@tangle-network/agent-eval 0.128.2 → 0.130.0

package/dist/fuzz.js

@@ -1,198 +1,207 @@
-import {
-  varianceBasedCurriculum
-} from "./chunk-VZSRQ272.js";
-import {
-  CostLedger,
-  CostReceiptCaptureError
-} from "./chunk-WS3NZZQQ.js";
-import "./chunk-VI2UW6B6.js";
-import {
-  ValidationError
-} from "./chunk-ONWEPEDO.js";
-import "./chunk-PZ5AY32C.js";
-
-// src/fuzz/cube.ts
+import { s as ValidationError } from "./errors-8YnH8WlF.js";
+import { i as CostLedger, o as CostReceiptCaptureError } from "./cost-ledger-DIgQUFZZ.js";
+import { r as varianceBasedCurriculum } from "./active-curriculum-C4mk67HP.js";
+//#region src/fuzz/cube.ts
+/** Enumerate every input cell (cartesian product of the axes), in stable order. */
 function enumerateCells(space) {
-  if (space.axes.length === 0) return [];
-  let partials = [{}];
-  for (const axis of space.axes) {
-    const next = [];
-    for (const partial of partials) {
-      for (const value of axis.values) next.push({ ...partial, [axis.name]: value });
-    }
-    partials = next;
-  }
-  return partials.map((coords) => ({ id: cellId(space, coords), coords }));
+	if (space.axes.length === 0) return [];
+	let partials = [{}];
+	for (const axis of space.axes) {
+		const next = [];
+		for (const partial of partials) for (const value of axis.values) next.push({
+			...partial,
+			[axis.name]: value
+		});
+		partials = next;
+	}
+	return partials.map((coords) => ({
+		id: cellId(space, coords),
+		coords
+	}));
 }
+/** Deterministic id for a coordinate map, e.g. `matterType=nda|difficulty=hard`. */
 function cellId(space, coords) {
-  return space.axes.map((a) => `${a.name}=${coords[a.name]}`).join("|");
+	return space.axes.map((a) => `${a.name}=${coords[a.name]}`).join("|");
 }
+/** Nearest-rank percentile on a pre-sorted ascending sample. */
 function percentile(sorted, p) {
-  const idx = Math.min(sorted.length - 1, Math.max(0, Math.ceil(p * sorted.length) - 1));
-  return sorted[idx];
+	return sorted[Math.min(sorted.length - 1, Math.max(0, Math.ceil(p * sorted.length) - 1))];
 }
+/** Summarize a sample. Throws on an empty sample — callers represent "no data"
+*  as `null`, never as a zeroed distribution. */
 function distribution(values) {
-  if (values.length === 0)
-    throw new Error("distribution: empty sample \u2014 represent missing data as null, not zeros");
-  const sorted = [...values].sort((a, b) => a - b);
-  const mean = sorted.reduce((a, b) => a + b, 0) / sorted.length;
-  return {
-    mean,
-    median: percentile(sorted, 0.5),
-    p90: percentile(sorted, 0.9),
-    min: sorted[0],
-    max: sorted[sorted.length - 1],
-    n: sorted.length
-  };
+	if (values.length === 0) throw new Error("distribution: empty sample — represent missing data as null, not zeros");
+	const sorted = [...values].sort((a, b) => a - b);
+	return {
+		mean: sorted.reduce((a, b) => a + b, 0) / sorted.length,
+		median: percentile(sorted, .5),
+		p90: percentile(sorted, .9),
+		min: sorted[0],
+		max: sorted[sorted.length - 1],
+		n: sorted.length
+	};
 }
+/**
+* Project the evaluation log into the per-input-cell coverage map. A cell with
+* no evaluations reports `score: null` (honestly uncovered), never zeros.
+*/
 function buildCoverage(cells, log, threshold) {
-  const byCell = /* @__PURE__ */ new Map();
-  for (const r of log) {
-    const arr = byCell.get(r.cell.id) ?? [];
-    arr.push(r);
-    byCell.set(r.cell.id, arr);
-  }
-  return cells.map((cell) => {
-    const recs = byCell.get(cell.id) ?? [];
-    const runs = recs.length;
-    if (runs === 0)
-      return { cell, runs: 0, score: null, findingRate: 0, dimensions: {}, latencyMs: null };
-    const score = distribution(recs.map((r) => r.ev.score));
-    const latencyMs = distribution(recs.map((r) => r.latencyMs));
-    const findingRate = recs.filter((r) => r.interest >= threshold).length / runs;
-    const dimSamples = {};
-    for (const r of recs) {
-      for (const [k, v] of Object.entries(r.ev.scores ?? {})) {
-        const samples = dimSamples[k] ?? [];
-        samples.push(v);
-        dimSamples[k] = samples;
-      }
-    }
-    const dimensions = {};
-    for (const [k, xs] of Object.entries(dimSamples)) dimensions[k] = distribution(xs);
-    const tracked = recs.filter((r) => r.costUsd !== void 0);
-    const known = tracked.filter((r) => r.costUsd !== null);
-    const cost = tracked.length > 0 ? {
-      costUsd: known.reduce((a, r) => a + r.costUsd, 0),
-      ...tracked.length > known.length ? { costUnknownRuns: tracked.length - known.length } : {}
-    } : {};
-    return { cell, runs, score, findingRate, dimensions, latencyMs, ...cost };
-  });
+	const byCell = /* @__PURE__ */ new Map();
+	for (const r of log) {
+		const arr = byCell.get(r.cell.id) ?? [];
+		arr.push(r);
+		byCell.set(r.cell.id, arr);
+	}
+	return cells.map((cell) => {
+		const recs = byCell.get(cell.id) ?? [];
+		const runs = recs.length;
+		if (runs === 0) return {
+			cell,
+			runs: 0,
+			score: null,
+			findingRate: 0,
+			dimensions: {},
+			latencyMs: null
+		};
+		const score = distribution(recs.map((r) => r.ev.score));
+		const latencyMs = distribution(recs.map((r) => r.latencyMs));
+		const findingRate = recs.filter((r) => r.interest >= threshold).length / runs;
+		const dimSamples = {};
+		for (const r of recs) for (const [k, v] of Object.entries(r.ev.scores ?? {})) {
+			const samples = dimSamples[k] ?? [];
+			samples.push(v);
+			dimSamples[k] = samples;
+		}
+		const dimensions = {};
+		for (const [k, xs] of Object.entries(dimSamples)) dimensions[k] = distribution(xs);
+		const tracked = recs.filter((r) => r.costUsd !== void 0);
+		const known = tracked.filter((r) => r.costUsd !== null);
+		return {
+			cell,
+			runs,
+			score,
+			findingRate,
+			dimensions,
+			latencyMs,
+			...tracked.length > 0 ? {
+				costUsd: known.reduce((a, r) => a + r.costUsd, 0),
+				...tracked.length > known.length ? { costUnknownRuns: tracked.length - known.length } : {}
+			} : {}
+		};
+	});
 }
-
-// src/fuzz/capsule.ts
+//#endregion
+//#region src/fuzz/capsule.ts
 function buildCapsule(input) {
-  const coverage = buildCoverage(input.cells, input.log, input.threshold);
-  const covered = coverage.filter((c) => c.runs > 0);
-  const robustness = covered.length === 0 ? null : distribution(covered.map((c) => c.score.mean));
-  const latencyMs = input.log.length === 0 ? null : distribution(input.log.map((r) => r.latencyMs));
-  const behaviorBinsObserved = input.archive.filter((e) => e.binId !== e.cell.id).length;
-  return {
-    target: input.target,
-    objective: input.objective,
-    coverage,
-    findings: [...input.findings].sort((a, b) => b.interest - a.interest),
-    archive: [...input.archive].sort((a, b) => b.interest - a.interest),
-    stats: {
-      totalRuns: input.runsUsed,
-      cellsTotal: input.cells.length,
-      cellsCovered: covered.length,
-      behaviorBinsObserved,
-      candidateFindings: input.candidateFindings,
-      verifiedFindings: input.findings.length,
-      robustness,
-      latencyMs,
-      ...input.cost ? { costUsd: input.cost.costUsd, costUnknownRuns: input.cost.costUnknownRuns } : {},
-      evalErrors: input.evalErrors,
-      ...input.stoppedEarly ? { stoppedEarly: input.stoppedEarly } : {}
-    }
-  };
+	const coverage = buildCoverage(input.cells, input.log, input.threshold);
+	const covered = coverage.filter((c) => c.runs > 0);
+	const robustness = covered.length === 0 ? null : distribution(covered.map((c) => c.score.mean));
+	const latencyMs = input.log.length === 0 ? null : distribution(input.log.map((r) => r.latencyMs));
+	const behaviorBinsObserved = input.archive.filter((e) => e.binId !== e.cell.id).length;
+	return {
+		target: input.target,
+		objective: input.objective,
+		coverage,
+		findings: [...input.findings].sort((a, b) => b.interest - a.interest),
+		archive: [...input.archive].sort((a, b) => b.interest - a.interest),
+		stats: {
+			totalRuns: input.runsUsed,
+			cellsTotal: input.cells.length,
+			cellsCovered: covered.length,
+			behaviorBinsObserved,
+			candidateFindings: input.candidateFindings,
+			verifiedFindings: input.findings.length,
+			robustness,
+			latencyMs,
+			...input.cost ? {
+				costUsd: input.cost.costUsd,
+				costUnknownRuns: input.cost.costUnknownRuns
+			} : {},
+			evalErrors: input.evalErrors,
+			...input.stoppedEarly ? { stoppedEarly: input.stoppedEarly } : {}
+		}
+	};
 }
 function esc(s) {
-  return s.replace(
-    /[&<>"]/g,
-    (c) => ({ "&": "&amp;", "<": "&lt;", ">": "&gt;", '"': "&quot;" })[c]
-  );
+	return s.replace(/[&<>"]/g, (c) => ({
+		"&": "&amp;",
+		"<": "&lt;",
+		">": "&gt;",
+		"\"": "&quot;"
+	})[c]);
 }
+/** red (0) → amber (.5) → green (1); uncovered cells render gray. */
 function robustnessColor(r) {
-  if (r == null) return "#2a2a2e";
-  return `hsl(${Math.round(r * 120)} 70% 42%)`;
+	if (r == null) return "#2a2a2e";
+	return `hsl(${Math.round(r * 120)} 70% 42%)`;
 }
 function pct(x) {
-  return `${Math.round(x * 100)}%`;
+	return `${Math.round(x * 100)}%`;
 }
 function deriveAxes(coverage) {
-  const order = [];
-  const seen = /* @__PURE__ */ new Map();
-  for (const c of coverage) {
-    for (const [k, v] of Object.entries(c.cell.coords)) {
-      if (!seen.has(k)) {
-        seen.set(k, /* @__PURE__ */ new Set());
-        order.push(k);
-      }
-      seen.get(k)?.add(v);
-    }
-  }
-  return order.map((name) => ({ name, values: [...seen.get(name) ?? []] }));
+	const order = [];
+	const seen = /* @__PURE__ */ new Map();
+	for (const c of coverage) for (const [k, v] of Object.entries(c.cell.coords)) {
+		if (!seen.has(k)) {
+			seen.set(k, /* @__PURE__ */ new Set());
+			order.push(k);
+		}
+		seen.get(k)?.add(v);
+	}
+	return order.map((name) => ({
+		name,
+		values: [...seen.get(name) ?? []]
+	}));
 }
+/** The weakest dimension chip for a cell, e.g. `safety 32%` — shown when scores exist. */
 function weakestDim(c) {
-  const entries = Object.entries(c.dimensions).map(([k, d]) => [k, d.mean]);
-  if (entries.length === 0) return "";
-  const sorted = entries.sort((a, b) => a[1] - b[1]);
-  const w = sorted[0];
-  if (!w) return "";
-  return `<span class="dim">${esc(w[0])} ${pct(w[1])}</span>`;
+	const entries = Object.entries(c.dimensions).map(([k, d]) => [k, d.mean]);
+	if (entries.length === 0) return "";
+	const w = entries.sort((a, b) => a[1] - b[1])[0];
+	if (!w) return "";
+	return `<span class="dim">${esc(w[0])} ${pct(w[1])}</span>`;
 }
 function heatmapHtml(coverage) {
-  const axes = deriveAxes(coverage);
-  const byId = new Map(coverage.map((c) => [c.cell.id, c]));
-  const tile = (c, label) => {
-    const r = c?.score?.mean ?? null;
-    const title = c?.score ? `${pct(c.score.mean)} robust (median ${pct(c.score.median)}, min ${pct(c.score.min)}) \xB7 ${c.runs} runs \xB7 ${pct(c.findingRate)} flagged${c.latencyMs ? ` \xB7 ${(c.latencyMs.median / 1e3).toFixed(1)}s median` : ""}${c.costUsd !== void 0 ? ` \xB7 $${c.costUsd.toFixed(2)}` : ""}` : "not covered";
-    return `<div class="tile" style="background:${robustnessColor(r)}" title="${esc(title)}">${label ? `<span class="tl">${esc(label)}</span>` : ""}<span class="tv">${c && r != null ? pct(r) : "\u2014"}</span>${c ? weakestDim(c) : ""}</div>`;
-  };
-  const rowAxis = axes[0];
-  const colAxis = axes[1];
-  if (axes.length === 2 && rowAxis && colAxis) {
-    const head = `<tr><th></th>${colAxis.values.map((v) => `<th>${esc(v)}</th>`).join("")}</tr>`;
-    const rows = rowAxis.values.map((rv) => {
-      const cells = colAxis.values.map((cv) => {
-        const id = `${rowAxis.name}=${rv}|${colAxis.name}=${cv}`;
-        return `<td>${tile(byId.get(id), "")}</td>`;
-      }).join("");
-      return `<tr><th class="rh">${esc(rv)}</th>${cells}</tr>`;
-    }).join("");
-    return `<div class="axis-label">rows: <b>${esc(rowAxis.name)}</b> \xB7 cols: <b>${esc(colAxis.name)}</b></div><table class="heat">${head}${rows}</table>`;
-  }
-  const sorted = [...coverage].sort((a, b) => (a.score?.mean ?? 2) - (b.score?.mean ?? 2));
-  return `<div class="grid">${sorted.map((c) => tile(c, Object.values(c.cell.coords).join(" \xB7 "))).join("")}</div>`;
+	const axes = deriveAxes(coverage);
+	const byId = new Map(coverage.map((c) => [c.cell.id, c]));
+	const tile = (c, label) => {
+		const r = c?.score?.mean ?? null;
+		const title = c?.score ? `${pct(c.score.mean)} robust (median ${pct(c.score.median)}, min ${pct(c.score.min)}) · ${c.runs} runs · ${pct(c.findingRate)} flagged${c.latencyMs ? ` · ${(c.latencyMs.median / 1e3).toFixed(1)}s median` : ""}${c.costUsd !== void 0 ? ` · $${c.costUsd.toFixed(2)}` : ""}` : "not covered";
+		return `<div class="tile" style="background:${robustnessColor(r)}" title="${esc(title)}">${label ? `<span class="tl">${esc(label)}</span>` : ""}<span class="tv">${c && r != null ? pct(r) : "—"}</span>${c ? weakestDim(c) : ""}</div>`;
+	};
+	const rowAxis = axes[0];
+	const colAxis = axes[1];
+	if (axes.length === 2 && rowAxis && colAxis) {
+		const head = `<tr><th></th>${colAxis.values.map((v) => `<th>${esc(v)}</th>`).join("")}</tr>`;
+		const rows = rowAxis.values.map((rv) => {
+			const cells = colAxis.values.map((cv) => {
+				const id = `${rowAxis.name}=${rv}|${colAxis.name}=${cv}`;
+				return `<td>${tile(byId.get(id), "")}</td>`;
+			}).join("");
+			return `<tr><th class="rh">${esc(rv)}</th>${cells}</tr>`;
+		}).join("");
+		return `<div class="axis-label">rows: <b>${esc(rowAxis.name)}</b> · cols: <b>${esc(colAxis.name)}</b></div><table class="heat">${head}${rows}</table>`;
+	}
+	return `<div class="grid">${[...coverage].sort((a, b) => (a.score?.mean ?? 2) - (b.score?.mean ?? 2)).map((c) => tile(c, Object.values(c.cell.coords).join(" · "))).join("")}</div>`;
 }
 function findingsHtml(findings, limit) {
-  if (findings.length === 0)
-    return `<p class="none">No verified findings \u2014 the target held across every covered cell.</p>`;
-  return findings.slice(0, limit).map((f) => {
-    const coords = Object.entries(f.cell.coords).map(([k, v]) => `${k}:${v}`).join(" \xB7 ");
-    const labels = (f.evaluation.labels ?? []).map((l) => `<span class="fclass">${esc(l)}</span>`).join("");
-    const dims = Object.entries(f.evaluation.scores ?? {}).sort((a, b) => a[1] - b[1]).slice(0, 3).map(([k, v]) => `<span class="dim">${esc(k)} ${pct(v)}</span>`).join("");
-    return `<div class="fail"><div class="fmeta"><span class="fcell">${esc(coords)}</span>${labels}<span class="fsev">interest ${pct(f.interest)}</span></div><div class="sevbar"><div class="sevfill" style="width:${pct(f.interest)}"></div></div><div class="ftext">${esc(f.text ?? "(scenario text not captured)")}</div>${dims ? `<div class="fdims">${dims}</div>` : ""}</div>`;
-  }).join("");
+	if (findings.length === 0) return `<p class="none">No verified findings — the target held across every covered cell.</p>`;
+	return findings.slice(0, limit).map((f) => {
+		const coords = Object.entries(f.cell.coords).map(([k, v]) => `${k}:${v}`).join(" · ");
+		const labels = (f.evaluation.labels ?? []).map((l) => `<span class="fclass">${esc(l)}</span>`).join("");
+		const dims = Object.entries(f.evaluation.scores ?? {}).sort((a, b) => a[1] - b[1]).slice(0, 3).map(([k, v]) => `<span class="dim">${esc(k)} ${pct(v)}</span>`).join("");
+		return `<div class="fail"><div class="fmeta"><span class="fcell">${esc(coords)}</span>${labels}<span class="fsev">interest ${pct(f.interest)}</span></div><div class="sevbar"><div class="sevfill" style="width:${pct(f.interest)}"></div></div><div class="ftext">${esc(f.text ?? "(scenario text not captured)")}</div>${dims ? `<div class="fdims">${dims}</div>` : ""}</div>`;
+	}).join("");
 }
+/** Render a self-contained HTML capsule — heat-map + per-dimension chips + verified findings. */
 function renderCapsuleHtml(capsule, opts = {}) {
-  const s = capsule.stats;
-  const kpi = (label, value, accent = "#e6e6e6") => `<div class="kpi"><div class="kv" style="color:${accent}">${esc(value)}</div><div class="kl">${esc(label)}</div></div>`;
-  const lift = capsule.lift ? kpi(
-    "hardening lift",
-    `${capsule.lift.before.toFixed(2)} \u2192 ${capsule.lift.after.toFixed(2)}`,
-    "#5ad17a"
-  ) : "";
-  const cost = s.costUsd !== void 0 ? kpi(
-    s.costUnknownRuns ? `cost \xB7 ${s.costUnknownRuns} runs unpriced` : "cost",
-    `$${s.costUsd.toFixed(2)}`,
-    s.costUnknownRuns ? "#e5c07b" : "#e6e6e6"
-  ) : "";
-  const stamp = opts.generatedAt ?? capsule.generatedAt ?? "";
-  return `<!doctype html><html><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
-<title>${esc(capsule.objective)} capsule \u2014 ${esc(capsule.target)}</title>
+	const s = capsule.stats;
+	const kpi = (label, value, accent = "#e6e6e6") => `<div class="kpi"><div class="kv" style="color:${accent}">${esc(value)}</div><div class="kl">${esc(label)}</div></div>`;
+	const lift = capsule.lift ? kpi("hardening lift", `${capsule.lift.before.toFixed(2)} → ${capsule.lift.after.toFixed(2)}`, "#5ad17a") : "";
+	const cost = s.costUsd !== void 0 ? kpi(s.costUnknownRuns ? `cost · ${s.costUnknownRuns} runs unpriced` : "cost", `$${s.costUsd.toFixed(2)}`, s.costUnknownRuns ? "#e5c07b" : "#e6e6e6") : "";
+	const stamp = opts.generatedAt ?? capsule.generatedAt ?? "";
+	return `<!doctype html><html><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
+<title>${esc(capsule.objective)} capsule — ${esc(capsule.target)}</title>
 <style>
 :root{color-scheme:dark}
 *{box-sizing:border-box}
@@ -228,11 +237,11 @@ table.heat th.rh{text-align:right}
 .none{color:#5ad17a}
 .foot{color:#5a5a63;font-size:11px;margin-top:36px}
 </style></head><body><div class="wrap">
-<h1>${esc(capsule.objective)} exploration \xB7 ${esc(capsule.target)}</h1>
-<div class="sub">${s.totalRuns} scenarios across ${s.cellsCovered}/${s.cellsTotal} planned cells${s.behaviorBinsObserved > 0 ? ` \xB7 ${s.behaviorBinsObserved} measured behavior bins` : ""}${stamp ? ` \xB7 ${esc(stamp)}` : ""}</div>
+<h1>${esc(capsule.objective)} exploration · ${esc(capsule.target)}</h1>
+<div class="sub">${s.totalRuns} scenarios across ${s.cellsCovered}/${s.cellsTotal} planned cells${s.behaviorBinsObserved > 0 ? ` · ${s.behaviorBinsObserved} measured behavior bins` : ""}${stamp ? ` · ${esc(stamp)}` : ""}</div>
 <div class="kpis">
-${s.robustness ? kpi("robustness", `${pct(s.robustness.mean)}`, s.robustness.mean < 0.6 ? "#e58a96" : "#5ad17a") : ""}
-${s.robustness ? kpi("cell spread", `${pct(s.robustness.min)}\u2013${pct(s.robustness.max)}`) : ""}
+${s.robustness ? kpi("robustness", `${pct(s.robustness.mean)}`, s.robustness.mean < .6 ? "#e58a96" : "#5ad17a") : ""}
+${s.robustness ? kpi("cell spread", `${pct(s.robustness.min)}–${pct(s.robustness.max)}`) : ""}
 ${s.latencyMs ? kpi("median latency", `${(s.latencyMs.median / 1e3).toFixed(1)}s`, s.latencyMs.p90 > 4 * s.latencyMs.median ? "#e5b566" : "#e6e6e6") : ""}
 ${kpi("verified findings", String(s.verifiedFindings), s.verifiedFindings > 0 ? "#e58a96" : "#5ad17a")}
 ${kpi("cells covered", `${s.cellsCovered}/${s.cellsTotal}`)}
@@ -241,475 +250,519 @@ ${cost}
 ${s.evalErrors > 0 ? kpi("eval errors", String(s.evalErrors), "#e5b566") : ""}
 ${lift}
 </div>
-${s.stoppedEarly ? `<div class="sub" style="color:#e5b566">stopped early: ${esc(s.stoppedEarly.detail)} \u2014 coverage below reflects the completed portion only</div>` : ""}
+${s.stoppedEarly ? `<div class="sub" style="color:#e5b566">stopped early: ${esc(s.stoppedEarly.detail)} — coverage below reflects the completed portion only</div>` : ""}
 <h2>Coverage map</h2>
 ${heatmapHtml(capsule.coverage)}
-<h2>Verified findings${s.candidateFindings > s.verifiedFindings ? ` \xB7 ${s.verifiedFindings} of ${s.candidateFindings} candidates passed the validity gates` : ""}</h2>
+<h2>Verified findings${s.candidateFindings > s.verifiedFindings ? ` · ${s.verifiedFindings} of ${s.candidateFindings} candidates passed the validity gates` : ""}</h2>
 ${findingsHtml(capsule.findings, opts.maxFindings ?? 8)}
 <div class="foot">Findings are gate-verified: each is a fair, answerable task that reproduces under a meaning-preserving rephrase, minimized to the smallest trigger.</div>
 </div></body></html>`;
 }
-
-// src/fuzz/policies.ts
-var clamp01 = (x) => Math.max(0, Math.min(1, x));
+//#endregion
+//#region src/fuzz/policies.ts
+const clamp01 = (x) => Math.max(0, Math.min(1, x));
+/**
+* Perturbation-based search: apply the cell's mutation operators to the current
+* elites + seeds, deduping by id. Elites first — mutating the most interesting
+* scenario found so far is what makes the search deepen across rounds.
+*/
 function mutationProposer(opts) {
-  return async (ctx) => {
-    const mutations = opts.mutationsFor(ctx.cell);
-    const parents = [...ctx.elites, ...ctx.seeds];
-    const seen = new Set(parents.map(opts.scenarioId));
-    const out = [];
-    for (const parent of parents) {
-      if (out.length >= ctx.count) break;
-      for (const m of mutations) {
-        const children = await m.mutate(parent, ctx.rng);
-        for (const child of children) {
-          const id = opts.scenarioId(child);
-          if (seen.has(id)) continue;
-          seen.add(id);
-          out.push(child);
-          if (out.length >= ctx.count) break;
-        }
-        if (out.length >= ctx.count) break;
-      }
-    }
-    return out;
-  };
+	return async (ctx) => {
+		const mutations = opts.mutationsFor(ctx.cell);
+		const parents = [...ctx.elites, ...ctx.seeds];
+		const seen = new Set(parents.map(opts.scenarioId));
+		const out = [];
+		for (const parent of parents) {
+			if (out.length >= ctx.count) break;
+			for (const m of mutations) {
+				const children = await m.mutate(parent, ctx.rng);
+				for (const child of children) {
+					const id = opts.scenarioId(child);
+					if (seen.has(id)) continue;
+					seen.add(id);
+					out.push(child);
+					if (out.length >= ctx.count) break;
+				}
+				if (out.length >= ctx.count) break;
+			}
+		}
+		return out;
+	};
 }
-function adversarialObjective(threshold = 0.5) {
-  return { kind: "adversarial", threshold, interest: (ev) => clamp01(1 - ev.score) };
+/** Adversarial: a low headline score is interesting — find where the agent fails. */
+function adversarialObjective(threshold = .5) {
+	return {
+		kind: "adversarial",
+		threshold,
+		interest: (ev) => clamp01(1 - ev.score)
+	};
 }
 function hamming(a, b) {
-  if (!a || !b) return 1;
-  const keys = /* @__PURE__ */ new Set([...Object.keys(a), ...Object.keys(b)]);
-  if (keys.size === 0) return 0;
-  let diff = 0;
-  for (const k of keys) if (a[k] !== b[k]) diff++;
-  return diff / keys.size;
+	if (!a || !b) return 1;
+	const keys = /* @__PURE__ */ new Set([...Object.keys(a), ...Object.keys(b)]);
+	if (keys.size === 0) return 0;
+	let diff = 0;
+	for (const k of keys) if (a[k] !== b[k]) diff++;
+	return diff / keys.size;
 }
-function noveltyObjective(threshold = 0.3) {
-  return {
-    kind: "novelty",
-    threshold,
-    interest: (ev, ctx) => {
-      const scoreNovelty = ctx.archiveScores.length === 0 ? 1 : Math.min(...ctx.archiveScores.map((s) => Math.abs(s - ev.score)));
-      const descNovelty = ctx.archiveDescriptors.length === 0 ? 1 : Math.min(...ctx.archiveDescriptors.map((d) => hamming(d, ev.descriptor)));
-      return clamp01(0.5 * scoreNovelty + 0.5 * descNovelty);
-    }
-  };
+/**
+* Novelty: interesting when far from the archive in score AND measured behavior
+* descriptor — quality-diversity's diversity pressure; drives corpus growth
+* rather than re-finding the same hole.
+*/
+function noveltyObjective(threshold = .3) {
+	return {
+		kind: "novelty",
+		threshold,
+		interest: (ev, ctx) => {
+			const scoreNovelty = ctx.archiveScores.length === 0 ? 1 : Math.min(...ctx.archiveScores.map((s) => Math.abs(s - ev.score)));
+			const descNovelty = ctx.archiveDescriptors.length === 0 ? 1 : Math.min(...ctx.archiveDescriptors.map((d) => hamming(d, ev.descriptor)));
+			return clamp01(.5 * scoreNovelty + .5 * descNovelty);
+		}
+	};
 }
-
-// src/fuzz/explorer.ts
+//#endregion
+//#region src/fuzz/explorer.ts
+/**
+* The exploration engine — a stateful session over a behavior space.
+*
+* Each `step()`: allocate budget across INPUT cells (floor first, then variance
+* steering toward the least-certain cells), propose candidates (the proposer
+* reads current elites + findings, so the search deepens generationally),
+* evaluate with bounded concurrency, archive the most interesting scenario per
+* input×measured bin, and admit notable candidates that pass the validity gates.
+* `run()` loops to budget. `coverage()`/`findings()`/`capsule()` read live state —
+* the surface `makeExploreTools` exposes so an agent can drive the session.
+*
+* One evaluation log (`EvalRecord[]`) is the source of truth; allocation
+* observations and coverage are projections of it.
+*/
 async function pMap(items, fn, concurrency, signal) {
-  let i = 0;
-  const workers = Array.from(
-    { length: Math.max(1, Math.min(concurrency, items.length)) },
-    async () => {
-      while (i < items.length) {
-        if (signal?.aborted) return;
-        const item = items[i++];
-        if (item === void 0) return;
-        await fn(item);
-      }
-    }
-  );
-  await Promise.all(workers);
+	let i = 0;
+	const workers = Array.from({ length: Math.max(1, Math.min(concurrency, items.length)) }, async () => {
+		while (i < items.length) {
+			if (signal?.aborted) return;
+			const item = items[i++];
+			if (item === void 0) return;
+			await fn(item);
+		}
+	});
+	await Promise.all(workers);
 }
 var BehaviorExplorer = class {
-  constructor(opts) {
-    this.opts = opts;
-    this.cells = enumerateCells(opts.space);
-    if (this.cells.length === 0)
-      throw new Error("BehaviorExplorer: space has no cells \u2014 every axis needs \u22651 value");
-    if (opts.costBudgetUsd !== void 0) {
-      if (typeof opts.costBudgetUsd !== "number" || !Number.isFinite(opts.costBudgetUsd) || opts.costBudgetUsd < 0) {
-        throw new RangeError(
-          `BehaviorExplorer: costBudgetUsd must be a nonnegative finite number, got ${String(opts.costBudgetUsd)}`
-        );
-      }
-    }
-    if (!opts.costOf && (opts.costBudgetUsd !== void 0 || opts.ledger || opts.onCost)) {
-      throw new ValidationError(
-        "BehaviorExplorer: costBudgetUsd/ledger/onCost require costOf \u2014 the explorer cannot know run cost without it; supply costOf or drop the cost options"
-      );
-    }
-    if ((opts.costBudgetUsd !== void 0 || opts.ledger?.costCeilingUsd !== void 0) && !opts.maximumChargeOf) {
-      throw new ValidationError(
-        "BehaviorExplorer: capped cost tracking requires maximumChargeOf before evaluation"
-      );
-    }
-    if (opts.ledger && opts.costBudgetUsd !== void 0 && opts.ledger.costCeilingUsd !== opts.costBudgetUsd) {
-      throw new ValidationError(
-        "BehaviorExplorer: costBudgetUsd must match the supplied CostLedger ceiling"
-      );
-    }
-    if (opts.costOf) this.costLedger = opts.ledger ?? new CostLedger(opts.costBudgetUsd);
-    this.cellById = new Map(this.cells.map((c) => [c.id, c]));
-    this.objective = opts.objective ?? adversarialObjective(0.5);
-    this.threshold = this.objective.threshold ?? 0.5;
-    this.floorPerCell = opts.floorPerCell ?? 2;
-    this.perRoundBudget = Math.max(
-      this.cells.length * this.floorPerCell,
-      Math.ceil(opts.budget / 4)
-    );
-    this.rngState = (opts.seed ?? 1) >>> 0;
-  }
-  opts;
-  cells;
-  cellById;
-  objective;
-  threshold;
-  floorPerCell;
-  perRoundBudget;
-  /** The single evaluation log — coverage + allocation are projections of it. */
-  log = [];
-  /** binId (input × measured coords) → the most interesting entry seen. */
-  archiveByBin = /* @__PURE__ */ new Map();
-  _findings = [];
-  runsUsed = 0;
-  candidateFindings = 0;
-  evalErrors = 0;
-  consecutiveEvalErrors = 0;
-  stoppedEarly;
-  rngState;
-  costLedger;
-  costPhase = "fuzz.explore";
-  // mulberry32 — deterministic per session.
-  rng = () => {
-    this.rngState = this.rngState + 1831565813 | 0;
-    let t = this.rngState;
-    t = Math.imul(t ^ t >>> 15, t | 1);
-    t ^= t + Math.imul(t ^ t >>> 7, t | 61);
-    return ((t ^ t >>> 14) >>> 0) / 4294967296;
-  };
-  binId(cell, descriptor) {
-    if (!descriptor || Object.keys(descriptor).length === 0) return cell.id;
-    const measured = Object.keys(descriptor).sort().map((k) => `${k}=${descriptor[k]}`).join("|");
-    return `${cell.id}|${measured}`;
-  }
-  allocate(budget) {
-    if ((this.opts.allocation ?? "variance") === "uniform") {
-      const per = Math.max(this.floorPerCell, Math.floor(budget / this.cells.length));
-      return this.cells.map((c) => ({ cellId: c.id, count: per }));
-    }
-    return varianceBasedCurriculum(
-      this.log.map((r) => ({
-        variantId: r.cell.id,
-        scenarioId: r.scenarioId,
-        score: r.ev.score,
-        pass: r.ev.valid && r.ev.score >= 0.5
-      })),
-      this.cells.map((c) => ({ variantId: c.id, scenarioId: "*" })),
-      { budget, floorPerCell: this.floorPerCell }
-    ).map((a) => ({ cellId: a.variantId, count: a.count }));
-  }
-  objectiveContext() {
-    const entries = [...this.archiveByBin.values()];
-    return {
-      archiveScores: entries.map((e) => e.evaluation.score),
-      archiveDescriptors: entries.map((e) => e.evaluation.descriptor)
-    };
-  }
-  /** Elites whose INPUT cell matches — what the proposer mutates/deepens from. */
-  elitesFor(cellId2) {
-    const out = [];
-    for (const e of this.archiveByBin.values()) if (e.cell.id === cellId2) out.push(e.scenario);
-    return out;
-  }
-  /** One allocate → propose → evaluate → gate → archive round. */
-  async step() {
-    const remaining = this.opts.budget - this.runsUsed;
-    if (remaining <= 0 || this.opts.signal?.aborted) return { runs: 0, findings: [] };
-    const allocations = this.allocate(Math.min(this.perRoundBudget, remaining));
-    const newFindings = [];
-    let runsThisStep = 0;
-    for (const alloc of allocations) {
-      if (this.runsUsed >= this.opts.budget || this.stoppedEarly !== void 0 || this.opts.signal?.aborted)
-        break;
-      const cell = this.cellById.get(alloc.cellId);
-      if (!cell) continue;
-      const cap = Math.min(alloc.count, this.opts.budget - this.runsUsed);
-      if (cap <= 0) continue;
-      this.opts.onProgress?.({ type: "cell-allocated", cell, count: cap });
-      const seeds = await this.opts.seedsFor(cell);
-      const elites = this.elitesFor(cell.id);
-      const proposed = await this.opts.proposer({
-        cell,
-        seeds,
-        elites,
-        findings: this._findings,
-        count: cap,
-        rng: this.rng
-      });
-      const cold = this.log.every((r) => r.cell.id !== cell.id);
-      const toEval = [...cold ? seeds : [], ...proposed].slice(0, cap);
-      await pMap(
-        toEval,
-        async (scenario) => {
-          if (this.runsUsed >= this.opts.budget || this.stoppedEarly !== void 0 || this.opts.signal?.aborted)
-            return;
-          try {
-            const startedAt = performance.now();
-            const paid = this.costLedger ? await this.costLedger.runPaidCall({
-              channel: "agent",
-              phase: this.costPhase,
-              actor: "fuzz.evaluate",
-              tags: { target: this.opts.target, cell: cell.id },
-              signal: this.opts.signal,
-              maximumCharge: this.opts.maximumChargeOf?.(scenario, cell),
-              execute: () => this.opts.evaluate(scenario, cell),
-              receipt: (evaluation) => {
-                const cost = this.opts.costOf(scenario, cell, evaluation);
-                if (cost === null) {
-                  return {
-                    model: "unattributed",
-                    inputTokens: 0,
-                    outputTokens: 0,
-                    costUnknown: true
-                  };
-                }
-                if (!Number.isFinite(cost.usd) || cost.usd < 0) {
-                  throw new RangeError(
-                    `BehaviorExplorer: costOf returned an invalid usd (${String(cost.usd)}) \u2014 return null when cost is unknown, never a fabricated number`
-                  );
-                }
-                return {
-                  model: cost.model ?? "unattributed",
-                  inputTokens: 0,
-                  outputTokens: 0,
-                  actualCostUsd: cost.usd
-                };
-              }
-            }) : void 0;
-            if (paid && !paid.succeeded) {
-              if (paid.error instanceof CostReceiptCaptureError && paid.error.receiptError instanceof RangeError) {
-                throw paid.error.receiptError;
-              }
-              throw paid.error;
-            }
-            const ev = paid ? paid.value : await this.opts.evaluate(scenario, cell);
-            const latencyMs = ev.latencyMs ?? performance.now() - startedAt;
-            this.runsUsed++;
-            runsThisStep++;
-            this.consecutiveEvalErrors = 0;
-            const costUsd = paid ? paid.receipt.costUnknown ? null : paid.receipt.costUsd : void 0;
-            if (paid?.receipt.actualCostUsd !== void 0) {
-              this.opts.onCost?.({ usd: paid.receipt.actualCostUsd, channel: "agent" });
-            }
-            const interest = this.objective.interest(ev, this.objectiveContext());
-            this.log.push({
-              cell,
-              ev,
-              interest,
-              latencyMs,
-              ...costUsd !== void 0 ? { costUsd } : {},
-              scenarioId: this.opts.scenarioId(scenario)
-            });
-            this.opts.onProgress?.({ type: "evaluated", cell, scenario, evaluation: ev });
-            const bin = this.binId(cell, ev.descriptor);
-            const cur = this.archiveByBin.get(bin);
-            if (!cur || interest > cur.interest)
-              this.archiveByBin.set(bin, { binId: bin, cell, scenario, evaluation: ev, interest });
-            if (interest < this.threshold) return;
-            this.candidateFindings++;
-            if (this.opts.gates?.isValid && !await this.opts.gates.isValid(scenario, ev, cell))
-              return;
-            if (this.opts.gates?.isUncontaminated && !await this.opts.gates.isUncontaminated(scenario, ev, cell))
-              return;
-            const minimized = this.opts.minimize ? await this.opts.minimize(scenario, this.opts.evaluate, cell) : scenario;
-            const finding = {
-              id: this.opts.scenarioId(scenario),
-              cell,
-              scenario,
-              minimized,
-              text: this.opts.scenarioText?.(minimized),
-              evaluation: ev,
-              interest,
-              objective: this.objective.kind
-            };
-            this._findings.push(finding);
-            newFindings.push(finding);
-            this.opts.onProgress?.({ type: "finding", finding });
-          } catch (err) {
-            if (err instanceof RangeError) throw err;
-            this.evalErrors++;
-            this.consecutiveEvalErrors++;
-            const message = err instanceof Error ? err.message : String(err);
-            this.opts.onProgress?.({
-              type: "eval-error",
-              cell,
-              scenarioId: this.opts.scenarioId(scenario),
-              message
-            });
-            const limit = this.opts.maxConsecutiveEvalErrors ?? 5;
-            if (this.consecutiveEvalErrors >= limit) {
-              this.stoppedEarly = {
-                reason: "eval-errors",
-                detail: `${this.consecutiveEvalErrors} consecutive eval errors (last: ${message})`
-              };
-            }
-          }
-        },
-        this.opts.concurrency ?? 1,
-        this.opts.signal
-      );
-    }
-    this.opts.onProgress?.({ type: "round", runsUsed: this.runsUsed, budget: this.opts.budget });
-    return { runs: runsThisStep, findings: newFindings };
-  }
-  /** Loop `step()` until the run or dollar budget is spent, the signal aborts,
-   *  or no progress is made. */
-  async run() {
-    while (this.runsUsed < this.opts.budget && this.stoppedEarly === void 0 && !this.opts.signal?.aborted) {
-      const { runs } = await this.step();
-      if (runs === 0 && this.stoppedEarly === void 0) break;
-    }
-    return this.capsule();
-  }
-  coverage() {
-    return this.capsule().coverage;
-  }
-  findings() {
-    return [...this._findings].sort((a, b) => b.interest - a.interest);
-  }
-  capsule() {
-    const cost = this.costLedger?.summary({
-      phase: this.costPhase,
-      tags: { target: this.opts.target }
-    });
-    return buildCapsule({
-      target: this.opts.target,
-      objective: this.objective.kind,
-      cells: this.cells,
-      log: this.log,
-      threshold: this.threshold,
-      archive: [...this.archiveByBin.values()],
-      findings: this._findings,
-      candidateFindings: this.candidateFindings,
-      runsUsed: this.runsUsed,
-      cost: cost ? {
-        costUsd: cost.totalCostUsd,
-        costUnknownRuns: cost.byChannel.reduce((sum, row) => sum + row.unpricedCalls, 0)
-      } : void 0,
-      evalErrors: this.evalErrors,
-      stoppedEarly: this.stoppedEarly
-    });
-  }
+	opts;
+	cells;
+	cellById;
+	objective;
+	threshold;
+	floorPerCell;
+	perRoundBudget;
+	/** The single evaluation log — coverage + allocation are projections of it. */
+	log = [];
+	/** binId (input × measured coords) → the most interesting entry seen. */
+	archiveByBin = /* @__PURE__ */ new Map();
+	_findings = [];
+	runsUsed = 0;
+	candidateFindings = 0;
+	evalErrors = 0;
+	consecutiveEvalErrors = 0;
+	stoppedEarly;
+	rngState;
+	costLedger;
+	costPhase = "fuzz.explore";
+	constructor(opts) {
+		this.opts = opts;
+		this.cells = enumerateCells(opts.space);
+		if (this.cells.length === 0) throw new Error("BehaviorExplorer: space has no cells — every axis needs ≥1 value");
+		if (opts.costBudgetUsd !== void 0) {
+			if (typeof opts.costBudgetUsd !== "number" || !Number.isFinite(opts.costBudgetUsd) || opts.costBudgetUsd < 0) throw new RangeError(`BehaviorExplorer: costBudgetUsd must be a nonnegative finite number, got ${String(opts.costBudgetUsd)}`);
+		}
+		if (!opts.costOf && (opts.costBudgetUsd !== void 0 || opts.ledger || opts.onCost)) throw new ValidationError("BehaviorExplorer: costBudgetUsd/ledger/onCost require costOf — the explorer cannot know run cost without it; supply costOf or drop the cost options");
+		if ((opts.costBudgetUsd !== void 0 || opts.ledger?.costCeilingUsd !== void 0) && !opts.maximumChargeOf) throw new ValidationError("BehaviorExplorer: capped cost tracking requires maximumChargeOf before evaluation");
+		if (opts.ledger && opts.costBudgetUsd !== void 0 && opts.ledger.costCeilingUsd !== opts.costBudgetUsd) throw new ValidationError("BehaviorExplorer: costBudgetUsd must match the supplied CostLedger ceiling");
+		if (opts.costOf) this.costLedger = opts.ledger ?? new CostLedger(opts.costBudgetUsd);
+		this.cellById = new Map(this.cells.map((c) => [c.id, c]));
+		this.objective = opts.objective ?? adversarialObjective(.5);
+		this.threshold = this.objective.threshold ?? .5;
+		this.floorPerCell = opts.floorPerCell ?? 2;
+		this.perRoundBudget = Math.max(this.cells.length * this.floorPerCell, Math.ceil(opts.budget / 4));
+		this.rngState = (opts.seed ?? 1) >>> 0;
+	}
+	rng = () => {
+		this.rngState = this.rngState + 1831565813 | 0;
+		let t = this.rngState;
+		t = Math.imul(t ^ t >>> 15, t | 1);
+		t ^= t + Math.imul(t ^ t >>> 7, t | 61);
+		return ((t ^ t >>> 14) >>> 0) / 4294967296;
+	};
+	binId(cell, descriptor) {
+		if (!descriptor || Object.keys(descriptor).length === 0) return cell.id;
+		const measured = Object.keys(descriptor).sort().map((k) => `${k}=${descriptor[k]}`).join("|");
+		return `${cell.id}|${measured}`;
+	}
+	allocate(budget) {
+		if ((this.opts.allocation ?? "variance") === "uniform") {
+			const per = Math.max(this.floorPerCell, Math.floor(budget / this.cells.length));
+			return this.cells.map((c) => ({
+				cellId: c.id,
+				count: per
+			}));
+		}
+		return varianceBasedCurriculum(this.log.map((r) => ({
+			variantId: r.cell.id,
+			scenarioId: r.scenarioId,
+			score: r.ev.score,
+			pass: r.ev.valid && r.ev.score >= .5
+		})), this.cells.map((c) => ({
+			variantId: c.id,
+			scenarioId: "*"
+		})), {
+			budget,
+			floorPerCell: this.floorPerCell
+		}).map((a) => ({
+			cellId: a.variantId,
+			count: a.count
+		}));
+	}
+	objectiveContext() {
+		const entries = [...this.archiveByBin.values()];
+		return {
+			archiveScores: entries.map((e) => e.evaluation.score),
+			archiveDescriptors: entries.map((e) => e.evaluation.descriptor)
+		};
+	}
+	/** Elites whose INPUT cell matches — what the proposer mutates/deepens from. */
+	elitesFor(cellId) {
+		const out = [];
+		for (const e of this.archiveByBin.values()) if (e.cell.id === cellId) out.push(e.scenario);
+		return out;
+	}
+	/** One allocate → propose → evaluate → gate → archive round. */
+	async step() {
+		const remaining = this.opts.budget - this.runsUsed;
+		if (remaining <= 0 || this.opts.signal?.aborted) return {
+			runs: 0,
+			findings: []
+		};
+		const allocations = this.allocate(Math.min(this.perRoundBudget, remaining));
+		const newFindings = [];
+		let runsThisStep = 0;
+		for (const alloc of allocations) {
+			if (this.runsUsed >= this.opts.budget || this.stoppedEarly !== void 0 || this.opts.signal?.aborted) break;
+			const cell = this.cellById.get(alloc.cellId);
+			if (!cell) continue;
+			const cap = Math.min(alloc.count, this.opts.budget - this.runsUsed);
+			if (cap <= 0) continue;
+			this.opts.onProgress?.({
+				type: "cell-allocated",
+				cell,
+				count: cap
+			});
+			const seeds = await this.opts.seedsFor(cell);
+			const elites = this.elitesFor(cell.id);
+			const proposed = await this.opts.proposer({
+				cell,
+				seeds,
+				elites,
+				findings: this._findings,
+				count: cap,
+				rng: this.rng
+			});
+			await pMap([...this.log.every((r) => r.cell.id !== cell.id) ? seeds : [], ...proposed].slice(0, cap), async (scenario) => {
+				if (this.runsUsed >= this.opts.budget || this.stoppedEarly !== void 0 || this.opts.signal?.aborted) return;
+				try {
+					const startedAt = performance.now();
+					const paid = this.costLedger ? await this.costLedger.runPaidCall({
+						channel: "agent",
+						phase: this.costPhase,
+						actor: "fuzz.evaluate",
+						tags: {
+							target: this.opts.target,
+							cell: cell.id
+						},
+						signal: this.opts.signal,
+						maximumCharge: this.opts.maximumChargeOf?.(scenario, cell),
+						execute: () => this.opts.evaluate(scenario, cell),
+						receipt: (evaluation) => {
+							const cost = this.opts.costOf(scenario, cell, evaluation);
+							if (cost === null) return {
+								model: "unattributed",
+								inputTokens: 0,
+								outputTokens: 0,
+								costUnknown: true
+							};
+							if (!Number.isFinite(cost.usd) || cost.usd < 0) throw new RangeError(`BehaviorExplorer: costOf returned an invalid usd (${String(cost.usd)}) — return null when cost is unknown, never a fabricated number`);
+							return {
+								model: cost.model ?? "unattributed",
+								inputTokens: 0,
+								outputTokens: 0,
+								actualCostUsd: cost.usd
+							};
+						}
+					}) : void 0;
+					if (paid && !paid.succeeded) {
+						if (paid.error instanceof CostReceiptCaptureError && paid.error.receiptError instanceof RangeError) throw paid.error.receiptError;
+						throw paid.error;
+					}
+					const ev = paid ? paid.value : await this.opts.evaluate(scenario, cell);
+					const latencyMs = ev.latencyMs ?? performance.now() - startedAt;
+					this.runsUsed++;
+					runsThisStep++;
+					this.consecutiveEvalErrors = 0;
+					const costUsd = paid ? paid.receipt.costUnknown ? null : paid.receipt.costUsd : void 0;
+					if (paid?.receipt.actualCostUsd !== void 0) this.opts.onCost?.({
+						usd: paid.receipt.actualCostUsd,
+						channel: "agent"
+					});
+					const interest = this.objective.interest(ev, this.objectiveContext());
+					this.log.push({
+						cell,
+						ev,
+						interest,
+						latencyMs,
+						...costUsd !== void 0 ? { costUsd } : {},
+						scenarioId: this.opts.scenarioId(scenario)
+					});
+					this.opts.onProgress?.({
+						type: "evaluated",
+						cell,
+						scenario,
+						evaluation: ev
+					});
+					const bin = this.binId(cell, ev.descriptor);
+					const cur = this.archiveByBin.get(bin);
+					if (!cur || interest > cur.interest) this.archiveByBin.set(bin, {
+						binId: bin,
+						cell,
+						scenario,
+						evaluation: ev,
+						interest
+					});
+					if (interest < this.threshold) return;
+					this.candidateFindings++;
+					if (this.opts.gates?.isValid && !await this.opts.gates.isValid(scenario, ev, cell)) return;
+					if (this.opts.gates?.isUncontaminated && !await this.opts.gates.isUncontaminated(scenario, ev, cell)) return;
+					const minimized = this.opts.minimize ? await this.opts.minimize(scenario, this.opts.evaluate, cell) : scenario;
+					const finding = {
+						id: this.opts.scenarioId(scenario),
+						cell,
+						scenario,
+						minimized,
+						text: this.opts.scenarioText?.(minimized),
+						evaluation: ev,
+						interest,
+						objective: this.objective.kind
+					};
+					this._findings.push(finding);
+					newFindings.push(finding);
+					this.opts.onProgress?.({
+						type: "finding",
+						finding
+					});
+				} catch (err) {
+					if (err instanceof RangeError) throw err;
+					this.evalErrors++;
+					this.consecutiveEvalErrors++;
+					const message = err instanceof Error ? err.message : String(err);
+					this.opts.onProgress?.({
+						type: "eval-error",
+						cell,
+						scenarioId: this.opts.scenarioId(scenario),
+						message
+					});
+					const limit = this.opts.maxConsecutiveEvalErrors ?? 5;
+					if (this.consecutiveEvalErrors >= limit) this.stoppedEarly = {
+						reason: "eval-errors",
+						detail: `${this.consecutiveEvalErrors} consecutive eval errors (last: ${message})`
+					};
+				}
+			}, this.opts.concurrency ?? 1, this.opts.signal);
+		}
+		this.opts.onProgress?.({
+			type: "round",
+			runsUsed: this.runsUsed,
+			budget: this.opts.budget
+		});
+		return {
+			runs: runsThisStep,
+			findings: newFindings
+		};
+	}
+	/** Loop `step()` until the run or dollar budget is spent, the signal aborts,
+	*  or no progress is made. */
+	async run() {
+		while (this.runsUsed < this.opts.budget && this.stoppedEarly === void 0 && !this.opts.signal?.aborted) {
+			const { runs } = await this.step();
+			if (runs === 0 && this.stoppedEarly === void 0) break;
+		}
+		return this.capsule();
+	}
+	coverage() {
+		return this.capsule().coverage;
+	}
+	findings() {
+		return [...this._findings].sort((a, b) => b.interest - a.interest);
+	}
+	capsule() {
+		const cost = this.costLedger?.summary({
+			phase: this.costPhase,
+			tags: { target: this.opts.target }
+		});
+		return buildCapsule({
+			target: this.opts.target,
+			objective: this.objective.kind,
+			cells: this.cells,
+			log: this.log,
+			threshold: this.threshold,
+			archive: [...this.archiveByBin.values()],
+			findings: this._findings,
+			candidateFindings: this.candidateFindings,
+			runsUsed: this.runsUsed,
+			cost: cost ? {
+				costUsd: cost.totalCostUsd,
+				costUnknownRuns: cost.byChannel.reduce((sum, row) => sum + row.unpricedCalls, 0)
+			} : void 0,
+			evalErrors: this.evalErrors,
+			stoppedEarly: this.stoppedEarly
+		});
+	}
 };
-
-// src/fuzz/fuzz-agent.ts
+//#endregion
+//#region src/fuzz/fuzz-agent.ts
+/**
+* `fuzzAgent` — the adversarial batch preset over `BehaviorExplorer`.
+*
+* One call: explore the space to budget with the adversarial objective and
+* return the capsule. For agent-driven, incremental, or multi-objective use,
+* construct a `BehaviorExplorer` and drive it via `makeExploreTools`.
+*/
 async function fuzzAgent(opts) {
-  const { failureThreshold, ...rest } = opts;
-  const explorer = new BehaviorExplorer({
-    ...rest,
-    objective: adversarialObjective(failureThreshold ?? 0.5)
-  });
-  return { capsule: await explorer.run() };
+	const { failureThreshold, ...rest } = opts;
+	return { capsule: await new BehaviorExplorer({
+		...rest,
+		objective: adversarialObjective(failureThreshold ?? .5)
+	}).run() };
 }
-
-// src/fuzz/gates.ts
+//#endregion
+//#region src/fuzz/gates.ts
+/** Combine gate sets; a candidate must pass every gate in every set. */
 function composeGates(...sets) {
-  const present = sets.filter((s) => s != null);
-  return {
-    isValid: async (scenario, ev, cell) => {
-      for (const g of present) {
-        if (g.isValid && !await g.isValid(scenario, ev, cell)) return false;
-      }
-      return true;
-    },
-    isUncontaminated: async (scenario, ev, cell) => {
-      for (const g of present) {
-        if (g.isUncontaminated && !await g.isUncontaminated(scenario, ev, cell)) return false;
-      }
-      return true;
-    }
-  };
+	const present = sets.filter((s) => s != null);
+	return {
+		isValid: async (scenario, ev, cell) => {
+			for (const g of present) if (g.isValid && !await g.isValid(scenario, ev, cell)) return false;
+			return true;
+		},
+		isUncontaminated: async (scenario, ev, cell) => {
+			for (const g of present) if (g.isUncontaminated && !await g.isUncontaminated(scenario, ev, cell)) return false;
+			return true;
+		}
+	};
 }
+/**
+* Reproducibility gate. Re-run the target on a meaning-preserving rephrase of the
+* flagged scenario; keep the finding only when the rephrase ALSO scores below the
+* threshold. A finding that flips under a cosmetic rewrite was keyed to surface
+* form, not the task — a false signal we must not report. Costs one extra
+* evaluation per candidate (candidates are rare, so cheap).
+*/
 function perturbationStabilityGate(opts) {
-  const threshold = opts.failureThreshold ?? 0.5;
-  return {
-    isUncontaminated: async (scenario, _ev, cell) => {
-      const rephrased = opts.perturb(scenario);
-      if (rephrased == null) return true;
-      const re = await opts.evaluate(rephrased, cell);
-      return re.score < threshold;
-    }
-  };
+	const threshold = opts.failureThreshold ?? .5;
+	return { isUncontaminated: async (scenario, _ev, cell) => {
+		const rephrased = opts.perturb(scenario);
+		if (rephrased == null) return true;
+		return (await opts.evaluate(rephrased, cell)).score < threshold;
+	} };
 }
+/**
+* Severity-floor gate. Reject borderline candidates whose score sits in a band
+* just under the threshold — judge noise, not a real defect.
+*/
 function severityFloorGate(opts) {
-  const threshold = opts.failureThreshold ?? 0.5;
-  const margin = opts.margin ?? 0.1;
-  return {
-    isValid: (_scenario, ev) => ev.score <= threshold - margin
-  };
+	const threshold = opts.failureThreshold ?? .5;
+	const margin = opts.margin ?? .1;
+	return { isValid: (_scenario, ev) => ev.score <= threshold - margin };
 }
-
-// src/fuzz/tools.ts
-var NO_ARGS = { type: "object", properties: {}, additionalProperties: false };
+//#endregion
+//#region src/fuzz/tools.ts
+/**
+* Agent-drivable surface over a live exploration session.
+*
+* Framework-neutral tool defs ({name, description, parameters: JSON Schema,
+* handler}) so the on-demand agent — not a batch script — drives the search:
+* step it, read coverage, inspect findings, render the capsule. Transport
+* encodings (OpenAI function shape, MCP) are one-line mappings the host owns.
+*/
+const NO_ARGS = {
+	type: "object",
+	properties: {},
+	additionalProperties: false
+};
 function makeExploreTools(explorer) {
-  return [
-    {
-      name: "explore_step",
-      description: "Run one exploration round: allocate budget across cells, propose + evaluate scenarios, archive elites, admit gate-verified findings. Returns runs spent and new findings.",
-      parameters: NO_ARGS,
-      handler: async () => {
-        const { runs, findings } = await explorer.step();
-        return { runs, newFindings: findings.length, findings };
-      }
-    },
-    {
-      name: "explore_coverage",
-      description: "Read the live coverage map: per planned cell \u2014 runs, robustness (null = uncovered), finding rate, per-dimension means.",
-      parameters: NO_ARGS,
-      handler: async () => explorer.coverage()
-    },
-    {
-      name: "explore_findings",
-      description: "List gate-verified findings so far, sorted by descending interest.",
-      parameters: {
-        type: "object",
-        properties: { limit: { type: "number", description: "max findings to return" } },
-        additionalProperties: false
-      },
-      handler: async (args) => {
-        const limit = args?.limit;
-        const all = explorer.findings();
-        return typeof limit === "number" ? all.slice(0, limit) : all;
-      }
-    },
-    {
-      name: "explore_capsule",
-      description: 'Build the capsule artifact from the current session state. format "data" returns the structured CapsuleData; "html" returns the standalone page.',
-      parameters: {
-        type: "object",
-        properties: {
-          format: { type: "string", enum: ["data", "html"] },
-          generatedAt: { type: "string", description: "ISO timestamp to stamp into the page" }
-        },
-        additionalProperties: false
-      },
-      handler: async (args) => {
-        const a = args ?? {};
-        const capsule = explorer.capsule();
-        if (a.format === "html") return renderCapsuleHtml(capsule, { generatedAt: a.generatedAt });
-        return capsule;
-      }
-    }
-  ];
+	return [
+		{
+			name: "explore_step",
+			description: "Run one exploration round: allocate budget across cells, propose + evaluate scenarios, archive elites, admit gate-verified findings. Returns runs spent and new findings.",
+			parameters: NO_ARGS,
+			handler: async () => {
+				const { runs, findings } = await explorer.step();
+				return {
+					runs,
+					newFindings: findings.length,
+					findings
+				};
+			}
+		},
+		{
+			name: "explore_coverage",
+			description: "Read the live coverage map: per planned cell — runs, robustness (null = uncovered), finding rate, per-dimension means.",
+			parameters: NO_ARGS,
+			handler: async () => explorer.coverage()
+		},
+		{
+			name: "explore_findings",
+			description: "List gate-verified findings so far, sorted by descending interest.",
+			parameters: {
+				type: "object",
+				properties: { limit: {
+					type: "number",
+					description: "max findings to return"
+				} },
+				additionalProperties: false
+			},
+			handler: async (args) => {
+				const limit = args?.limit;
+				const all = explorer.findings();
+				return typeof limit === "number" ? all.slice(0, limit) : all;
+			}
+		},
+		{
+			name: "explore_capsule",
+			description: "Build the capsule artifact from the current session state. format \"data\" returns the structured CapsuleData; \"html\" returns the standalone page.",
+			parameters: {
+				type: "object",
+				properties: {
+					format: {
+						type: "string",
+						enum: ["data", "html"]
+					},
+					generatedAt: {
+						type: "string",
+						description: "ISO timestamp to stamp into the page"
+					}
+				},
+				additionalProperties: false
+			},
+			handler: async (args) => {
+				const a = args ?? {};
+				const capsule = explorer.capsule();
+				if (a.format === "html") return renderCapsuleHtml(capsule, { generatedAt: a.generatedAt });
+				return capsule;
+			}
+		}
+	];
 }
-export {
-  BehaviorExplorer,
-  adversarialObjective,
-  buildCapsule,
-  buildCoverage,
-  cellId,
-  composeGates,
-  enumerateCells,
-  fuzzAgent,
-  makeExploreTools,
-  mutationProposer,
-  noveltyObjective,
-  perturbationStabilityGate,
-  renderCapsuleHtml,
-  severityFloorGate
-};
+//#endregion
+export { BehaviorExplorer, adversarialObjective, buildCapsule, buildCoverage, cellId, composeGates, enumerateCells, fuzzAgent, makeExploreTools, mutationProposer, noveltyObjective, perturbationStabilityGate, renderCapsuleHtml, severityFloorGate };
+
 //# sourceMappingURL=fuzz.js.map