAAI,qBAAqB,OAAO,KAAK,KAAK,GAAG,CAAC,YAAY,UAAU,OAAO,GAAG,CAAC,IAAI,GAAG;AAAA,EAC9F;AACF;AAGA,SAAS,KAAK,OAAuB;AACnC,SACE,MACG,YAAY,EACZ,QAAQ,eAAe,GAAG,EAC1B,QAAQ,YAAY,EAAE,EACtB,MAAM,GAAG,EAAE,KAAK;AAEvB;AAEO,SAAS,mBAAmB,MAAkD;AACnF,QAAM,MAAM,KAAK,OAAO;AACxB,QAAM,cAAc,KAAK,eAAeA,MAAK,KAAK,UAAU,YAAY;AACxE,QAAM,eAAe,KAAK,gBAAgB;AAE1C,SAAO;AAAA,IACL,MAAM,OAAO,EAAE,SAAS,MAAM,GAAG;AAC/B,YAAM,KAAK,GAAG,KAAK,KAAK,CAAC,IAAI,KAAK,IAAI,EAAE,SAAS,EAAE,CAAC,IAAI,KAAK,OAAO,EAAE,SAAS,EAAE,EAAE,MAAM,GAAG,CAAC,CAAC;AAC9F,YAAM,SAAS,GAAG,YAAY,IAAI,EAAE;AACpC,YAAM,OAAOA,MAAK,aAAa,EAAE;AACjC,UAAI,CAAC,YAAY,OAAO,MAAM,QAAQ,MAAM,OAAO,GAAG,KAAK,QAAQ;AACnE,aAAO,EAAE,MAAM,QAAQ,QAAQ;AAAA,IACjC;AAAA,IAEA,MAAM,SAAS,UAAU,SAAS;AAGhC,YAAM,SAAS,IAAI,CAAC,UAAU,aAAa,GAAG,SAAS,IAAI;AAC3D,UAAI,OAAO,SAAS,GAAG;AACrB,YAAI,CAAC,OAAO,IAAI,GAAG,SAAS,IAAI;AAChC,YAAI,CAAC,UAAU,MAAM,OAAO,GAAG,SAAS,IAAI;AAAA,MAC9C;AACA,aAAO;AAAA,QACL,MAAM;AAAA,QACN,aAAa,SAAS;AAAA,QACtB,SAAS,SAAS;AAAA,QAClB;AAAA,MACF;AAAA,IACF;AAAA,IAEA,MAAM,QAAQ,UAAU;AAGtB,UAAI,CAAC,YAAY,UAAU,WAAW,SAAS,IAAI,GAAG,KAAK,QAAQ;AACnE,UAAI,CAAC,UAAU,MAAM,SAAS,MAAM,GAAG,KAAK,QAAQ;AAAA,IACtD;AAAA,EACF;AACF;AAKO,SAAS,oBAAoB,SAAsB,aAA8B;AACtF,MAAI,WAAW,QAAQ,WAAW,KAAKD,YAAW,QAAQ,WAAW,EAAG,QAAO,QAAQ;AACvF,MAAI,YAAa,QAAOC,MAAK,aAAa,SAAS,QAAQ,WAAW,CAAC;AACvE,SAAO,QAAQ;AACjB;","names":["existsSync","join"]}

package/dist/contract/index.d.ts

@@ -3,8 +3,8 @@ export { f as CampaignAggregates, g as CampaignArtifactWriter, h as CampaignCell
 import { C as CampaignStorage, d as RunImprovementLoopResult } from '../run-improvement-loop-BPMjNKMJ.js';
 export { D as DefaultProductionGateOptions, E as EvolutionaryDriverOptions, G as GepaDriverOptions, H as HeldOutGateOptions, R as RunCampaignOptions, b as RunEvalOptions, c as RunImprovementLoopOptions, g as composeGate, h as defaultProductionGate, i as evolutionaryDriver, j as fsCampaignStorage, k as gepaDriver, l as heldOutGate, m as inMemoryCampaignStorage, r as runCampaign, n as runEval, p as runImprovementLoop } from '../run-improvement-loop-BPMjNKMJ.js';
 export { D as DeploymentOutcome, F as FileSystemOutcomeStore, b as FileSystemOutcomeStoreOptions, I as InMemoryOutcomeStore, O as OutcomeStore } from '../outcome-store-D6KWmYvj.js';
-import { a as HostedTenant, I as InsightReport, T as TraceSpanEvent } from '../index-BRxz6qov.js';
-export { F as FailureClusterInsight, b as InterRaterInsight, J as JudgeInsight, L as LiftInsight, O as OutcomeCorrelationInsight, R as Recommendation, c as ReleaseSummary, S as ScalarDistribution } from '../index-BRxz6qov.js';
+import { a as HostedTenant, I as InsightReport, T as TraceSpanEvent } from '../index-DQHtWQ57.js';
+export { F as FailureClusterInsight, b as InterRaterInsight, J as JudgeInsight, L as LiftInsight, O as OutcomeCorrelationInsight, R as Recommendation, c as ReleaseSummary, S as ScalarDistribution } from '../index-DQHtWQ57.js';
 import { A as AnalystRegistry } from '../registry-8KAs18kY.js';
 import { a as DatasetScenario } from '../dataset-BlwAtYYf.js';
 import { R as RunRecord, a as RunSplitTag } from '../run-record-BGY6bHRh.js';

package/dist/contract/index.js

@@ -49,14 +49,27 @@ async function analyzeRuns(opts) {
   const bins = opts.histogramBins ?? 12;
   const threshold = opts.decisionThreshold ?? 0.02;
   const split = resolveSplit(runs, opts.split ?? "auto");
+  const compositeWithIds = runs.map((r) => ({ runId: r.runId, score: compositeOf(r, split) })).filter((p) => Number.isFinite(p.score));
   const composite = distributionOf(
-    runs.map((r) => compositeOf(r, split)).filter(Number.isFinite),
-    bins
+    compositeWithIds.map((p) => p.score),
+    bins,
+    compositeWithIds
   );
   const perDimension = computePerDimension(runs, bins);
+  const costs = runs.map((r) => r.costUsd).filter(Number.isFinite);
+  const costDist = distributionOf(costs, bins);
+  const pareto = paretoChart(runs, { split });
+  const degraded = {};
+  if (costs.length === 0 || costs.every((c) => c === 0)) {
+    degraded.cost = "no costUsd values recorded \u2014 cost axis carries no signal";
+  }
+  if (pareto.points.length < 2) {
+    degraded.pareto = pareto.points.length === 0 ? "no candidates \u2014 Pareto unavailable" : "single candidate \u2014 Pareto is a single point, not a frontier";
+  }
   const costQuality = {
-    cost: distributionOf(runs.map((r) => r.costUsd).filter(Number.isFinite), bins),
-    pareto: paretoChart(runs, { split })
+    cost: costDist,
+    pareto,
+    ...degraded.cost || degraded.pareto ? { degraded } : {}
   };
   const judges = computeJudgeInsights(runs);
   const interRater = opts.raterScores ? computeInterRater(opts.raterScores) : void 0;
@@ -101,7 +114,7 @@ function compositeOf(run, split) {
   const alt = split === "holdout" ? run.outcome.searchScore : run.outcome.holdoutScore;
   return Number.isFinite(alt) ? alt : Number.NaN;
 }
-function distributionOf(values, bins) {
+function distributionOf(values, bins, withIds) {
   if (values.length === 0) {
     return {
       n: 0,
@@ -119,6 +132,7 @@ function distributionOf(values, bins) {
   const mean2 = sorted.reduce((s, v) => s + v, 0) / n;
   const variance = sorted.reduce((s, v) => s + (v - mean2) ** 2, 0) / n;
   const stddev = Math.sqrt(variance);
+  const tailRuns = withIds ? [...withIds].sort((a, b) => a.score - b.score).slice(0, Math.min(5, withIds.length)) : void 0;
   return {
     n,
     mean: mean2,
@@ -127,7 +141,8 @@ function distributionOf(values, bins) {
     stddev,
     min: sorted[0],
     max: sorted[n - 1],
-    histogram: histogram(sorted, bins)
+    histogram: histogram(sorted, bins),
+    ...tailRuns ? { tailRuns } : {}
   };
 }
 function percentile(sorted, q) {
@@ -477,6 +492,38 @@ function buildReleaseScorecard(composite, lift, contamination) {
 }
 function buildRecommendations(ctx) {
   const out = [];
+  if (ctx.composite.n > 0) {
+    if (ctx.composite.mean < 0.3) {
+      const tail = ctx.composite.tailRuns ?? [];
+      const names = tail.slice(0, 5).map((t) => `${t.runId}=${t.score.toFixed(3)}`).join(", ");
+      out.push({
+        priority: "critical",
+        kind: "investigate",
+        title: `Composite mean ${ctx.composite.mean.toFixed(3)} is below the 0.3 floor \u2014 the agent is broken on this corpus`,
+        detail: tail.length > 0 ? `Worst ${tail.length} run${tail.length === 1 ? "" : "s"} to inspect first: ${names}. Histogram p50=${ctx.composite.p50.toFixed(3)}, p95=${ctx.composite.p95.toFixed(3)}.` : `Histogram p50=${ctx.composite.p50.toFixed(3)}, p95=${ctx.composite.p95.toFixed(3)}.`,
+        evidencePath: "composite.tailRuns"
+      });
+    } else if (ctx.composite.mean < 0.5) {
+      const tail = ctx.composite.tailRuns ?? [];
+      const names = tail.slice(0, 3).map((t) => `${t.runId}=${t.score.toFixed(3)}`).join(", ");
+      out.push({
+        priority: "high",
+        kind: "investigate",
+        title: `Composite mean ${ctx.composite.mean.toFixed(3)} is below 0.5 \u2014 investigate the lower tail before claiming the agent is healthy`,
+        detail: tail.length > 0 ? `Worst ${tail.length} run${tail.length === 1 ? "" : "s"}: ${names}. Histogram p50=${ctx.composite.p50.toFixed(3)}, p95=${ctx.composite.p95.toFixed(3)}.` : `Histogram p50=${ctx.composite.p50.toFixed(3)}, p95=${ctx.composite.p95.toFixed(3)}.`,
+        evidencePath: "composite.tailRuns"
+      });
+    }
+  }
+  if (Object.keys(ctx.judges).length === 0 && ctx.composite.n > 0) {
+    out.push({
+      priority: "medium",
+      kind: "expand-corpus",
+      title: "No judge scores recorded \u2014 per-dimension + calibration insights unavailable",
+      detail: "Records have no `outcome.judgeScores`. To unlock perDimension, judges, and calibration, attach a Judge run during your eval pass and populate `outcome.judgeScores.perJudge[judgeName][dimension] = score`. See `docs/insight-report.md` for the expected shape.",
+      evidencePath: "judges"
+    });
+  }
   if (ctx.lift) {
     const decisive = ctx.lift.ci95[0] > ctx.threshold;
     const inconclusive = ctx.lift.ci95[0] <= ctx.threshold && ctx.lift.ci95[1] > ctx.threshold;

package/dist/contract/index.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../../src/contract/analyze-runs.ts","../../src/contract/self-improve.ts","../../src/contract/intake/feedback-table.ts","../../src/contract/intake/otel-spans.ts"],"sourcesContent":["/**\n * # `analyzeRuns()` — turn a set of agent runs into an actionable decision packet.\n *\n * Wires the substrate's statistical, calibration, clustering, Pareto, and\n * release-confidence primitives into one `InsightReport`. Two top-level\n * entry points use this function:\n *\n *   - `selfImprove()` calls it on the campaign output to attach a packet\n *     to every run.\n *   - Consumers with observed `RunRecord[]` (production traces, gold\n *     corpora, approve/reject tables) call it directly via `analyzeRuns()`\n *     for analysis without a closed loop.\n *\n * Every section is opt-in based on what the input data supports — the\n * function never invents signal. If runs carry no judge scores, `judges`\n * is empty. If there's no baseline/candidate split, `lift` is undefined.\n * If no `analyst` is wired, `failureClusters` is undefined.\n *\n * The `recommendations` array is the human-readable layer; everything\n * else is the evidence backing each recommendation.\n */\n\nimport type { AnalystRegistry } from '../analyst/registry'\nimport type { AnalystFinding } from '../analyst/types'\nimport { checkCanaries } from '../contamination-guard'\nimport type { DatasetScenario } from '../dataset'\nimport type { RunRecord } from '../run-record'\nimport { cohensD, pairedBootstrap, pairedMde, pairedTTest, requiredSampleSize } from '../statistics'\nimport { type ParetoFigureSpec, paretoChart } from '../summary-report'\n\nimport type {\n  FailureClusterInsight,\n  InsightReport,\n  InterRaterInsight,\n  JudgeInsight,\n  LiftInsight,\n  OutcomeCorrelationInsight,\n  Recommendation,\n  ScalarDistribution,\n} from './insight-report'\n\n// ── Public API ───────────────────────────────────────────────────────\n\nexport interface AnalyzeRunsOptions {\n  /** The runs to analyze. */\n  runs: RunRecord[]\n  /** Which split to score against when reading composite from RunOutcome.\n   *  Default: holdout when ANY run has a `holdoutScore`, else search. */\n  split?: 'search' | 'holdout' | 'auto'\n  /** Pairwise analysis configuration. When both `baselineCandidateId` and\n   *  `candidateCandidateId` are present, lift is computed on paired\n   *  (experimentId, seed) tuples shared between the two sides. */\n  baselineCandidateId?: string\n  candidateCandidateId?: string\n  /** Canary scenarios — checked against every run's raw output for\n   *  holdout contamination. */\n  canaryScenarios?: DatasetScenario[]\n  /** Analyst registry for failure clustering. When omitted, the\n   *  `failureClusters` section is left undefined. */\n  analyst?: AnalystRegistry\n  /** Downstream outcome metric per run (e.g. engagement rate, approval\n   *  rate, downstream pass rate). When present, the report includes\n   *  `outcomeCorrelation` + a simple linear reward model fit. */\n  outcomeSignal?: {\n    metric: string\n    valueByRunId: Record<string, number>\n  }\n  /** Multi-rater feedback for inter-rater agreement. Each entry is one\n   *  rater's score for one run. Two or more raters → kappa + disagreement\n   *  triage list. */\n  raterScores?: Array<{ runId: string; rater: string; score: number }>\n  /** Number of histogram bins for distributional summaries. Default 12. */\n  histogramBins?: number\n  /** Decision threshold — the smallest composite lift the caller cares\n   *  about. Used by the recommendations engine to call ship vs hold.\n   *  Default 0.02. */\n  decisionThreshold?: number\n}\n\nexport async function analyzeRuns(opts: AnalyzeRunsOptions): Promise<InsightReport> {\n  const runs = opts.runs\n  const bins = opts.histogramBins ?? 12\n  const threshold = opts.decisionThreshold ?? 0.02\n  const split = resolveSplit(runs, opts.split ?? 'auto')\n\n  const composite = distributionOf(\n    runs.map((r) => compositeOf(r, split)).filter(Number.isFinite) as number[],\n    bins,\n  )\n\n  const perDimension = computePerDimension(runs, bins)\n\n  const costQuality = {\n    cost: distributionOf(runs.map((r) => r.costUsd).filter(Number.isFinite), bins),\n    pareto: paretoChart(runs, { split }),\n  }\n\n  const judges = computeJudgeInsights(runs)\n\n  const interRater = opts.raterScores ? computeInterRater(opts.raterScores) : undefined\n\n  const lift = computeLift(runs, opts.baselineCandidateId, opts.candidateCandidateId, split)\n\n  const failureClusters = opts.analyst\n    ? await computeFailureClusters(runs, opts.analyst, split)\n    : undefined\n\n  const contamination = opts.canaryScenarios\n    ? computeContamination(runs, opts.canaryScenarios)\n    : undefined\n\n  const outcomeCorrelation = opts.outcomeSignal\n    ? computeOutcomeCorrelation(runs, opts.outcomeSignal, split)\n    : undefined\n\n  const release = buildReleaseScorecard(composite, lift, contamination)\n\n  const recommendations = buildRecommendations({\n    composite,\n    judges,\n    interRater,\n    lift,\n    failureClusters,\n    contamination,\n    outcomeCorrelation,\n    threshold,\n  })\n\n  return {\n    n: runs.length,\n    composite,\n    perDimension,\n    costQuality,\n    judges,\n    interRater,\n    lift,\n    failureClusters,\n    contamination,\n    outcomeCorrelation,\n    release,\n    recommendations,\n  }\n}\n\n// ── Composite + split selection ─────────────────────────────────────\n\nfunction resolveSplit(\n  runs: RunRecord[],\n  pref: 'search' | 'holdout' | 'auto',\n): 'search' | 'holdout' {\n  if (pref !== 'auto') return pref\n  const hasHoldout = runs.some((r) => Number.isFinite(r.outcome.holdoutScore))\n  return hasHoldout ? 'holdout' : 'search'\n}\n\nfunction compositeOf(run: RunRecord, split: 'search' | 'holdout'): number {\n  const primary = split === 'holdout' ? run.outcome.holdoutScore : run.outcome.searchScore\n  if (Number.isFinite(primary)) return primary as number\n  // Fall through to the other split if the preferred one is missing —\n  // analyzeRuns shouldn't refuse to summarise a run just because the\n  // caller asked for the split that wasn't recorded.\n  const alt = split === 'holdout' ? run.outcome.searchScore : run.outcome.holdoutScore\n  return Number.isFinite(alt) ? (alt as number) : Number.NaN\n}\n\n// ── Distribution helpers ────────────────────────────────────────────\n\nfunction distributionOf(values: number[], bins: number): ScalarDistribution {\n  if (values.length === 0) {\n    return {\n      n: 0,\n      mean: 0,\n      p50: 0,\n      p95: 0,\n      stddev: 0,\n      min: 0,\n      max: 0,\n      histogram: [],\n    }\n  }\n  const sorted = [...values].sort((a, b) => a - b)\n  const n = sorted.length\n  const mean = sorted.reduce((s, v) => s + v, 0) / n\n  const variance = sorted.reduce((s, v) => s + (v - mean) ** 2, 0) / n\n  const stddev = Math.sqrt(variance)\n  return {\n    n,\n    mean,\n    p50: percentile(sorted, 0.5),\n    p95: percentile(sorted, 0.95),\n    stddev,\n    min: sorted[0]!,\n    max: sorted[n - 1]!,\n    histogram: histogram(sorted, bins),\n  }\n}\n\nfunction percentile(sorted: number[], q: number): number {\n  if (sorted.length === 0) return 0\n  if (sorted.length === 1) return sorted[0]!\n  const idx = (sorted.length - 1) * q\n  const lo = Math.floor(idx)\n  const hi = Math.ceil(idx)\n  if (lo === hi) return sorted[lo]!\n  const w = idx - lo\n  return sorted[lo]! * (1 - w) + sorted[hi]! * w\n}\n\n/** Even-width histogram over the value range. Returns inclusive-lo /\n *  exclusive-hi bins (closed on right for the last bin) compatible with\n *  the substrate's `GainDistributionBin` shape. */\nfunction histogram(sorted: number[], bins: number): ScalarDistribution['histogram'] {\n  if (sorted.length === 0 || bins < 1) return []\n  const min = sorted[0]!\n  const max = sorted[sorted.length - 1]!\n  if (min === max) return [{ lo: min, hi: max, count: sorted.length }]\n  const width = (max - min) / bins\n  const out: ScalarDistribution['histogram'] = []\n  for (let i = 0; i < bins; i++) {\n    const lo = min + i * width\n    const hi = i === bins - 1 ? max : lo + width\n    out.push({ lo, hi, count: 0 })\n  }\n  for (const v of sorted) {\n    const idx = Math.min(bins - 1, Math.floor((v - min) / width))\n    out[idx]!.count++\n  }\n  return out\n}\n\nfunction computePerDimension(runs: RunRecord[], bins: number): Record<string, ScalarDistribution> {\n  // JudgeScoresRecord pre-aggregates `perDimMean` (mean across judges per\n  // dimension). We collect those means across runs to produce a per-dim\n  // distribution at the corpus level. Consumers who want per-judge\n  // dimension values reach into `perJudge[judgeId][dim]` themselves.\n  const byDim = new Map<string, number[]>()\n  for (const run of runs) {\n    const scores = run.outcome.judgeScores\n    if (!scores) continue\n    for (const [dim, value] of Object.entries(scores.perDimMean ?? {})) {\n      if (!Number.isFinite(value)) continue\n      const arr = byDim.get(dim) ?? []\n      arr.push(value)\n      byDim.set(dim, arr)\n    }\n  }\n  const out: Record<string, ScalarDistribution> = {}\n  for (const [dim, values] of byDim) out[dim] = distributionOf(values, bins)\n  return out\n}\n\n// ── Judge insights ──────────────────────────────────────────────────\n\nfunction computeJudgeInsights(runs: RunRecord[]): Record<string, JudgeInsight> {\n  // Each judge's per-run mean is the average of its per-dimension scores\n  // for that run. We aggregate those means across all runs each judge\n  // scored — giving consumers a \"this judge's typical verdict\" reading.\n  const out: Record<string, JudgeInsight> = {}\n  const byJudge = new Map<string, number[]>()\n  for (const run of runs) {\n    const scores = run.outcome.judgeScores\n    if (!scores?.perJudge) continue\n    for (const [judgeId, dims] of Object.entries(scores.perJudge)) {\n      const dimValues = Object.values(dims).filter(Number.isFinite) as number[]\n      if (dimValues.length === 0) continue\n      const judgeMean = dimValues.reduce((s, v) => s + v, 0) / dimValues.length\n      const arr = byJudge.get(judgeId) ?? []\n      arr.push(judgeMean)\n      byJudge.set(judgeId, arr)\n    }\n  }\n  for (const [judgeId, values] of byJudge) {\n    out[judgeId] = {\n      n: values.length,\n      meanScore: values.reduce((s, v) => s + v, 0) / values.length,\n    }\n  }\n  return out\n}\n\n// ── Inter-rater agreement ───────────────────────────────────────────\n\nfunction computeInterRater(\n  ratings: Array<{ runId: string; rater: string; score: number }>,\n): InterRaterInsight | undefined {\n  const byRun = new Map<string, Array<{ rater: string; score: number }>>()\n  for (const r of ratings) {\n    if (!Number.isFinite(r.score)) continue\n    const list = byRun.get(r.runId) ?? []\n    list.push({ rater: r.rater, score: r.score })\n    byRun.set(r.runId, list)\n  }\n  const raters = new Set(ratings.map((r) => r.rater))\n  const jointlyRated: string[] = []\n  for (const [runId, ratersForRun] of byRun) {\n    const seen = new Set(ratersForRun.map((r) => r.rater))\n    let all = true\n    for (const r of raters) if (!seen.has(r)) all = false\n    if (all) jointlyRated.push(runId)\n  }\n  if (raters.size < 2 || jointlyRated.length === 0) return undefined\n\n  const raterList = [...raters].sort()\n  const perPair: Record<string, number> = {}\n  for (let i = 0; i < raterList.length; i++) {\n    for (let j = i + 1; j < raterList.length; j++) {\n      const a = raterList[i]!\n      const b = raterList[j]!\n      const aScores: number[] = []\n      const bScores: number[] = []\n      for (const runId of jointlyRated) {\n        const ratersForRun = byRun.get(runId)!\n        const sa = ratersForRun.find((r) => r.rater === a)?.score\n        const sb = ratersForRun.find((r) => r.rater === b)?.score\n        if (sa !== undefined && sb !== undefined) {\n          aScores.push(sa)\n          bScores.push(sb)\n        }\n      }\n      perPair[`${a}::${b}`] = pearson(aScores, bScores)\n    }\n  }\n  const pairKappas = Object.values(perPair)\n  const kappa =\n    pairKappas.length === 0 ? 0 : pairKappas.reduce((s, v) => s + v, 0) / pairKappas.length\n\n  const disagreementCases = jointlyRated\n    .map((runId) => {\n      const ratersForRun = byRun.get(runId)!\n      const scores = ratersForRun.map((r) => r.score)\n      const range = Math.max(...scores) - Math.min(...scores)\n      return { runId, ratings: ratersForRun, range }\n    })\n    .sort((a, b) => b.range - a.range)\n    .slice(0, 20)\n\n  return {\n    raters: raters.size,\n    jointlyRated: jointlyRated.length,\n    kappa,\n    perPair,\n    disagreementCases,\n  }\n}\n\nfunction pearson(a: number[], b: number[]): number {\n  if (a.length !== b.length || a.length === 0) return 0\n  const n = a.length\n  const meanA = a.reduce((s, v) => s + v, 0) / n\n  const meanB = b.reduce((s, v) => s + v, 0) / n\n  let num = 0\n  let denomA = 0\n  let denomB = 0\n  for (let i = 0; i < n; i++) {\n    const da = a[i]! - meanA\n    const db = b[i]! - meanB\n    num += da * db\n    denomA += da * da\n    denomB += db * db\n  }\n  const denom = Math.sqrt(denomA * denomB)\n  return denom === 0 ? 0 : num / denom\n}\n\n// ── Lift ────────────────────────────────────────────────────────────\n\nfunction computeLift(\n  runs: RunRecord[],\n  baselineId: string | undefined,\n  candidateId: string | undefined,\n  split: 'search' | 'holdout',\n): LiftInsight | undefined {\n  let bId = baselineId\n  let cId = candidateId\n  if (!bId || !cId) {\n    // Auto-detect: when exactly two distinct candidateIds appear, treat the\n    // lower-mean side as baseline.\n    const ids = [...new Set(runs.map((r) => r.candidateId))]\n    if (ids.length !== 2) return undefined\n    const [idA, idB] = ids as [string, string]\n    const meanA = mean(runs.filter((r) => r.candidateId === idA).map((r) => compositeOf(r, split)))\n    const meanB = mean(runs.filter((r) => r.candidateId === idB).map((r) => compositeOf(r, split)))\n    bId = meanA <= meanB ? idA : idB\n    cId = meanA <= meanB ? idB : idA\n  }\n\n  const baseline = runs.filter((r) => r.candidateId === bId)\n  const candidate = runs.filter((r) => r.candidateId === cId)\n  if (baseline.length === 0 || candidate.length === 0) return undefined\n\n  // Pair on (experimentId, seed). When that key doesn't match, fall back\n  // to ordinal pairing — common for fresh runs from the same scenario list.\n  const baselineByKey = new Map(baseline.map((r) => [pairingKey(r), r]))\n  const pairedBaseline: number[] = []\n  const pairedCandidate: number[] = []\n  let usedKeyPairing = false\n  for (const cand of candidate) {\n    const b = baselineByKey.get(pairingKey(cand))\n    if (b) {\n      const bC = compositeOf(b, split)\n      const cC = compositeOf(cand, split)\n      if (Number.isFinite(bC) && Number.isFinite(cC)) {\n        pairedBaseline.push(bC)\n        pairedCandidate.push(cC)\n        usedKeyPairing = true\n      }\n    }\n  }\n  if (!usedKeyPairing) {\n    const n = Math.min(baseline.length, candidate.length)\n    for (let i = 0; i < n; i++) {\n      const bC = compositeOf(baseline[i]!, split)\n      const cC = compositeOf(candidate[i]!, split)\n      if (Number.isFinite(bC) && Number.isFinite(cC)) {\n        pairedBaseline.push(bC)\n        pairedCandidate.push(cC)\n      }\n    }\n  }\n  if (pairedBaseline.length === 0) return undefined\n\n  const baselineMean = mean(pairedBaseline)\n  const candidateMean = mean(pairedCandidate)\n  const delta = candidateMean - baselineMean\n\n  const bootstrap = pairedBootstrap(pairedBaseline, pairedCandidate, {\n    confidence: 0.95,\n    resamples: 2000,\n    statistic: 'mean',\n  })\n  const tTest = pairedTTest(pairedBaseline, pairedCandidate)\n  const d = cohensD(pairedBaseline, pairedCandidate)\n  const mde = pairedMde({ nPaired: pairedBaseline.length, power: 0.8, alpha: 0.05 })\n  const requiredN = requiredSampleSize({\n    effect: Math.max(Math.abs(delta), 1e-6),\n    power: 0.8,\n    alpha: 0.05,\n  })\n\n  return {\n    baselineMean,\n    candidateMean,\n    delta,\n    ci95: [bootstrap.low, bootstrap.high],\n    pValue: tTest.p,\n    n: pairedBaseline.length,\n    cohensD: d,\n    mde,\n    requiredN,\n  }\n}\n\nfunction pairingKey(r: RunRecord): string {\n  return `${r.experimentId}::${r.seed}`\n}\n\nfunction mean(arr: number[]): number {\n  return arr.length === 0 ? 0 : arr.reduce((s, v) => s + v, 0) / arr.length\n}\n\n// ── Failure clustering ──────────────────────────────────────────────\n\nasync function computeFailureClusters(\n  runs: RunRecord[],\n  analyst: AnalystRegistry,\n  split: 'search' | 'holdout',\n): Promise<FailureClusterInsight | undefined> {\n  const failed = runs.filter((r) => compositeOf(r, split) < 0.5 || r.failureMode !== undefined)\n  if (failed.length === 0) return { clusters: [], totalFailures: 0 }\n\n  const clusters = new Map<string, { exemplars: string[]; share: number }>()\n  for (const run of failed) {\n    try {\n      const result = await analyst.run(run.runId, {\n        kind: 'run-record',\n        run,\n      } as Parameters<typeof analyst.run>[1])\n      for (const finding of result.findings as AnalystFinding[]) {\n        const key = finding.area || finding.analyst_id || 'unclassified'\n        const c = clusters.get(key) ?? { exemplars: [], share: 0 }\n        if (c.exemplars.length < 5) c.exemplars.push(run.runId)\n        clusters.set(key, c)\n      }\n    } catch {\n      const c = clusters.get('analyst-error') ?? { exemplars: [], share: 0 }\n      if (c.exemplars.length < 5) c.exemplars.push(run.runId)\n      clusters.set('analyst-error', c)\n    }\n  }\n  const clusterList = [...clusters.entries()].map(([id, c]) => ({\n    id,\n    name: id,\n    share: c.exemplars.length / failed.length,\n    exemplars: c.exemplars,\n  }))\n  clusterList.sort((a, b) => b.share - a.share)\n  return { clusters: clusterList, totalFailures: failed.length }\n}\n\n// ── Contamination ──────────────────────────────────────────────────\n\nfunction computeContamination(\n  runs: RunRecord[],\n  canaries: DatasetScenario[],\n): InsightReport['contamination'] {\n  let leaks = 0\n  const details: Array<{ runId: string; canary: string; matched: string }> = []\n  for (const run of runs) {\n    const output = stringifyOutput(run)\n    if (!output) continue\n    const leaksHere = checkCanaries(output, canaries)\n    for (const leak of leaksHere) {\n      leaks++\n      details.push({ runId: run.runId, canary: leak.canary, matched: leak.evidence })\n    }\n  }\n  return { leaks, holdoutAuditPassed: leaks === 0, details }\n}\n\nfunction stringifyOutput(run: RunRecord): string | undefined {\n  // RunRecord doesn't fix where \"the agent's output\" lives — different\n  // consumers stash it differently. We probe the common shapes: the\n  // outcome.raw map (numeric only by design — unlikely to contain text),\n  // and any string-valued fields tucked under metadata via type casting.\n  // Consumers with bespoke shapes pass canaryScenarios only when they\n  // know their runs carry a stringifiable surface.\n  const metadata = (run as unknown as { metadata?: Record<string, unknown> }).metadata\n  if (typeof metadata?.output === 'string') return metadata.output\n  if (typeof metadata?.text === 'string') return metadata.text\n  return undefined\n}\n\n// ── Outcome correlation + linear reward model ──────────────────────\n\nfunction computeOutcomeCorrelation(\n  runs: RunRecord[],\n  outcome: { metric: string; valueByRunId: Record<string, number> },\n  split: 'search' | 'holdout',\n): OutcomeCorrelationInsight | undefined {\n  const xs: number[] = []\n  const ys: number[] = []\n  for (const run of runs) {\n    const y = outcome.valueByRunId[run.runId]\n    if (y === undefined || !Number.isFinite(y)) continue\n    const x = compositeOf(run, split)\n    if (!Number.isFinite(x)) continue\n    xs.push(x)\n    ys.push(y)\n  }\n  if (xs.length < 3) return undefined\n\n  const p = pearson(xs, ys)\n  const s = spearman(xs, ys)\n  const meanX = mean(xs)\n  const meanY = mean(ys)\n  let num = 0\n  let denom = 0\n  for (let i = 0; i < xs.length; i++) {\n    num += (xs[i]! - meanX) * (ys[i]! - meanY)\n    denom += (xs[i]! - meanX) ** 2\n  }\n  const slope = denom === 0 ? 0 : num / denom\n  const intercept = meanY - slope * meanX\n  const ssTot = ys.reduce((a, y) => a + (y - meanY) ** 2, 0)\n  const ssRes = ys.reduce((a, y, i) => a + (y - (intercept + slope * xs[i]!)) ** 2, 0)\n  const r2 = ssTot === 0 ? 0 : 1 - ssRes / ssTot\n\n  return {\n    metric: outcome.metric,\n    n: xs.length,\n    pearson: p,\n    spearman: s,\n    rewardModel: { intercept, slope, r2 },\n  }\n}\n\nfunction spearman(a: number[], b: number[]): number {\n  if (a.length !== b.length || a.length === 0) return 0\n  return pearson(rank(a), rank(b))\n}\n\nfunction rank(arr: number[]): number[] {\n  const indexed = arr.map((v, i) => ({ v, i }))\n  indexed.sort((x, y) => x.v - y.v)\n  const ranks = new Array(arr.length).fill(0)\n  let i = 0\n  while (i < indexed.length) {\n    let j = i\n    while (j + 1 < indexed.length && indexed[j + 1]!.v === indexed[i]!.v) j++\n    const avg = (i + j + 2) / 2\n    for (let k = i; k <= j; k++) ranks[indexed[k]!.i] = avg\n    i = j + 1\n  }\n  return ranks\n}\n\n// ── Release confidence scorecard ───────────────────────────────────\n\nfunction buildReleaseScorecard(\n  composite: ScalarDistribution,\n  lift: LiftInsight | undefined,\n  contamination: InsightReport['contamination'],\n): InsightReport['release'] {\n  // Synthesise a minimal scorecard from the rolled-up signal. The\n  // substrate's `evaluateReleaseConfidence` primitive consumes a richer\n  // input shape that callers can produce by wiring SLO definitions; the\n  // shape here is the contract `selfImprove`/`analyzeRuns` consumers\n  // receive automatically. They can call `evaluateReleaseConfidence`\n  // directly when they want SLO-based axis evaluation.\n  const axes: InsightReport['release']['axes'] = []\n  const liftPass =\n    lift === undefined || lift.ci95[0] > 0\n      ? ('pass' as const)\n      : lift.delta > 0\n        ? ('warn' as const)\n        : ('fail' as const)\n  axes.push({\n    name: 'quality-lift',\n    status: liftPass,\n    detail: lift\n      ? `delta=${lift.delta.toFixed(3)}, CI95=[${lift.ci95[0].toFixed(3)}, ${lift.ci95[1].toFixed(3)}], n=${lift.n}`\n      : 'no baseline/candidate pair available',\n  })\n  const contamPass =\n    contamination === undefined || contamination.leaks === 0 ? ('pass' as const) : ('fail' as const)\n  axes.push({\n    name: 'contamination',\n    status: contamPass,\n    detail: contamination ? `${contamination.leaks} canary leak(s)` : 'no canaries supplied',\n  })\n  axes.push({\n    name: 'composite-distribution',\n    status: composite.mean >= 0.5 ? 'pass' : composite.mean >= 0.3 ? 'warn' : 'fail',\n    detail: `mean=${composite.mean.toFixed(3)}, p50=${composite.p50.toFixed(3)}, p95=${composite.p95.toFixed(3)} over n=${composite.n}`,\n  })\n  const status = axes.some((a) => a.status === 'fail')\n    ? 'fail'\n    : axes.some((a) => a.status === 'warn')\n      ? 'warn'\n      : 'pass'\n  return {\n    status,\n    axes,\n    issues: [],\n  }\n}\n\n// ── Recommendations engine ─────────────────────────────────────────\n\ninterface RecommendationContext {\n  composite: ScalarDistribution\n  judges: Record<string, JudgeInsight>\n  interRater?: InterRaterInsight\n  lift?: LiftInsight\n  failureClusters?: FailureClusterInsight\n  contamination?: InsightReport['contamination']\n  outcomeCorrelation?: OutcomeCorrelationInsight\n  threshold: number\n}\n\nfunction buildRecommendations(ctx: RecommendationContext): Recommendation[] {\n  const out: Recommendation[] = []\n\n  if (ctx.lift) {\n    const decisive = ctx.lift.ci95[0] > ctx.threshold\n    const inconclusive = ctx.lift.ci95[0] <= ctx.threshold && ctx.lift.ci95[1] > ctx.threshold\n    if (decisive) {\n      out.push({\n        priority: 'critical',\n        kind: 'ship',\n        title: `Ship — lift ${ctx.lift.delta.toFixed(3)} (95% CI ${ctx.lift.ci95[0].toFixed(3)}..${ctx.lift.ci95[1].toFixed(3)})`,\n        detail: `Holdout lift exceeds threshold ${ctx.threshold} with 95% bootstrap confidence (n=${ctx.lift.n}, p=${ctx.lift.pValue.toFixed(4)}, d=${ctx.lift.cohensD.toFixed(2)}).`,\n        evidencePath: 'lift',\n      })\n    } else if (inconclusive) {\n      out.push({\n        priority: 'high',\n        kind: 'expand-corpus',\n        title: `Inconclusive — need ~${ctx.lift.requiredN} paired runs (have ${ctx.lift.n}) at current effect size`,\n        detail: `CI straddles threshold. Current MDE at 80% power is ${ctx.lift.mde.toFixed(3)}; observed delta is ${ctx.lift.delta.toFixed(3)}.`,\n        evidencePath: 'lift',\n      })\n    } else {\n      out.push({\n        priority: 'critical',\n        kind: 'hold',\n        title: `Hold — lift CI lower bound ${ctx.lift.ci95[0].toFixed(3)} is at or below threshold ${ctx.threshold}`,\n        detail: `Bootstrap CI provides no statistical evidence the candidate is better. Consider tightening the mutation or expanding the holdout.`,\n        evidencePath: 'lift',\n      })\n    }\n  }\n\n  if (ctx.contamination && ctx.contamination.leaks > 0) {\n    out.push({\n      priority: 'critical',\n      kind: 'fix',\n      title: `${ctx.contamination.leaks} canary leak${ctx.contamination.leaks === 1 ? '' : 's'} detected`,\n      detail: `Holdout integrity is compromised. The lift number is unreliable until you investigate.`,\n      evidencePath: 'contamination',\n    })\n  }\n\n  if (ctx.interRater && ctx.interRater.kappa < 0.5) {\n    out.push({\n      priority: 'high',\n      kind: 'recalibrate',\n      title: `Inter-rater agreement κ=${ctx.interRater.kappa.toFixed(2)} is below 0.5`,\n      detail: `Raters disagree on what 'good' looks like. Top disagreement cases listed in interRater.disagreementCases — consider a triage meeting or refining the rubric.`,\n      evidencePath: 'interRater',\n    })\n  }\n\n  if (ctx.failureClusters && ctx.failureClusters.clusters.length > 0) {\n    const top = ctx.failureClusters.clusters[0]!\n    out.push({\n      priority: 'high',\n      kind: 'investigate',\n      title: `Top failure cluster: ${top.name} (${(top.share * 100).toFixed(0)}% of failures)`,\n      detail: `${ctx.failureClusters.totalFailures} runs failed. The largest cluster groups ${top.exemplars.length} exemplars under '${top.name}'.`,\n      evidencePath: 'failureClusters.clusters[0]',\n    })\n  }\n\n  if (ctx.outcomeCorrelation && Math.abs(ctx.outcomeCorrelation.spearman) < 0.3) {\n    out.push({\n      priority: 'medium',\n      kind: 'recalibrate',\n      title: `Judge scores decoupled from ${ctx.outcomeCorrelation.metric} (Spearman ρ=${ctx.outcomeCorrelation.spearman.toFixed(2)})`,\n      detail: `Your judges score what they were trained to score, but it isn't predicting downstream ${ctx.outcomeCorrelation.metric}. Consider retraining the judge against ${ctx.outcomeCorrelation.metric} as the gold signal.`,\n      evidencePath: 'outcomeCorrelation',\n    })\n  }\n\n  return out\n}\n\n// ── Re-export pareto figure spec for hosted-side rendering ─────────\n\nexport type { ParetoFigureSpec }\n","/**\n * # `selfImprove()` — the LAND-tier one-shot.\n *\n * The cheapest possible call site to run a real closed-loop self-\n * improvement over your agent. Wraps `runImprovementLoop` with smart\n * defaults and a budget-shaped options API; every escape hatch the\n * substrate exposes is reachable from here without losing the\n * one-function feel.\n *\n * Defaults picked to match the LAND-tier story:\n *   - In-memory storage (no filesystem touch).\n *   - `gepaDriver` reflective mutation with copywriting-flavored primitives\n *     (override `driver` or `mutationPrimitives` for any domain).\n *   - `defaultProductionGate` with `deltaThreshold: 0.05`.\n *   - Held-out split = 25% of scenarios, deterministic by id hash.\n *   - 3 generations × population 2 (raise via `budget` for more search).\n *   - `autoOnPromote: 'none'` (we don't open PRs unless you ask).\n *\n * Want one-click? Provide `agent` + `scenarios` + `judge`. Done.\n * Want distributed? Pass `cellPlacement` + an `httpDispatch`-backed\n * agent. Want a code-tier surface? Pass a `MutableSurface` + your own\n * `driver`. Same function.\n */\n\nimport { gepaDriver } from '../campaign/drivers/gepa'\nimport { defaultProductionGate } from '../campaign/gates/default-production-gate'\nimport {\n  type RunImprovementLoopResult,\n  runImprovementLoop,\n} from '../campaign/presets/run-improvement-loop'\nimport { type CampaignStorage, inMemoryCampaignStorage } from '../campaign/storage'\nimport type {\n  CampaignCellResult,\n  DispatchContext,\n  Gate,\n  ImprovementDriver,\n  JudgeConfig,\n  MutableSurface,\n  Scenario,\n} from '../campaign/types'\nimport { createHostedClient, type HostedTenant } from '../hosted/client'\nimport type { EvalRunCellScore, EvalRunEvent, EvalRunGenerationSnapshot } from '../hosted/types'\nimport type { JudgeScoresRecord, RunRecord } from '../run-record'\nimport { analyzeRuns } from './analyze-runs'\nimport type { InsightReport } from './insight-report'\n\nexport interface SelfImproveBudget {\n  /** Hard $ ceiling across all cells in baseline + every generation. Cells\n   *  beyond the ceiling are skipped (cost-aware, not aborted). */\n  dollars?: number\n  /** How many improvement generations to explore. Default 3. Set 0 to\n   *  skip improvement entirely (selfImprove becomes a baseline-only run). */\n  generations?: number\n  /** Candidates the driver proposes per generation. Default 2. */\n  populationSize?: number\n  /** Max concurrent cells across the loop. Default 2. */\n  maxConcurrency?: number\n  /** Fraction of `scenarios` held out from training, used for the gate.\n   *  Default 0.25. Ignored when `holdoutScenarios` is set explicitly. */\n  holdoutFraction?: number\n  /** Explicit held-out scenarios; overrides `holdoutFraction`. */\n  holdoutScenarios?: Scenario[]\n}\n\nexport interface SelfImproveLlm {\n  /** Endpoint base URL. Default Tangle Router. */\n  baseUrl?: string\n  /** Bearer token. Default `process.env.OPENAI_API_KEY`. */\n  apiKey?: string\n  /** Model id used by `gepaDriver` reflection. Default\n   *  `anthropic/claude-sonnet-4.6`. */\n  model?: string\n}\n\nexport type SelfImproveProgressEvent =\n  | { kind: 'baseline.started'; scenarios: number }\n  | { kind: 'baseline.completed'; compositeMean: number; durationMs: number }\n  | { kind: 'generation.started'; index: number; populationSize: number }\n  | { kind: 'generation.completed'; index: number; bestComposite: number; durationMs: number }\n  | { kind: 'gate.decided'; decision: string; lift: number }\n\nexport interface SelfImproveOptions<TScenario extends Scenario, TArtifact> {\n  /**\n   * Your agent — a function that takes the current `MutableSurface`\n   * (typically a system prompt the loop is optimizing) plus the\n   * scenario + cell ctx, and returns the artifact your judge scores.\n   *\n   * Same shape as `RunOptimizationOptions.dispatchWithSurface`. Wrap a\n   * plain `Dispatch` if you don't have a surface seam:\n   *\n   *   agent: (_surface, scenario, ctx) => yourPlainDispatch(scenario, ctx)\n   *\n   * That mode evaluates without mutating any surface — useful as a\n   * baseline-only run (set `budget.generations = 0`).\n   */\n  agent: (surface: MutableSurface, scenario: TScenario, ctx: DispatchContext) => Promise<TArtifact>\n\n  /** Scenarios to evaluate against. Train/holdout split is computed from\n   *  these unless `budget.holdoutScenarios` is set explicitly. */\n  scenarios: TScenario[]\n\n  /** Judge that scores artifacts. Bring your own; use `langchainJudge`\n   *  from `/adapters/langchain` for a Runnable-shaped one. */\n  judge: JudgeConfig<TArtifact, TScenario>\n\n  /** Starting surface — system prompt, JSON config, anything `MutableSurface`\n   *  accepts. The driver mutates this each generation. */\n  baselineSurface: MutableSurface\n\n  /** Budget + loop shape. All fields optional; defaults pick the LAND-tier\n   *  story. */\n  budget?: SelfImproveBudget\n\n  /** Custom driver. Default is `gepaDriver` configured from `llm` +\n   *  `mutationPrimitives`. */\n  driver?: ImprovementDriver\n\n  /** Default-driver overrides — used when `driver` is unset. */\n  mutationPrimitives?: string[]\n  driverTarget?: string\n\n  /** Custom gate. Default is `defaultProductionGate` with\n   *  `deltaThreshold: 0.05` on the held-out split. */\n  gate?: Gate<TArtifact, TScenario>\n\n  /** LLM config consumed by the default `gepaDriver`. Ignored if you pass\n   *  your own `driver`. */\n  llm?: SelfImproveLlm\n\n  /** Storage backend. Default `inMemoryCampaignStorage()` — nothing\n   *  persists past the call. Pass `fsCampaignStorage()` to write to disk. */\n  storage?: CampaignStorage\n\n  /** Run directory (logical for in-memory storage, real path for fs).\n   *  Default `mem://selfImprove-<timestamp>`. */\n  runDir?: string\n\n  /** Distributed-driver seam — same as `RunCampaignOptions.cellPlacement`.\n   *  Returns an opaque placement key the substrate forwards to your agent\n   *  as `ctx.placement`. Combined with `httpDispatch` from\n   *  `/adapters/http`, fans cells across regions. */\n  cellPlacement?: (input: {\n    scenario: TScenario\n    rep: number\n    generation?: number\n  }) => string | undefined\n\n  /** Streaming hook — fires on baseline + each generation + gate decision.\n   *  Consumer routes events wherever (UI, dashboard, logs). */\n  onProgress?: (event: SelfImproveProgressEvent) => void\n\n  /** Auto-promotion behavior on a ship decision. Default `'none'` — we\n   *  return the winner; you ship it however you ship. `'pr'` opens a\n   *  GitHub PR via `openAutoPr`; requires `ghOwner` + `ghRepo`. */\n  autoOnPromote?: 'pr' | 'none'\n  ghOwner?: string\n  ghRepo?: string\n\n  /**\n   * Opt-in: ship eval-run events to a hosted orchestrator (ours, your\n   * self-hosted one, or any compatible implementation of the\n   * `docs/hosted-ingest-spec.md` wire format). When set, the substrate\n   * POSTs the final `EvalRunEvent` to `${endpoint}/v1/ingest/eval-runs`\n   * after the loop completes. Failures are logged but do not fail the\n   * loop — local result is always returned.\n   *\n   * For our orchestrator: `{ endpoint: 'https://orchestrator.tangle.tools/v1', apiKey, tenantId }`.\n   *\n   * For your self-hosted: any URL serving the wire format. See\n   * `examples/hosted-ingest-server/` for the reference receiver.\n   */\n  hostedTenant?: HostedTenant\n\n  /** Free-form labels attached to the hosted event (env, branch, model id,\n   *  etc.). Ignored when `hostedTenant` is unset. */\n  hostedLabels?: Record<string, string>\n}\n\nexport interface SelfImproveResult<TScenario extends Scenario, TArtifact> {\n  /** Composite mean across all scenarios, baseline run. */\n  baseline: {\n    compositeMean: number\n    perScenario: Record<string, number>\n  }\n  /** Composite mean on the held-out set, winner run. */\n  winner: {\n    compositeMean: number\n    perScenario: Record<string, number>\n    surface: MutableSurface\n  }\n  /** `winner.compositeMean - baselineOnHoldout.compositeMean`. Positive\n   *  means the gate observed improvement. */\n  lift: number\n  /** `defaultProductionGate.decide()` result. */\n  gateDecision: 'ship' | 'hold' | 'need_more_work' | 'model_ceiling' | 'arch_ceiling'\n  /** Number of generations actually explored (may be less than the\n   *  budget if the driver gave up early). */\n  generationsExplored: number\n  /** Wall-clock total. */\n  durationMs: number\n  /** Total cost across baseline + every generation. */\n  totalCostUsd: number\n  /**\n   * Rigor packet: distributional summary, paired-bootstrap lift CI,\n   * judge stats, contamination check, recommendations. Wired through\n   * `analyzeRuns()` on the baseline + winner cells of the campaign.\n   * Hosted-tier dashboards render this as the v3-vs-v4 decision view.\n   */\n  insight: InsightReport\n  /**\n   * Raw substrate result for advanced inspection — full per-generation\n   * candidates, full campaign artifacts, all judge scores. Useful for\n   * debugging or reporting beyond the summary.\n   */\n  raw: RunImprovementLoopResult<TArtifact, TScenario>\n}\n\n/**\n * Deterministic train/holdout split by a stable hash of `scenario.id`,\n * so the same scenario set always splits the same way across runs.\n */\nfunction splitTrainHoldout<TScenario extends Scenario>(\n  scenarios: TScenario[],\n  fraction: number,\n): { train: TScenario[]; holdout: TScenario[] } {\n  // Stable fnv-1a-ish hash of the id for ordering.\n  function hash(s: string): number {\n    let h = 2166136261 >>> 0\n    for (let i = 0; i < s.length; i++) {\n      h ^= s.charCodeAt(i)\n      h = Math.imul(h, 16777619) >>> 0\n    }\n    return h\n  }\n  const sorted = [...scenarios].sort((a, b) => hash(a.id) - hash(b.id))\n  const nHoldout = Math.max(1, Math.min(sorted.length - 1, Math.round(sorted.length * fraction)))\n  return {\n    holdout: sorted.slice(0, nHoldout),\n    train: sorted.slice(nHoldout),\n  }\n}\n\nfunction meanComposite(byScenario: Record<string, { meanComposite: number }>): {\n  compositeMean: number\n  perScenario: Record<string, number>\n} {\n  const perScenario: Record<string, number> = {}\n  const values: number[] = []\n  for (const [id, agg] of Object.entries(byScenario)) {\n    perScenario[id] = agg.meanComposite\n    values.push(agg.meanComposite)\n  }\n  return {\n    compositeMean: values.length === 0 ? 0 : values.reduce((s, v) => s + v, 0) / values.length,\n    perScenario,\n  }\n}\n\nconst DEFAULT_MUTATION_PRIMITIVES = [\n  'Tighten the hook: lead with the specific user outcome.',\n  'Replace generic adjectives with specific verbs or proof numbers.',\n  \"Anchor every claim in something the scenario's brief literally supports.\",\n  'Honor the surface-shape constraint (length, register, audience vocabulary).',\n]\n\n/**\n * One-shot self-improvement loop. See module docstring for defaults +\n * extension points.\n *\n * @example Minimum (LAND tier):\n *\n *   const result = await selfImprove({\n *     agent: (surface, scenario, ctx) => myAgent(surface, scenario, ctx.signal),\n *     scenarios,\n *     judge,\n *     baselineSurface: DEFAULT_PROMPT,\n *   })\n *   console.log(`lift: ${result.lift.toFixed(3)} (${result.gateDecision})`)\n *\n * @example Distributed (workers in three regions):\n *\n *   await selfImprove({\n *     agent: httpDispatch({ resolveUrl: ({ placement }) => REGION_URLS[placement!] }),\n *     scenarios,\n *     judge,\n *     baselineSurface: DEFAULT_PROMPT,\n *     cellPlacement: ({ scenario }) => scenario.region,\n *     budget: { maxConcurrency: 12 },\n *   })\n */\nexport async function selfImprove<TScenario extends Scenario, TArtifact>(\n  opts: SelfImproveOptions<TScenario, TArtifact>,\n): Promise<SelfImproveResult<TScenario, TArtifact>> {\n  const startedAt = Date.now()\n\n  const budget = opts.budget ?? {}\n  const generations = budget.generations ?? 3\n  const populationSize = budget.populationSize ?? 2\n  const maxConcurrency = budget.maxConcurrency ?? 2\n  const holdoutFraction = budget.holdoutFraction ?? 0.25\n  const costCeiling = budget.dollars\n\n  const explicitHoldout = budget.holdoutScenarios\n  const { train, holdout } = explicitHoldout\n    ? {\n        train: opts.scenarios.filter((s) => !explicitHoldout.some((h) => h.id === s.id)),\n        holdout: explicitHoldout as TScenario[],\n      }\n    : splitTrainHoldout(opts.scenarios, holdoutFraction)\n\n  if (train.length === 0) {\n    throw new Error(\n      'selfImprove: train split is empty. Reduce holdoutFraction or pass more scenarios.',\n    )\n  }\n  if (holdout.length === 0) {\n    throw new Error('selfImprove: holdout split is empty. Pass more scenarios.')\n  }\n\n  const driver: ImprovementDriver =\n    opts.driver ??\n    gepaDriver({\n      llm: {\n        baseUrl: opts.llm?.baseUrl ?? 'https://router.tangle.tools/v1',\n        apiKey: opts.llm?.apiKey ?? process.env.OPENAI_API_KEY ?? '',\n      },\n      model: opts.llm?.model ?? 'anthropic/claude-sonnet-4.6',\n      target:\n        opts.driverTarget ??\n        'agent surface (system prompt or config) being optimized by selfImprove',\n      mutationPrimitives: opts.mutationPrimitives ?? DEFAULT_MUTATION_PRIMITIVES,\n    })\n\n  const gate: Gate<TArtifact, TScenario> =\n    opts.gate ??\n    defaultProductionGate<TArtifact, TScenario>({\n      holdoutScenarios: holdout,\n      deltaThreshold: 0.05,\n    })\n\n  const storage = opts.storage ?? inMemoryCampaignStorage()\n  const runDir = opts.runDir ?? `mem://selfImprove-${startedAt}`\n\n  if (opts.onProgress) {\n    opts.onProgress({ kind: 'baseline.started', scenarios: opts.scenarios.length })\n  }\n\n  const result = await runImprovementLoop<TScenario, TArtifact>({\n    scenarios: train,\n    baselineSurface: opts.baselineSurface,\n    dispatchWithSurface: opts.agent,\n    driver,\n    judges: [opts.judge],\n    populationSize,\n    maxGenerations: generations,\n    holdoutScenarios: holdout,\n    gate,\n    autoOnPromote: opts.autoOnPromote ?? 'none',\n    ghOwner: opts.ghOwner,\n    ghRepo: opts.ghRepo,\n    storage,\n    runDir,\n    maxConcurrency,\n    cellPlacement: opts.cellPlacement,\n    costCeiling,\n  })\n\n  const baseline = meanComposite(result.baselineOnHoldout.aggregates.byScenario)\n  const winnerStats = meanComposite(result.winnerOnHoldout.aggregates.byScenario)\n\n  if (opts.onProgress) {\n    opts.onProgress({\n      kind: 'baseline.completed',\n      compositeMean: baseline.compositeMean,\n      durationMs: Date.now() - startedAt,\n    })\n    opts.onProgress({\n      kind: 'gate.decided',\n      decision: result.gateResult.decision,\n      lift: winnerStats.compositeMean - baseline.compositeMean,\n    })\n  }\n\n  const totalCost =\n    result.baselineCampaign.aggregates.totalCostUsd +\n    result.generations.reduce(\n      (sum, gen) =>\n        sum + gen.surfaces.reduce((s, sf) => s + sf.campaign.aggregates.totalCostUsd, 0),\n      0,\n    )\n\n  // Rigor packet: feed baseline + winner cells through analyzeRuns().\n  // The two candidates (`baseline` / `winner`) give the lift section a\n  // clean paired comparison; per-judge / per-dimension / cost-quality\n  // sections populate from the cells' judgeScores.\n  const insight = await analyzeRuns({\n    runs: [\n      ...cellsToRunRecords(result.baselineCampaign.cells, 'baseline', runDir),\n      ...cellsToRunRecords(result.winnerOnHoldout.cells, 'winner', runDir),\n    ],\n    baselineCandidateId: 'baseline',\n    candidateCandidateId: 'winner',\n  })\n\n  const summary: SelfImproveResult<TScenario, TArtifact> = {\n    baseline,\n    winner: {\n      ...winnerStats,\n      surface: result.winnerSurface,\n    },\n    lift: winnerStats.compositeMean - baseline.compositeMean,\n    gateDecision: result.gateResult.decision,\n    generationsExplored: result.generations.length,\n    durationMs: Date.now() - startedAt,\n    totalCostUsd: totalCost,\n    insight,\n    raw: result,\n  }\n\n  // Opt-in hosted ingest. Failures logged but never fail the loop — the\n  // local result is always returned. This matches the wedge-doc invariant\n  // that LAND-tier never blocks on EXPAND-tier infra.\n  if (opts.hostedTenant) {\n    try {\n      await shipEvalRunToHosted(opts.hostedTenant, opts, summary, result, runDir)\n    } catch (err) {\n      const msg = err instanceof Error ? err.message : String(err)\n      // eslint-disable-next-line no-console -- intentional: hosted-ingest is best-effort\n      console.warn(`[agent-eval] hosted ingest failed (continuing): ${msg}`)\n    }\n  }\n\n  return summary\n}\n\nasync function shipEvalRunToHosted<TScenario extends Scenario, TArtifact>(\n  tenant: HostedTenant,\n  opts: SelfImproveOptions<TScenario, TArtifact>,\n  summary: SelfImproveResult<TScenario, TArtifact>,\n  raw: RunImprovementLoopResult<TArtifact, TScenario>,\n  runDir: string,\n): Promise<void> {\n  const client = createHostedClient(tenant)\n\n  function snapshotFromCampaign(\n    index: number,\n    surface: MutableSurface | undefined,\n    campaign: RunImprovementLoopResult<TArtifact, TScenario>['baselineCampaign'],\n    durationMs: number,\n  ): EvalRunGenerationSnapshot {\n    const cells: EvalRunCellScore[] = campaign.cells.map((cell) => {\n      const judgeScores = Object.values(cell.judgeScores)\n      const composite =\n        judgeScores.length === 0\n          ? 0\n          : judgeScores.reduce((s, j) => s + j.composite, 0) / judgeScores.length\n      return {\n        scenarioId: cell.scenarioId,\n        rep: cell.rep,\n        compositeMean: composite,\n        dimensions: Object.fromEntries(\n          Object.entries(cell.judgeScores).map(([name, score]) => [name, score.dimensions]),\n        ),\n        errorMessage: cell.error ?? undefined,\n      }\n    })\n    const compositeMean =\n      cells.length === 0 ? 0 : cells.reduce((s, c) => s + c.compositeMean, 0) / cells.length\n    return {\n      index,\n      surfaceHash:\n        typeof surface === 'string'\n          ? hashString(surface)\n          : hashString(JSON.stringify(surface ?? '')),\n      surface,\n      cells,\n      compositeMean,\n      costUsd: campaign.aggregates.totalCostUsd,\n      durationMs,\n    }\n  }\n\n  const generations: EvalRunGenerationSnapshot[] = []\n  // Baseline as generation 0.\n  generations.push(snapshotFromCampaign(0, opts.baselineSurface, raw.baselineCampaign, 0))\n  // Improvement generations as 1..N. Substrate stores per-surface campaigns\n  // per generation — we summarize the WINNING surface per generation here.\n  for (const gen of raw.generations) {\n    const winner = gen.surfaces.reduce(\n      (best, s) =>\n        s.campaign.aggregates.cellsExecuted > 0 &&\n        (best === undefined || averageComposite(s.campaign) > averageComposite(best.campaign))\n          ? s\n          : best,\n      gen.surfaces[0],\n    )\n    if (!winner) continue\n    generations.push(\n      snapshotFromCampaign(gen.record.generationIndex + 1, winner.surface, winner.campaign, 0),\n    )\n  }\n\n  const event: EvalRunEvent = {\n    runId: `${runDir}#${Date.now()}`,\n    runDir,\n    timestamp: new Date().toISOString(),\n    status: 'finished',\n    labels: opts.hostedLabels ?? {},\n    baseline: generations[0],\n    generations,\n    gateDecision: summary.gateDecision,\n    holdoutLift: summary.lift,\n    totalCostUsd: summary.totalCostUsd,\n    totalDurationMs: summary.durationMs,\n    insightReport: summary.insight,\n  }\n\n  await client.ingestEvalRun(event)\n}\n\nfunction averageComposite(\n  campaign: RunImprovementLoopResult<unknown, Scenario>['baselineCampaign'],\n): number {\n  const aggs = Object.values(campaign.aggregates.byScenario)\n  return aggs.length === 0 ? 0 : aggs.reduce((s, a) => s + a.meanComposite, 0) / aggs.length\n}\n\nfunction hashString(s: string): string {\n  let h = 2166136261 >>> 0\n  for (let i = 0; i < s.length; i++) {\n    h ^= s.charCodeAt(i)\n    h = Math.imul(h, 16777619) >>> 0\n  }\n  return h.toString(16).padStart(8, '0')\n}\n\n/**\n * Adapt campaign cells into the `RunRecord` shape `analyzeRuns()` consumes.\n * Each cell becomes one run; `candidateId` is the caller-supplied label so\n * baseline + winner pair cleanly on `(experimentId, seed)`.\n */\nfunction cellsToRunRecords<TArtifact>(\n  cells: ReadonlyArray<CampaignCellResult<TArtifact>>,\n  candidateId: 'baseline' | 'winner',\n  runId: string,\n): RunRecord[] {\n  return cells.map((cell) => {\n    const perJudge: Record<string, Record<string, number>> = {}\n    const perDimMeanAccum: Record<string, { sum: number; n: number }> = {}\n    let compositeSum = 0\n    let compositeCount = 0\n    for (const [judgeId, score] of Object.entries(cell.judgeScores)) {\n      perJudge[judgeId] = { ...score.dimensions }\n      for (const [dim, value] of Object.entries(score.dimensions)) {\n        if (!Number.isFinite(value)) continue\n        const accum = perDimMeanAccum[dim] ?? { sum: 0, n: 0 }\n        accum.sum += value\n        accum.n += 1\n        perDimMeanAccum[dim] = accum\n      }\n      if (Number.isFinite(score.composite)) {\n        compositeSum += score.composite\n        compositeCount += 1\n      }\n    }\n    const perDimMean: Record<string, number> = {}\n    for (const [dim, { sum, n }] of Object.entries(perDimMeanAccum)) {\n      perDimMean[dim] = n === 0 ? 0 : sum / n\n    }\n    const composite = compositeCount === 0 ? 0 : compositeSum / compositeCount\n    const judgeScores: JudgeScoresRecord = {\n      perJudge,\n      perDimMean,\n      composite,\n    }\n    return {\n      runId: `${runId}::${candidateId}::${cell.cellId}`,\n      experimentId: runId,\n      candidateId,\n      // Pair on (scenarioId, rep) — analyzeRuns pairs on (experimentId, seed).\n      // Synthesize a stable seed for that pairing.\n      seed:\n        cell.rep * 1_000_000 +\n        hashString(cell.scenarioId)\n          .slice(0, 6)\n          .split('')\n          .reduce((a, c) => (a * 31 + c.charCodeAt(0)) >>> 0, 0),\n      model: 'campaign-cell',\n      promptHash: 'sha256:cell',\n      configHash: 'sha256:cell',\n      commitSha: 'cell',\n      wallMs: cell.durationMs,\n      costUsd: cell.costUsd,\n      tokenUsage: { input: 0, output: 0 },\n      outcome: {\n        holdoutScore: composite,\n        raw: {},\n        judgeScores,\n      },\n      splitTag: 'holdout',\n      ...(cell.error ? { failureMode: cell.error } : {}),\n    } satisfies RunRecord\n  })\n}\n","/**\n * # `intake/feedback-table` — multi-rater approve/reject corpus → `RunRecord[]`.\n *\n * The generic shape behind Obsidian's `#approved` / `#rejected` tags, a\n * Google Sheet, a Postgres `feedback` table, or any CSV with ratings.\n *\n * Caller supplies one row per (run, rater) tuple plus per-run metadata; the\n * adapter rolls them up into the substrate-canonical `RunRecord` shape so\n * `analyzeRuns({ runs, raterScores })` can produce inter-rater agreement,\n * disagreement triage, and downstream recommendations.\n *\n * Per-run `RunRecord.outcome.searchScore` is the rater-mean rating\n * (normalised to 0..1 when scale is supplied); `outcome.raw` carries the\n * per-rater scores keyed by rater id for downstream attribution.\n */\n\nimport type { JudgeScoresRecord, RunOutcome, RunRecord, RunSplitTag } from '../../run-record'\n\nexport interface FeedbackTableRow {\n  /** Stable id for this run — the unit a rater scored. Drives pairing\n   *  across analysis primitives. */\n  runId: string\n  /** Identifier of the rater that produced this rating. */\n  rater: string\n  /** The rating itself. Accepts boolean (approve/reject), 0..1 scalar,\n   *  or any numeric scale — see `scale`. */\n  rating: number | boolean\n  /** Optional metadata carried through to `RunRecord.outcome.raw` and the\n   *  custom-shape metadata bag. */\n  metadata?: Record<string, unknown>\n}\n\nexport interface FeedbackTableMeta {\n  runId: string\n  /** When omitted, defaults to `'feedback-corpus'`. Used to group related\n   *  runs in `analyzeRuns()` lift analysis. */\n  experimentId?: string\n  /** When omitted, defaults to `runId` — each run is its own candidate. */\n  candidateId?: string\n  /** Cost in USD, when available. Set to 0 when unknown — the consumer's\n   *  cost analysis sections will collapse gracefully. */\n  costUsd?: number\n  /** Wall-clock ms, when available. Defaults to 0. */\n  wallMs?: number\n  /** Model identifier including snapshot. Default `unknown@unknown`. */\n  model?: string\n  /** Optional sha256 of the prompt; default `'sha256:unknown'`. */\n  promptHash?: string\n  /** Default `'sha256:unknown'`. */\n  configHash?: string\n  /** Default `'unknown'`. */\n  commitSha?: string\n  /** Default `'holdout'` — feedback corpora are by nature the holdout\n   *  signal a closed-loop improvement aims at. */\n  splitTag?: RunSplitTag\n  /** Free-form metadata available to consumers via the cast-out path on\n   *  the resulting RunRecord. */\n  extras?: Record<string, unknown>\n}\n\nexport interface FromFeedbackTableOptions {\n  /** Per-(run, rater) ratings. */\n  ratings: FeedbackTableRow[]\n  /** Per-run metadata. When a runId appears in `ratings` but not here, the\n   *  adapter synthesises minimal metadata with defaults documented above. */\n  meta?: FeedbackTableMeta[]\n  /** Rating scale. Provide `{ min, max }` for non-0..1 numeric scales.\n   *  Booleans are normalised: true → 1, false → 0. Default: assumes\n   *  ratings are already 0..1. */\n  scale?: { min: number; max: number }\n  /** When true, the rater scores are emitted into `raterScores` (a sibling\n   *  array `analyzeRuns()` accepts) instead of being averaged into the\n   *  run's `outcome.searchScore`. Default `true` — preserves rater-level\n   *  signal for inter-rater analysis. */\n  emitRaterScores?: boolean\n}\n\nexport interface FromFeedbackTableResult {\n  runs: RunRecord[]\n  /** Rater-level scores ready to pass into `analyzeRuns({ raterScores })`\n   *  for inter-rater agreement + disagreement triage. */\n  raterScores: Array<{ runId: string; rater: string; score: number }>\n}\n\nexport function fromFeedbackTable(opts: FromFeedbackTableOptions): FromFeedbackTableResult {\n  const { ratings, meta = [], scale, emitRaterScores = true } = opts\n  const metaByRun = new Map(meta.map((m) => [m.runId, m]))\n\n  // Normalise per-rating to a 0..1 score.\n  const normalise = (rating: number | boolean): number => {\n    if (typeof rating === 'boolean') return rating ? 1 : 0\n    if (!Number.isFinite(rating)) return Number.NaN\n    if (!scale) return rating\n    const { min, max } = scale\n    if (max === min) return rating\n    return (rating - min) / (max - min)\n  }\n\n  // Group ratings by runId.\n  const byRun = new Map<string, FeedbackTableRow[]>()\n  for (const row of ratings) {\n    const list = byRun.get(row.runId) ?? []\n    list.push(row)\n    byRun.set(row.runId, list)\n  }\n\n  const runs: RunRecord[] = []\n  const raterScores: FromFeedbackTableResult['raterScores'] = []\n\n  for (const [runId, rowsForRun] of byRun) {\n    const normalised = rowsForRun\n      .map((r) => ({ rater: r.rater, score: normalise(r.rating) }))\n      .filter((r) => Number.isFinite(r.score))\n    if (normalised.length === 0) continue\n\n    const meanScore = normalised.reduce((s, r) => s + r.score, 0) / normalised.length\n\n    const runMeta = metaByRun.get(runId) ?? ({ runId } as FeedbackTableMeta)\n\n    const judgeScores: JudgeScoresRecord = {\n      perJudge: Object.fromEntries(normalised.map((r) => [r.rater, { rating: r.score }])),\n      perDimMean: { rating: meanScore },\n      composite: meanScore,\n    }\n\n    const outcome: RunOutcome = {\n      // Feedback corpora ARE the holdout signal — score lands on\n      // `holdoutScore` so downstream substrate primitives (`paretoChart`,\n      // promotion gates) read it correctly by default.\n      holdoutScore: meanScore,\n      raw: Object.fromEntries(normalised.map((r) => [`rater:${r.rater}`, r.score])),\n      judgeScores,\n    }\n\n    runs.push({\n      runId,\n      experimentId: runMeta.experimentId ?? 'feedback-corpus',\n      candidateId: runMeta.candidateId ?? runId,\n      seed: 0,\n      model: runMeta.model ?? 'unknown@unknown',\n      promptHash: runMeta.promptHash ?? 'sha256:unknown',\n      configHash: runMeta.configHash ?? 'sha256:unknown',\n      commitSha: runMeta.commitSha ?? 'unknown',\n      wallMs: runMeta.wallMs ?? 0,\n      costUsd: runMeta.costUsd ?? 0,\n      tokenUsage: { input: 0, output: 0 },\n      outcome,\n      splitTag: runMeta.splitTag ?? 'holdout',\n    } as RunRecord)\n\n    if (emitRaterScores) {\n      for (const r of normalised) raterScores.push({ runId, rater: r.rater, score: r.score })\n    }\n  }\n\n  return { runs, raterScores }\n}\n","/**\n * # `intake/otel-spans` — OTel `TraceSpanEvent[]` → `RunRecord[]`.\n *\n * Turns an existing observability stream into the substrate-canonical\n * `RunRecord` shape so consumers with logs but no eval discipline can\n * call `analyzeRuns()` against their production traffic immediately.\n *\n * Pivot rule: spans are grouped by `tangle.runId` (the same attribute the\n * hosted-tier wire format uses) or, when absent, by `traceId`. One group\n * becomes one `RunRecord`. The root span (no `parentSpanId`) supplies:\n *\n *   - `runId` (the group key)\n *   - `wallMs` from `endTimeUnixNano - startTimeUnixNano`\n *   - `model` from `gen_ai.request.model` / `llm.model` / `tangle.model`\n *   - cost from `cost.usd` / `gen_ai.usage.cost_usd` / `tangle.cost.usd`\n *   - token usage from `gen_ai.usage.{input,output}_tokens`\n *   - `outcome.searchScore` from `tangle.score` / `eval.score` when\n *     present; `outcome.raw` collects every numeric attribute.\n *\n * Spans that ERRORed (`status.code === 'ERROR'`) populate `failureMode`\n * with their `name` so `analyzeRuns()`'s failure clustering sees them.\n */\n\nimport type { TraceSpanEvent } from '../../hosted/types'\nimport type {\n  JudgeScoresRecord,\n  RunOutcome,\n  RunRecord,\n  RunSplitTag,\n  RunTokenUsage,\n} from '../../run-record'\n\nconst SCORE_KEYS = ['tangle.score', 'eval.score', 'score']\nconst MODEL_KEYS = ['tangle.model', 'gen_ai.request.model', 'llm.model', 'model']\nconst COST_KEYS = ['tangle.cost.usd', 'gen_ai.usage.cost_usd', 'cost.usd', 'cost']\nconst INPUT_TOKEN_KEYS = ['gen_ai.usage.input_tokens', 'tangle.tokens.in', 'tokens.in']\nconst OUTPUT_TOKEN_KEYS = ['gen_ai.usage.output_tokens', 'tangle.tokens.out', 'tokens.out']\nconst PROMPT_HASH_KEYS = ['tangle.prompt_hash', 'prompt.hash']\nconst CONFIG_HASH_KEYS = ['tangle.config_hash', 'config.hash']\n\nexport interface FromOtelSpansOptions {\n  spans: TraceSpanEvent[]\n  /** Default split tag for synthesized records. Defaults to `'holdout'`. */\n  defaultSplit?: RunSplitTag\n  /** Default `experimentId` when not present on any span. */\n  experimentId?: string\n}\n\nexport function fromOtelSpans(opts: FromOtelSpansOptions): RunRecord[] {\n  const { spans, defaultSplit = 'holdout', experimentId = 'otel-corpus' } = opts\n  const grouped = groupSpans(spans)\n\n  const runs: RunRecord[] = []\n  for (const [groupKey, groupSpans] of grouped) {\n    const root = findRoot(groupSpans)\n    if (!root) continue\n\n    const wallMs = Math.max(0, (root.endTimeUnixNano - root.startTimeUnixNano) / 1_000_000)\n    const model = readAttrString(groupSpans, MODEL_KEYS) ?? 'unknown@unknown'\n    const costUsd = readAttrNumber(groupSpans, COST_KEYS) ?? 0\n    const inputTokens = readAttrNumber(groupSpans, INPUT_TOKEN_KEYS) ?? 0\n    const outputTokens = readAttrNumber(groupSpans, OUTPUT_TOKEN_KEYS) ?? 0\n    const promptHash = readAttrString(groupSpans, PROMPT_HASH_KEYS) ?? 'sha256:unknown'\n    const configHash = readAttrString(groupSpans, CONFIG_HASH_KEYS) ?? 'sha256:unknown'\n    const score = readAttrNumber(groupSpans, SCORE_KEYS)\n\n    const rawNumeric = collectNumericAttrs(groupSpans)\n    const tokenUsage: RunTokenUsage = {\n      input: inputTokens,\n      output: outputTokens,\n    }\n\n    const judgeScores: JudgeScoresRecord | undefined =\n      score !== undefined\n        ? {\n            perJudge: { 'otel-derived': { score } },\n            perDimMean: { score },\n            composite: score,\n          }\n        : undefined\n\n    const errorSpan = groupSpans.find((s) => s.status?.code === 'ERROR')\n    const outcome: RunOutcome = {\n      ...(opts.defaultSplit === 'search' ? { searchScore: score } : { holdoutScore: score }),\n      raw: rawNumeric,\n      ...(judgeScores ? { judgeScores } : {}),\n    }\n\n    runs.push({\n      runId: groupKey,\n      experimentId,\n      candidateId: (root.attributes['tangle.candidateId'] as string | undefined) ?? 'otel-default',\n      seed: 0,\n      model,\n      promptHash,\n      configHash,\n      commitSha: (root.attributes['tangle.commit_sha'] as string | undefined) ?? 'unknown',\n      wallMs,\n      costUsd,\n      tokenUsage,\n      outcome,\n      splitTag: defaultSplit,\n      ...(errorSpan ? { failureMode: errorSpan.name } : {}),\n    } as RunRecord)\n  }\n  return runs\n}\n\n// ── Internal helpers ────────────────────────────────────────────────\n\nfunction groupSpans(spans: TraceSpanEvent[]): Map<string, TraceSpanEvent[]> {\n  const m = new Map<string, TraceSpanEvent[]>()\n  for (const span of spans) {\n    const key = (span['tangle.runId'] as string | undefined) ?? span.traceId\n    const list = m.get(key) ?? []\n    list.push(span)\n    m.set(key, list)\n  }\n  return m\n}\n\nfunction findRoot(group: TraceSpanEvent[]): TraceSpanEvent | undefined {\n  return group.find((s) => !s.parentSpanId) ?? group[0]\n}\n\nfunction readAttrString(spans: TraceSpanEvent[], keys: string[]): string | undefined {\n  for (const span of spans) {\n    for (const key of keys) {\n      const v = span.attributes[key]\n      if (typeof v === 'string' && v.length > 0) return v\n    }\n  }\n  return undefined\n}\n\nfunction readAttrNumber(spans: TraceSpanEvent[], keys: string[]): number | undefined {\n  for (const span of spans) {\n    for (const key of keys) {\n      const v = span.attributes[key]\n      if (typeof v === 'number' && Number.isFinite(v)) return v\n      if (typeof v === 'string') {\n        const parsed = Number(v)\n        if (Number.isFinite(parsed)) return parsed\n      }\n    }\n  }\n  return undefined\n}\n\nfunction collectNumericAttrs(spans: TraceSpanEvent[]): Record<string, number> {\n  const raw: Record<string, number> = {}\n  for (const span of spans) {\n    for (const [k, v] of Object.entries(span.attributes)) {\n      if (typeof v === 'number' && Number.isFinite(v)) raw[k] = v\n    }\n  }\n  return raw\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA+EA,eAAsB,YAAY,MAAkD;AAClF,QAAM,OAAO,KAAK;AAClB,QAAM,OAAO,KAAK,iBAAiB;AACnC,QAAM,YAAY,KAAK,qBAAqB;AAC5C,QAAM,QAAQ,aAAa,MAAM,KAAK,SAAS,MAAM;AAErD,QAAM,YAAY;AAAA,IAChB,KAAK,IAAI,CAAC,MAAM,YAAY,GAAG,KAAK,CAAC,EAAE,OAAO,OAAO,QAAQ;AAAA,IAC7D;AAAA,EACF;AAEA,QAAM,eAAe,oBAAoB,MAAM,IAAI;AAEnD,QAAM,cAAc;AAAA,IAClB,MAAM,eAAe,KAAK,IAAI,CAAC,MAAM,EAAE,OAAO,EAAE,OAAO,OAAO,QAAQ,GAAG,IAAI;AAAA,IAC7E,QAAQ,YAAY,MAAM,EAAE,MAAM,CAAC;AAAA,EACrC;AAEA,QAAM,SAAS,qBAAqB,IAAI;AAExC,QAAM,aAAa,KAAK,cAAc,kBAAkB,KAAK,WAAW,IAAI;AAE5E,QAAM,OAAO,YAAY,MAAM,KAAK,qBAAqB,KAAK,sBAAsB,KAAK;AAEzF,QAAM,kBAAkB,KAAK,UACzB,MAAM,uBAAuB,MAAM,KAAK,SAAS,KAAK,IACtD;AAEJ,QAAM,gBAAgB,KAAK,kBACvB,qBAAqB,MAAM,KAAK,eAAe,IAC/C;AAEJ,QAAM,qBAAqB,KAAK,gBAC5B,0BAA0B,MAAM,KAAK,eAAe,KAAK,IACzD;AAEJ,QAAM,UAAU,sBAAsB,WAAW,MAAM,aAAa;AAEpE,QAAM,kBAAkB,qBAAqB;AAAA,IAC3C;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF,CAAC;AAED,SAAO;AAAA,IACL,GAAG,KAAK;AAAA,IACR;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF;AACF;AAIA,SAAS,aACP,MACA,MACsB;AACtB,MAAI,SAAS,OAAQ,QAAO;AAC5B,QAAM,aAAa,KAAK,KAAK,CAAC,MAAM,OAAO,SAAS,EAAE,QAAQ,YAAY,CAAC;AAC3E,SAAO,aAAa,YAAY;AAClC;AAEA,SAAS,YAAY,KAAgB,OAAqC;AACxE,QAAM,UAAU,UAAU,YAAY,IAAI,QAAQ,eAAe,IAAI,QAAQ;AAC7E,MAAI,OAAO,SAAS,OAAO,EAAG,QAAO;AAIrC,QAAM,MAAM,UAAU,YAAY,IAAI,QAAQ,cAAc,IAAI,QAAQ;AACxE,SAAO,OAAO,SAAS,GAAG,IAAK,MAAiB,OAAO;AACzD;AAIA,SAAS,eAAe,QAAkB,MAAkC;AAC1E,MAAI,OAAO,WAAW,GAAG;AACvB,WAAO;AAAA,MACL,GAAG;AAAA,MACH,MAAM;AAAA,MACN,KAAK;AAAA,MACL,KAAK;AAAA,MACL,QAAQ;AAAA,MACR,KAAK;AAAA,MACL,KAAK;AAAA,MACL,WAAW,CAAC;AAAA,IACd;AAAA,EACF;AACA,QAAM,SAAS,CAAC,GAAG,MAAM,EAAE,KAAK,CAAC,GAAG,MAAM,IAAI,CAAC;AAC/C,QAAM,IAAI,OAAO;AACjB,QAAMA,QAAO,OAAO,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI;AACjD,QAAM,WAAW,OAAO,OAAO,CAAC,GAAG,MAAM,KAAK,IAAIA,UAAS,GAAG,CAAC,IAAI;AACnE,QAAM,SAAS,KAAK,KAAK,QAAQ;AACjC,SAAO;AAAA,IACL;AAAA,IACA,MAAAA;AAAA,IACA,KAAK,WAAW,QAAQ,GAAG;AAAA,IAC3B,KAAK,WAAW,QAAQ,IAAI;AAAA,IAC5B;AAAA,IACA,KAAK,OAAO,CAAC;AAAA,IACb,KAAK,OAAO,IAAI,CAAC;AAAA,IACjB,WAAW,UAAU,QAAQ,IAAI;AAAA,EACnC;AACF;AAEA,SAAS,WAAW,QAAkB,GAAmB;AACvD,MAAI,OAAO,WAAW,EAAG,QAAO;AAChC,MAAI,OAAO,WAAW,EAAG,QAAO,OAAO,CAAC;AACxC,QAAM,OAAO,OAAO,SAAS,KAAK;AAClC,QAAM,KAAK,KAAK,MAAM,GAAG;AACzB,QAAM,KAAK,KAAK,KAAK,GAAG;AACxB,MAAI,OAAO,GAAI,QAAO,OAAO,EAAE;AAC/B,QAAM,IAAI,MAAM;AAChB,SAAO,OAAO,EAAE,KAAM,IAAI,KAAK,OAAO,EAAE,IAAK;AAC/C;AAKA,SAAS,UAAU,QAAkB,MAA+C;AAClF,MAAI,OAAO,WAAW,KAAK,OAAO,EAAG,QAAO,CAAC;AAC7C,QAAM,MAAM,OAAO,CAAC;AACpB,QAAM,MAAM,OAAO,OAAO,SAAS,CAAC;AACpC,MAAI,QAAQ,IAAK,QAAO,CAAC,EAAE,IAAI,KAAK,IAAI,KAAK,OAAO,OAAO,OAAO,CAAC;AACnE,QAAM,SAAS,MAAM,OAAO;AAC5B,QAAM,MAAuC,CAAC;AAC9C,WAAS,IAAI,GAAG,IAAI,MAAM,KAAK;AAC7B,UAAM,KAAK,MAAM,IAAI;AACrB,UAAM,KAAK,MAAM,OAAO,IAAI,MAAM,KAAK;AACvC,QAAI,KAAK,EAAE,IAAI,IAAI,OAAO,EAAE,CAAC;AAAA,EAC/B;AACA,aAAW,KAAK,QAAQ;AACtB,UAAM,MAAM,KAAK,IAAI,OAAO,GAAG,KAAK,OAAO,IAAI,OAAO,KAAK,CAAC;AAC5D,QAAI,GAAG,EAAG;AAAA,EACZ;AACA,SAAO;AACT;AAEA,SAAS,oBAAoB,MAAmB,MAAkD;AAKhG,QAAM,QAAQ,oBAAI,IAAsB;AACxC,aAAW,OAAO,MAAM;AACtB,UAAM,SAAS,IAAI,QAAQ;AAC3B,QAAI,CAAC,OAAQ;AACb,eAAW,CAAC,KAAK,KAAK,KAAK,OAAO,QAAQ,OAAO,cAAc,CAAC,CAAC,GAAG;AAClE,UAAI,CAAC,OAAO,SAAS,KAAK,EAAG;AAC7B,YAAM,MAAM,MAAM,IAAI,GAAG,KAAK,CAAC;AAC/B,UAAI,KAAK,KAAK;AACd,YAAM,IAAI,KAAK,GAAG;AAAA,IACpB;AAAA,EACF;AACA,QAAM,MAA0C,CAAC;AACjD,aAAW,CAAC,KAAK,MAAM,KAAK,MAAO,KAAI,GAAG,IAAI,eAAe,QAAQ,IAAI;AACzE,SAAO;AACT;AAIA,SAAS,qBAAqB,MAAiD;AAI7E,QAAM,MAAoC,CAAC;AAC3C,QAAM,UAAU,oBAAI,IAAsB;AAC1C,aAAW,OAAO,MAAM;AACtB,UAAM,SAAS,IAAI,QAAQ;AAC3B,QAAI,CAAC,QAAQ,SAAU;AACvB,eAAW,CAAC,SAAS,IAAI,KAAK,OAAO,QAAQ,OAAO,QAAQ,GAAG;AAC7D,YAAM,YAAY,OAAO,OAAO,IAAI,EAAE,OAAO,OAAO,QAAQ;AAC5D,UAAI,UAAU,WAAW,EAAG;AAC5B,YAAM,YAAY,UAAU,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,UAAU;AACnE,YAAM,MAAM,QAAQ,IAAI,OAAO,KAAK,CAAC;AACrC,UAAI,KAAK,SAAS;AAClB,cAAQ,IAAI,SAAS,GAAG;AAAA,IAC1B;AAAA,EACF;AACA,aAAW,CAAC,SAAS,MAAM,KAAK,SAAS;AACvC,QAAI,OAAO,IAAI;AAAA,MACb,GAAG,OAAO;AAAA,MACV,WAAW,OAAO,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,OAAO;AAAA,IACxD;AAAA,EACF;AACA,SAAO;AACT;AAIA,SAAS,kBACP,SAC+B;AAC/B,QAAM,QAAQ,oBAAI,IAAqD;AACvE,aAAW,KAAK,SAAS;AACvB,QAAI,CAAC,OAAO,SAAS,EAAE,KAAK,EAAG;AAC/B,UAAM,OAAO,MAAM,IAAI,EAAE,KAAK,KAAK,CAAC;AACpC,SAAK,KAAK,EAAE,OAAO,EAAE,OAAO,OAAO,EAAE,MAAM,CAAC;AAC5C,UAAM,IAAI,EAAE,OAAO,IAAI;AAAA,EACzB;AACA,QAAM,SAAS,IAAI,IAAI,QAAQ,IAAI,CAAC,MAAM,EAAE,KAAK,CAAC;AAClD,QAAM,eAAyB,CAAC;AAChC,aAAW,CAAC,OAAO,YAAY,KAAK,OAAO;AACzC,UAAM,OAAO,IAAI,IAAI,aAAa,IAAI,CAAC,MAAM,EAAE,KAAK,CAAC;AACrD,QAAI,MAAM;AACV,eAAW,KAAK,OAAQ,KAAI,CAAC,KAAK,IAAI,CAAC,EAAG,OAAM;AAChD,QAAI,IAAK,cAAa,KAAK,KAAK;AAAA,EAClC;AACA,MAAI,OAAO,OAAO,KAAK,aAAa,WAAW,EAAG,QAAO;AAEzD,QAAM,YAAY,CAAC,GAAG,MAAM,EAAE,KAAK;AACnC,QAAM,UAAkC,CAAC;AACzC,WAAS,IAAI,GAAG,IAAI,UAAU,QAAQ,KAAK;AACzC,aAAS,IAAI,IAAI,GAAG,IAAI,UAAU,QAAQ,KAAK;AAC7C,YAAM,IAAI,UAAU,CAAC;AACrB,YAAM,IAAI,UAAU,CAAC;AACrB,YAAM,UAAoB,CAAC;AAC3B,YAAM,UAAoB,CAAC;AAC3B,iBAAW,SAAS,cAAc;AAChC,cAAM,eAAe,MAAM,IAAI,KAAK;AACpC,cAAM,KAAK,aAAa,KAAK,CAAC,MAAM,EAAE,UAAU,CAAC,GAAG;AACpD,cAAM,KAAK,aAAa,KAAK,CAAC,MAAM,EAAE,UAAU,CAAC,GAAG;AACpD,YAAI,OAAO,UAAa,OAAO,QAAW;AACxC,kBAAQ,KAAK,EAAE;AACf,kBAAQ,KAAK,EAAE;AAAA,QACjB;AAAA,MACF;AACA,cAAQ,GAAG,CAAC,KAAK,CAAC,EAAE,IAAI,QAAQ,SAAS,OAAO;AAAA,IAClD;AAAA,EACF;AACA,QAAM,aAAa,OAAO,OAAO,OAAO;AACxC,QAAM,QACJ,WAAW,WAAW,IAAI,IAAI,WAAW,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,WAAW;AAEnF,QAAM,oBAAoB,aACvB,IAAI,CAAC,UAAU;AACd,UAAM,eAAe,MAAM,IAAI,KAAK;AACpC,UAAM,SAAS,aAAa,IAAI,CAAC,MAAM,EAAE,KAAK;AAC9C,UAAM,QAAQ,KAAK,IAAI,GAAG,MAAM,IAAI,KAAK,IAAI,GAAG,MAAM;AACtD,WAAO,EAAE,OAAO,SAAS,cAAc,MAAM;AAAA,EAC/C,CAAC,EACA,KAAK,CAAC,GAAG,MAAM,EAAE,QAAQ,EAAE,KAAK,EAChC,MAAM,GAAG,EAAE;AAEd,SAAO;AAAA,IACL,QAAQ,OAAO;AAAA,IACf,cAAc,aAAa;AAAA,IAC3B;AAAA,IACA;AAAA,IACA;AAAA,EACF;AACF;AAEA,SAAS,QAAQ,GAAa,GAAqB;AACjD,MAAI,EAAE,WAAW,EAAE,UAAU,EAAE,WAAW,EAAG,QAAO;AACpD,QAAM,IAAI,EAAE;AACZ,QAAM,QAAQ,EAAE,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI;AAC7C,QAAM,QAAQ,EAAE,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI;AAC7C,MAAI,MAAM;AACV,MAAI,SAAS;AACb,MAAI,SAAS;AACb,WAAS,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,UAAM,KAAK,EAAE,CAAC,IAAK;AACnB,UAAM,KAAK,EAAE,CAAC,IAAK;AACnB,WAAO,KAAK;AACZ,cAAU,KAAK;AACf,cAAU,KAAK;AAAA,EACjB;AACA,QAAM,QAAQ,KAAK,KAAK,SAAS,MAAM;AACvC,SAAO,UAAU,IAAI,IAAI,MAAM;AACjC;AAIA,SAAS,YACP,MACA,YACA,aACA,OACyB;AACzB,MAAI,MAAM;AACV,MAAI,MAAM;AACV,MAAI,CAAC,OAAO,CAAC,KAAK;AAGhB,UAAM,MAAM,CAAC,GAAG,IAAI,IAAI,KAAK,IAAI,CAAC,MAAM,EAAE,WAAW,CAAC,CAAC;AACvD,QAAI,IAAI,WAAW,EAAG,QAAO;AAC7B,UAAM,CAAC,KAAK,GAAG,IAAI;AACnB,UAAM,QAAQ,KAAK,KAAK,OAAO,CAAC,MAAM,EAAE,gBAAgB,GAAG,EAAE,IAAI,CAAC,MAAM,YAAY,GAAG,KAAK,CAAC,CAAC;AAC9F,UAAM,QAAQ,KAAK,KAAK,OAAO,CAAC,MAAM,EAAE,gBAAgB,GAAG,EAAE,IAAI,CAAC,MAAM,YAAY,GAAG,KAAK,CAAC,CAAC;AAC9F,UAAM,SAAS,QAAQ,MAAM;AAC7B,UAAM,SAAS,QAAQ,MAAM;AAAA,EAC/B;AAEA,QAAM,WAAW,KAAK,OAAO,CAAC,MAAM,EAAE,gBAAgB,GAAG;AACzD,QAAM,YAAY,KAAK,OAAO,CAAC,MAAM,EAAE,gBAAgB,GAAG;AAC1D,MAAI,SAAS,WAAW,KAAK,UAAU,WAAW,EAAG,QAAO;AAI5D,QAAM,gBAAgB,IAAI,IAAI,SAAS,IAAI,CAAC,MAAM,CAAC,WAAW,CAAC,GAAG,CAAC,CAAC,CAAC;AACrE,QAAM,iBAA2B,CAAC;AAClC,QAAM,kBAA4B,CAAC;AACnC,MAAI,iBAAiB;AACrB,aAAW,QAAQ,WAAW;AAC5B,UAAM,IAAI,cAAc,IAAI,WAAW,IAAI,CAAC;AAC5C,QAAI,GAAG;AACL,YAAM,KAAK,YAAY,GAAG,KAAK;AAC/B,YAAM,KAAK,YAAY,MAAM,KAAK;AAClC,UAAI,OAAO,SAAS,EAAE,KAAK,OAAO,SAAS,EAAE,GAAG;AAC9C,uBAAe,KAAK,EAAE;AACtB,wBAAgB,KAAK,EAAE;AACvB,yBAAiB;AAAA,MACnB;AAAA,IACF;AAAA,EACF;AACA,MAAI,CAAC,gBAAgB;AACnB,UAAM,IAAI,KAAK,IAAI,SAAS,QAAQ,UAAU,MAAM;AACpD,aAAS,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,YAAM,KAAK,YAAY,SAAS,CAAC,GAAI,KAAK;AAC1C,YAAM,KAAK,YAAY,UAAU,CAAC,GAAI,KAAK;AAC3C,UAAI,OAAO,SAAS,EAAE,KAAK,OAAO,SAAS,EAAE,GAAG;AAC9C,uBAAe,KAAK,EAAE;AACtB,wBAAgB,KAAK,EAAE;AAAA,MACzB;AAAA,IACF;AAAA,EACF;AACA,MAAI,eAAe,WAAW,EAAG,QAAO;AAExC,QAAM,eAAe,KAAK,cAAc;AACxC,QAAM,gBAAgB,KAAK,eAAe;AAC1C,QAAM,QAAQ,gBAAgB;AAE9B,QAAM,YAAY,gBAAgB,gBAAgB,iBAAiB;AAAA,IACjE,YAAY;AAAA,IACZ,WAAW;AAAA,IACX,WAAW;AAAA,EACb,CAAC;AACD,QAAM,QAAQ,YAAY,gBAAgB,eAAe;AACzD,QAAM,IAAI,QAAQ,gBAAgB,eAAe;AACjD,QAAM,MAAM,UAAU,EAAE,SAAS,eAAe,QAAQ,OAAO,KAAK,OAAO,KAAK,CAAC;AACjF,QAAM,YAAY,mBAAmB;AAAA,IACnC,QAAQ,KAAK,IAAI,KAAK,IAAI,KAAK,GAAG,IAAI;AAAA,IACtC,OAAO;AAAA,IACP,OAAO;AAAA,EACT,CAAC;AAED,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA;AAAA,IACA,MAAM,CAAC,UAAU,KAAK,UAAU,IAAI;AAAA,IACpC,QAAQ,MAAM;AAAA,IACd,GAAG,eAAe;AAAA,IAClB,SAAS;AAAA,IACT;AAAA,IACA;AAAA,EACF;AACF;AAEA,SAAS,WAAW,GAAsB;AACxC,SAAO,GAAG,EAAE,YAAY,KAAK,EAAE,IAAI;AACrC;AAEA,SAAS,KAAK,KAAuB;AACnC,SAAO,IAAI,WAAW,IAAI,IAAI,IAAI,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,IAAI;AACrE;AAIA,eAAe,uBACb,MACA,SACA,OAC4C;AAC5C,QAAM,SAAS,KAAK,OAAO,CAAC,MAAM,YAAY,GAAG,KAAK,IAAI,OAAO,EAAE,gBAAgB,MAAS;AAC5F,MAAI,OAAO,WAAW,EAAG,QAAO,EAAE,UAAU,CAAC,GAAG,eAAe,EAAE;AAEjE,QAAM,WAAW,oBAAI,IAAoD;AACzE,aAAW,OAAO,QAAQ;AACxB,QAAI;AACF,YAAM,SAAS,MAAM,QAAQ,IAAI,IAAI,OAAO;AAAA,QAC1C,MAAM;AAAA,QACN;AAAA,MACF,CAAsC;AACtC,iBAAW,WAAW,OAAO,UAA8B;AACzD,cAAM,MAAM,QAAQ,QAAQ,QAAQ,cAAc;AAClD,cAAM,IAAI,SAAS,IAAI,GAAG,KAAK,EAAE,WAAW,CAAC,GAAG,OAAO,EAAE;AACzD,YAAI,EAAE,UAAU,SAAS,EAAG,GAAE,UAAU,KAAK,IAAI,KAAK;AACtD,iBAAS,IAAI,KAAK,CAAC;AAAA,MACrB;AAAA,IACF,QAAQ;AACN,YAAM,IAAI,SAAS,IAAI,eAAe,KAAK,EAAE,WAAW,CAAC,GAAG,OAAO,EAAE;AACrE,UAAI,EAAE,UAAU,SAAS,EAAG,GAAE,UAAU,KAAK,IAAI,KAAK;AACtD,eAAS,IAAI,iBAAiB,CAAC;AAAA,IACjC;AAAA,EACF;AACA,QAAM,cAAc,CAAC,GAAG,SAAS,QAAQ,CAAC,EAAE,IAAI,CAAC,CAAC,IAAI,CAAC,OAAO;AAAA,IAC5D;AAAA,IACA,MAAM;AAAA,IACN,OAAO,EAAE,UAAU,SAAS,OAAO;AAAA,IACnC,WAAW,EAAE;AAAA,EACf,EAAE;AACF,cAAY,KAAK,CAAC,GAAG,MAAM,EAAE,QAAQ,EAAE,KAAK;AAC5C,SAAO,EAAE,UAAU,aAAa,eAAe,OAAO,OAAO;AAC/D;AAIA,SAAS,qBACP,MACA,UACgC;AAChC,MAAI,QAAQ;AACZ,QAAM,UAAqE,CAAC;AAC5E,aAAW,OAAO,MAAM;AACtB,UAAM,SAAS,gBAAgB,GAAG;AAClC,QAAI,CAAC,OAAQ;AACb,UAAM,YAAY,cAAc,QAAQ,QAAQ;AAChD,eAAW,QAAQ,WAAW;AAC5B;AACA,cAAQ,KAAK,EAAE,OAAO,IAAI,OAAO,QAAQ,KAAK,QAAQ,SAAS,KAAK,SAAS,CAAC;AAAA,IAChF;AAAA,EACF;AACA,SAAO,EAAE,OAAO,oBAAoB,UAAU,GAAG,QAAQ;AAC3D;AAEA,SAAS,gBAAgB,KAAoC;AAO3D,QAAM,WAAY,IAA0D;AAC5E,MAAI,OAAO,UAAU,WAAW,SAAU,QAAO,SAAS;AAC1D,MAAI,OAAO,UAAU,SAAS,SAAU,QAAO,SAAS;AACxD,SAAO;AACT;AAIA,SAAS,0BACP,MACA,SACA,OACuC;AACvC,QAAM,KAAe,CAAC;AACtB,QAAM,KAAe,CAAC;AACtB,aAAW,OAAO,MAAM;AACtB,UAAM,IAAI,QAAQ,aAAa,IAAI,KAAK;AACxC,QAAI,MAAM,UAAa,CAAC,OAAO,SAAS,CAAC,EAAG;AAC5C,UAAM,IAAI,YAAY,KAAK,KAAK;AAChC,QAAI,CAAC,OAAO,SAAS,CAAC,EAAG;AACzB,OAAG,KAAK,CAAC;AACT,OAAG,KAAK,CAAC;AAAA,EACX;AACA,MAAI,GAAG,SAAS,EAAG,QAAO;AAE1B,QAAM,IAAI,QAAQ,IAAI,EAAE;AACxB,QAAM,IAAI,SAAS,IAAI,EAAE;AACzB,QAAM,QAAQ,KAAK,EAAE;AACrB,QAAM,QAAQ,KAAK,EAAE;AACrB,MAAI,MAAM;AACV,MAAI,QAAQ;AACZ,WAAS,IAAI,GAAG,IAAI,GAAG,QAAQ,KAAK;AAClC,YAAQ,GAAG,CAAC,IAAK,UAAU,GAAG,CAAC,IAAK;AACpC,cAAU,GAAG,CAAC,IAAK,UAAU;AAAA,EAC/B;AACA,QAAM,QAAQ,UAAU,IAAI,IAAI,MAAM;AACtC,QAAM,YAAY,QAAQ,QAAQ;AAClC,QAAM,QAAQ,GAAG,OAAO,CAAC,GAAG,MAAM,KAAK,IAAI,UAAU,GAAG,CAAC;AACzD,QAAM,QAAQ,GAAG,OAAO,CAAC,GAAG,GAAG,MAAM,KAAK,KAAK,YAAY,QAAQ,GAAG,CAAC,OAAQ,GAAG,CAAC;AACnF,QAAM,KAAK,UAAU,IAAI,IAAI,IAAI,QAAQ;AAEzC,SAAO;AAAA,IACL,QAAQ,QAAQ;AAAA,IAChB,GAAG,GAAG;AAAA,IACN,SAAS;AAAA,IACT,UAAU;AAAA,IACV,aAAa,EAAE,WAAW,OAAO,GAAG;AAAA,EACtC;AACF;AAEA,SAAS,SAAS,GAAa,GAAqB;AAClD,MAAI,EAAE,WAAW,EAAE,UAAU,EAAE,WAAW,EAAG,QAAO;AACpD,SAAO,QAAQ,KAAK,CAAC,GAAG,KAAK,CAAC,CAAC;AACjC;AAEA,SAAS,KAAK,KAAyB;AACrC,QAAM,UAAU,IAAI,IAAI,CAAC,GAAGC,QAAO,EAAE,GAAG,GAAAA,GAAE,EAAE;AAC5C,UAAQ,KAAK,CAAC,GAAG,MAAM,EAAE,IAAI,EAAE,CAAC;AAChC,QAAM,QAAQ,IAAI,MAAM,IAAI,MAAM,EAAE,KAAK,CAAC;AAC1C,MAAI,IAAI;AACR,SAAO,IAAI,QAAQ,QAAQ;AACzB,QAAI,IAAI;AACR,WAAO,IAAI,IAAI,QAAQ,UAAU,QAAQ,IAAI,CAAC,EAAG,MAAM,QAAQ,CAAC,EAAG,EAAG;AACtE,UAAM,OAAO,IAAI,IAAI,KAAK;AAC1B,aAAS,IAAI,GAAG,KAAK,GAAG,IAAK,OAAM,QAAQ,CAAC,EAAG,CAAC,IAAI;AACpD,QAAI,IAAI;AAAA,EACV;AACA,SAAO;AACT;AAIA,SAAS,sBACP,WACA,MACA,eAC0B;AAO1B,QAAM,OAAyC,CAAC;AAChD,QAAM,WACJ,SAAS,UAAa,KAAK,KAAK,CAAC,IAAI,IAChC,SACD,KAAK,QAAQ,IACV,SACA;AACT,OAAK,KAAK;AAAA,IACR,MAAM;AAAA,IACN,QAAQ;AAAA,IACR,QAAQ,OACJ,SAAS,KAAK,MAAM,QAAQ,CAAC,CAAC,WAAW,KAAK,KAAK,CAAC,EAAE,QAAQ,CAAC,CAAC,KAAK,KAAK,KAAK,CAAC,EAAE,QAAQ,CAAC,CAAC,QAAQ,KAAK,CAAC,KAC1G;AAAA,EACN,CAAC;AACD,QAAM,aACJ,kBAAkB,UAAa,cAAc,UAAU,IAAK,SAAoB;AAClF,OAAK,KAAK;AAAA,IACR,MAAM;AAAA,IACN,QAAQ;AAAA,IACR,QAAQ,gBAAgB,GAAG,cAAc,KAAK,oBAAoB;AAAA,EACpE,CAAC;AACD,OAAK,KAAK;AAAA,IACR,MAAM;AAAA,IACN,QAAQ,UAAU,QAAQ,MAAM,SAAS,UAAU,QAAQ,MAAM,SAAS;AAAA,IAC1E,QAAQ,QAAQ,UAAU,KAAK,QAAQ,CAAC,CAAC,SAAS,UAAU,IAAI,QAAQ,CAAC,CAAC,SAAS,UAAU,IAAI,QAAQ,CAAC,CAAC,WAAW,UAAU,CAAC;AAAA,EACnI,CAAC;AACD,QAAM,SAAS,KAAK,KAAK,CAAC,MAAM,EAAE,WAAW,MAAM,IAC/C,SACA,KAAK,KAAK,CAAC,MAAM,EAAE,WAAW,MAAM,IAClC,SACA;AACN,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA,QAAQ,CAAC;AAAA,EACX;AACF;AAeA,SAAS,qBAAqB,KAA8C;AAC1E,QAAM,MAAwB,CAAC;AAE/B,MAAI,IAAI,MAAM;AACZ,UAAM,WAAW,IAAI,KAAK,KAAK,CAAC,IAAI,IAAI;AACxC,UAAM,eAAe,IAAI,KAAK,KAAK,CAAC,KAAK,IAAI,aAAa,IAAI,KAAK,KAAK,CAAC,IAAI,IAAI;AACjF,QAAI,UAAU;AACZ,UAAI,KAAK;AAAA,QACP,UAAU;AAAA,QACV,MAAM;AAAA,QACN,OAAO,oBAAe,IAAI,KAAK,MAAM,QAAQ,CAAC,CAAC,YAAY,IAAI,KAAK,KAAK,CAAC,EAAE,QAAQ,CAAC,CAAC,KAAK,IAAI,KAAK,KAAK,CAAC,EAAE,QAAQ,CAAC,CAAC;AAAA,QACtH,QAAQ,kCAAkC,IAAI,SAAS,qCAAqC,IAAI,KAAK,CAAC,OAAO,IAAI,KAAK,OAAO,QAAQ,CAAC,CAAC,OAAO,IAAI,KAAK,QAAQ,QAAQ,CAAC,CAAC;AAAA,QACzK,cAAc;AAAA,MAChB,CAAC;AAAA,IACH,WAAW,cAAc;AACvB,UAAI,KAAK;AAAA,QACP,UAAU;AAAA,QACV,MAAM;AAAA,QACN,OAAO,6BAAwB,IAAI,KAAK,SAAS,sBAAsB,IAAI,KAAK,CAAC;AAAA,QACjF,QAAQ,uDAAuD,IAAI,KAAK,IAAI,QAAQ,CAAC,CAAC,uBAAuB,IAAI,KAAK,MAAM,QAAQ,CAAC,CAAC;AAAA,QACtI,cAAc;AAAA,MAChB,CAAC;AAAA,IACH,OAAO;AACL,UAAI,KAAK;AAAA,QACP,UAAU;AAAA,QACV,MAAM;AAAA,QACN,OAAO,mCAA8B,IAAI,KAAK,KAAK,CAAC,EAAE,QAAQ,CAAC,CAAC,6BAA6B,IAAI,SAAS;AAAA,QAC1G,QAAQ;AAAA,QACR,cAAc;AAAA,MAChB,CAAC;AAAA,IACH;AAAA,EACF;AAEA,MAAI,IAAI,iBAAiB,IAAI,cAAc,QAAQ,GAAG;AACpD,QAAI,KAAK;AAAA,MACP,UAAU;AAAA,MACV,MAAM;AAAA,MACN,OAAO,GAAG,IAAI,cAAc,KAAK,eAAe,IAAI,cAAc,UAAU,IAAI,KAAK,GAAG;AAAA,MACxF,QAAQ;AAAA,MACR,cAAc;AAAA,IAChB,CAAC;AAAA,EACH;AAEA,MAAI,IAAI,cAAc,IAAI,WAAW,QAAQ,KAAK;AAChD,QAAI,KAAK;AAAA,MACP,UAAU;AAAA,MACV,MAAM;AAAA,MACN,OAAO,gCAA2B,IAAI,WAAW,MAAM,QAAQ,CAAC,CAAC;AAAA,MACjE,QAAQ;AAAA,MACR,cAAc;AAAA,IAChB,CAAC;AAAA,EACH;AAEA,MAAI,IAAI,mBAAmB,IAAI,gBAAgB,SAAS,SAAS,GAAG;AAClE,UAAM,MAAM,IAAI,gBAAgB,SAAS,CAAC;AAC1C,QAAI,KAAK;AAAA,MACP,UAAU;AAAA,MACV,MAAM;AAAA,MACN,OAAO,wBAAwB,IAAI,IAAI,MAAM,IAAI,QAAQ,KAAK,QAAQ,CAAC,CAAC;AAAA,MACxE,QAAQ,GAAG,IAAI,gBAAgB,aAAa,4CAA4C,IAAI,UAAU,MAAM,qBAAqB,IAAI,IAAI;AAAA,MACzI,cAAc;AAAA,IAChB,CAAC;AAAA,EACH;AAEA,MAAI,IAAI,sBAAsB,KAAK,IAAI,IAAI,mBAAmB,QAAQ,IAAI,KAAK;AAC7E,QAAI,KAAK;AAAA,MACP,UAAU;AAAA,MACV,MAAM;AAAA,MACN,OAAO,+BAA+B,IAAI,mBAAmB,MAAM,qBAAgB,IAAI,mBAAmB,SAAS,QAAQ,CAAC,CAAC;AAAA,MAC7H,QAAQ,yFAAyF,IAAI,mBAAmB,MAAM,2CAA2C,IAAI,mBAAmB,MAAM;AAAA,MACtM,cAAc;AAAA,IAChB,CAAC;AAAA,EACH;AAEA,SAAO;AACT;;;AClgBA,SAAS,kBACP,WACA,UAC8C;AAE9C,WAAS,KAAK,GAAmB;AAC/B,QAAI,IAAI,eAAe;AACvB,aAAS,IAAI,GAAG,IAAI,EAAE,QAAQ,KAAK;AACjC,WAAK,EAAE,WAAW,CAAC;AACnB,UAAI,KAAK,KAAK,GAAG,QAAQ,MAAM;AAAA,IACjC;AACA,WAAO;AAAA,EACT;AACA,QAAM,SAAS,CAAC,GAAG,SAAS,EAAE,KAAK,CAAC,GAAG,MAAM,KAAK,EAAE,EAAE,IAAI,KAAK,EAAE,EAAE,CAAC;AACpE,QAAM,WAAW,KAAK,IAAI,GAAG,KAAK,IAAI,OAAO,SAAS,GAAG,KAAK,MAAM,OAAO,SAAS,QAAQ,CAAC,CAAC;AAC9F,SAAO;AAAA,IACL,SAAS,OAAO,MAAM,GAAG,QAAQ;AAAA,IACjC,OAAO,OAAO,MAAM,QAAQ;AAAA,EAC9B;AACF;AAEA,SAAS,cAAc,YAGrB;AACA,QAAM,cAAsC,CAAC;AAC7C,QAAM,SAAmB,CAAC;AAC1B,aAAW,CAAC,IAAI,GAAG,KAAK,OAAO,QAAQ,UAAU,GAAG;AAClD,gBAAY,EAAE,IAAI,IAAI;AACtB,WAAO,KAAK,IAAI,aAAa;AAAA,EAC/B;AACA,SAAO;AAAA,IACL,eAAe,OAAO,WAAW,IAAI,IAAI,OAAO,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,OAAO;AAAA,IACpF;AAAA,EACF;AACF;AAEA,IAAM,8BAA8B;AAAA,EAClC;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF;AA2BA,eAAsB,YACpB,MACkD;AAClD,QAAM,YAAY,KAAK,IAAI;AAE3B,QAAM,SAAS,KAAK,UAAU,CAAC;AAC/B,QAAM,cAAc,OAAO,eAAe;AAC1C,QAAM,iBAAiB,OAAO,kBAAkB;AAChD,QAAM,iBAAiB,OAAO,kBAAkB;AAChD,QAAM,kBAAkB,OAAO,mBAAmB;AAClD,QAAM,cAAc,OAAO;AAE3B,QAAM,kBAAkB,OAAO;AAC/B,QAAM,EAAE,OAAO,QAAQ,IAAI,kBACvB;AAAA,IACE,OAAO,KAAK,UAAU,OAAO,CAAC,MAAM,CAAC,gBAAgB,KAAK,CAAC,MAAM,EAAE,OAAO,EAAE,EAAE,CAAC;AAAA,IAC/E,SAAS;AAAA,EACX,IACA,kBAAkB,KAAK,WAAW,eAAe;AAErD,MAAI,MAAM,WAAW,GAAG;AACtB,UAAM,IAAI;AAAA,MACR;AAAA,IACF;AAAA,EACF;AACA,MAAI,QAAQ,WAAW,GAAG;AACxB,UAAM,IAAI,MAAM,2DAA2D;AAAA,EAC7E;AAEA,QAAM,SACJ,KAAK,UACL,WAAW;AAAA,IACT,KAAK;AAAA,MACH,SAAS,KAAK,KAAK,WAAW;AAAA,MAC9B,QAAQ,KAAK,KAAK,UAAU,QAAQ,IAAI,kBAAkB;AAAA,IAC5D;AAAA,IACA,OAAO,KAAK,KAAK,SAAS;AAAA,IAC1B,QACE,KAAK,gBACL;AAAA,IACF,oBAAoB,KAAK,sBAAsB;AAAA,EACjD,CAAC;AAEH,QAAM,OACJ,KAAK,QACL,sBAA4C;AAAA,IAC1C,kBAAkB;AAAA,IAClB,gBAAgB;AAAA,EAClB,CAAC;AAEH,QAAM,UAAU,KAAK,WAAW,wBAAwB;AACxD,QAAM,SAAS,KAAK,UAAU,qBAAqB,SAAS;AAE5D,MAAI,KAAK,YAAY;AACnB,SAAK,WAAW,EAAE,MAAM,oBAAoB,WAAW,KAAK,UAAU,OAAO,CAAC;AAAA,EAChF;AAEA,QAAM,SAAS,MAAM,mBAAyC;AAAA,IAC5D,WAAW;AAAA,IACX,iBAAiB,KAAK;AAAA,IACtB,qBAAqB,KAAK;AAAA,IAC1B;AAAA,IACA,QAAQ,CAAC,KAAK,KAAK;AAAA,IACnB;AAAA,IACA,gBAAgB;AAAA,IAChB,kBAAkB;AAAA,IAClB;AAAA,IACA,eAAe,KAAK,iBAAiB;AAAA,IACrC,SAAS,KAAK;AAAA,IACd,QAAQ,KAAK;AAAA,IACb;AAAA,IACA;AAAA,IACA;AAAA,IACA,eAAe,KAAK;AAAA,IACpB;AAAA,EACF,CAAC;AAED,QAAM,WAAW,cAAc,OAAO,kBAAkB,WAAW,UAAU;AAC7E,QAAM,cAAc,cAAc,OAAO,gBAAgB,WAAW,UAAU;AAE9E,MAAI,KAAK,YAAY;AACnB,SAAK,WAAW;AAAA,MACd,MAAM;AAAA,MACN,eAAe,SAAS;AAAA,MACxB,YAAY,KAAK,IAAI,IAAI;AAAA,IAC3B,CAAC;AACD,SAAK,WAAW;AAAA,MACd,MAAM;AAAA,MACN,UAAU,OAAO,WAAW;AAAA,MAC5B,MAAM,YAAY,gBAAgB,SAAS;AAAA,IAC7C,CAAC;AAAA,EACH;AAEA,QAAM,YACJ,OAAO,iBAAiB,WAAW,eACnC,OAAO,YAAY;AAAA,IACjB,CAAC,KAAK,QACJ,MAAM,IAAI,SAAS,OAAO,CAAC,GAAG,OAAO,IAAI,GAAG,SAAS,WAAW,cAAc,CAAC;AAAA,IACjF;AAAA,EACF;AAMF,QAAM,UAAU,MAAM,YAAY;AAAA,IAChC,MAAM;AAAA,MACJ,GAAG,kBAAkB,OAAO,iBAAiB,OAAO,YAAY,MAAM;AAAA,MACtE,GAAG,kBAAkB,OAAO,gBAAgB,OAAO,UAAU,MAAM;AAAA,IACrE;AAAA,IACA,qBAAqB;AAAA,IACrB,sBAAsB;AAAA,EACxB,CAAC;AAED,QAAM,UAAmD;AAAA,IACvD;AAAA,IACA,QAAQ;AAAA,MACN,GAAG;AAAA,MACH,SAAS,OAAO;AAAA,IAClB;AAAA,IACA,MAAM,YAAY,gBAAgB,SAAS;AAAA,IAC3C,cAAc,OAAO,WAAW;AAAA,IAChC,qBAAqB,OAAO,YAAY;AAAA,IACxC,YAAY,KAAK,IAAI,IAAI;AAAA,IACzB,cAAc;AAAA,IACd;AAAA,IACA,KAAK;AAAA,EACP;AAKA,MAAI,KAAK,cAAc;AACrB,QAAI;AACF,YAAM,oBAAoB,KAAK,cAAc,MAAM,SAAS,QAAQ,MAAM;AAAA,IAC5E,SAAS,KAAK;AACZ,YAAM,MAAM,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAE3D,cAAQ,KAAK,mDAAmD,GAAG,EAAE;AAAA,IACvE;AAAA,EACF;AAEA,SAAO;AACT;AAEA,eAAe,oBACb,QACA,MACA,SACA,KACA,QACe;AACf,QAAM,SAAS,mBAAmB,MAAM;AAExC,WAAS,qBACP,OACA,SACA,UACA,YAC2B;AAC3B,UAAM,QAA4B,SAAS,MAAM,IAAI,CAAC,SAAS;AAC7D,YAAM,cAAc,OAAO,OAAO,KAAK,WAAW;AAClD,YAAM,YACJ,YAAY,WAAW,IACnB,IACA,YAAY,OAAO,CAAC,GAAG,MAAM,IAAI,EAAE,WAAW,CAAC,IAAI,YAAY;AACrE,aAAO;AAAA,QACL,YAAY,KAAK;AAAA,QACjB,KAAK,KAAK;AAAA,QACV,eAAe;AAAA,QACf,YAAY,OAAO;AAAA,UACjB,OAAO,QAAQ,KAAK,WAAW,EAAE,IAAI,CAAC,CAAC,MAAM,KAAK,MAAM,CAAC,MAAM,MAAM,UAAU,CAAC;AAAA,QAClF;AAAA,QACA,cAAc,KAAK,SAAS;AAAA,MAC9B;AAAA,IACF,CAAC;AACD,UAAM,gBACJ,MAAM,WAAW,IAAI,IAAI,MAAM,OAAO,CAAC,GAAG,MAAM,IAAI,EAAE,eAAe,CAAC,IAAI,MAAM;AAClF,WAAO;AAAA,MACL;AAAA,MACA,aACE,OAAO,YAAY,WACf,WAAW,OAAO,IAClB,WAAW,KAAK,UAAU,WAAW,EAAE,CAAC;AAAA,MAC9C;AAAA,MACA;AAAA,MACA;AAAA,MACA,SAAS,SAAS,WAAW;AAAA,MAC7B;AAAA,IACF;AAAA,EACF;AAEA,QAAM,cAA2C,CAAC;AAElD,cAAY,KAAK,qBAAqB,GAAG,KAAK,iBAAiB,IAAI,kBAAkB,CAAC,CAAC;AAGvF,aAAW,OAAO,IAAI,aAAa;AACjC,UAAM,SAAS,IAAI,SAAS;AAAA,MAC1B,CAAC,MAAM,MACL,EAAE,SAAS,WAAW,gBAAgB,MACrC,SAAS,UAAa,iBAAiB,EAAE,QAAQ,IAAI,iBAAiB,KAAK,QAAQ,KAChF,IACA;AAAA,MACN,IAAI,SAAS,CAAC;AAAA,IAChB;AACA,QAAI,CAAC,OAAQ;AACb,gBAAY;AAAA,MACV,qBAAqB,IAAI,OAAO,kBAAkB,GAAG,OAAO,SAAS,OAAO,UAAU,CAAC;AAAA,IACzF;AAAA,EACF;AAEA,QAAM,QAAsB;AAAA,IAC1B,OAAO,GAAG,MAAM,IAAI,KAAK,IAAI,CAAC;AAAA,IAC9B;AAAA,IACA,YAAW,oBAAI,KAAK,GAAE,YAAY;AAAA,IAClC,QAAQ;AAAA,IACR,QAAQ,KAAK,gBAAgB,CAAC;AAAA,IAC9B,UAAU,YAAY,CAAC;AAAA,IACvB;AAAA,IACA,cAAc,QAAQ;AAAA,IACtB,aAAa,QAAQ;AAAA,IACrB,cAAc,QAAQ;AAAA,IACtB,iBAAiB,QAAQ;AAAA,IACzB,eAAe,QAAQ;AAAA,EACzB;AAEA,QAAM,OAAO,cAAc,KAAK;AAClC;AAEA,SAAS,iBACP,UACQ;AACR,QAAM,OAAO,OAAO,OAAO,SAAS,WAAW,UAAU;AACzD,SAAO,KAAK,WAAW,IAAI,IAAI,KAAK,OAAO,CAAC,GAAG,MAAM,IAAI,EAAE,eAAe,CAAC,IAAI,KAAK;AACtF;AAEA,SAAS,WAAW,GAAmB;AACrC,MAAI,IAAI,eAAe;AACvB,WAAS,IAAI,GAAG,IAAI,EAAE,QAAQ,KAAK;AACjC,SAAK,EAAE,WAAW,CAAC;AACnB,QAAI,KAAK,KAAK,GAAG,QAAQ,MAAM;AAAA,EACjC;AACA,SAAO,EAAE,SAAS,EAAE,EAAE,SAAS,GAAG,GAAG;AACvC;AAOA,SAAS,kBACP,OACA,aACA,OACa;AACb,SAAO,MAAM,IAAI,CAAC,SAAS;AACzB,UAAM,WAAmD,CAAC;AAC1D,UAAM,kBAA8D,CAAC;AACrE,QAAI,eAAe;AACnB,QAAI,iBAAiB;AACrB,eAAW,CAAC,SAAS,KAAK,KAAK,OAAO,QAAQ,KAAK,WAAW,GAAG;AAC/D,eAAS,OAAO,IAAI,EAAE,GAAG,MAAM,WAAW;AAC1C,iBAAW,CAAC,KAAK,KAAK,KAAK,OAAO,QAAQ,MAAM,UAAU,GAAG;AAC3D,YAAI,CAAC,OAAO,SAAS,KAAK,EAAG;AAC7B,cAAM,QAAQ,gBAAgB,GAAG,KAAK,EAAE,KAAK,GAAG,GAAG,EAAE;AACrD,cAAM,OAAO;AACb,cAAM,KAAK;AACX,wBAAgB,GAAG,IAAI;AAAA,MACzB;AACA,UAAI,OAAO,SAAS,MAAM,SAAS,GAAG;AACpC,wBAAgB,MAAM;AACtB,0BAAkB;AAAA,MACpB;AAAA,IACF;AACA,UAAM,aAAqC,CAAC;AAC5C,eAAW,CAAC,KAAK,EAAE,KAAK,EAAE,CAAC,KAAK,OAAO,QAAQ,eAAe,GAAG;AAC/D,iBAAW,GAAG,IAAI,MAAM,IAAI,IAAI,MAAM;AAAA,IACxC;AACA,UAAM,YAAY,mBAAmB,IAAI,IAAI,eAAe;AAC5D,UAAM,cAAiC;AAAA,MACrC;AAAA,MACA;AAAA,MACA;AAAA,IACF;AACA,WAAO;AAAA,MACL,OAAO,GAAG,KAAK,KAAK,WAAW,KAAK,KAAK,MAAM;AAAA,MAC/C,cAAc;AAAA,MACd;AAAA;AAAA;AAAA,MAGA,MACE,KAAK,MAAM,MACX,WAAW,KAAK,UAAU,EACvB,MAAM,GAAG,CAAC,EACV,MAAM,EAAE,EACR,OAAO,CAAC,GAAG,MAAO,IAAI,KAAK,EAAE,WAAW,CAAC,MAAO,GAAG,CAAC;AAAA,MACzD,OAAO;AAAA,MACP,YAAY;AAAA,MACZ,YAAY;AAAA,MACZ,WAAW;AAAA,MACX,QAAQ,KAAK;AAAA,MACb,SAAS,KAAK;AAAA,MACd,YAAY,EAAE,OAAO,GAAG,QAAQ,EAAE;AAAA,MAClC,SAAS;AAAA,QACP,cAAc;AAAA,QACd,KAAK,CAAC;AAAA,QACN;AAAA,MACF;AAAA,MACA,UAAU;AAAA,MACV,GAAI,KAAK,QAAQ,EAAE,aAAa,KAAK,MAAM,IAAI,CAAC;AAAA,IAClD;AAAA,EACF,CAAC;AACH;;;ACvgBO,SAAS,kBAAkB,MAAyD;AACzF,QAAM,EAAE,SAAS,OAAO,CAAC,GAAG,OAAO,kBAAkB,KAAK,IAAI;AAC9D,QAAM,YAAY,IAAI,IAAI,KAAK,IAAI,CAAC,MAAM,CAAC,EAAE,OAAO,CAAC,CAAC,CAAC;AAGvD,QAAM,YAAY,CAAC,WAAqC;AACtD,QAAI,OAAO,WAAW,UAAW,QAAO,SAAS,IAAI;AACrD,QAAI,CAAC,OAAO,SAAS,MAAM,EAAG,QAAO,OAAO;AAC5C,QAAI,CAAC,MAAO,QAAO;AACnB,UAAM,EAAE,KAAK,IAAI,IAAI;AACrB,QAAI,QAAQ,IAAK,QAAO;AACxB,YAAQ,SAAS,QAAQ,MAAM;AAAA,EACjC;AAGA,QAAM,QAAQ,oBAAI,IAAgC;AAClD,aAAW,OAAO,SAAS;AACzB,UAAM,OAAO,MAAM,IAAI,IAAI,KAAK,KAAK,CAAC;AACtC,SAAK,KAAK,GAAG;AACb,UAAM,IAAI,IAAI,OAAO,IAAI;AAAA,EAC3B;AAEA,QAAM,OAAoB,CAAC;AAC3B,QAAM,cAAsD,CAAC;AAE7D,aAAW,CAAC,OAAO,UAAU,KAAK,OAAO;AACvC,UAAM,aAAa,WAChB,IAAI,CAAC,OAAO,EAAE,OAAO,EAAE,OAAO,OAAO,UAAU,EAAE,MAAM,EAAE,EAAE,EAC3D,OAAO,CAAC,MAAM,OAAO,SAAS,EAAE,KAAK,CAAC;AACzC,QAAI,WAAW,WAAW,EAAG;AAE7B,UAAM,YAAY,WAAW,OAAO,CAAC,GAAG,MAAM,IAAI,EAAE,OAAO,CAAC,IAAI,WAAW;AAE3E,UAAM,UAAU,UAAU,IAAI,KAAK,KAAM,EAAE,MAAM;AAEjD,UAAM,cAAiC;AAAA,MACrC,UAAU,OAAO,YAAY,WAAW,IAAI,CAAC,MAAM,CAAC,EAAE,OAAO,EAAE,QAAQ,EAAE,MAAM,CAAC,CAAC,CAAC;AAAA,MAClF,YAAY,EAAE,QAAQ,UAAU;AAAA,MAChC,WAAW;AAAA,IACb;AAEA,UAAM,UAAsB;AAAA;AAAA;AAAA;AAAA,MAI1B,cAAc;AAAA,MACd,KAAK,OAAO,YAAY,WAAW,IAAI,CAAC,MAAM,CAAC,SAAS,EAAE,KAAK,IAAI,EAAE,KAAK,CAAC,CAAC;AAAA,MAC5E;AAAA,IACF;AAEA,SAAK,KAAK;AAAA,MACR;AAAA,MACA,cAAc,QAAQ,gBAAgB;AAAA,MACtC,aAAa,QAAQ,eAAe;AAAA,MACpC,MAAM;AAAA,MACN,OAAO,QAAQ,SAAS;AAAA,MACxB,YAAY,QAAQ,cAAc;AAAA,MAClC,YAAY,QAAQ,cAAc;AAAA,MAClC,WAAW,QAAQ,aAAa;AAAA,MAChC,QAAQ,QAAQ,UAAU;AAAA,MAC1B,SAAS,QAAQ,WAAW;AAAA,MAC5B,YAAY,EAAE,OAAO,GAAG,QAAQ,EAAE;AAAA,MAClC;AAAA,MACA,UAAU,QAAQ,YAAY;AAAA,IAChC,CAAc;AAEd,QAAI,iBAAiB;AACnB,iBAAW,KAAK,WAAY,aAAY,KAAK,EAAE,OAAO,OAAO,EAAE,OAAO,OAAO,EAAE,MAAM,CAAC;AAAA,IACxF;AAAA,EACF;AAEA,SAAO,EAAE,MAAM,YAAY;AAC7B;;;AC5HA,IAAM,aAAa,CAAC,gBAAgB,cAAc,OAAO;AACzD,IAAM,aAAa,CAAC,gBAAgB,wBAAwB,aAAa,OAAO;AAChF,IAAM,YAAY,CAAC,mBAAmB,yBAAyB,YAAY,MAAM;AACjF,IAAM,mBAAmB,CAAC,6BAA6B,oBAAoB,WAAW;AACtF,IAAM,oBAAoB,CAAC,8BAA8B,qBAAqB,YAAY;AAC1F,IAAM,mBAAmB,CAAC,sBAAsB,aAAa;AAC7D,IAAM,mBAAmB,CAAC,sBAAsB,aAAa;AAUtD,SAAS,cAAc,MAAyC;AACrE,QAAM,EAAE,OAAO,eAAe,WAAW,eAAe,cAAc,IAAI;AAC1E,QAAM,UAAU,WAAW,KAAK;AAEhC,QAAM,OAAoB,CAAC;AAC3B,aAAW,CAAC,UAAUC,WAAU,KAAK,SAAS;AAC5C,UAAM,OAAO,SAASA,WAAU;AAChC,QAAI,CAAC,KAAM;AAEX,UAAM,SAAS,KAAK,IAAI,IAAI,KAAK,kBAAkB,KAAK,qBAAqB,GAAS;AACtF,UAAM,QAAQ,eAAeA,aAAY,UAAU,KAAK;AACxD,UAAM,UAAU,eAAeA,aAAY,SAAS,KAAK;AACzD,UAAM,cAAc,eAAeA,aAAY,gBAAgB,KAAK;AACpE,UAAM,eAAe,eAAeA,aAAY,iBAAiB,KAAK;AACtE,UAAM,aAAa,eAAeA,aAAY,gBAAgB,KAAK;AACnE,UAAM,aAAa,eAAeA,aAAY,gBAAgB,KAAK;AACnE,UAAM,QAAQ,eAAeA,aAAY,UAAU;AAEnD,UAAM,aAAa,oBAAoBA,WAAU;AACjD,UAAM,aAA4B;AAAA,MAChC,OAAO;AAAA,MACP,QAAQ;AAAA,IACV;AAEA,UAAM,cACJ,UAAU,SACN;AAAA,MACE,UAAU,EAAE,gBAAgB,EAAE,MAAM,EAAE;AAAA,MACtC,YAAY,EAAE,MAAM;AAAA,MACpB,WAAW;AAAA,IACb,IACA;AAEN,UAAM,YAAYA,YAAW,KAAK,CAAC,MAAM,EAAE,QAAQ,SAAS,OAAO;AACnE,UAAM,UAAsB;AAAA,MAC1B,GAAI,KAAK,iBAAiB,WAAW,EAAE,aAAa,MAAM,IAAI,EAAE,cAAc,MAAM;AAAA,MACpF,KAAK;AAAA,MACL,GAAI,cAAc,EAAE,YAAY,IAAI,CAAC;AAAA,IACvC;AAEA,SAAK,KAAK;AAAA,MACR,OAAO;AAAA,MACP;AAAA,MACA,aAAc,KAAK,WAAW,oBAAoB,KAA4B;AAAA,MAC9E,MAAM;AAAA,MACN;AAAA,MACA;AAAA,MACA;AAAA,MACA,WAAY,KAAK,WAAW,mBAAmB,KAA4B;AAAA,MAC3E;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA,UAAU;AAAA,MACV,GAAI,YAAY,EAAE,aAAa,UAAU,KAAK,IAAI,CAAC;AAAA,IACrD,CAAc;AAAA,EAChB;AACA,SAAO;AACT;AAIA,SAAS,WAAW,OAAwD;AAC1E,QAAM,IAAI,oBAAI,IAA8B;AAC5C,aAAW,QAAQ,OAAO;AACxB,UAAM,MAAO,KAAK,cAAc,KAA4B,KAAK;AACjE,UAAM,OAAO,EAAE,IAAI,GAAG,KAAK,CAAC;AAC5B,SAAK,KAAK,IAAI;AACd,MAAE,IAAI,KAAK,IAAI;AAAA,EACjB;AACA,SAAO;AACT;AAEA,SAAS,SAAS,OAAqD;AACrE,SAAO,MAAM,KAAK,CAAC,MAAM,CAAC,EAAE,YAAY,KAAK,MAAM,CAAC;AACtD;AAEA,SAAS,eAAe,OAAyB,MAAoC;AACnF,aAAW,QAAQ,OAAO;AACxB,eAAW,OAAO,MAAM;AACtB,YAAM,IAAI,KAAK,WAAW,GAAG;AAC7B,UAAI,OAAO,MAAM,YAAY,EAAE,SAAS,EAAG,QAAO;AAAA,IACpD;AAAA,EACF;AACA,SAAO;AACT;AAEA,SAAS,eAAe,OAAyB,MAAoC;AACnF,aAAW,QAAQ,OAAO;AACxB,eAAW,OAAO,MAAM;AACtB,YAAM,IAAI,KAAK,WAAW,GAAG;AAC7B,UAAI,OAAO,MAAM,YAAY,OAAO,SAAS,CAAC,EAAG,QAAO;AACxD,UAAI,OAAO,MAAM,UAAU;AACzB,cAAM,SAAS,OAAO,CAAC;AACvB,YAAI,OAAO,SAAS,MAAM,EAAG,QAAO;AAAA,MACtC;AAAA,IACF;AAAA,EACF;AACA,SAAO;AACT;AAEA,SAAS,oBAAoB,OAAiD;AAC5E,QAAM,MAA8B,CAAC;AACrC,aAAW,QAAQ,OAAO;AACxB,eAAW,CAAC,GAAG,CAAC,KAAK,OAAO,QAAQ,KAAK,UAAU,GAAG;AACpD,UAAI,OAAO,MAAM,YAAY,OAAO,SAAS,CAAC,EAAG,KAAI,CAAC,IAAI;AAAA,IAC5D;AAAA,EACF;AACA,SAAO;AACT;","names":["mean","i","groupSpans"]}
+{"version":3,"sources":["../../src/contract/analyze-runs.ts","../../src/contract/self-improve.ts","../../src/contract/intake/feedback-table.ts","../../src/contract/intake/otel-spans.ts"],"sourcesContent":["/**\n * # `analyzeRuns()` — turn a set of agent runs into an actionable decision packet.\n *\n * Wires the substrate's statistical, calibration, clustering, Pareto, and\n * release-confidence primitives into one `InsightReport`. Two top-level\n * entry points use this function:\n *\n *   - `selfImprove()` calls it on the campaign output to attach a packet\n *     to every run.\n *   - Consumers with observed `RunRecord[]` (production traces, gold\n *     corpora, approve/reject tables) call it directly via `analyzeRuns()`\n *     for analysis without a closed loop.\n *\n * Every section is opt-in based on what the input data supports — the\n * function never invents signal. If runs carry no judge scores, `judges`\n * is empty. If there's no baseline/candidate split, `lift` is undefined.\n * If no `analyst` is wired, `failureClusters` is undefined.\n *\n * The `recommendations` array is the human-readable layer; everything\n * else is the evidence backing each recommendation.\n */\n\nimport type { AnalystRegistry } from '../analyst/registry'\nimport type { AnalystFinding } from '../analyst/types'\nimport { checkCanaries } from '../contamination-guard'\nimport type { DatasetScenario } from '../dataset'\nimport type { RunRecord } from '../run-record'\nimport { cohensD, pairedBootstrap, pairedMde, pairedTTest, requiredSampleSize } from '../statistics'\nimport { type ParetoFigureSpec, paretoChart } from '../summary-report'\n\nimport type {\n  FailureClusterInsight,\n  InsightReport,\n  InterRaterInsight,\n  JudgeInsight,\n  LiftInsight,\n  OutcomeCorrelationInsight,\n  Recommendation,\n  ScalarDistribution,\n} from './insight-report'\n\n// ── Public API ───────────────────────────────────────────────────────\n\nexport interface AnalyzeRunsOptions {\n  /** The runs to analyze. */\n  runs: RunRecord[]\n  /** Which split to score against when reading composite from RunOutcome.\n   *  Default: holdout when ANY run has a `holdoutScore`, else search. */\n  split?: 'search' | 'holdout' | 'auto'\n  /** Pairwise analysis configuration. When both `baselineCandidateId` and\n   *  `candidateCandidateId` are present, lift is computed on paired\n   *  (experimentId, seed) tuples shared between the two sides. */\n  baselineCandidateId?: string\n  candidateCandidateId?: string\n  /** Canary scenarios — checked against every run's raw output for\n   *  holdout contamination. */\n  canaryScenarios?: DatasetScenario[]\n  /** Analyst registry for failure clustering. When omitted, the\n   *  `failureClusters` section is left undefined. */\n  analyst?: AnalystRegistry\n  /** Downstream outcome metric per run (e.g. engagement rate, approval\n   *  rate, downstream pass rate). When present, the report includes\n   *  `outcomeCorrelation` + a simple linear reward model fit. */\n  outcomeSignal?: {\n    metric: string\n    valueByRunId: Record<string, number>\n  }\n  /** Multi-rater feedback for inter-rater agreement. Each entry is one\n   *  rater's score for one run. Two or more raters → kappa + disagreement\n   *  triage list. */\n  raterScores?: Array<{ runId: string; rater: string; score: number }>\n  /** Number of histogram bins for distributional summaries. Default 12. */\n  histogramBins?: number\n  /** Decision threshold — the smallest composite lift the caller cares\n   *  about. Used by the recommendations engine to call ship vs hold.\n   *  Default 0.02. */\n  decisionThreshold?: number\n}\n\nexport async function analyzeRuns(opts: AnalyzeRunsOptions): Promise<InsightReport> {\n  const runs = opts.runs\n  const bins = opts.histogramBins ?? 12\n  const threshold = opts.decisionThreshold ?? 0.02\n  const split = resolveSplit(runs, opts.split ?? 'auto')\n\n  const compositeWithIds = runs\n    .map((r) => ({ runId: r.runId, score: compositeOf(r, split) }))\n    .filter((p) => Number.isFinite(p.score))\n  const composite = distributionOf(\n    compositeWithIds.map((p) => p.score),\n    bins,\n    compositeWithIds,\n  )\n\n  const perDimension = computePerDimension(runs, bins)\n\n  const costs = runs.map((r) => r.costUsd).filter(Number.isFinite)\n  const costDist = distributionOf(costs, bins)\n  const pareto = paretoChart(runs, { split })\n  const degraded: { cost?: string; pareto?: string } = {}\n  if (costs.length === 0 || costs.every((c) => c === 0)) {\n    degraded.cost = 'no costUsd values recorded — cost axis carries no signal'\n  }\n  if (pareto.points.length < 2) {\n    degraded.pareto =\n      pareto.points.length === 0\n        ? 'no candidates — Pareto unavailable'\n        : 'single candidate — Pareto is a single point, not a frontier'\n  }\n  const costQuality = {\n    cost: costDist,\n    pareto,\n    ...(degraded.cost || degraded.pareto ? { degraded } : {}),\n  }\n\n  const judges = computeJudgeInsights(runs)\n\n  const interRater = opts.raterScores ? computeInterRater(opts.raterScores) : undefined\n\n  const lift = computeLift(runs, opts.baselineCandidateId, opts.candidateCandidateId, split)\n\n  const failureClusters = opts.analyst\n    ? await computeFailureClusters(runs, opts.analyst, split)\n    : undefined\n\n  const contamination = opts.canaryScenarios\n    ? computeContamination(runs, opts.canaryScenarios)\n    : undefined\n\n  const outcomeCorrelation = opts.outcomeSignal\n    ? computeOutcomeCorrelation(runs, opts.outcomeSignal, split)\n    : undefined\n\n  const release = buildReleaseScorecard(composite, lift, contamination)\n\n  const recommendations = buildRecommendations({\n    composite,\n    judges,\n    interRater,\n    lift,\n    failureClusters,\n    contamination,\n    outcomeCorrelation,\n    threshold,\n  })\n\n  return {\n    n: runs.length,\n    composite,\n    perDimension,\n    costQuality,\n    judges,\n    interRater,\n    lift,\n    failureClusters,\n    contamination,\n    outcomeCorrelation,\n    release,\n    recommendations,\n  }\n}\n\n// ── Composite + split selection ─────────────────────────────────────\n\nfunction resolveSplit(\n  runs: RunRecord[],\n  pref: 'search' | 'holdout' | 'auto',\n): 'search' | 'holdout' {\n  if (pref !== 'auto') return pref\n  const hasHoldout = runs.some((r) => Number.isFinite(r.outcome.holdoutScore))\n  return hasHoldout ? 'holdout' : 'search'\n}\n\nfunction compositeOf(run: RunRecord, split: 'search' | 'holdout'): number {\n  const primary = split === 'holdout' ? run.outcome.holdoutScore : run.outcome.searchScore\n  if (Number.isFinite(primary)) return primary as number\n  // Fall through to the other split if the preferred one is missing —\n  // analyzeRuns shouldn't refuse to summarise a run just because the\n  // caller asked for the split that wasn't recorded.\n  const alt = split === 'holdout' ? run.outcome.searchScore : run.outcome.holdoutScore\n  return Number.isFinite(alt) ? (alt as number) : Number.NaN\n}\n\n// ── Distribution helpers ────────────────────────────────────────────\n\nfunction distributionOf(\n  values: number[],\n  bins: number,\n  withIds?: Array<{ runId: string; score: number }>,\n): ScalarDistribution {\n  if (values.length === 0) {\n    return {\n      n: 0,\n      mean: 0,\n      p50: 0,\n      p95: 0,\n      stddev: 0,\n      min: 0,\n      max: 0,\n      histogram: [],\n    }\n  }\n  const sorted = [...values].sort((a, b) => a - b)\n  const n = sorted.length\n  const mean = sorted.reduce((s, v) => s + v, 0) / n\n  const variance = sorted.reduce((s, v) => s + (v - mean) ** 2, 0) / n\n  const stddev = Math.sqrt(variance)\n  const tailRuns = withIds\n    ? [...withIds].sort((a, b) => a.score - b.score).slice(0, Math.min(5, withIds.length))\n    : undefined\n  return {\n    n,\n    mean,\n    p50: percentile(sorted, 0.5),\n    p95: percentile(sorted, 0.95),\n    stddev,\n    min: sorted[0]!,\n    max: sorted[n - 1]!,\n    histogram: histogram(sorted, bins),\n    ...(tailRuns ? { tailRuns } : {}),\n  }\n}\n\nfunction percentile(sorted: number[], q: number): number {\n  if (sorted.length === 0) return 0\n  if (sorted.length === 1) return sorted[0]!\n  const idx = (sorted.length - 1) * q\n  const lo = Math.floor(idx)\n  const hi = Math.ceil(idx)\n  if (lo === hi) return sorted[lo]!\n  const w = idx - lo\n  return sorted[lo]! * (1 - w) + sorted[hi]! * w\n}\n\n/** Even-width histogram over the value range. Returns inclusive-lo /\n *  exclusive-hi bins (closed on right for the last bin) compatible with\n *  the substrate's `GainDistributionBin` shape. */\nfunction histogram(sorted: number[], bins: number): ScalarDistribution['histogram'] {\n  if (sorted.length === 0 || bins < 1) return []\n  const min = sorted[0]!\n  const max = sorted[sorted.length - 1]!\n  if (min === max) return [{ lo: min, hi: max, count: sorted.length }]\n  const width = (max - min) / bins\n  const out: ScalarDistribution['histogram'] = []\n  for (let i = 0; i < bins; i++) {\n    const lo = min + i * width\n    const hi = i === bins - 1 ? max : lo + width\n    out.push({ lo, hi, count: 0 })\n  }\n  for (const v of sorted) {\n    const idx = Math.min(bins - 1, Math.floor((v - min) / width))\n    out[idx]!.count++\n  }\n  return out\n}\n\nfunction computePerDimension(runs: RunRecord[], bins: number): Record<string, ScalarDistribution> {\n  // JudgeScoresRecord pre-aggregates `perDimMean` (mean across judges per\n  // dimension). We collect those means across runs to produce a per-dim\n  // distribution at the corpus level. Consumers who want per-judge\n  // dimension values reach into `perJudge[judgeId][dim]` themselves.\n  const byDim = new Map<string, number[]>()\n  for (const run of runs) {\n    const scores = run.outcome.judgeScores\n    if (!scores) continue\n    for (const [dim, value] of Object.entries(scores.perDimMean ?? {})) {\n      if (!Number.isFinite(value)) continue\n      const arr = byDim.get(dim) ?? []\n      arr.push(value)\n      byDim.set(dim, arr)\n    }\n  }\n  const out: Record<string, ScalarDistribution> = {}\n  for (const [dim, values] of byDim) out[dim] = distributionOf(values, bins)\n  return out\n}\n\n// ── Judge insights ──────────────────────────────────────────────────\n\nfunction computeJudgeInsights(runs: RunRecord[]): Record<string, JudgeInsight> {\n  // Each judge's per-run mean is the average of its per-dimension scores\n  // for that run. We aggregate those means across all runs each judge\n  // scored — giving consumers a \"this judge's typical verdict\" reading.\n  const out: Record<string, JudgeInsight> = {}\n  const byJudge = new Map<string, number[]>()\n  for (const run of runs) {\n    const scores = run.outcome.judgeScores\n    if (!scores?.perJudge) continue\n    for (const [judgeId, dims] of Object.entries(scores.perJudge)) {\n      const dimValues = Object.values(dims).filter(Number.isFinite) as number[]\n      if (dimValues.length === 0) continue\n      const judgeMean = dimValues.reduce((s, v) => s + v, 0) / dimValues.length\n      const arr = byJudge.get(judgeId) ?? []\n      arr.push(judgeMean)\n      byJudge.set(judgeId, arr)\n    }\n  }\n  for (const [judgeId, values] of byJudge) {\n    out[judgeId] = {\n      n: values.length,\n      meanScore: values.reduce((s, v) => s + v, 0) / values.length,\n    }\n  }\n  return out\n}\n\n// ── Inter-rater agreement ───────────────────────────────────────────\n\nfunction computeInterRater(\n  ratings: Array<{ runId: string; rater: string; score: number }>,\n): InterRaterInsight | undefined {\n  const byRun = new Map<string, Array<{ rater: string; score: number }>>()\n  for (const r of ratings) {\n    if (!Number.isFinite(r.score)) continue\n    const list = byRun.get(r.runId) ?? []\n    list.push({ rater: r.rater, score: r.score })\n    byRun.set(r.runId, list)\n  }\n  const raters = new Set(ratings.map((r) => r.rater))\n  const jointlyRated: string[] = []\n  for (const [runId, ratersForRun] of byRun) {\n    const seen = new Set(ratersForRun.map((r) => r.rater))\n    let all = true\n    for (const r of raters) if (!seen.has(r)) all = false\n    if (all) jointlyRated.push(runId)\n  }\n  if (raters.size < 2 || jointlyRated.length === 0) return undefined\n\n  const raterList = [...raters].sort()\n  const perPair: Record<string, number> = {}\n  for (let i = 0; i < raterList.length; i++) {\n    for (let j = i + 1; j < raterList.length; j++) {\n      const a = raterList[i]!\n      const b = raterList[j]!\n      const aScores: number[] = []\n      const bScores: number[] = []\n      for (const runId of jointlyRated) {\n        const ratersForRun = byRun.get(runId)!\n        const sa = ratersForRun.find((r) => r.rater === a)?.score\n        const sb = ratersForRun.find((r) => r.rater === b)?.score\n        if (sa !== undefined && sb !== undefined) {\n          aScores.push(sa)\n          bScores.push(sb)\n        }\n      }\n      perPair[`${a}::${b}`] = pearson(aScores, bScores)\n    }\n  }\n  const pairKappas = Object.values(perPair)\n  const kappa =\n    pairKappas.length === 0 ? 0 : pairKappas.reduce((s, v) => s + v, 0) / pairKappas.length\n\n  const disagreementCases = jointlyRated\n    .map((runId) => {\n      const ratersForRun = byRun.get(runId)!\n      const scores = ratersForRun.map((r) => r.score)\n      const range = Math.max(...scores) - Math.min(...scores)\n      return { runId, ratings: ratersForRun, range }\n    })\n    .sort((a, b) => b.range - a.range)\n    .slice(0, 20)\n\n  return {\n    raters: raters.size,\n    jointlyRated: jointlyRated.length,\n    kappa,\n    perPair,\n    disagreementCases,\n  }\n}\n\nfunction pearson(a: number[], b: number[]): number {\n  if (a.length !== b.length || a.length === 0) return 0\n  const n = a.length\n  const meanA = a.reduce((s, v) => s + v, 0) / n\n  const meanB = b.reduce((s, v) => s + v, 0) / n\n  let num = 0\n  let denomA = 0\n  let denomB = 0\n  for (let i = 0; i < n; i++) {\n    const da = a[i]! - meanA\n    const db = b[i]! - meanB\n    num += da * db\n    denomA += da * da\n    denomB += db * db\n  }\n  const denom = Math.sqrt(denomA * denomB)\n  return denom === 0 ? 0 : num / denom\n}\n\n// ── Lift ────────────────────────────────────────────────────────────\n\nfunction computeLift(\n  runs: RunRecord[],\n  baselineId: string | undefined,\n  candidateId: string | undefined,\n  split: 'search' | 'holdout',\n): LiftInsight | undefined {\n  let bId = baselineId\n  let cId = candidateId\n  if (!bId || !cId) {\n    // Auto-detect: when exactly two distinct candidateIds appear, treat the\n    // lower-mean side as baseline.\n    const ids = [...new Set(runs.map((r) => r.candidateId))]\n    if (ids.length !== 2) return undefined\n    const [idA, idB] = ids as [string, string]\n    const meanA = mean(runs.filter((r) => r.candidateId === idA).map((r) => compositeOf(r, split)))\n    const meanB = mean(runs.filter((r) => r.candidateId === idB).map((r) => compositeOf(r, split)))\n    bId = meanA <= meanB ? idA : idB\n    cId = meanA <= meanB ? idB : idA\n  }\n\n  const baseline = runs.filter((r) => r.candidateId === bId)\n  const candidate = runs.filter((r) => r.candidateId === cId)\n  if (baseline.length === 0 || candidate.length === 0) return undefined\n\n  // Pair on (experimentId, seed). When that key doesn't match, fall back\n  // to ordinal pairing — common for fresh runs from the same scenario list.\n  const baselineByKey = new Map(baseline.map((r) => [pairingKey(r), r]))\n  const pairedBaseline: number[] = []\n  const pairedCandidate: number[] = []\n  let usedKeyPairing = false\n  for (const cand of candidate) {\n    const b = baselineByKey.get(pairingKey(cand))\n    if (b) {\n      const bC = compositeOf(b, split)\n      const cC = compositeOf(cand, split)\n      if (Number.isFinite(bC) && Number.isFinite(cC)) {\n        pairedBaseline.push(bC)\n        pairedCandidate.push(cC)\n        usedKeyPairing = true\n      }\n    }\n  }\n  if (!usedKeyPairing) {\n    const n = Math.min(baseline.length, candidate.length)\n    for (let i = 0; i < n; i++) {\n      const bC = compositeOf(baseline[i]!, split)\n      const cC = compositeOf(candidate[i]!, split)\n      if (Number.isFinite(bC) && Number.isFinite(cC)) {\n        pairedBaseline.push(bC)\n        pairedCandidate.push(cC)\n      }\n    }\n  }\n  if (pairedBaseline.length === 0) return undefined\n\n  const baselineMean = mean(pairedBaseline)\n  const candidateMean = mean(pairedCandidate)\n  const delta = candidateMean - baselineMean\n\n  const bootstrap = pairedBootstrap(pairedBaseline, pairedCandidate, {\n    confidence: 0.95,\n    resamples: 2000,\n    statistic: 'mean',\n  })\n  const tTest = pairedTTest(pairedBaseline, pairedCandidate)\n  const d = cohensD(pairedBaseline, pairedCandidate)\n  const mde = pairedMde({ nPaired: pairedBaseline.length, power: 0.8, alpha: 0.05 })\n  const requiredN = requiredSampleSize({\n    effect: Math.max(Math.abs(delta), 1e-6),\n    power: 0.8,\n    alpha: 0.05,\n  })\n\n  return {\n    baselineMean,\n    candidateMean,\n    delta,\n    ci95: [bootstrap.low, bootstrap.high],\n    pValue: tTest.p,\n    n: pairedBaseline.length,\n    cohensD: d,\n    mde,\n    requiredN,\n  }\n}\n\nfunction pairingKey(r: RunRecord): string {\n  return `${r.experimentId}::${r.seed}`\n}\n\nfunction mean(arr: number[]): number {\n  return arr.length === 0 ? 0 : arr.reduce((s, v) => s + v, 0) / arr.length\n}\n\n// ── Failure clustering ──────────────────────────────────────────────\n\nasync function computeFailureClusters(\n  runs: RunRecord[],\n  analyst: AnalystRegistry,\n  split: 'search' | 'holdout',\n): Promise<FailureClusterInsight | undefined> {\n  const failed = runs.filter((r) => compositeOf(r, split) < 0.5 || r.failureMode !== undefined)\n  if (failed.length === 0) return { clusters: [], totalFailures: 0 }\n\n  const clusters = new Map<string, { exemplars: string[]; share: number }>()\n  for (const run of failed) {\n    try {\n      const result = await analyst.run(run.runId, {\n        kind: 'run-record',\n        run,\n      } as Parameters<typeof analyst.run>[1])\n      for (const finding of result.findings as AnalystFinding[]) {\n        const key = finding.area || finding.analyst_id || 'unclassified'\n        const c = clusters.get(key) ?? { exemplars: [], share: 0 }\n        if (c.exemplars.length < 5) c.exemplars.push(run.runId)\n        clusters.set(key, c)\n      }\n    } catch {\n      const c = clusters.get('analyst-error') ?? { exemplars: [], share: 0 }\n      if (c.exemplars.length < 5) c.exemplars.push(run.runId)\n      clusters.set('analyst-error', c)\n    }\n  }\n  const clusterList = [...clusters.entries()].map(([id, c]) => ({\n    id,\n    name: id,\n    share: c.exemplars.length / failed.length,\n    exemplars: c.exemplars,\n  }))\n  clusterList.sort((a, b) => b.share - a.share)\n  return { clusters: clusterList, totalFailures: failed.length }\n}\n\n// ── Contamination ──────────────────────────────────────────────────\n\nfunction computeContamination(\n  runs: RunRecord[],\n  canaries: DatasetScenario[],\n): InsightReport['contamination'] {\n  let leaks = 0\n  const details: Array<{ runId: string; canary: string; matched: string }> = []\n  for (const run of runs) {\n    const output = stringifyOutput(run)\n    if (!output) continue\n    const leaksHere = checkCanaries(output, canaries)\n    for (const leak of leaksHere) {\n      leaks++\n      details.push({ runId: run.runId, canary: leak.canary, matched: leak.evidence })\n    }\n  }\n  return { leaks, holdoutAuditPassed: leaks === 0, details }\n}\n\nfunction stringifyOutput(run: RunRecord): string | undefined {\n  // RunRecord doesn't fix where \"the agent's output\" lives — different\n  // consumers stash it differently. We probe the common shapes: the\n  // outcome.raw map (numeric only by design — unlikely to contain text),\n  // and any string-valued fields tucked under metadata via type casting.\n  // Consumers with bespoke shapes pass canaryScenarios only when they\n  // know their runs carry a stringifiable surface.\n  const metadata = (run as unknown as { metadata?: Record<string, unknown> }).metadata\n  if (typeof metadata?.output === 'string') return metadata.output\n  if (typeof metadata?.text === 'string') return metadata.text\n  return undefined\n}\n\n// ── Outcome correlation + linear reward model ──────────────────────\n\nfunction computeOutcomeCorrelation(\n  runs: RunRecord[],\n  outcome: { metric: string; valueByRunId: Record<string, number> },\n  split: 'search' | 'holdout',\n): OutcomeCorrelationInsight | undefined {\n  const xs: number[] = []\n  const ys: number[] = []\n  for (const run of runs) {\n    const y = outcome.valueByRunId[run.runId]\n    if (y === undefined || !Number.isFinite(y)) continue\n    const x = compositeOf(run, split)\n    if (!Number.isFinite(x)) continue\n    xs.push(x)\n    ys.push(y)\n  }\n  if (xs.length < 3) return undefined\n\n  const p = pearson(xs, ys)\n  const s = spearman(xs, ys)\n  const meanX = mean(xs)\n  const meanY = mean(ys)\n  let num = 0\n  let denom = 0\n  for (let i = 0; i < xs.length; i++) {\n    num += (xs[i]! - meanX) * (ys[i]! - meanY)\n    denom += (xs[i]! - meanX) ** 2\n  }\n  const slope = denom === 0 ? 0 : num / denom\n  const intercept = meanY - slope * meanX\n  const ssTot = ys.reduce((a, y) => a + (y - meanY) ** 2, 0)\n  const ssRes = ys.reduce((a, y, i) => a + (y - (intercept + slope * xs[i]!)) ** 2, 0)\n  const r2 = ssTot === 0 ? 0 : 1 - ssRes / ssTot\n\n  return {\n    metric: outcome.metric,\n    n: xs.length,\n    pearson: p,\n    spearman: s,\n    rewardModel: { intercept, slope, r2 },\n  }\n}\n\nfunction spearman(a: number[], b: number[]): number {\n  if (a.length !== b.length || a.length === 0) return 0\n  return pearson(rank(a), rank(b))\n}\n\nfunction rank(arr: number[]): number[] {\n  const indexed = arr.map((v, i) => ({ v, i }))\n  indexed.sort((x, y) => x.v - y.v)\n  const ranks = new Array(arr.length).fill(0)\n  let i = 0\n  while (i < indexed.length) {\n    let j = i\n    while (j + 1 < indexed.length && indexed[j + 1]!.v === indexed[i]!.v) j++\n    const avg = (i + j + 2) / 2\n    for (let k = i; k <= j; k++) ranks[indexed[k]!.i] = avg\n    i = j + 1\n  }\n  return ranks\n}\n\n// ── Release confidence scorecard ───────────────────────────────────\n\nfunction buildReleaseScorecard(\n  composite: ScalarDistribution,\n  lift: LiftInsight | undefined,\n  contamination: InsightReport['contamination'],\n): InsightReport['release'] {\n  // Synthesise a minimal scorecard from the rolled-up signal. The\n  // substrate's `evaluateReleaseConfidence` primitive consumes a richer\n  // input shape that callers can produce by wiring SLO definitions; the\n  // shape here is the contract `selfImprove`/`analyzeRuns` consumers\n  // receive automatically. They can call `evaluateReleaseConfidence`\n  // directly when they want SLO-based axis evaluation.\n  const axes: InsightReport['release']['axes'] = []\n  const liftPass =\n    lift === undefined || lift.ci95[0] > 0\n      ? ('pass' as const)\n      : lift.delta > 0\n        ? ('warn' as const)\n        : ('fail' as const)\n  axes.push({\n    name: 'quality-lift',\n    status: liftPass,\n    detail: lift\n      ? `delta=${lift.delta.toFixed(3)}, CI95=[${lift.ci95[0].toFixed(3)}, ${lift.ci95[1].toFixed(3)}], n=${lift.n}`\n      : 'no baseline/candidate pair available',\n  })\n  const contamPass =\n    contamination === undefined || contamination.leaks === 0 ? ('pass' as const) : ('fail' as const)\n  axes.push({\n    name: 'contamination',\n    status: contamPass,\n    detail: contamination ? `${contamination.leaks} canary leak(s)` : 'no canaries supplied',\n  })\n  axes.push({\n    name: 'composite-distribution',\n    status: composite.mean >= 0.5 ? 'pass' : composite.mean >= 0.3 ? 'warn' : 'fail',\n    detail: `mean=${composite.mean.toFixed(3)}, p50=${composite.p50.toFixed(3)}, p95=${composite.p95.toFixed(3)} over n=${composite.n}`,\n  })\n  const status = axes.some((a) => a.status === 'fail')\n    ? 'fail'\n    : axes.some((a) => a.status === 'warn')\n      ? 'warn'\n      : 'pass'\n  return {\n    status,\n    axes,\n    issues: [],\n  }\n}\n\n// ── Recommendations engine ─────────────────────────────────────────\n\ninterface RecommendationContext {\n  composite: ScalarDistribution\n  judges: Record<string, JudgeInsight>\n  interRater?: InterRaterInsight\n  lift?: LiftInsight\n  failureClusters?: FailureClusterInsight\n  contamination?: InsightReport['contamination']\n  outcomeCorrelation?: OutcomeCorrelationInsight\n  threshold: number\n}\n\nfunction buildRecommendations(ctx: RecommendationContext): Recommendation[] {\n  const out: Recommendation[] = []\n\n  // Composite-distribution branch. Fires when the overall quality signal is\n  // poor regardless of lift / contamination / clusters — the customer needs\n  // to know they have a problem AND which specific runs to inspect.\n  if (ctx.composite.n > 0) {\n    if (ctx.composite.mean < 0.3) {\n      const tail = ctx.composite.tailRuns ?? []\n      const names = tail\n        .slice(0, 5)\n        .map((t) => `${t.runId}=${t.score.toFixed(3)}`)\n        .join(', ')\n      out.push({\n        priority: 'critical',\n        kind: 'investigate',\n        title: `Composite mean ${ctx.composite.mean.toFixed(3)} is below the 0.3 floor — the agent is broken on this corpus`,\n        detail:\n          tail.length > 0\n            ? `Worst ${tail.length} run${tail.length === 1 ? '' : 's'} to inspect first: ${names}. Histogram p50=${ctx.composite.p50.toFixed(3)}, p95=${ctx.composite.p95.toFixed(3)}.`\n            : `Histogram p50=${ctx.composite.p50.toFixed(3)}, p95=${ctx.composite.p95.toFixed(3)}.`,\n        evidencePath: 'composite.tailRuns',\n      })\n    } else if (ctx.composite.mean < 0.5) {\n      const tail = ctx.composite.tailRuns ?? []\n      const names = tail\n        .slice(0, 3)\n        .map((t) => `${t.runId}=${t.score.toFixed(3)}`)\n        .join(', ')\n      out.push({\n        priority: 'high',\n        kind: 'investigate',\n        title: `Composite mean ${ctx.composite.mean.toFixed(3)} is below 0.5 — investigate the lower tail before claiming the agent is healthy`,\n        detail:\n          tail.length > 0\n            ? `Worst ${tail.length} run${tail.length === 1 ? '' : 's'}: ${names}. Histogram p50=${ctx.composite.p50.toFixed(3)}, p95=${ctx.composite.p95.toFixed(3)}.`\n            : `Histogram p50=${ctx.composite.p50.toFixed(3)}, p95=${ctx.composite.p95.toFixed(3)}.`,\n        evidencePath: 'composite.tailRuns',\n      })\n    }\n  }\n\n  // Missing-judges branch. The report can't surface per-dimension or\n  // calibration signal when `outcome.judgeScores` is empty across the\n  // corpus. Tell the customer how to enrich.\n  if (Object.keys(ctx.judges).length === 0 && ctx.composite.n > 0) {\n    out.push({\n      priority: 'medium',\n      kind: 'expand-corpus',\n      title: 'No judge scores recorded — per-dimension + calibration insights unavailable',\n      detail:\n        'Records have no `outcome.judgeScores`. To unlock perDimension, judges, and calibration, attach a Judge run during your eval pass and populate `outcome.judgeScores.perJudge[judgeName][dimension] = score`. See `docs/insight-report.md` for the expected shape.',\n      evidencePath: 'judges',\n    })\n  }\n\n  if (ctx.lift) {\n    const decisive = ctx.lift.ci95[0] > ctx.threshold\n    const inconclusive = ctx.lift.ci95[0] <= ctx.threshold && ctx.lift.ci95[1] > ctx.threshold\n    if (decisive) {\n      out.push({\n        priority: 'critical',\n        kind: 'ship',\n        title: `Ship — lift ${ctx.lift.delta.toFixed(3)} (95% CI ${ctx.lift.ci95[0].toFixed(3)}..${ctx.lift.ci95[1].toFixed(3)})`,\n        detail: `Holdout lift exceeds threshold ${ctx.threshold} with 95% bootstrap confidence (n=${ctx.lift.n}, p=${ctx.lift.pValue.toFixed(4)}, d=${ctx.lift.cohensD.toFixed(2)}).`,\n        evidencePath: 'lift',\n      })\n    } else if (inconclusive) {\n      out.push({\n        priority: 'high',\n        kind: 'expand-corpus',\n        title: `Inconclusive — need ~${ctx.lift.requiredN} paired runs (have ${ctx.lift.n}) at current effect size`,\n        detail: `CI straddles threshold. Current MDE at 80% power is ${ctx.lift.mde.toFixed(3)}; observed delta is ${ctx.lift.delta.toFixed(3)}.`,\n        evidencePath: 'lift',\n      })\n    } else {\n      out.push({\n        priority: 'critical',\n        kind: 'hold',\n        title: `Hold — lift CI lower bound ${ctx.lift.ci95[0].toFixed(3)} is at or below threshold ${ctx.threshold}`,\n        detail: `Bootstrap CI provides no statistical evidence the candidate is better. Consider tightening the mutation or expanding the holdout.`,\n        evidencePath: 'lift',\n      })\n    }\n  }\n\n  if (ctx.contamination && ctx.contamination.leaks > 0) {\n    out.push({\n      priority: 'critical',\n      kind: 'fix',\n      title: `${ctx.contamination.leaks} canary leak${ctx.contamination.leaks === 1 ? '' : 's'} detected`,\n      detail: `Holdout integrity is compromised. The lift number is unreliable until you investigate.`,\n      evidencePath: 'contamination',\n    })\n  }\n\n  if (ctx.interRater && ctx.interRater.kappa < 0.5) {\n    out.push({\n      priority: 'high',\n      kind: 'recalibrate',\n      title: `Inter-rater agreement κ=${ctx.interRater.kappa.toFixed(2)} is below 0.5`,\n      detail: `Raters disagree on what 'good' looks like. Top disagreement cases listed in interRater.disagreementCases — consider a triage meeting or refining the rubric.`,\n      evidencePath: 'interRater',\n    })\n  }\n\n  if (ctx.failureClusters && ctx.failureClusters.clusters.length > 0) {\n    const top = ctx.failureClusters.clusters[0]!\n    out.push({\n      priority: 'high',\n      kind: 'investigate',\n      title: `Top failure cluster: ${top.name} (${(top.share * 100).toFixed(0)}% of failures)`,\n      detail: `${ctx.failureClusters.totalFailures} runs failed. The largest cluster groups ${top.exemplars.length} exemplars under '${top.name}'.`,\n      evidencePath: 'failureClusters.clusters[0]',\n    })\n  }\n\n  if (ctx.outcomeCorrelation && Math.abs(ctx.outcomeCorrelation.spearman) < 0.3) {\n    out.push({\n      priority: 'medium',\n      kind: 'recalibrate',\n      title: `Judge scores decoupled from ${ctx.outcomeCorrelation.metric} (Spearman ρ=${ctx.outcomeCorrelation.spearman.toFixed(2)})`,\n      detail: `Your judges score what they were trained to score, but it isn't predicting downstream ${ctx.outcomeCorrelation.metric}. Consider retraining the judge against ${ctx.outcomeCorrelation.metric} as the gold signal.`,\n      evidencePath: 'outcomeCorrelation',\n    })\n  }\n\n  return out\n}\n\n// ── Re-export pareto figure spec for hosted-side rendering ─────────\n\nexport type { ParetoFigureSpec }\n","/**\n * # `selfImprove()` — the LAND-tier one-shot.\n *\n * The cheapest possible call site to run a real closed-loop self-\n * improvement over your agent. Wraps `runImprovementLoop` with smart\n * defaults and a budget-shaped options API; every escape hatch the\n * substrate exposes is reachable from here without losing the\n * one-function feel.\n *\n * Defaults picked to match the LAND-tier story:\n *   - In-memory storage (no filesystem touch).\n *   - `gepaDriver` reflective mutation with copywriting-flavored primitives\n *     (override `driver` or `mutationPrimitives` for any domain).\n *   - `defaultProductionGate` with `deltaThreshold: 0.05`.\n *   - Held-out split = 25% of scenarios, deterministic by id hash.\n *   - 3 generations × population 2 (raise via `budget` for more search).\n *   - `autoOnPromote: 'none'` (we don't open PRs unless you ask).\n *\n * Want one-click? Provide `agent` + `scenarios` + `judge`. Done.\n * Want distributed? Pass `cellPlacement` + an `httpDispatch`-backed\n * agent. Want a code-tier surface? Pass a `MutableSurface` + your own\n * `driver`. Same function.\n */\n\nimport { gepaDriver } from '../campaign/drivers/gepa'\nimport { defaultProductionGate } from '../campaign/gates/default-production-gate'\nimport {\n  type RunImprovementLoopResult,\n  runImprovementLoop,\n} from '../campaign/presets/run-improvement-loop'\nimport { type CampaignStorage, inMemoryCampaignStorage } from '../campaign/storage'\nimport type {\n  CampaignCellResult,\n  DispatchContext,\n  Gate,\n  ImprovementDriver,\n  JudgeConfig,\n  MutableSurface,\n  Scenario,\n} from '../campaign/types'\nimport { createHostedClient, type HostedTenant } from '../hosted/client'\nimport type { EvalRunCellScore, EvalRunEvent, EvalRunGenerationSnapshot } from '../hosted/types'\nimport type { JudgeScoresRecord, RunRecord } from '../run-record'\nimport { analyzeRuns } from './analyze-runs'\nimport type { InsightReport } from './insight-report'\n\nexport interface SelfImproveBudget {\n  /** Hard $ ceiling across all cells in baseline + every generation. Cells\n   *  beyond the ceiling are skipped (cost-aware, not aborted). */\n  dollars?: number\n  /** How many improvement generations to explore. Default 3. Set 0 to\n   *  skip improvement entirely (selfImprove becomes a baseline-only run). */\n  generations?: number\n  /** Candidates the driver proposes per generation. Default 2. */\n  populationSize?: number\n  /** Max concurrent cells across the loop. Default 2. */\n  maxConcurrency?: number\n  /** Fraction of `scenarios` held out from training, used for the gate.\n   *  Default 0.25. Ignored when `holdoutScenarios` is set explicitly. */\n  holdoutFraction?: number\n  /** Explicit held-out scenarios; overrides `holdoutFraction`. */\n  holdoutScenarios?: Scenario[]\n}\n\nexport interface SelfImproveLlm {\n  /** Endpoint base URL. Default Tangle Router. */\n  baseUrl?: string\n  /** Bearer token. Default `process.env.OPENAI_API_KEY`. */\n  apiKey?: string\n  /** Model id used by `gepaDriver` reflection. Default\n   *  `anthropic/claude-sonnet-4.6`. */\n  model?: string\n}\n\nexport type SelfImproveProgressEvent =\n  | { kind: 'baseline.started'; scenarios: number }\n  | { kind: 'baseline.completed'; compositeMean: number; durationMs: number }\n  | { kind: 'generation.started'; index: number; populationSize: number }\n  | { kind: 'generation.completed'; index: number; bestComposite: number; durationMs: number }\n  | { kind: 'gate.decided'; decision: string; lift: number }\n\nexport interface SelfImproveOptions<TScenario extends Scenario, TArtifact> {\n  /**\n   * Your agent — a function that takes the current `MutableSurface`\n   * (typically a system prompt the loop is optimizing) plus the\n   * scenario + cell ctx, and returns the artifact your judge scores.\n   *\n   * Same shape as `RunOptimizationOptions.dispatchWithSurface`. Wrap a\n   * plain `Dispatch` if you don't have a surface seam:\n   *\n   *   agent: (_surface, scenario, ctx) => yourPlainDispatch(scenario, ctx)\n   *\n   * That mode evaluates without mutating any surface — useful as a\n   * baseline-only run (set `budget.generations = 0`).\n   */\n  agent: (surface: MutableSurface, scenario: TScenario, ctx: DispatchContext) => Promise<TArtifact>\n\n  /** Scenarios to evaluate against. Train/holdout split is computed from\n   *  these unless `budget.holdoutScenarios` is set explicitly. */\n  scenarios: TScenario[]\n\n  /** Judge that scores artifacts. Bring your own; use `langchainJudge`\n   *  from `/adapters/langchain` for a Runnable-shaped one. */\n  judge: JudgeConfig<TArtifact, TScenario>\n\n  /** Starting surface — system prompt, JSON config, anything `MutableSurface`\n   *  accepts. The driver mutates this each generation. */\n  baselineSurface: MutableSurface\n\n  /** Budget + loop shape. All fields optional; defaults pick the LAND-tier\n   *  story. */\n  budget?: SelfImproveBudget\n\n  /** Custom driver. Default is `gepaDriver` configured from `llm` +\n   *  `mutationPrimitives`. */\n  driver?: ImprovementDriver\n\n  /** Default-driver overrides — used when `driver` is unset. */\n  mutationPrimitives?: string[]\n  driverTarget?: string\n\n  /** Custom gate. Default is `defaultProductionGate` with\n   *  `deltaThreshold: 0.05` on the held-out split. */\n  gate?: Gate<TArtifact, TScenario>\n\n  /** LLM config consumed by the default `gepaDriver`. Ignored if you pass\n   *  your own `driver`. */\n  llm?: SelfImproveLlm\n\n  /** Storage backend. Default `inMemoryCampaignStorage()` — nothing\n   *  persists past the call. Pass `fsCampaignStorage()` to write to disk. */\n  storage?: CampaignStorage\n\n  /** Run directory (logical for in-memory storage, real path for fs).\n   *  Default `mem://selfImprove-<timestamp>`. */\n  runDir?: string\n\n  /** Distributed-driver seam — same as `RunCampaignOptions.cellPlacement`.\n   *  Returns an opaque placement key the substrate forwards to your agent\n   *  as `ctx.placement`. Combined with `httpDispatch` from\n   *  `/adapters/http`, fans cells across regions. */\n  cellPlacement?: (input: {\n    scenario: TScenario\n    rep: number\n    generation?: number\n  }) => string | undefined\n\n  /** Streaming hook — fires on baseline + each generation + gate decision.\n   *  Consumer routes events wherever (UI, dashboard, logs). */\n  onProgress?: (event: SelfImproveProgressEvent) => void\n\n  /** Auto-promotion behavior on a ship decision. Default `'none'` — we\n   *  return the winner; you ship it however you ship. `'pr'` opens a\n   *  GitHub PR via `openAutoPr`; requires `ghOwner` + `ghRepo`. */\n  autoOnPromote?: 'pr' | 'none'\n  ghOwner?: string\n  ghRepo?: string\n\n  /**\n   * Opt-in: ship eval-run events to a hosted orchestrator (ours, your\n   * self-hosted one, or any compatible implementation of the\n   * `docs/hosted-ingest-spec.md` wire format). When set, the substrate\n   * POSTs the final `EvalRunEvent` to `${endpoint}/v1/ingest/eval-runs`\n   * after the loop completes. Failures are logged but do not fail the\n   * loop — local result is always returned.\n   *\n   * For our orchestrator: `{ endpoint: 'https://orchestrator.tangle.tools/v1', apiKey, tenantId }`.\n   *\n   * For your self-hosted: any URL serving the wire format. See\n   * `examples/hosted-ingest-server/` for the reference receiver.\n   */\n  hostedTenant?: HostedTenant\n\n  /** Free-form labels attached to the hosted event (env, branch, model id,\n   *  etc.). Ignored when `hostedTenant` is unset. */\n  hostedLabels?: Record<string, string>\n}\n\nexport interface SelfImproveResult<TScenario extends Scenario, TArtifact> {\n  /** Composite mean across all scenarios, baseline run. */\n  baseline: {\n    compositeMean: number\n    perScenario: Record<string, number>\n  }\n  /** Composite mean on the held-out set, winner run. */\n  winner: {\n    compositeMean: number\n    perScenario: Record<string, number>\n    surface: MutableSurface\n  }\n  /** `winner.compositeMean - baselineOnHoldout.compositeMean`. Positive\n   *  means the gate observed improvement. */\n  lift: number\n  /** `defaultProductionGate.decide()` result. */\n  gateDecision: 'ship' | 'hold' | 'need_more_work' | 'model_ceiling' | 'arch_ceiling'\n  /** Number of generations actually explored (may be less than the\n   *  budget if the driver gave up early). */\n  generationsExplored: number\n  /** Wall-clock total. */\n  durationMs: number\n  /** Total cost across baseline + every generation. */\n  totalCostUsd: number\n  /**\n   * Rigor packet: distributional summary, paired-bootstrap lift CI,\n   * judge stats, contamination check, recommendations. Wired through\n   * `analyzeRuns()` on the baseline + winner cells of the campaign.\n   * Hosted-tier dashboards render this as the v3-vs-v4 decision view.\n   */\n  insight: InsightReport\n  /**\n   * Raw substrate result for advanced inspection — full per-generation\n   * candidates, full campaign artifacts, all judge scores. Useful for\n   * debugging or reporting beyond the summary.\n   */\n  raw: RunImprovementLoopResult<TArtifact, TScenario>\n}\n\n/**\n * Deterministic train/holdout split by a stable hash of `scenario.id`,\n * so the same scenario set always splits the same way across runs.\n */\nfunction splitTrainHoldout<TScenario extends Scenario>(\n  scenarios: TScenario[],\n  fraction: number,\n): { train: TScenario[]; holdout: TScenario[] } {\n  // Stable fnv-1a-ish hash of the id for ordering.\n  function hash(s: string): number {\n    let h = 2166136261 >>> 0\n    for (let i = 0; i < s.length; i++) {\n      h ^= s.charCodeAt(i)\n      h = Math.imul(h, 16777619) >>> 0\n    }\n    return h\n  }\n  const sorted = [...scenarios].sort((a, b) => hash(a.id) - hash(b.id))\n  const nHoldout = Math.max(1, Math.min(sorted.length - 1, Math.round(sorted.length * fraction)))\n  return {\n    holdout: sorted.slice(0, nHoldout),\n    train: sorted.slice(nHoldout),\n  }\n}\n\nfunction meanComposite(byScenario: Record<string, { meanComposite: number }>): {\n  compositeMean: number\n  perScenario: Record<string, number>\n} {\n  const perScenario: Record<string, number> = {}\n  const values: number[] = []\n  for (const [id, agg] of Object.entries(byScenario)) {\n    perScenario[id] = agg.meanComposite\n    values.push(agg.meanComposite)\n  }\n  return {\n    compositeMean: values.length === 0 ? 0 : values.reduce((s, v) => s + v, 0) / values.length,\n    perScenario,\n  }\n}\n\nconst DEFAULT_MUTATION_PRIMITIVES = [\n  'Tighten the hook: lead with the specific user outcome.',\n  'Replace generic adjectives with specific verbs or proof numbers.',\n  \"Anchor every claim in something the scenario's brief literally supports.\",\n  'Honor the surface-shape constraint (length, register, audience vocabulary).',\n]\n\n/**\n * One-shot self-improvement loop. See module docstring for defaults +\n * extension points.\n *\n * @example Minimum (LAND tier):\n *\n *   const result = await selfImprove({\n *     agent: (surface, scenario, ctx) => myAgent(surface, scenario, ctx.signal),\n *     scenarios,\n *     judge,\n *     baselineSurface: DEFAULT_PROMPT,\n *   })\n *   console.log(`lift: ${result.lift.toFixed(3)} (${result.gateDecision})`)\n *\n * @example Distributed (workers in three regions):\n *\n *   await selfImprove({\n *     agent: httpDispatch({ resolveUrl: ({ placement }) => REGION_URLS[placement!] }),\n *     scenarios,\n *     judge,\n *     baselineSurface: DEFAULT_PROMPT,\n *     cellPlacement: ({ scenario }) => scenario.region,\n *     budget: { maxConcurrency: 12 },\n *   })\n */\nexport async function selfImprove<TScenario extends Scenario, TArtifact>(\n  opts: SelfImproveOptions<TScenario, TArtifact>,\n): Promise<SelfImproveResult<TScenario, TArtifact>> {\n  const startedAt = Date.now()\n\n  const budget = opts.budget ?? {}\n  const generations = budget.generations ?? 3\n  const populationSize = budget.populationSize ?? 2\n  const maxConcurrency = budget.maxConcurrency ?? 2\n  const holdoutFraction = budget.holdoutFraction ?? 0.25\n  const costCeiling = budget.dollars\n\n  const explicitHoldout = budget.holdoutScenarios\n  const { train, holdout } = explicitHoldout\n    ? {\n        train: opts.scenarios.filter((s) => !explicitHoldout.some((h) => h.id === s.id)),\n        holdout: explicitHoldout as TScenario[],\n      }\n    : splitTrainHoldout(opts.scenarios, holdoutFraction)\n\n  if (train.length === 0) {\n    throw new Error(\n      'selfImprove: train split is empty. Reduce holdoutFraction or pass more scenarios.',\n    )\n  }\n  if (holdout.length === 0) {\n    throw new Error('selfImprove: holdout split is empty. Pass more scenarios.')\n  }\n\n  const driver: ImprovementDriver =\n    opts.driver ??\n    gepaDriver({\n      llm: {\n        baseUrl: opts.llm?.baseUrl ?? 'https://router.tangle.tools/v1',\n        apiKey: opts.llm?.apiKey ?? process.env.OPENAI_API_KEY ?? '',\n      },\n      model: opts.llm?.model ?? 'anthropic/claude-sonnet-4.6',\n      target:\n        opts.driverTarget ??\n        'agent surface (system prompt or config) being optimized by selfImprove',\n      mutationPrimitives: opts.mutationPrimitives ?? DEFAULT_MUTATION_PRIMITIVES,\n    })\n\n  const gate: Gate<TArtifact, TScenario> =\n    opts.gate ??\n    defaultProductionGate<TArtifact, TScenario>({\n      holdoutScenarios: holdout,\n      deltaThreshold: 0.05,\n    })\n\n  const storage = opts.storage ?? inMemoryCampaignStorage()\n  const runDir = opts.runDir ?? `mem://selfImprove-${startedAt}`\n\n  if (opts.onProgress) {\n    opts.onProgress({ kind: 'baseline.started', scenarios: opts.scenarios.length })\n  }\n\n  const result = await runImprovementLoop<TScenario, TArtifact>({\n    scenarios: train,\n    baselineSurface: opts.baselineSurface,\n    dispatchWithSurface: opts.agent,\n    driver,\n    judges: [opts.judge],\n    populationSize,\n    maxGenerations: generations,\n    holdoutScenarios: holdout,\n    gate,\n    autoOnPromote: opts.autoOnPromote ?? 'none',\n    ghOwner: opts.ghOwner,\n    ghRepo: opts.ghRepo,\n    storage,\n    runDir,\n    maxConcurrency,\n    cellPlacement: opts.cellPlacement,\n    costCeiling,\n  })\n\n  const baseline = meanComposite(result.baselineOnHoldout.aggregates.byScenario)\n  const winnerStats = meanComposite(result.winnerOnHoldout.aggregates.byScenario)\n\n  if (opts.onProgress) {\n    opts.onProgress({\n      kind: 'baseline.completed',\n      compositeMean: baseline.compositeMean,\n      durationMs: Date.now() - startedAt,\n    })\n    opts.onProgress({\n      kind: 'gate.decided',\n      decision: result.gateResult.decision,\n      lift: winnerStats.compositeMean - baseline.compositeMean,\n    })\n  }\n\n  const totalCost =\n    result.baselineCampaign.aggregates.totalCostUsd +\n    result.generations.reduce(\n      (sum, gen) =>\n        sum + gen.surfaces.reduce((s, sf) => s + sf.campaign.aggregates.totalCostUsd, 0),\n      0,\n    )\n\n  // Rigor packet: feed baseline + winner cells through analyzeRuns().\n  // The two candidates (`baseline` / `winner`) give the lift section a\n  // clean paired comparison; per-judge / per-dimension / cost-quality\n  // sections populate from the cells' judgeScores.\n  const insight = await analyzeRuns({\n    runs: [\n      ...cellsToRunRecords(result.baselineCampaign.cells, 'baseline', runDir),\n      ...cellsToRunRecords(result.winnerOnHoldout.cells, 'winner', runDir),\n    ],\n    baselineCandidateId: 'baseline',\n    candidateCandidateId: 'winner',\n  })\n\n  const summary: SelfImproveResult<TScenario, TArtifact> = {\n    baseline,\n    winner: {\n      ...winnerStats,\n      surface: result.winnerSurface,\n    },\n    lift: winnerStats.compositeMean - baseline.compositeMean,\n    gateDecision: result.gateResult.decision,\n    generationsExplored: result.generations.length,\n    durationMs: Date.now() - startedAt,\n    totalCostUsd: totalCost,\n    insight,\n    raw: result,\n  }\n\n  // Opt-in hosted ingest. Failures logged but never fail the loop — the\n  // local result is always returned. This matches the wedge-doc invariant\n  // that LAND-tier never blocks on EXPAND-tier infra.\n  if (opts.hostedTenant) {\n    try {\n      await shipEvalRunToHosted(opts.hostedTenant, opts, summary, result, runDir)\n    } catch (err) {\n      const msg = err instanceof Error ? err.message : String(err)\n      // eslint-disable-next-line no-console -- intentional: hosted-ingest is best-effort\n      console.warn(`[agent-eval] hosted ingest failed (continuing): ${msg}`)\n    }\n  }\n\n  return summary\n}\n\nasync function shipEvalRunToHosted<TScenario extends Scenario, TArtifact>(\n  tenant: HostedTenant,\n  opts: SelfImproveOptions<TScenario, TArtifact>,\n  summary: SelfImproveResult<TScenario, TArtifact>,\n  raw: RunImprovementLoopResult<TArtifact, TScenario>,\n  runDir: string,\n): Promise<void> {\n  const client = createHostedClient(tenant)\n\n  function snapshotFromCampaign(\n    index: number,\n    surface: MutableSurface | undefined,\n    campaign: RunImprovementLoopResult<TArtifact, TScenario>['baselineCampaign'],\n    durationMs: number,\n  ): EvalRunGenerationSnapshot {\n    const cells: EvalRunCellScore[] = campaign.cells.map((cell) => {\n      const judgeScores = Object.values(cell.judgeScores)\n      const composite =\n        judgeScores.length === 0\n          ? 0\n          : judgeScores.reduce((s, j) => s + j.composite, 0) / judgeScores.length\n      return {\n        scenarioId: cell.scenarioId,\n        rep: cell.rep,\n        compositeMean: composite,\n        dimensions: Object.fromEntries(\n          Object.entries(cell.judgeScores).map(([name, score]) => [name, score.dimensions]),\n        ),\n        errorMessage: cell.error ?? undefined,\n      }\n    })\n    const compositeMean =\n      cells.length === 0 ? 0 : cells.reduce((s, c) => s + c.compositeMean, 0) / cells.length\n    return {\n      index,\n      surfaceHash:\n        typeof surface === 'string'\n          ? hashString(surface)\n          : hashString(JSON.stringify(surface ?? '')),\n      surface,\n      cells,\n      compositeMean,\n      costUsd: campaign.aggregates.totalCostUsd,\n      durationMs,\n    }\n  }\n\n  const generations: EvalRunGenerationSnapshot[] = []\n  // Baseline as generation 0.\n  generations.push(snapshotFromCampaign(0, opts.baselineSurface, raw.baselineCampaign, 0))\n  // Improvement generations as 1..N. Substrate stores per-surface campaigns\n  // per generation — we summarize the WINNING surface per generation here.\n  for (const gen of raw.generations) {\n    const winner = gen.surfaces.reduce(\n      (best, s) =>\n        s.campaign.aggregates.cellsExecuted > 0 &&\n        (best === undefined || averageComposite(s.campaign) > averageComposite(best.campaign))\n          ? s\n          : best,\n      gen.surfaces[0],\n    )\n    if (!winner) continue\n    generations.push(\n      snapshotFromCampaign(gen.record.generationIndex + 1, winner.surface, winner.campaign, 0),\n    )\n  }\n\n  const event: EvalRunEvent = {\n    runId: `${runDir}#${Date.now()}`,\n    runDir,\n    timestamp: new Date().toISOString(),\n    status: 'finished',\n    labels: opts.hostedLabels ?? {},\n    baseline: generations[0],\n    generations,\n    gateDecision: summary.gateDecision,\n    holdoutLift: summary.lift,\n    totalCostUsd: summary.totalCostUsd,\n    totalDurationMs: summary.durationMs,\n    insightReport: summary.insight,\n  }\n\n  await client.ingestEvalRun(event)\n}\n\nfunction averageComposite(\n  campaign: RunImprovementLoopResult<unknown, Scenario>['baselineCampaign'],\n): number {\n  const aggs = Object.values(campaign.aggregates.byScenario)\n  return aggs.length === 0 ? 0 : aggs.reduce((s, a) => s + a.meanComposite, 0) / aggs.length\n}\n\nfunction hashString(s: string): string {\n  let h = 2166136261 >>> 0\n  for (let i = 0; i < s.length; i++) {\n    h ^= s.charCodeAt(i)\n    h = Math.imul(h, 16777619) >>> 0\n  }\n  return h.toString(16).padStart(8, '0')\n}\n\n/**\n * Adapt campaign cells into the `RunRecord` shape `analyzeRuns()` consumes.\n * Each cell becomes one run; `candidateId` is the caller-supplied label so\n * baseline + winner pair cleanly on `(experimentId, seed)`.\n */\nfunction cellsToRunRecords<TArtifact>(\n  cells: ReadonlyArray<CampaignCellResult<TArtifact>>,\n  candidateId: 'baseline' | 'winner',\n  runId: string,\n): RunRecord[] {\n  return cells.map((cell) => {\n    const perJudge: Record<string, Record<string, number>> = {}\n    const perDimMeanAccum: Record<string, { sum: number; n: number }> = {}\n    let compositeSum = 0\n    let compositeCount = 0\n    for (const [judgeId, score] of Object.entries(cell.judgeScores)) {\n      perJudge[judgeId] = { ...score.dimensions }\n      for (const [dim, value] of Object.entries(score.dimensions)) {\n        if (!Number.isFinite(value)) continue\n        const accum = perDimMeanAccum[dim] ?? { sum: 0, n: 0 }\n        accum.sum += value\n        accum.n += 1\n        perDimMeanAccum[dim] = accum\n      }\n      if (Number.isFinite(score.composite)) {\n        compositeSum += score.composite\n        compositeCount += 1\n      }\n    }\n    const perDimMean: Record<string, number> = {}\n    for (const [dim, { sum, n }] of Object.entries(perDimMeanAccum)) {\n      perDimMean[dim] = n === 0 ? 0 : sum / n\n    }\n    const composite = compositeCount === 0 ? 0 : compositeSum / compositeCount\n    const judgeScores: JudgeScoresRecord = {\n      perJudge,\n      perDimMean,\n      composite,\n    }\n    return {\n      runId: `${runId}::${candidateId}::${cell.cellId}`,\n      experimentId: runId,\n      candidateId,\n      // Pair on (scenarioId, rep) — analyzeRuns pairs on (experimentId, seed).\n      // Synthesize a stable seed for that pairing.\n      seed:\n        cell.rep * 1_000_000 +\n        hashString(cell.scenarioId)\n          .slice(0, 6)\n          .split('')\n          .reduce((a, c) => (a * 31 + c.charCodeAt(0)) >>> 0, 0),\n      model: 'campaign-cell',\n      promptHash: 'sha256:cell',\n      configHash: 'sha256:cell',\n      commitSha: 'cell',\n      wallMs: cell.durationMs,\n      costUsd: cell.costUsd,\n      tokenUsage: { input: 0, output: 0 },\n      outcome: {\n        holdoutScore: composite,\n        raw: {},\n        judgeScores,\n      },\n      splitTag: 'holdout',\n      ...(cell.error ? { failureMode: cell.error } : {}),\n    } satisfies RunRecord\n  })\n}\n","/**\n * # `intake/feedback-table` — multi-rater approve/reject corpus → `RunRecord[]`.\n *\n * The generic shape behind Obsidian's `#approved` / `#rejected` tags, a\n * Google Sheet, a Postgres `feedback` table, or any CSV with ratings.\n *\n * Caller supplies one row per (run, rater) tuple plus per-run metadata; the\n * adapter rolls them up into the substrate-canonical `RunRecord` shape so\n * `analyzeRuns({ runs, raterScores })` can produce inter-rater agreement,\n * disagreement triage, and downstream recommendations.\n *\n * Per-run `RunRecord.outcome.searchScore` is the rater-mean rating\n * (normalised to 0..1 when scale is supplied); `outcome.raw` carries the\n * per-rater scores keyed by rater id for downstream attribution.\n */\n\nimport type { JudgeScoresRecord, RunOutcome, RunRecord, RunSplitTag } from '../../run-record'\n\nexport interface FeedbackTableRow {\n  /** Stable id for this run — the unit a rater scored. Drives pairing\n   *  across analysis primitives. */\n  runId: string\n  /** Identifier of the rater that produced this rating. */\n  rater: string\n  /** The rating itself. Accepts boolean (approve/reject), 0..1 scalar,\n   *  or any numeric scale — see `scale`. */\n  rating: number | boolean\n  /** Optional metadata carried through to `RunRecord.outcome.raw` and the\n   *  custom-shape metadata bag. */\n  metadata?: Record<string, unknown>\n}\n\nexport interface FeedbackTableMeta {\n  runId: string\n  /** When omitted, defaults to `'feedback-corpus'`. Used to group related\n   *  runs in `analyzeRuns()` lift analysis. */\n  experimentId?: string\n  /** When omitted, defaults to `runId` — each run is its own candidate. */\n  candidateId?: string\n  /** Cost in USD, when available. Set to 0 when unknown — the consumer's\n   *  cost analysis sections will collapse gracefully. */\n  costUsd?: number\n  /** Wall-clock ms, when available. Defaults to 0. */\n  wallMs?: number\n  /** Model identifier including snapshot. Default `unknown@unknown`. */\n  model?: string\n  /** Optional sha256 of the prompt; default `'sha256:unknown'`. */\n  promptHash?: string\n  /** Default `'sha256:unknown'`. */\n  configHash?: string\n  /** Default `'unknown'`. */\n  commitSha?: string\n  /** Default `'holdout'` — feedback corpora are by nature the holdout\n   *  signal a closed-loop improvement aims at. */\n  splitTag?: RunSplitTag\n  /** Free-form metadata available to consumers via the cast-out path on\n   *  the resulting RunRecord. */\n  extras?: Record<string, unknown>\n}\n\nexport interface FromFeedbackTableOptions {\n  /** Per-(run, rater) ratings. */\n  ratings: FeedbackTableRow[]\n  /** Per-run metadata. When a runId appears in `ratings` but not here, the\n   *  adapter synthesises minimal metadata with defaults documented above. */\n  meta?: FeedbackTableMeta[]\n  /** Rating scale. Provide `{ min, max }` for non-0..1 numeric scales.\n   *  Booleans are normalised: true → 1, false → 0. Default: assumes\n   *  ratings are already 0..1. */\n  scale?: { min: number; max: number }\n  /** When true, the rater scores are emitted into `raterScores` (a sibling\n   *  array `analyzeRuns()` accepts) instead of being averaged into the\n   *  run's `outcome.searchScore`. Default `true` — preserves rater-level\n   *  signal for inter-rater analysis. */\n  emitRaterScores?: boolean\n}\n\nexport interface FromFeedbackTableResult {\n  runs: RunRecord[]\n  /** Rater-level scores ready to pass into `analyzeRuns({ raterScores })`\n   *  for inter-rater agreement + disagreement triage. */\n  raterScores: Array<{ runId: string; rater: string; score: number }>\n}\n\nexport function fromFeedbackTable(opts: FromFeedbackTableOptions): FromFeedbackTableResult {\n  const { ratings, meta = [], scale, emitRaterScores = true } = opts\n  const metaByRun = new Map(meta.map((m) => [m.runId, m]))\n\n  // Normalise per-rating to a 0..1 score.\n  const normalise = (rating: number | boolean): number => {\n    if (typeof rating === 'boolean') return rating ? 1 : 0\n    if (!Number.isFinite(rating)) return Number.NaN\n    if (!scale) return rating\n    const { min, max } = scale\n    if (max === min) return rating\n    return (rating - min) / (max - min)\n  }\n\n  // Group ratings by runId.\n  const byRun = new Map<string, FeedbackTableRow[]>()\n  for (const row of ratings) {\n    const list = byRun.get(row.runId) ?? []\n    list.push(row)\n    byRun.set(row.runId, list)\n  }\n\n  const runs: RunRecord[] = []\n  const raterScores: FromFeedbackTableResult['raterScores'] = []\n\n  for (const [runId, rowsForRun] of byRun) {\n    const normalised = rowsForRun\n      .map((r) => ({ rater: r.rater, score: normalise(r.rating) }))\n      .filter((r) => Number.isFinite(r.score))\n    if (normalised.length === 0) continue\n\n    const meanScore = normalised.reduce((s, r) => s + r.score, 0) / normalised.length\n\n    const runMeta = metaByRun.get(runId) ?? ({ runId } as FeedbackTableMeta)\n\n    const judgeScores: JudgeScoresRecord = {\n      perJudge: Object.fromEntries(normalised.map((r) => [r.rater, { rating: r.score }])),\n      perDimMean: { rating: meanScore },\n      composite: meanScore,\n    }\n\n    const outcome: RunOutcome = {\n      // Feedback corpora ARE the holdout signal — score lands on\n      // `holdoutScore` so downstream substrate primitives (`paretoChart`,\n      // promotion gates) read it correctly by default.\n      holdoutScore: meanScore,\n      raw: Object.fromEntries(normalised.map((r) => [`rater:${r.rater}`, r.score])),\n      judgeScores,\n    }\n\n    runs.push({\n      runId,\n      experimentId: runMeta.experimentId ?? 'feedback-corpus',\n      candidateId: runMeta.candidateId ?? runId,\n      seed: 0,\n      model: runMeta.model ?? 'unknown@unknown',\n      promptHash: runMeta.promptHash ?? 'sha256:unknown',\n      configHash: runMeta.configHash ?? 'sha256:unknown',\n      commitSha: runMeta.commitSha ?? 'unknown',\n      wallMs: runMeta.wallMs ?? 0,\n      costUsd: runMeta.costUsd ?? 0,\n      tokenUsage: { input: 0, output: 0 },\n      outcome,\n      splitTag: runMeta.splitTag ?? 'holdout',\n    } as RunRecord)\n\n    if (emitRaterScores) {\n      for (const r of normalised) raterScores.push({ runId, rater: r.rater, score: r.score })\n    }\n  }\n\n  return { runs, raterScores }\n}\n","/**\n * # `intake/otel-spans` — OTel `TraceSpanEvent[]` → `RunRecord[]`.\n *\n * Turns an existing observability stream into the substrate-canonical\n * `RunRecord` shape so consumers with logs but no eval discipline can\n * call `analyzeRuns()` against their production traffic immediately.\n *\n * Pivot rule: spans are grouped by `tangle.runId` (the same attribute the\n * hosted-tier wire format uses) or, when absent, by `traceId`. One group\n * becomes one `RunRecord`. The root span (no `parentSpanId`) supplies:\n *\n *   - `runId` (the group key)\n *   - `wallMs` from `endTimeUnixNano - startTimeUnixNano`\n *   - `model` from `gen_ai.request.model` / `llm.model` / `tangle.model`\n *   - cost from `cost.usd` / `gen_ai.usage.cost_usd` / `tangle.cost.usd`\n *   - token usage from `gen_ai.usage.{input,output}_tokens`\n *   - `outcome.searchScore` from `tangle.score` / `eval.score` when\n *     present; `outcome.raw` collects every numeric attribute.\n *\n * Spans that ERRORed (`status.code === 'ERROR'`) populate `failureMode`\n * with their `name` so `analyzeRuns()`'s failure clustering sees them.\n */\n\nimport type { TraceSpanEvent } from '../../hosted/types'\nimport type {\n  JudgeScoresRecord,\n  RunOutcome,\n  RunRecord,\n  RunSplitTag,\n  RunTokenUsage,\n} from '../../run-record'\n\nconst SCORE_KEYS = ['tangle.score', 'eval.score', 'score']\nconst MODEL_KEYS = ['tangle.model', 'gen_ai.request.model', 'llm.model', 'model']\nconst COST_KEYS = ['tangle.cost.usd', 'gen_ai.usage.cost_usd', 'cost.usd', 'cost']\nconst INPUT_TOKEN_KEYS = ['gen_ai.usage.input_tokens', 'tangle.tokens.in', 'tokens.in']\nconst OUTPUT_TOKEN_KEYS = ['gen_ai.usage.output_tokens', 'tangle.tokens.out', 'tokens.out']\nconst PROMPT_HASH_KEYS = ['tangle.prompt_hash', 'prompt.hash']\nconst CONFIG_HASH_KEYS = ['tangle.config_hash', 'config.hash']\n\nexport interface FromOtelSpansOptions {\n  spans: TraceSpanEvent[]\n  /** Default split tag for synthesized records. Defaults to `'holdout'`. */\n  defaultSplit?: RunSplitTag\n  /** Default `experimentId` when not present on any span. */\n  experimentId?: string\n}\n\nexport function fromOtelSpans(opts: FromOtelSpansOptions): RunRecord[] {\n  const { spans, defaultSplit = 'holdout', experimentId = 'otel-corpus' } = opts\n  const grouped = groupSpans(spans)\n\n  const runs: RunRecord[] = []\n  for (const [groupKey, groupSpans] of grouped) {\n    const root = findRoot(groupSpans)\n    if (!root) continue\n\n    const wallMs = Math.max(0, (root.endTimeUnixNano - root.startTimeUnixNano) / 1_000_000)\n    const model = readAttrString(groupSpans, MODEL_KEYS) ?? 'unknown@unknown'\n    const costUsd = readAttrNumber(groupSpans, COST_KEYS) ?? 0\n    const inputTokens = readAttrNumber(groupSpans, INPUT_TOKEN_KEYS) ?? 0\n    const outputTokens = readAttrNumber(groupSpans, OUTPUT_TOKEN_KEYS) ?? 0\n    const promptHash = readAttrString(groupSpans, PROMPT_HASH_KEYS) ?? 'sha256:unknown'\n    const configHash = readAttrString(groupSpans, CONFIG_HASH_KEYS) ?? 'sha256:unknown'\n    const score = readAttrNumber(groupSpans, SCORE_KEYS)\n\n    const rawNumeric = collectNumericAttrs(groupSpans)\n    const tokenUsage: RunTokenUsage = {\n      input: inputTokens,\n      output: outputTokens,\n    }\n\n    const judgeScores: JudgeScoresRecord | undefined =\n      score !== undefined\n        ? {\n            perJudge: { 'otel-derived': { score } },\n            perDimMean: { score },\n            composite: score,\n          }\n        : undefined\n\n    const errorSpan = groupSpans.find((s) => s.status?.code === 'ERROR')\n    const outcome: RunOutcome = {\n      ...(opts.defaultSplit === 'search' ? { searchScore: score } : { holdoutScore: score }),\n      raw: rawNumeric,\n      ...(judgeScores ? { judgeScores } : {}),\n    }\n\n    runs.push({\n      runId: groupKey,\n      experimentId,\n      candidateId: (root.attributes['tangle.candidateId'] as string | undefined) ?? 'otel-default',\n      seed: 0,\n      model,\n      promptHash,\n      configHash,\n      commitSha: (root.attributes['tangle.commit_sha'] as string | undefined) ?? 'unknown',\n      wallMs,\n      costUsd,\n      tokenUsage,\n      outcome,\n      splitTag: defaultSplit,\n      ...(errorSpan ? { failureMode: errorSpan.name } : {}),\n    } as RunRecord)\n  }\n  return runs\n}\n\n// ── Internal helpers ────────────────────────────────────────────────\n\nfunction groupSpans(spans: TraceSpanEvent[]): Map<string, TraceSpanEvent[]> {\n  const m = new Map<string, TraceSpanEvent[]>()\n  for (const span of spans) {\n    const key = (span['tangle.runId'] as string | undefined) ?? span.traceId\n    const list = m.get(key) ?? []\n    list.push(span)\n    m.set(key, list)\n  }\n  return m\n}\n\nfunction findRoot(group: TraceSpanEvent[]): TraceSpanEvent | undefined {\n  return group.find((s) => !s.parentSpanId) ?? group[0]\n}\n\nfunction readAttrString(spans: TraceSpanEvent[], keys: string[]): string | undefined {\n  for (const span of spans) {\n    for (const key of keys) {\n      const v = span.attributes[key]\n      if (typeof v === 'string' && v.length > 0) return v\n    }\n  }\n  return undefined\n}\n\nfunction readAttrNumber(spans: TraceSpanEvent[], keys: string[]): number | undefined {\n  for (const span of spans) {\n    for (const key of keys) {\n      const v = span.attributes[key]\n      if (typeof v === 'number' && Number.isFinite(v)) return v\n      if (typeof v === 'string') {\n        const parsed = Number(v)\n        if (Number.isFinite(parsed)) return parsed\n      }\n    }\n  }\n  return undefined\n}\n\nfunction collectNumericAttrs(spans: TraceSpanEvent[]): Record<string, number> {\n  const raw: Record<string, number> = {}\n  for (const span of spans) {\n    for (const [k, v] of Object.entries(span.attributes)) {\n      if (typeof v === 'number' && Number.isFinite(v)) raw[k] = v\n    }\n  }\n  return raw\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA+EA,eAAsB,YAAY,MAAkD;AAClF,QAAM,OAAO,KAAK;AAClB,QAAM,OAAO,KAAK,iBAAiB;AACnC,QAAM,YAAY,KAAK,qBAAqB;AAC5C,QAAM,QAAQ,aAAa,MAAM,KAAK,SAAS,MAAM;AAErD,QAAM,mBAAmB,KACtB,IAAI,CAAC,OAAO,EAAE,OAAO,EAAE,OAAO,OAAO,YAAY,GAAG,KAAK,EAAE,EAAE,EAC7D,OAAO,CAAC,MAAM,OAAO,SAAS,EAAE,KAAK,CAAC;AACzC,QAAM,YAAY;AAAA,IAChB,iBAAiB,IAAI,CAAC,MAAM,EAAE,KAAK;AAAA,IACnC;AAAA,IACA;AAAA,EACF;AAEA,QAAM,eAAe,oBAAoB,MAAM,IAAI;AAEnD,QAAM,QAAQ,KAAK,IAAI,CAAC,MAAM,EAAE,OAAO,EAAE,OAAO,OAAO,QAAQ;AAC/D,QAAM,WAAW,eAAe,OAAO,IAAI;AAC3C,QAAM,SAAS,YAAY,MAAM,EAAE,MAAM,CAAC;AAC1C,QAAM,WAA+C,CAAC;AACtD,MAAI,MAAM,WAAW,KAAK,MAAM,MAAM,CAAC,MAAM,MAAM,CAAC,GAAG;AACrD,aAAS,OAAO;AAAA,EAClB;AACA,MAAI,OAAO,OAAO,SAAS,GAAG;AAC5B,aAAS,SACP,OAAO,OAAO,WAAW,IACrB,4CACA;AAAA,EACR;AACA,QAAM,cAAc;AAAA,IAClB,MAAM;AAAA,IACN;AAAA,IACA,GAAI,SAAS,QAAQ,SAAS,SAAS,EAAE,SAAS,IAAI,CAAC;AAAA,EACzD;AAEA,QAAM,SAAS,qBAAqB,IAAI;AAExC,QAAM,aAAa,KAAK,cAAc,kBAAkB,KAAK,WAAW,IAAI;AAE5E,QAAM,OAAO,YAAY,MAAM,KAAK,qBAAqB,KAAK,sBAAsB,KAAK;AAEzF,QAAM,kBAAkB,KAAK,UACzB,MAAM,uBAAuB,MAAM,KAAK,SAAS,KAAK,IACtD;AAEJ,QAAM,gBAAgB,KAAK,kBACvB,qBAAqB,MAAM,KAAK,eAAe,IAC/C;AAEJ,QAAM,qBAAqB,KAAK,gBAC5B,0BAA0B,MAAM,KAAK,eAAe,KAAK,IACzD;AAEJ,QAAM,UAAU,sBAAsB,WAAW,MAAM,aAAa;AAEpE,QAAM,kBAAkB,qBAAqB;AAAA,IAC3C;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF,CAAC;AAED,SAAO;AAAA,IACL,GAAG,KAAK;AAAA,IACR;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF;AACF;AAIA,SAAS,aACP,MACA,MACsB;AACtB,MAAI,SAAS,OAAQ,QAAO;AAC5B,QAAM,aAAa,KAAK,KAAK,CAAC,MAAM,OAAO,SAAS,EAAE,QAAQ,YAAY,CAAC;AAC3E,SAAO,aAAa,YAAY;AAClC;AAEA,SAAS,YAAY,KAAgB,OAAqC;AACxE,QAAM,UAAU,UAAU,YAAY,IAAI,QAAQ,eAAe,IAAI,QAAQ;AAC7E,MAAI,OAAO,SAAS,OAAO,EAAG,QAAO;AAIrC,QAAM,MAAM,UAAU,YAAY,IAAI,QAAQ,cAAc,IAAI,QAAQ;AACxE,SAAO,OAAO,SAAS,GAAG,IAAK,MAAiB,OAAO;AACzD;AAIA,SAAS,eACP,QACA,MACA,SACoB;AACpB,MAAI,OAAO,WAAW,GAAG;AACvB,WAAO;AAAA,MACL,GAAG;AAAA,MACH,MAAM;AAAA,MACN,KAAK;AAAA,MACL,KAAK;AAAA,MACL,QAAQ;AAAA,MACR,KAAK;AAAA,MACL,KAAK;AAAA,MACL,WAAW,CAAC;AAAA,IACd;AAAA,EACF;AACA,QAAM,SAAS,CAAC,GAAG,MAAM,EAAE,KAAK,CAAC,GAAG,MAAM,IAAI,CAAC;AAC/C,QAAM,IAAI,OAAO;AACjB,QAAMA,QAAO,OAAO,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI;AACjD,QAAM,WAAW,OAAO,OAAO,CAAC,GAAG,MAAM,KAAK,IAAIA,UAAS,GAAG,CAAC,IAAI;AACnE,QAAM,SAAS,KAAK,KAAK,QAAQ;AACjC,QAAM,WAAW,UACb,CAAC,GAAG,OAAO,EAAE,KAAK,CAAC,GAAG,MAAM,EAAE,QAAQ,EAAE,KAAK,EAAE,MAAM,GAAG,KAAK,IAAI,GAAG,QAAQ,MAAM,CAAC,IACnF;AACJ,SAAO;AAAA,IACL;AAAA,IACA,MAAAA;AAAA,IACA,KAAK,WAAW,QAAQ,GAAG;AAAA,IAC3B,KAAK,WAAW,QAAQ,IAAI;AAAA,IAC5B;AAAA,IACA,KAAK,OAAO,CAAC;AAAA,IACb,KAAK,OAAO,IAAI,CAAC;AAAA,IACjB,WAAW,UAAU,QAAQ,IAAI;AAAA,IACjC,GAAI,WAAW,EAAE,SAAS,IAAI,CAAC;AAAA,EACjC;AACF;AAEA,SAAS,WAAW,QAAkB,GAAmB;AACvD,MAAI,OAAO,WAAW,EAAG,QAAO;AAChC,MAAI,OAAO,WAAW,EAAG,QAAO,OAAO,CAAC;AACxC,QAAM,OAAO,OAAO,SAAS,KAAK;AAClC,QAAM,KAAK,KAAK,MAAM,GAAG;AACzB,QAAM,KAAK,KAAK,KAAK,GAAG;AACxB,MAAI,OAAO,GAAI,QAAO,OAAO,EAAE;AAC/B,QAAM,IAAI,MAAM;AAChB,SAAO,OAAO,EAAE,KAAM,IAAI,KAAK,OAAO,EAAE,IAAK;AAC/C;AAKA,SAAS,UAAU,QAAkB,MAA+C;AAClF,MAAI,OAAO,WAAW,KAAK,OAAO,EAAG,QAAO,CAAC;AAC7C,QAAM,MAAM,OAAO,CAAC;AACpB,QAAM,MAAM,OAAO,OAAO,SAAS,CAAC;AACpC,MAAI,QAAQ,IAAK,QAAO,CAAC,EAAE,IAAI,KAAK,IAAI,KAAK,OAAO,OAAO,OAAO,CAAC;AACnE,QAAM,SAAS,MAAM,OAAO;AAC5B,QAAM,MAAuC,CAAC;AAC9C,WAAS,IAAI,GAAG,IAAI,MAAM,KAAK;AAC7B,UAAM,KAAK,MAAM,IAAI;AACrB,UAAM,KAAK,MAAM,OAAO,IAAI,MAAM,KAAK;AACvC,QAAI,KAAK,EAAE,IAAI,IAAI,OAAO,EAAE,CAAC;AAAA,EAC/B;AACA,aAAW,KAAK,QAAQ;AACtB,UAAM,MAAM,KAAK,IAAI,OAAO,GAAG,KAAK,OAAO,IAAI,OAAO,KAAK,CAAC;AAC5D,QAAI,GAAG,EAAG;AAAA,EACZ;AACA,SAAO;AACT;AAEA,SAAS,oBAAoB,MAAmB,MAAkD;AAKhG,QAAM,QAAQ,oBAAI,IAAsB;AACxC,aAAW,OAAO,MAAM;AACtB,UAAM,SAAS,IAAI,QAAQ;AAC3B,QAAI,CAAC,OAAQ;AACb,eAAW,CAAC,KAAK,KAAK,KAAK,OAAO,QAAQ,OAAO,cAAc,CAAC,CAAC,GAAG;AAClE,UAAI,CAAC,OAAO,SAAS,KAAK,EAAG;AAC7B,YAAM,MAAM,MAAM,IAAI,GAAG,KAAK,CAAC;AAC/B,UAAI,KAAK,KAAK;AACd,YAAM,IAAI,KAAK,GAAG;AAAA,IACpB;AAAA,EACF;AACA,QAAM,MAA0C,CAAC;AACjD,aAAW,CAAC,KAAK,MAAM,KAAK,MAAO,KAAI,GAAG,IAAI,eAAe,QAAQ,IAAI;AACzE,SAAO;AACT;AAIA,SAAS,qBAAqB,MAAiD;AAI7E,QAAM,MAAoC,CAAC;AAC3C,QAAM,UAAU,oBAAI,IAAsB;AAC1C,aAAW,OAAO,MAAM;AACtB,UAAM,SAAS,IAAI,QAAQ;AAC3B,QAAI,CAAC,QAAQ,SAAU;AACvB,eAAW,CAAC,SAAS,IAAI,KAAK,OAAO,QAAQ,OAAO,QAAQ,GAAG;AAC7D,YAAM,YAAY,OAAO,OAAO,IAAI,EAAE,OAAO,OAAO,QAAQ;AAC5D,UAAI,UAAU,WAAW,EAAG;AAC5B,YAAM,YAAY,UAAU,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,UAAU;AACnE,YAAM,MAAM,QAAQ,IAAI,OAAO,KAAK,CAAC;AACrC,UAAI,KAAK,SAAS;AAClB,cAAQ,IAAI,SAAS,GAAG;AAAA,IAC1B;AAAA,EACF;AACA,aAAW,CAAC,SAAS,MAAM,KAAK,SAAS;AACvC,QAAI,OAAO,IAAI;AAAA,MACb,GAAG,OAAO;AAAA,MACV,WAAW,OAAO,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,OAAO;AAAA,IACxD;AAAA,EACF;AACA,SAAO;AACT;AAIA,SAAS,kBACP,SAC+B;AAC/B,QAAM,QAAQ,oBAAI,IAAqD;AACvE,aAAW,KAAK,SAAS;AACvB,QAAI,CAAC,OAAO,SAAS,EAAE,KAAK,EAAG;AAC/B,UAAM,OAAO,MAAM,IAAI,EAAE,KAAK,KAAK,CAAC;AACpC,SAAK,KAAK,EAAE,OAAO,EAAE,OAAO,OAAO,EAAE,MAAM,CAAC;AAC5C,UAAM,IAAI,EAAE,OAAO,IAAI;AAAA,EACzB;AACA,QAAM,SAAS,IAAI,IAAI,QAAQ,IAAI,CAAC,MAAM,EAAE,KAAK,CAAC;AAClD,QAAM,eAAyB,CAAC;AAChC,aAAW,CAAC,OAAO,YAAY,KAAK,OAAO;AACzC,UAAM,OAAO,IAAI,IAAI,aAAa,IAAI,CAAC,MAAM,EAAE,KAAK,CAAC;AACrD,QAAI,MAAM;AACV,eAAW,KAAK,OAAQ,KAAI,CAAC,KAAK,IAAI,CAAC,EAAG,OAAM;AAChD,QAAI,IAAK,cAAa,KAAK,KAAK;AAAA,EAClC;AACA,MAAI,OAAO,OAAO,KAAK,aAAa,WAAW,EAAG,QAAO;AAEzD,QAAM,YAAY,CAAC,GAAG,MAAM,EAAE,KAAK;AACnC,QAAM,UAAkC,CAAC;AACzC,WAAS,IAAI,GAAG,IAAI,UAAU,QAAQ,KAAK;AACzC,aAAS,IAAI,IAAI,GAAG,IAAI,UAAU,QAAQ,KAAK;AAC7C,YAAM,IAAI,UAAU,CAAC;AACrB,YAAM,IAAI,UAAU,CAAC;AACrB,YAAM,UAAoB,CAAC;AAC3B,YAAM,UAAoB,CAAC;AAC3B,iBAAW,SAAS,cAAc;AAChC,cAAM,eAAe,MAAM,IAAI,KAAK;AACpC,cAAM,KAAK,aAAa,KAAK,CAAC,MAAM,EAAE,UAAU,CAAC,GAAG;AACpD,cAAM,KAAK,aAAa,KAAK,CAAC,MAAM,EAAE,UAAU,CAAC,GAAG;AACpD,YAAI,OAAO,UAAa,OAAO,QAAW;AACxC,kBAAQ,KAAK,EAAE;AACf,kBAAQ,KAAK,EAAE;AAAA,QACjB;AAAA,MACF;AACA,cAAQ,GAAG,CAAC,KAAK,CAAC,EAAE,IAAI,QAAQ,SAAS,OAAO;AAAA,IAClD;AAAA,EACF;AACA,QAAM,aAAa,OAAO,OAAO,OAAO;AACxC,QAAM,QACJ,WAAW,WAAW,IAAI,IAAI,WAAW,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,WAAW;AAEnF,QAAM,oBAAoB,aACvB,IAAI,CAAC,UAAU;AACd,UAAM,eAAe,MAAM,IAAI,KAAK;AACpC,UAAM,SAAS,aAAa,IAAI,CAAC,MAAM,EAAE,KAAK;AAC9C,UAAM,QAAQ,KAAK,IAAI,GAAG,MAAM,IAAI,KAAK,IAAI,GAAG,MAAM;AACtD,WAAO,EAAE,OAAO,SAAS,cAAc,MAAM;AAAA,EAC/C,CAAC,EACA,KAAK,CAAC,GAAG,MAAM,EAAE,QAAQ,EAAE,KAAK,EAChC,MAAM,GAAG,EAAE;AAEd,SAAO;AAAA,IACL,QAAQ,OAAO;AAAA,IACf,cAAc,aAAa;AAAA,IAC3B;AAAA,IACA;AAAA,IACA;AAAA,EACF;AACF;AAEA,SAAS,QAAQ,GAAa,GAAqB;AACjD,MAAI,EAAE,WAAW,EAAE,UAAU,EAAE,WAAW,EAAG,QAAO;AACpD,QAAM,IAAI,EAAE;AACZ,QAAM,QAAQ,EAAE,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI;AAC7C,QAAM,QAAQ,EAAE,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI;AAC7C,MAAI,MAAM;AACV,MAAI,SAAS;AACb,MAAI,SAAS;AACb,WAAS,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,UAAM,KAAK,EAAE,CAAC,IAAK;AACnB,UAAM,KAAK,EAAE,CAAC,IAAK;AACnB,WAAO,KAAK;AACZ,cAAU,KAAK;AACf,cAAU,KAAK;AAAA,EACjB;AACA,QAAM,QAAQ,KAAK,KAAK,SAAS,MAAM;AACvC,SAAO,UAAU,IAAI,IAAI,MAAM;AACjC;AAIA,SAAS,YACP,MACA,YACA,aACA,OACyB;AACzB,MAAI,MAAM;AACV,MAAI,MAAM;AACV,MAAI,CAAC,OAAO,CAAC,KAAK;AAGhB,UAAM,MAAM,CAAC,GAAG,IAAI,IAAI,KAAK,IAAI,CAAC,MAAM,EAAE,WAAW,CAAC,CAAC;AACvD,QAAI,IAAI,WAAW,EAAG,QAAO;AAC7B,UAAM,CAAC,KAAK,GAAG,IAAI;AACnB,UAAM,QAAQ,KAAK,KAAK,OAAO,CAAC,MAAM,EAAE,gBAAgB,GAAG,EAAE,IAAI,CAAC,MAAM,YAAY,GAAG,KAAK,CAAC,CAAC;AAC9F,UAAM,QAAQ,KAAK,KAAK,OAAO,CAAC,MAAM,EAAE,gBAAgB,GAAG,EAAE,IAAI,CAAC,MAAM,YAAY,GAAG,KAAK,CAAC,CAAC;AAC9F,UAAM,SAAS,QAAQ,MAAM;AAC7B,UAAM,SAAS,QAAQ,MAAM;AAAA,EAC/B;AAEA,QAAM,WAAW,KAAK,OAAO,CAAC,MAAM,EAAE,gBAAgB,GAAG;AACzD,QAAM,YAAY,KAAK,OAAO,CAAC,MAAM,EAAE,gBAAgB,GAAG;AAC1D,MAAI,SAAS,WAAW,KAAK,UAAU,WAAW,EAAG,QAAO;AAI5D,QAAM,gBAAgB,IAAI,IAAI,SAAS,IAAI,CAAC,MAAM,CAAC,WAAW,CAAC,GAAG,CAAC,CAAC,CAAC;AACrE,QAAM,iBAA2B,CAAC;AAClC,QAAM,kBAA4B,CAAC;AACnC,MAAI,iBAAiB;AACrB,aAAW,QAAQ,WAAW;AAC5B,UAAM,IAAI,cAAc,IAAI,WAAW,IAAI,CAAC;AAC5C,QAAI,GAAG;AACL,YAAM,KAAK,YAAY,GAAG,KAAK;AAC/B,YAAM,KAAK,YAAY,MAAM,KAAK;AAClC,UAAI,OAAO,SAAS,EAAE,KAAK,OAAO,SAAS,EAAE,GAAG;AAC9C,uBAAe,KAAK,EAAE;AACtB,wBAAgB,KAAK,EAAE;AACvB,yBAAiB;AAAA,MACnB;AAAA,IACF;AAAA,EACF;AACA,MAAI,CAAC,gBAAgB;AACnB,UAAM,IAAI,KAAK,IAAI,SAAS,QAAQ,UAAU,MAAM;AACpD,aAAS,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,YAAM,KAAK,YAAY,SAAS,CAAC,GAAI,KAAK;AAC1C,YAAM,KAAK,YAAY,UAAU,CAAC,GAAI,KAAK;AAC3C,UAAI,OAAO,SAAS,EAAE,KAAK,OAAO,SAAS,EAAE,GAAG;AAC9C,uBAAe,KAAK,EAAE;AACtB,wBAAgB,KAAK,EAAE;AAAA,MACzB;AAAA,IACF;AAAA,EACF;AACA,MAAI,eAAe,WAAW,EAAG,QAAO;AAExC,QAAM,eAAe,KAAK,cAAc;AACxC,QAAM,gBAAgB,KAAK,eAAe;AAC1C,QAAM,QAAQ,gBAAgB;AAE9B,QAAM,YAAY,gBAAgB,gBAAgB,iBAAiB;AAAA,IACjE,YAAY;AAAA,IACZ,WAAW;AAAA,IACX,WAAW;AAAA,EACb,CAAC;AACD,QAAM,QAAQ,YAAY,gBAAgB,eAAe;AACzD,QAAM,IAAI,QAAQ,gBAAgB,eAAe;AACjD,QAAM,MAAM,UAAU,EAAE,SAAS,eAAe,QAAQ,OAAO,KAAK,OAAO,KAAK,CAAC;AACjF,QAAM,YAAY,mBAAmB;AAAA,IACnC,QAAQ,KAAK,IAAI,KAAK,IAAI,KAAK,GAAG,IAAI;AAAA,IACtC,OAAO;AAAA,IACP,OAAO;AAAA,EACT,CAAC;AAED,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA;AAAA,IACA,MAAM,CAAC,UAAU,KAAK,UAAU,IAAI;AAAA,IACpC,QAAQ,MAAM;AAAA,IACd,GAAG,eAAe;AAAA,IAClB,SAAS;AAAA,IACT;AAAA,IACA;AAAA,EACF;AACF;AAEA,SAAS,WAAW,GAAsB;AACxC,SAAO,GAAG,EAAE,YAAY,KAAK,EAAE,IAAI;AACrC;AAEA,SAAS,KAAK,KAAuB;AACnC,SAAO,IAAI,WAAW,IAAI,IAAI,IAAI,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,IAAI;AACrE;AAIA,eAAe,uBACb,MACA,SACA,OAC4C;AAC5C,QAAM,SAAS,KAAK,OAAO,CAAC,MAAM,YAAY,GAAG,KAAK,IAAI,OAAO,EAAE,gBAAgB,MAAS;AAC5F,MAAI,OAAO,WAAW,EAAG,QAAO,EAAE,UAAU,CAAC,GAAG,eAAe,EAAE;AAEjE,QAAM,WAAW,oBAAI,IAAoD;AACzE,aAAW,OAAO,QAAQ;AACxB,QAAI;AACF,YAAM,SAAS,MAAM,QAAQ,IAAI,IAAI,OAAO;AAAA,QAC1C,MAAM;AAAA,QACN;AAAA,MACF,CAAsC;AACtC,iBAAW,WAAW,OAAO,UAA8B;AACzD,cAAM,MAAM,QAAQ,QAAQ,QAAQ,cAAc;AAClD,cAAM,IAAI,SAAS,IAAI,GAAG,KAAK,EAAE,WAAW,CAAC,GAAG,OAAO,EAAE;AACzD,YAAI,EAAE,UAAU,SAAS,EAAG,GAAE,UAAU,KAAK,IAAI,KAAK;AACtD,iBAAS,IAAI,KAAK,CAAC;AAAA,MACrB;AAAA,IACF,QAAQ;AACN,YAAM,IAAI,SAAS,IAAI,eAAe,KAAK,EAAE,WAAW,CAAC,GAAG,OAAO,EAAE;AACrE,UAAI,EAAE,UAAU,SAAS,EAAG,GAAE,UAAU,KAAK,IAAI,KAAK;AACtD,eAAS,IAAI,iBAAiB,CAAC;AAAA,IACjC;AAAA,EACF;AACA,QAAM,cAAc,CAAC,GAAG,SAAS,QAAQ,CAAC,EAAE,IAAI,CAAC,CAAC,IAAI,CAAC,OAAO;AAAA,IAC5D;AAAA,IACA,MAAM;AAAA,IACN,OAAO,EAAE,UAAU,SAAS,OAAO;AAAA,IACnC,WAAW,EAAE;AAAA,EACf,EAAE;AACF,cAAY,KAAK,CAAC,GAAG,MAAM,EAAE,QAAQ,EAAE,KAAK;AAC5C,SAAO,EAAE,UAAU,aAAa,eAAe,OAAO,OAAO;AAC/D;AAIA,SAAS,qBACP,MACA,UACgC;AAChC,MAAI,QAAQ;AACZ,QAAM,UAAqE,CAAC;AAC5E,aAAW,OAAO,MAAM;AACtB,UAAM,SAAS,gBAAgB,GAAG;AAClC,QAAI,CAAC,OAAQ;AACb,UAAM,YAAY,cAAc,QAAQ,QAAQ;AAChD,eAAW,QAAQ,WAAW;AAC5B;AACA,cAAQ,KAAK,EAAE,OAAO,IAAI,OAAO,QAAQ,KAAK,QAAQ,SAAS,KAAK,SAAS,CAAC;AAAA,IAChF;AAAA,EACF;AACA,SAAO,EAAE,OAAO,oBAAoB,UAAU,GAAG,QAAQ;AAC3D;AAEA,SAAS,gBAAgB,KAAoC;AAO3D,QAAM,WAAY,IAA0D;AAC5E,MAAI,OAAO,UAAU,WAAW,SAAU,QAAO,SAAS;AAC1D,MAAI,OAAO,UAAU,SAAS,SAAU,QAAO,SAAS;AACxD,SAAO;AACT;AAIA,SAAS,0BACP,MACA,SACA,OACuC;AACvC,QAAM,KAAe,CAAC;AACtB,QAAM,KAAe,CAAC;AACtB,aAAW,OAAO,MAAM;AACtB,UAAM,IAAI,QAAQ,aAAa,IAAI,KAAK;AACxC,QAAI,MAAM,UAAa,CAAC,OAAO,SAAS,CAAC,EAAG;AAC5C,UAAM,IAAI,YAAY,KAAK,KAAK;AAChC,QAAI,CAAC,OAAO,SAAS,CAAC,EAAG;AACzB,OAAG,KAAK,CAAC;AACT,OAAG,KAAK,CAAC;AAAA,EACX;AACA,MAAI,GAAG,SAAS,EAAG,QAAO;AAE1B,QAAM,IAAI,QAAQ,IAAI,EAAE;AACxB,QAAM,IAAI,SAAS,IAAI,EAAE;AACzB,QAAM,QAAQ,KAAK,EAAE;AACrB,QAAM,QAAQ,KAAK,EAAE;AACrB,MAAI,MAAM;AACV,MAAI,QAAQ;AACZ,WAAS,IAAI,GAAG,IAAI,GAAG,QAAQ,KAAK;AAClC,YAAQ,GAAG,CAAC,IAAK,UAAU,GAAG,CAAC,IAAK;AACpC,cAAU,GAAG,CAAC,IAAK,UAAU;AAAA,EAC/B;AACA,QAAM,QAAQ,UAAU,IAAI,IAAI,MAAM;AACtC,QAAM,YAAY,QAAQ,QAAQ;AAClC,QAAM,QAAQ,GAAG,OAAO,CAAC,GAAG,MAAM,KAAK,IAAI,UAAU,GAAG,CAAC;AACzD,QAAM,QAAQ,GAAG,OAAO,CAAC,GAAG,GAAG,MAAM,KAAK,KAAK,YAAY,QAAQ,GAAG,CAAC,OAAQ,GAAG,CAAC;AACnF,QAAM,KAAK,UAAU,IAAI,IAAI,IAAI,QAAQ;AAEzC,SAAO;AAAA,IACL,QAAQ,QAAQ;AAAA,IAChB,GAAG,GAAG;AAAA,IACN,SAAS;AAAA,IACT,UAAU;AAAA,IACV,aAAa,EAAE,WAAW,OAAO,GAAG;AAAA,EACtC;AACF;AAEA,SAAS,SAAS,GAAa,GAAqB;AAClD,MAAI,EAAE,WAAW,EAAE,UAAU,EAAE,WAAW,EAAG,QAAO;AACpD,SAAO,QAAQ,KAAK,CAAC,GAAG,KAAK,CAAC,CAAC;AACjC;AAEA,SAAS,KAAK,KAAyB;AACrC,QAAM,UAAU,IAAI,IAAI,CAAC,GAAGC,QAAO,EAAE,GAAG,GAAAA,GAAE,EAAE;AAC5C,UAAQ,KAAK,CAAC,GAAG,MAAM,EAAE,IAAI,EAAE,CAAC;AAChC,QAAM,QAAQ,IAAI,MAAM,IAAI,MAAM,EAAE,KAAK,CAAC;AAC1C,MAAI,IAAI;AACR,SAAO,IAAI,QAAQ,QAAQ;AACzB,QAAI,IAAI;AACR,WAAO,IAAI,IAAI,QAAQ,UAAU,QAAQ,IAAI,CAAC,EAAG,MAAM,QAAQ,CAAC,EAAG,EAAG;AACtE,UAAM,OAAO,IAAI,IAAI,KAAK;AAC1B,aAAS,IAAI,GAAG,KAAK,GAAG,IAAK,OAAM,QAAQ,CAAC,EAAG,CAAC,IAAI;AACpD,QAAI,IAAI;AAAA,EACV;AACA,SAAO;AACT;AAIA,SAAS,sBACP,WACA,MACA,eAC0B;AAO1B,QAAM,OAAyC,CAAC;AAChD,QAAM,WACJ,SAAS,UAAa,KAAK,KAAK,CAAC,IAAI,IAChC,SACD,KAAK,QAAQ,IACV,SACA;AACT,OAAK,KAAK;AAAA,IACR,MAAM;AAAA,IACN,QAAQ;AAAA,IACR,QAAQ,OACJ,SAAS,KAAK,MAAM,QAAQ,CAAC,CAAC,WAAW,KAAK,KAAK,CAAC,EAAE,QAAQ,CAAC,CAAC,KAAK,KAAK,KAAK,CAAC,EAAE,QAAQ,CAAC,CAAC,QAAQ,KAAK,CAAC,KAC1G;AAAA,EACN,CAAC;AACD,QAAM,aACJ,kBAAkB,UAAa,cAAc,UAAU,IAAK,SAAoB;AAClF,OAAK,KAAK;AAAA,IACR,MAAM;AAAA,IACN,QAAQ;AAAA,IACR,QAAQ,gBAAgB,GAAG,cAAc,KAAK,oBAAoB;AAAA,EACpE,CAAC;AACD,OAAK,KAAK;AAAA,IACR,MAAM;AAAA,IACN,QAAQ,UAAU,QAAQ,MAAM,SAAS,UAAU,QAAQ,MAAM,SAAS;AAAA,IAC1E,QAAQ,QAAQ,UAAU,KAAK,QAAQ,CAAC,CAAC,SAAS,UAAU,IAAI,QAAQ,CAAC,CAAC,SAAS,UAAU,IAAI,QAAQ,CAAC,CAAC,WAAW,UAAU,CAAC;AAAA,EACnI,CAAC;AACD,QAAM,SAAS,KAAK,KAAK,CAAC,MAAM,EAAE,WAAW,MAAM,IAC/C,SACA,KAAK,KAAK,CAAC,MAAM,EAAE,WAAW,MAAM,IAClC,SACA;AACN,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA,QAAQ,CAAC;AAAA,EACX;AACF;AAeA,SAAS,qBAAqB,KAA8C;AAC1E,QAAM,MAAwB,CAAC;AAK/B,MAAI,IAAI,UAAU,IAAI,GAAG;AACvB,QAAI,IAAI,UAAU,OAAO,KAAK;AAC5B,YAAM,OAAO,IAAI,UAAU,YAAY,CAAC;AACxC,YAAM,QAAQ,KACX,MAAM,GAAG,CAAC,EACV,IAAI,CAAC,MAAM,GAAG,EAAE,KAAK,IAAI,EAAE,MAAM,QAAQ,CAAC,CAAC,EAAE,EAC7C,KAAK,IAAI;AACZ,UAAI,KAAK;AAAA,QACP,UAAU;AAAA,QACV,MAAM;AAAA,QACN,OAAO,kBAAkB,IAAI,UAAU,KAAK,QAAQ,CAAC,CAAC;AAAA,QACtD,QACE,KAAK,SAAS,IACV,SAAS,KAAK,MAAM,OAAO,KAAK,WAAW,IAAI,KAAK,GAAG,sBAAsB,KAAK,mBAAmB,IAAI,UAAU,IAAI,QAAQ,CAAC,CAAC,SAAS,IAAI,UAAU,IAAI,QAAQ,CAAC,CAAC,MACtK,iBAAiB,IAAI,UAAU,IAAI,QAAQ,CAAC,CAAC,SAAS,IAAI,UAAU,IAAI,QAAQ,CAAC,CAAC;AAAA,QACxF,cAAc;AAAA,MAChB,CAAC;AAAA,IACH,WAAW,IAAI,UAAU,OAAO,KAAK;AACnC,YAAM,OAAO,IAAI,UAAU,YAAY,CAAC;AACxC,YAAM,QAAQ,KACX,MAAM,GAAG,CAAC,EACV,IAAI,CAAC,MAAM,GAAG,EAAE,KAAK,IAAI,EAAE,MAAM,QAAQ,CAAC,CAAC,EAAE,EAC7C,KAAK,IAAI;AACZ,UAAI,KAAK;AAAA,QACP,UAAU;AAAA,QACV,MAAM;AAAA,QACN,OAAO,kBAAkB,IAAI,UAAU,KAAK,QAAQ,CAAC,CAAC;AAAA,QACtD,QACE,KAAK,SAAS,IACV,SAAS,KAAK,MAAM,OAAO,KAAK,WAAW,IAAI,KAAK,GAAG,KAAK,KAAK,mBAAmB,IAAI,UAAU,IAAI,QAAQ,CAAC,CAAC,SAAS,IAAI,UAAU,IAAI,QAAQ,CAAC,CAAC,MACrJ,iBAAiB,IAAI,UAAU,IAAI,QAAQ,CAAC,CAAC,SAAS,IAAI,UAAU,IAAI,QAAQ,CAAC,CAAC;AAAA,QACxF,cAAc;AAAA,MAChB,CAAC;AAAA,IACH;AAAA,EACF;AAKA,MAAI,OAAO,KAAK,IAAI,MAAM,EAAE,WAAW,KAAK,IAAI,UAAU,IAAI,GAAG;AAC/D,QAAI,KAAK;AAAA,MACP,UAAU;AAAA,MACV,MAAM;AAAA,MACN,OAAO;AAAA,MACP,QACE;AAAA,MACF,cAAc;AAAA,IAChB,CAAC;AAAA,EACH;AAEA,MAAI,IAAI,MAAM;AACZ,UAAM,WAAW,IAAI,KAAK,KAAK,CAAC,IAAI,IAAI;AACxC,UAAM,eAAe,IAAI,KAAK,KAAK,CAAC,KAAK,IAAI,aAAa,IAAI,KAAK,KAAK,CAAC,IAAI,IAAI;AACjF,QAAI,UAAU;AACZ,UAAI,KAAK;AAAA,QACP,UAAU;AAAA,QACV,MAAM;AAAA,QACN,OAAO,oBAAe,IAAI,KAAK,MAAM,QAAQ,CAAC,CAAC,YAAY,IAAI,KAAK,KAAK,CAAC,EAAE,QAAQ,CAAC,CAAC,KAAK,IAAI,KAAK,KAAK,CAAC,EAAE,QAAQ,CAAC,CAAC;AAAA,QACtH,QAAQ,kCAAkC,IAAI,SAAS,qCAAqC,IAAI,KAAK,CAAC,OAAO,IAAI,KAAK,OAAO,QAAQ,CAAC,CAAC,OAAO,IAAI,KAAK,QAAQ,QAAQ,CAAC,CAAC;AAAA,QACzK,cAAc;AAAA,MAChB,CAAC;AAAA,IACH,WAAW,cAAc;AACvB,UAAI,KAAK;AAAA,QACP,UAAU;AAAA,QACV,MAAM;AAAA,QACN,OAAO,6BAAwB,IAAI,KAAK,SAAS,sBAAsB,IAAI,KAAK,CAAC;AAAA,QACjF,QAAQ,uDAAuD,IAAI,KAAK,IAAI,QAAQ,CAAC,CAAC,uBAAuB,IAAI,KAAK,MAAM,QAAQ,CAAC,CAAC;AAAA,QACtI,cAAc;AAAA,MAChB,CAAC;AAAA,IACH,OAAO;AACL,UAAI,KAAK;AAAA,QACP,UAAU;AAAA,QACV,MAAM;AAAA,QACN,OAAO,mCAA8B,IAAI,KAAK,KAAK,CAAC,EAAE,QAAQ,CAAC,CAAC,6BAA6B,IAAI,SAAS;AAAA,QAC1G,QAAQ;AAAA,QACR,cAAc;AAAA,MAChB,CAAC;AAAA,IACH;AAAA,EACF;AAEA,MAAI,IAAI,iBAAiB,IAAI,cAAc,QAAQ,GAAG;AACpD,QAAI,KAAK;AAAA,MACP,UAAU;AAAA,MACV,MAAM;AAAA,MACN,OAAO,GAAG,IAAI,cAAc,KAAK,eAAe,IAAI,cAAc,UAAU,IAAI,KAAK,GAAG;AAAA,MACxF,QAAQ;AAAA,MACR,cAAc;AAAA,IAChB,CAAC;AAAA,EACH;AAEA,MAAI,IAAI,cAAc,IAAI,WAAW,QAAQ,KAAK;AAChD,QAAI,KAAK;AAAA,MACP,UAAU;AAAA,MACV,MAAM;AAAA,MACN,OAAO,gCAA2B,IAAI,WAAW,MAAM,QAAQ,CAAC,CAAC;AAAA,MACjE,QAAQ;AAAA,MACR,cAAc;AAAA,IAChB,CAAC;AAAA,EACH;AAEA,MAAI,IAAI,mBAAmB,IAAI,gBAAgB,SAAS,SAAS,GAAG;AAClE,UAAM,MAAM,IAAI,gBAAgB,SAAS,CAAC;AAC1C,QAAI,KAAK;AAAA,MACP,UAAU;AAAA,MACV,MAAM;AAAA,MACN,OAAO,wBAAwB,IAAI,IAAI,MAAM,IAAI,QAAQ,KAAK,QAAQ,CAAC,CAAC;AAAA,MACxE,QAAQ,GAAG,IAAI,gBAAgB,aAAa,4CAA4C,IAAI,UAAU,MAAM,qBAAqB,IAAI,IAAI;AAAA,MACzI,cAAc;AAAA,IAChB,CAAC;AAAA,EACH;AAEA,MAAI,IAAI,sBAAsB,KAAK,IAAI,IAAI,mBAAmB,QAAQ,IAAI,KAAK;AAC7E,QAAI,KAAK;AAAA,MACP,UAAU;AAAA,MACV,MAAM;AAAA,MACN,OAAO,+BAA+B,IAAI,mBAAmB,MAAM,qBAAgB,IAAI,mBAAmB,SAAS,QAAQ,CAAC,CAAC;AAAA,MAC7H,QAAQ,yFAAyF,IAAI,mBAAmB,MAAM,2CAA2C,IAAI,mBAAmB,MAAM;AAAA,MACtM,cAAc;AAAA,IAChB,CAAC;AAAA,EACH;AAEA,SAAO;AACT;;;ACjlBA,SAAS,kBACP,WACA,UAC8C;AAE9C,WAAS,KAAK,GAAmB;AAC/B,QAAI,IAAI,eAAe;AACvB,aAAS,IAAI,GAAG,IAAI,EAAE,QAAQ,KAAK;AACjC,WAAK,EAAE,WAAW,CAAC;AACnB,UAAI,KAAK,KAAK,GAAG,QAAQ,MAAM;AAAA,IACjC;AACA,WAAO;AAAA,EACT;AACA,QAAM,SAAS,CAAC,GAAG,SAAS,EAAE,KAAK,CAAC,GAAG,MAAM,KAAK,EAAE,EAAE,IAAI,KAAK,EAAE,EAAE,CAAC;AACpE,QAAM,WAAW,KAAK,IAAI,GAAG,KAAK,IAAI,OAAO,SAAS,GAAG,KAAK,MAAM,OAAO,SAAS,QAAQ,CAAC,CAAC;AAC9F,SAAO;AAAA,IACL,SAAS,OAAO,MAAM,GAAG,QAAQ;AAAA,IACjC,OAAO,OAAO,MAAM,QAAQ;AAAA,EAC9B;AACF;AAEA,SAAS,cAAc,YAGrB;AACA,QAAM,cAAsC,CAAC;AAC7C,QAAM,SAAmB,CAAC;AAC1B,aAAW,CAAC,IAAI,GAAG,KAAK,OAAO,QAAQ,UAAU,GAAG;AAClD,gBAAY,EAAE,IAAI,IAAI;AACtB,WAAO,KAAK,IAAI,aAAa;AAAA,EAC/B;AACA,SAAO;AAAA,IACL,eAAe,OAAO,WAAW,IAAI,IAAI,OAAO,OAAO,CAAC,GAAG,MAAM,IAAI,GAAG,CAAC,IAAI,OAAO;AAAA,IACpF;AAAA,EACF;AACF;AAEA,IAAM,8BAA8B;AAAA,EAClC;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF;AA2BA,eAAsB,YACpB,MACkD;AAClD,QAAM,YAAY,KAAK,IAAI;AAE3B,QAAM,SAAS,KAAK,UAAU,CAAC;AAC/B,QAAM,cAAc,OAAO,eAAe;AAC1C,QAAM,iBAAiB,OAAO,kBAAkB;AAChD,QAAM,iBAAiB,OAAO,kBAAkB;AAChD,QAAM,kBAAkB,OAAO,mBAAmB;AAClD,QAAM,cAAc,OAAO;AAE3B,QAAM,kBAAkB,OAAO;AAC/B,QAAM,EAAE,OAAO,QAAQ,IAAI,kBACvB;AAAA,IACE,OAAO,KAAK,UAAU,OAAO,CAAC,MAAM,CAAC,gBAAgB,KAAK,CAAC,MAAM,EAAE,OAAO,EAAE,EAAE,CAAC;AAAA,IAC/E,SAAS;AAAA,EACX,IACA,kBAAkB,KAAK,WAAW,eAAe;AAErD,MAAI,MAAM,WAAW,GAAG;AACtB,UAAM,IAAI;AAAA,MACR;AAAA,IACF;AAAA,EACF;AACA,MAAI,QAAQ,WAAW,GAAG;AACxB,UAAM,IAAI,MAAM,2DAA2D;AAAA,EAC7E;AAEA,QAAM,SACJ,KAAK,UACL,WAAW;AAAA,IACT,KAAK;AAAA,MACH,SAAS,KAAK,KAAK,WAAW;AAAA,MAC9B,QAAQ,KAAK,KAAK,UAAU,QAAQ,IAAI,kBAAkB;AAAA,IAC5D;AAAA,IACA,OAAO,KAAK,KAAK,SAAS;AAAA,IAC1B,QACE,KAAK,gBACL;AAAA,IACF,oBAAoB,KAAK,sBAAsB;AAAA,EACjD,CAAC;AAEH,QAAM,OACJ,KAAK,QACL,sBAA4C;AAAA,IAC1C,kBAAkB;AAAA,IAClB,gBAAgB;AAAA,EAClB,CAAC;AAEH,QAAM,UAAU,KAAK,WAAW,wBAAwB;AACxD,QAAM,SAAS,KAAK,UAAU,qBAAqB,SAAS;AAE5D,MAAI,KAAK,YAAY;AACnB,SAAK,WAAW,EAAE,MAAM,oBAAoB,WAAW,KAAK,UAAU,OAAO,CAAC;AAAA,EAChF;AAEA,QAAM,SAAS,MAAM,mBAAyC;AAAA,IAC5D,WAAW;AAAA,IACX,iBAAiB,KAAK;AAAA,IACtB,qBAAqB,KAAK;AAAA,IAC1B;AAAA,IACA,QAAQ,CAAC,KAAK,KAAK;AAAA,IACnB;AAAA,IACA,gBAAgB;AAAA,IAChB,kBAAkB;AAAA,IAClB;AAAA,IACA,eAAe,KAAK,iBAAiB;AAAA,IACrC,SAAS,KAAK;AAAA,IACd,QAAQ,KAAK;AAAA,IACb;AAAA,IACA;AAAA,IACA;AAAA,IACA,eAAe,KAAK;AAAA,IACpB;AAAA,EACF,CAAC;AAED,QAAM,WAAW,cAAc,OAAO,kBAAkB,WAAW,UAAU;AAC7E,QAAM,cAAc,cAAc,OAAO,gBAAgB,WAAW,UAAU;AAE9E,MAAI,KAAK,YAAY;AACnB,SAAK,WAAW;AAAA,MACd,MAAM;AAAA,MACN,eAAe,SAAS;AAAA,MACxB,YAAY,KAAK,IAAI,IAAI;AAAA,IAC3B,CAAC;AACD,SAAK,WAAW;AAAA,MACd,MAAM;AAAA,MACN,UAAU,OAAO,WAAW;AAAA,MAC5B,MAAM,YAAY,gBAAgB,SAAS;AAAA,IAC7C,CAAC;AAAA,EACH;AAEA,QAAM,YACJ,OAAO,iBAAiB,WAAW,eACnC,OAAO,YAAY;AAAA,IACjB,CAAC,KAAK,QACJ,MAAM,IAAI,SAAS,OAAO,CAAC,GAAG,OAAO,IAAI,GAAG,SAAS,WAAW,cAAc,CAAC;AAAA,IACjF;AAAA,EACF;AAMF,QAAM,UAAU,MAAM,YAAY;AAAA,IAChC,MAAM;AAAA,MACJ,GAAG,kBAAkB,OAAO,iBAAiB,OAAO,YAAY,MAAM;AAAA,MACtE,GAAG,kBAAkB,OAAO,gBAAgB,OAAO,UAAU,MAAM;AAAA,IACrE;AAAA,IACA,qBAAqB;AAAA,IACrB,sBAAsB;AAAA,EACxB,CAAC;AAED,QAAM,UAAmD;AAAA,IACvD;AAAA,IACA,QAAQ;AAAA,MACN,GAAG;AAAA,MACH,SAAS,OAAO;AAAA,IAClB;AAAA,IACA,MAAM,YAAY,gBAAgB,SAAS;AAAA,IAC3C,cAAc,OAAO,WAAW;AAAA,IAChC,qBAAqB,OAAO,YAAY;AAAA,IACxC,YAAY,KAAK,IAAI,IAAI;AAAA,IACzB,cAAc;AAAA,IACd;AAAA,IACA,KAAK;AAAA,EACP;AAKA,MAAI,KAAK,cAAc;AACrB,QAAI;AACF,YAAM,oBAAoB,KAAK,cAAc,MAAM,SAAS,QAAQ,MAAM;AAAA,IAC5E,SAAS,KAAK;AACZ,YAAM,MAAM,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAE3D,cAAQ,KAAK,mDAAmD,GAAG,EAAE;AAAA,IACvE;AAAA,EACF;AAEA,SAAO;AACT;AAEA,eAAe,oBACb,QACA,MACA,SACA,KACA,QACe;AACf,QAAM,SAAS,mBAAmB,MAAM;AAExC,WAAS,qBACP,OACA,SACA,UACA,YAC2B;AAC3B,UAAM,QAA4B,SAAS,MAAM,IAAI,CAAC,SAAS;AAC7D,YAAM,cAAc,OAAO,OAAO,KAAK,WAAW;AAClD,YAAM,YACJ,YAAY,WAAW,IACnB,IACA,YAAY,OAAO,CAAC,GAAG,MAAM,IAAI,EAAE,WAAW,CAAC,IAAI,YAAY;AACrE,aAAO;AAAA,QACL,YAAY,KAAK;AAAA,QACjB,KAAK,KAAK;AAAA,QACV,eAAe;AAAA,QACf,YAAY,OAAO;AAAA,UACjB,OAAO,QAAQ,KAAK,WAAW,EAAE,IAAI,CAAC,CAAC,MAAM,KAAK,MAAM,CAAC,MAAM,MAAM,UAAU,CAAC;AAAA,QAClF;AAAA,QACA,cAAc,KAAK,SAAS;AAAA,MAC9B;AAAA,IACF,CAAC;AACD,UAAM,gBACJ,MAAM,WAAW,IAAI,IAAI,MAAM,OAAO,CAAC,GAAG,MAAM,IAAI,EAAE,eAAe,CAAC,IAAI,MAAM;AAClF,WAAO;AAAA,MACL;AAAA,MACA,aACE,OAAO,YAAY,WACf,WAAW,OAAO,IAClB,WAAW,KAAK,UAAU,WAAW,EAAE,CAAC;AAAA,MAC9C;AAAA,MACA;AAAA,MACA;AAAA,MACA,SAAS,SAAS,WAAW;AAAA,MAC7B;AAAA,IACF;AAAA,EACF;AAEA,QAAM,cAA2C,CAAC;AAElD,cAAY,KAAK,qBAAqB,GAAG,KAAK,iBAAiB,IAAI,kBAAkB,CAAC,CAAC;AAGvF,aAAW,OAAO,IAAI,aAAa;AACjC,UAAM,SAAS,IAAI,SAAS;AAAA,MAC1B,CAAC,MAAM,MACL,EAAE,SAAS,WAAW,gBAAgB,MACrC,SAAS,UAAa,iBAAiB,EAAE,QAAQ,IAAI,iBAAiB,KAAK,QAAQ,KAChF,IACA;AAAA,MACN,IAAI,SAAS,CAAC;AAAA,IAChB;AACA,QAAI,CAAC,OAAQ;AACb,gBAAY;AAAA,MACV,qBAAqB,IAAI,OAAO,kBAAkB,GAAG,OAAO,SAAS,OAAO,UAAU,CAAC;AAAA,IACzF;AAAA,EACF;AAEA,QAAM,QAAsB;AAAA,IAC1B,OAAO,GAAG,MAAM,IAAI,KAAK,IAAI,CAAC;AAAA,IAC9B;AAAA,IACA,YAAW,oBAAI,KAAK,GAAE,YAAY;AAAA,IAClC,QAAQ;AAAA,IACR,QAAQ,KAAK,gBAAgB,CAAC;AAAA,IAC9B,UAAU,YAAY,CAAC;AAAA,IACvB;AAAA,IACA,cAAc,QAAQ;AAAA,IACtB,aAAa,QAAQ;AAAA,IACrB,cAAc,QAAQ;AAAA,IACtB,iBAAiB,QAAQ;AAAA,IACzB,eAAe,QAAQ;AAAA,EACzB;AAEA,QAAM,OAAO,cAAc,KAAK;AAClC;AAEA,SAAS,iBACP,UACQ;AACR,QAAM,OAAO,OAAO,OAAO,SAAS,WAAW,UAAU;AACzD,SAAO,KAAK,WAAW,IAAI,IAAI,KAAK,OAAO,CAAC,GAAG,MAAM,IAAI,EAAE,eAAe,CAAC,IAAI,KAAK;AACtF;AAEA,SAAS,WAAW,GAAmB;AACrC,MAAI,IAAI,eAAe;AACvB,WAAS,IAAI,GAAG,IAAI,EAAE,QAAQ,KAAK;AACjC,SAAK,EAAE,WAAW,CAAC;AACnB,QAAI,KAAK,KAAK,GAAG,QAAQ,MAAM;AAAA,EACjC;AACA,SAAO,EAAE,SAAS,EAAE,EAAE,SAAS,GAAG,GAAG;AACvC;AAOA,SAAS,kBACP,OACA,aACA,OACa;AACb,SAAO,MAAM,IAAI,CAAC,SAAS;AACzB,UAAM,WAAmD,CAAC;AAC1D,UAAM,kBAA8D,CAAC;AACrE,QAAI,eAAe;AACnB,QAAI,iBAAiB;AACrB,eAAW,CAAC,SAAS,KAAK,KAAK,OAAO,QAAQ,KAAK,WAAW,GAAG;AAC/D,eAAS,OAAO,IAAI,EAAE,GAAG,MAAM,WAAW;AAC1C,iBAAW,CAAC,KAAK,KAAK,KAAK,OAAO,QAAQ,MAAM,UAAU,GAAG;AAC3D,YAAI,CAAC,OAAO,SAAS,KAAK,EAAG;AAC7B,cAAM,QAAQ,gBAAgB,GAAG,KAAK,EAAE,KAAK,GAAG,GAAG,EAAE;AACrD,cAAM,OAAO;AACb,cAAM,KAAK;AACX,wBAAgB,GAAG,IAAI;AAAA,MACzB;AACA,UAAI,OAAO,SAAS,MAAM,SAAS,GAAG;AACpC,wBAAgB,MAAM;AACtB,0BAAkB;AAAA,MACpB;AAAA,IACF;AACA,UAAM,aAAqC,CAAC;AAC5C,eAAW,CAAC,KAAK,EAAE,KAAK,EAAE,CAAC,KAAK,OAAO,QAAQ,eAAe,GAAG;AAC/D,iBAAW,GAAG,IAAI,MAAM,IAAI,IAAI,MAAM;AAAA,IACxC;AACA,UAAM,YAAY,mBAAmB,IAAI,IAAI,eAAe;AAC5D,UAAM,cAAiC;AAAA,MACrC;AAAA,MACA;AAAA,MACA;AAAA,IACF;AACA,WAAO;AAAA,MACL,OAAO,GAAG,KAAK,KAAK,WAAW,KAAK,KAAK,MAAM;AAAA,MAC/C,cAAc;AAAA,MACd;AAAA;AAAA;AAAA,MAGA,MACE,KAAK,MAAM,MACX,WAAW,KAAK,UAAU,EACvB,MAAM,GAAG,CAAC,EACV,MAAM,EAAE,EACR,OAAO,CAAC,GAAG,MAAO,IAAI,KAAK,EAAE,WAAW,CAAC,MAAO,GAAG,CAAC;AAAA,MACzD,OAAO;AAAA,MACP,YAAY;AAAA,MACZ,YAAY;AAAA,MACZ,WAAW;AAAA,MACX,QAAQ,KAAK;AAAA,MACb,SAAS,KAAK;AAAA,MACd,YAAY,EAAE,OAAO,GAAG,QAAQ,EAAE;AAAA,MAClC,SAAS;AAAA,QACP,cAAc;AAAA,QACd,KAAK,CAAC;AAAA,QACN;AAAA,MACF;AAAA,MACA,UAAU;AAAA,MACV,GAAI,KAAK,QAAQ,EAAE,aAAa,KAAK,MAAM,IAAI,CAAC;AAAA,IAClD;AAAA,EACF,CAAC;AACH;;;ACvgBO,SAAS,kBAAkB,MAAyD;AACzF,QAAM,EAAE,SAAS,OAAO,CAAC,GAAG,OAAO,kBAAkB,KAAK,IAAI;AAC9D,QAAM,YAAY,IAAI,IAAI,KAAK,IAAI,CAAC,MAAM,CAAC,EAAE,OAAO,CAAC,CAAC,CAAC;AAGvD,QAAM,YAAY,CAAC,WAAqC;AACtD,QAAI,OAAO,WAAW,UAAW,QAAO,SAAS,IAAI;AACrD,QAAI,CAAC,OAAO,SAAS,MAAM,EAAG,QAAO,OAAO;AAC5C,QAAI,CAAC,MAAO,QAAO;AACnB,UAAM,EAAE,KAAK,IAAI,IAAI;AACrB,QAAI,QAAQ,IAAK,QAAO;AACxB,YAAQ,SAAS,QAAQ,MAAM;AAAA,EACjC;AAGA,QAAM,QAAQ,oBAAI,IAAgC;AAClD,aAAW,OAAO,SAAS;AACzB,UAAM,OAAO,MAAM,IAAI,IAAI,KAAK,KAAK,CAAC;AACtC,SAAK,KAAK,GAAG;AACb,UAAM,IAAI,IAAI,OAAO,IAAI;AAAA,EAC3B;AAEA,QAAM,OAAoB,CAAC;AAC3B,QAAM,cAAsD,CAAC;AAE7D,aAAW,CAAC,OAAO,UAAU,KAAK,OAAO;AACvC,UAAM,aAAa,WAChB,IAAI,CAAC,OAAO,EAAE,OAAO,EAAE,OAAO,OAAO,UAAU,EAAE,MAAM,EAAE,EAAE,EAC3D,OAAO,CAAC,MAAM,OAAO,SAAS,EAAE,KAAK,CAAC;AACzC,QAAI,WAAW,WAAW,EAAG;AAE7B,UAAM,YAAY,WAAW,OAAO,CAAC,GAAG,MAAM,IAAI,EAAE,OAAO,CAAC,IAAI,WAAW;AAE3E,UAAM,UAAU,UAAU,IAAI,KAAK,KAAM,EAAE,MAAM;AAEjD,UAAM,cAAiC;AAAA,MACrC,UAAU,OAAO,YAAY,WAAW,IAAI,CAAC,MAAM,CAAC,EAAE,OAAO,EAAE,QAAQ,EAAE,MAAM,CAAC,CAAC,CAAC;AAAA,MAClF,YAAY,EAAE,QAAQ,UAAU;AAAA,MAChC,WAAW;AAAA,IACb;AAEA,UAAM,UAAsB;AAAA;AAAA;AAAA;AAAA,MAI1B,cAAc;AAAA,MACd,KAAK,OAAO,YAAY,WAAW,IAAI,CAAC,MAAM,CAAC,SAAS,EAAE,KAAK,IAAI,EAAE,KAAK,CAAC,CAAC;AAAA,MAC5E;AAAA,IACF;AAEA,SAAK,KAAK;AAAA,MACR;AAAA,MACA,cAAc,QAAQ,gBAAgB;AAAA,MACtC,aAAa,QAAQ,eAAe;AAAA,MACpC,MAAM;AAAA,MACN,OAAO,QAAQ,SAAS;AAAA,MACxB,YAAY,QAAQ,cAAc;AAAA,MAClC,YAAY,QAAQ,cAAc;AAAA,MAClC,WAAW,QAAQ,aAAa;AAAA,MAChC,QAAQ,QAAQ,UAAU;AAAA,MAC1B,SAAS,QAAQ,WAAW;AAAA,MAC5B,YAAY,EAAE,OAAO,GAAG,QAAQ,EAAE;AAAA,MAClC;AAAA,MACA,UAAU,QAAQ,YAAY;AAAA,IAChC,CAAc;AAEd,QAAI,iBAAiB;AACnB,iBAAW,KAAK,WAAY,aAAY,KAAK,EAAE,OAAO,OAAO,EAAE,OAAO,OAAO,EAAE,MAAM,CAAC;AAAA,IACxF;AAAA,EACF;AAEA,SAAO,EAAE,MAAM,YAAY;AAC7B;;;AC5HA,IAAM,aAAa,CAAC,gBAAgB,cAAc,OAAO;AACzD,IAAM,aAAa,CAAC,gBAAgB,wBAAwB,aAAa,OAAO;AAChF,IAAM,YAAY,CAAC,mBAAmB,yBAAyB,YAAY,MAAM;AACjF,IAAM,mBAAmB,CAAC,6BAA6B,oBAAoB,WAAW;AACtF,IAAM,oBAAoB,CAAC,8BAA8B,qBAAqB,YAAY;AAC1F,IAAM,mBAAmB,CAAC,sBAAsB,aAAa;AAC7D,IAAM,mBAAmB,CAAC,sBAAsB,aAAa;AAUtD,SAAS,cAAc,MAAyC;AACrE,QAAM,EAAE,OAAO,eAAe,WAAW,eAAe,cAAc,IAAI;AAC1E,QAAM,UAAU,WAAW,KAAK;AAEhC,QAAM,OAAoB,CAAC;AAC3B,aAAW,CAAC,UAAUC,WAAU,KAAK,SAAS;AAC5C,UAAM,OAAO,SAASA,WAAU;AAChC,QAAI,CAAC,KAAM;AAEX,UAAM,SAAS,KAAK,IAAI,IAAI,KAAK,kBAAkB,KAAK,qBAAqB,GAAS;AACtF,UAAM,QAAQ,eAAeA,aAAY,UAAU,KAAK;AACxD,UAAM,UAAU,eAAeA,aAAY,SAAS,KAAK;AACzD,UAAM,cAAc,eAAeA,aAAY,gBAAgB,KAAK;AACpE,UAAM,eAAe,eAAeA,aAAY,iBAAiB,KAAK;AACtE,UAAM,aAAa,eAAeA,aAAY,gBAAgB,KAAK;AACnE,UAAM,aAAa,eAAeA,aAAY,gBAAgB,KAAK;AACnE,UAAM,QAAQ,eAAeA,aAAY,UAAU;AAEnD,UAAM,aAAa,oBAAoBA,WAAU;AACjD,UAAM,aAA4B;AAAA,MAChC,OAAO;AAAA,MACP,QAAQ;AAAA,IACV;AAEA,UAAM,cACJ,UAAU,SACN;AAAA,MACE,UAAU,EAAE,gBAAgB,EAAE,MAAM,EAAE;AAAA,MACtC,YAAY,EAAE,MAAM;AAAA,MACpB,WAAW;AAAA,IACb,IACA;AAEN,UAAM,YAAYA,YAAW,KAAK,CAAC,MAAM,EAAE,QAAQ,SAAS,OAAO;AACnE,UAAM,UAAsB;AAAA,MAC1B,GAAI,KAAK,iBAAiB,WAAW,EAAE,aAAa,MAAM,IAAI,EAAE,cAAc,MAAM;AAAA,MACpF,KAAK;AAAA,MACL,GAAI,cAAc,EAAE,YAAY,IAAI,CAAC;AAAA,IACvC;AAEA,SAAK,KAAK;AAAA,MACR,OAAO;AAAA,MACP;AAAA,MACA,aAAc,KAAK,WAAW,oBAAoB,KAA4B;AAAA,MAC9E,MAAM;AAAA,MACN;AAAA,MACA;AAAA,MACA;AAAA,MACA,WAAY,KAAK,WAAW,mBAAmB,KAA4B;AAAA,MAC3E;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA,UAAU;AAAA,MACV,GAAI,YAAY,EAAE,aAAa,UAAU,KAAK,IAAI,CAAC;AAAA,IACrD,CAAc;AAAA,EAChB;AACA,SAAO;AACT;AAIA,SAAS,WAAW,OAAwD;AAC1E,QAAM,IAAI,oBAAI,IAA8B;AAC5C,aAAW,QAAQ,OAAO;AACxB,UAAM,MAAO,KAAK,cAAc,KAA4B,KAAK;AACjE,UAAM,OAAO,EAAE,IAAI,GAAG,KAAK,CAAC;AAC5B,SAAK,KAAK,IAAI;AACd,MAAE,IAAI,KAAK,IAAI;AAAA,EACjB;AACA,SAAO;AACT;AAEA,SAAS,SAAS,OAAqD;AACrE,SAAO,MAAM,KAAK,CAAC,MAAM,CAAC,EAAE,YAAY,KAAK,MAAM,CAAC;AACtD;AAEA,SAAS,eAAe,OAAyB,MAAoC;AACnF,aAAW,QAAQ,OAAO;AACxB,eAAW,OAAO,MAAM;AACtB,YAAM,IAAI,KAAK,WAAW,GAAG;AAC7B,UAAI,OAAO,MAAM,YAAY,EAAE,SAAS,EAAG,QAAO;AAAA,IACpD;AAAA,EACF;AACA,SAAO;AACT;AAEA,SAAS,eAAe,OAAyB,MAAoC;AACnF,aAAW,QAAQ,OAAO;AACxB,eAAW,OAAO,MAAM;AACtB,YAAM,IAAI,KAAK,WAAW,GAAG;AAC7B,UAAI,OAAO,MAAM,YAAY,OAAO,SAAS,CAAC,EAAG,QAAO;AACxD,UAAI,OAAO,MAAM,UAAU;AACzB,cAAM,SAAS,OAAO,CAAC;AACvB,YAAI,OAAO,SAAS,MAAM,EAAG,QAAO;AAAA,MACtC;AAAA,IACF;AAAA,EACF;AACA,SAAO;AACT;AAEA,SAAS,oBAAoB,OAAiD;AAC5E,QAAM,MAA8B,CAAC;AACrC,aAAW,QAAQ,OAAO;AACxB,eAAW,CAAC,GAAG,CAAC,KAAK,OAAO,QAAQ,KAAK,UAAU,GAAG;AACpD,UAAI,OAAO,MAAM,YAAY,OAAO,SAAS,CAAC,EAAG,KAAI,CAAC,IAAI;AAAA,IAC5D;AAAA,EACF;AACA,SAAO;AACT;","names":["mean","i","groupSpans"]}

package/dist/hosted/index.d.ts

@@ -1,4 +1,4 @@
-export { E as EvalRunCellScore, d as EvalRunEvent, e as EvalRunGenerationSnapshot, f as EvalRunStatus, g as HOSTED_WIRE_VERSION, H as HostedClient, h as HostedIngestHeaders, a as HostedTenant, i as HostedWireVersion, j as IngestEvalRunsRequest, k as IngestResponse, l as IngestTracesRequest, T as TraceSpanEvent, m as createHostedClient } from '../index-BRxz6qov.js';
+export { E as EvalRunCellScore, d as EvalRunEvent, e as EvalRunGenerationSnapshot, f as EvalRunStatus, g as HOSTED_WIRE_VERSION, H as HostedClient, h as HostedIngestHeaders, a as HostedTenant, i as HostedWireVersion, j as IngestEvalRunsRequest, k as IngestResponse, l as IngestTracesRequest, T as TraceSpanEvent, m as createHostedClient } from '../index-DQHtWQ57.js';
 import '../types-Dbj5gu8n.js';
 import '../summary-report-B7gNRX-r.js';
 import '../run-record-BGY6bHRh.js';

package/dist/{index-BRxz6qov.d.ts → index-DQHtWQ57.d.ts}

@@ -44,6 +44,14 @@ interface InsightReport {
     costQuality: {
         cost: ScalarDistribution;
         pareto: ParetoFigureSpec;
+        /** Set when the cost/quality view is degraded because the input data
+         *  doesn't fully support it — e.g. all `costUsd` were zero, or only a
+         *  single candidate appears (so the Pareto is a single point). The
+         *  named fields name the degraded sub-view, free-text the reason. */
+        degraded?: {
+            cost?: string;
+            pareto?: string;
+        };
     };
     /** Per-judge calibration + bias detection. Populated for every judge name
      *  that appears in `outcome.judgeScores`. Bias fields require either a
@@ -89,6 +97,14 @@ interface ScalarDistribution {
     max: number;
     /** Histogram bins using `agent-eval`'s `gainHistogram` primitive. */
     histogram: GainDistributionBin[];
+    /** Worst-N runs by score, ascending. Populated for the composite
+     *  distribution so the report names the runs a customer should
+     *  inspect first. Undefined when the distribution was computed from a
+     *  raw value list with no run identity (e.g. cost). */
+    tailRuns?: Array<{
+        runId: string;
+        score: number;
+    }>;
 }
 interface JudgeInsight {
     /** Number of times this judge scored a run. */

package/dist/openapi.json

@@ -2,7 +2,7 @@
   "openapi": "3.1.0",
   "info": {
     "title": "@tangle-network/agent-eval — wire protocol",
-    "version": "0.50.1",
+    "version": "0.51.0",
     "description": "HTTP and stdio RPC interface to agent-eval. The TypeScript runtime is the source of truth; this spec is the contract that cross-language clients (Python, Rust, Go) generate from.\n\nWire-protocol version: 1.0.0. Bumps on breaking changes to request/response schemas.",
     "contact": {
       "name": "Tangle Network",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tangle-network/agent-eval",
-  "version": "0.50.1",
+  "version": "0.51.0",
   "description": "Substrate for self-improving agents: traces, verifiable rewards, preferences, GEPA / reflective mutation, auto-research, replay, sequential anytime-valid stats, and release gates.",
   "homepage": "https://github.com/tangle-network/agent-eval#readme",
   "repository": {