@ggui-ai/ui-gen 0.25.0 → 0.27.0

package/dist/evaluation/index.js

@@ -415,6 +415,13 @@ function buildCodingCriteriaSummary(designMode = "constrained") {
   }
   return lines.join("\n");
 }
+
+// src/adapters/claude/sdk-tool-pins.ts
+function sdkToolPins(input) {
+  const builtins = [...input.builtinTools ?? []];
+  const approved = input.allowedTools !== void 0 ? [...input.allowedTools] : Object.keys(input.mcpServers ?? {}).map((name) => `mcp__${name}`);
+  return { tools: builtins, settingSources: [], allowedTools: [...approved, ...builtins] };
+}
 function createAnthropicClient(rawKey) {
   if (rawKey === void 0) {
     return new Anthropic({
@@ -816,6 +823,26 @@ async function collectLoginTurn(stream, opts) {
 }
 
 // src/harness/llm-router.ts
+function servedModelOf(name) {
+  return typeof name === "string" && name.length > 0 ? { servedModel: name } : {};
+}
+function anthropicFinishReason(stopReason) {
+  if (stopReason === null || stopReason === void 0) return void 0;
+  if (stopReason === "end_turn" || stopReason === "stop_sequence") return "stop";
+  if (stopReason === "max_tokens") return "length";
+  if (stopReason === "refusal") return "content-filter";
+  return "other";
+}
+function googleFinishReason(finishReason) {
+  if (finishReason === null || finishReason === void 0) return void 0;
+  if (finishReason === "STOP") return "stop";
+  if (finishReason === "MAX_TOKENS") return "length";
+  if (["SAFETY", "RECITATION", "BLOCKLIST", "PROHIBITED_CONTENT", "SPII", "IMAGE_SAFETY"].includes(finishReason)) return "content-filter";
+  return "other";
+}
+function visionFinish(finishReason) {
+  return finishReason !== void 0 ? { finishReason } : {};
+}
 function addStrictSchemaConstraints(schema) {
   const result = { ...schema };
   if (result.type === "object" && result.properties) {
@@ -1156,7 +1183,9 @@ var AnthropicAgent = class extends LLMAgent {
       return {
         text,
         inputTokens: response.usage.input_tokens,
-        outputTokens: response.usage.output_tokens
+        outputTokens: response.usage.output_tokens,
+        ...visionFinish(anthropicFinishReason(response.stop_reason)),
+        ...servedModelOf(response.model)
       };
     } catch (e) {
       emitLlmTraceEvent({
@@ -1767,7 +1796,9 @@ var GoogleAgent = class extends LLMAgent {
         text,
         // ggui#1186 (third site): `promptTokenCount` includes `cachedContentTokenCount` — report the non-cached input.
         inputTokens: splitCacheInclusiveUsage(usage?.promptTokenCount ?? 0, usage?.candidatesTokenCount ?? 0, usage?.cachedContentTokenCount ?? 0).tokens.input,
-        outputTokens: usage?.candidatesTokenCount ?? 0
+        outputTokens: usage?.candidatesTokenCount ?? 0,
+        ...visionFinish(googleFinishReason(response.candidates?.[0]?.finishReason)),
+        ...servedModelOf(response.modelVersion)
       };
       const endedAt = Date.now();
       emitLlmTraceEvent({
@@ -2323,6 +2354,16 @@ function getDefaultEvalModel(provider) {
   }
 }
 
+// src/adapters/claude/source-capture.ts
+var COMPILE_TOOL_SUFFIX = "__compile_component";
+function sourceCodeFromToolUse(name, input) {
+  if (typeof name !== "string") return void 0;
+  const field2 = name === "Write" ? "content" : name.endsWith(COMPILE_TOOL_SUFFIX) ? "code" : void 0;
+  if (field2 === void 0 || typeof input !== "object" || input === null) return void 0;
+  const value = Reflect.get(input, field2);
+  return typeof value === "string" && value.length > 0 ? value : void 0;
+}
+
 // src/evaluation/message-parsing.ts
 function extractToolResultTexts(message) {
   if (message.type !== "user") return [];
@@ -2392,12 +2433,9 @@ function extractSourceCodeFromMessage(message) {
   const content = betaMessage?.content;
   if (!Array.isArray(content)) return void 0;
   for (const block of content) {
-    if (block.type === "tool_use" && block.name === "Write") {
-      const input = block.input;
-      if (input?.content) {
-        sourceCode = input.content;
-      }
-    }
+    if (block.type !== "tool_use") continue;
+    const value = sourceCodeFromToolUse(block.name, block.input);
+    if (value !== void 0) sourceCode = value;
   }
   return sourceCode;
 }
@@ -2514,12 +2552,18 @@ async function runEvaluationLoop(options) {
   const evaluationResults = [];
   let currentCode = context.compiledCode;
   let currentSourceCode = context.sourceCode;
-  let round = 0;
+  let round2 = 0;
   const maxRounds = Math.min(config.maxRounds ?? 3, MAX_EVAL_ROUNDS_HARD_LIMIT);
-  while (round < maxRounds) {
-    round++;
-    onProgress?.({ type: "evaluating", round });
-    console.log(`[eval] Round ${round}: evaluating...`);
+  const offersATool = Object.keys(generatorOptions?.mcpServers ?? {}).length > 0 || (generatorOptions?.builtinTools?.length ?? 0) > 0;
+  if (maxRounds > 1 && !offersATool) {
+    throw new Error(
+      "runEvaluationLoop: a fix round needs a tool that can return code \u2014 pass generatorOptions.mcpServers (e.g. a server with compile_component) or opt into a built-in with generatorOptions.builtinTools (e.g. ['Write'])."
+    );
+  }
+  while (round2 < maxRounds) {
+    round2++;
+    onProgress?.({ type: "evaluating", round: round2 });
+    console.log(`[eval] Round ${round2}: evaluating...`);
     const evalContext = {
       ...context,
       sourceCode: currentSourceCode || context.sourceCode
@@ -2527,13 +2571,13 @@ async function runEvaluationLoop(options) {
     const evalResult = await runEvaluation(evalContext, config);
     evaluationResults.push(evalResult);
     console.log(
-      `[eval] Round ${round}: score=${evalResult.finalScore}, passed=${evalResult.passed}, issues=${evalResult.issues.length}`
+      `[eval] Round ${round2}: score=${evalResult.finalScore}, passed=${evalResult.passed}, issues=${evalResult.issues.length}`
     );
-    if (evalResult.passed || round >= maxRounds) {
+    if (evalResult.passed || round2 >= maxRounds) {
       break;
     }
-    onProgress?.({ type: "fixing", round });
-    console.log(`[eval] Round ${round}: fixing...`);
+    onProgress?.({ type: "fixing", round: round2 });
+    console.log(`[eval] Round ${round2}: fixing...`);
     const fixPrompt = buildFixPrompt(evalResult, context.originalPrompt);
     const cliPath = process.env.CLAUDE_WRAPPER_PATH;
     const env = {};
@@ -2555,12 +2599,15 @@ async function runEvaluationLoop(options) {
         maxTurns: 15,
         maxBudgetUsd: config.maxBudgetPerFix,
         env,
-        permissionMode: "bypassPermissions",
-        allowDangerouslySkipPermissions: true,
         ...cliPath && { pathToClaudeCodeExecutable: cliPath },
         ...generatorOptions?.cwd && { cwd: generatorOptions.cwd },
         ...generatorOptions?.mcpServers && { mcpServers: generatorOptions.mcpServers },
-        ...generatorOptions?.allowedTools && { allowedTools: generatorOptions.allowedTools },
+        // ggui#1714 — no CLI built-ins unless asked for, no host settings, no permission bypass.
+        ...sdkToolPins({
+          mcpServers: generatorOptions?.mcpServers,
+          allowedTools: generatorOptions?.allowedTools,
+          builtinTools: generatorOptions?.builtinTools
+        }),
         ...generatorOptions?.model && { model: generatorOptions.model },
         ...generatorOptions?.stderr && { stderr: generatorOptions.stderr }
       }
@@ -2573,7 +2620,7 @@ async function runEvaluationLoop(options) {
     }
     if (fixedCompiledCode) {
       currentCode = fixedCompiledCode;
-      console.log(`[eval] Round ${round}: fixed code captured (${currentCode.length} bytes)`);
+      console.log(`[eval] Round ${round2}: fixed code captured (${currentCode.length} bytes)`);
     }
     if (fixedSourceCode) {
       currentSourceCode = fixedSourceCode;
@@ -4692,9 +4739,19 @@ var INSTRUMENT_BY_ID = {
   "floor.fit": "fit",
   "finish.canvas.inhabited": "fill"
 };
-function majority(id, answers) {
+var INSTRUMENT_BESIDE_JUDGE = {
+  "space.edge": "edge"
+};
+var REQUIRES_POST_ACTION_CAPTURE = /* @__PURE__ */ new Set(["state.feedback"]);
+var NO_POST_ACTION = "not evaluated: no post-action capture";
+function besideRead(kind, m) {
+  if (m.edge === void 0) return { verdict: "n/a", evidence: "not measured: this capture had no page to read" };
+  if (m.edge === null) return { verdict: "n/a", evidence: "edge probe failed on this capture" };
+  return { verdict: m.edge.verdict, evidence: m.edge.evidence };
+}
+function majority(id, answers, unansweredReason) {
   const votes = answers.flatMap((a) => a.filter((x) => x.id === id));
-  if (votes.length === 0) return { verdict: "n/a", evidence: "not answered" };
+  if (votes.length === 0) return { verdict: "n/a", evidence: unansweredReason ?? "not answered" };
   const counts = /* @__PURE__ */ new Map();
   for (const v of votes) counts.set(v.verdict, (counts.get(v.verdict) ?? 0) + 1);
   const ranked = [...counts.entries()].sort((a, b) => b[1] - a[1]);
@@ -4711,8 +4768,9 @@ function instrumentRead(kind, m) {
   if (m.inkRatio === null) return { verdict: "n/a", evidence: "capture unreadable" };
   return { verdict: "n/a", evidence: `fill ${m.inkRatio.toFixed(3)} reported; no floor ruled` };
 }
-function readRow(row, answers, m) {
-  if (row.evaluation === "judge") return majority(row.id, answers);
+function readRow(row, answers, m, unansweredReason) {
+  if (row.evaluation === "judge" && REQUIRES_POST_ACTION_CAPTURE.has(row.id) && m.postActionCapture !== true) return { verdict: "n/a", evidence: NO_POST_ACTION };
+  if (row.evaluation === "judge") return majority(row.id, answers, unansweredReason);
   if (row.evaluation === "human") return { verdict: "n/a", evidence: `reader's column (status ${row.status})` };
   if (row.status !== "live") return { verdict: "n/a", evidence: `not implemented (status ${row.status})` };
   const kind = INSTRUMENT_BY_ID[row.id];
@@ -4724,8 +4782,18 @@ function resolveCriteriaBlock(args) {
   for (const s of args.selected.selection) {
     const row = byId.get(s.id);
     if (row === void 0) continue;
-    const read = readRow(row, args.answers, args.measurements);
-    verdicts.push({ id: row.id, severity: row.severity, method: row.evaluation, status: row.status, source: s.source, verdict: read.verdict, evidence: read.evidence });
+    const read = readRow(row, args.answers, args.measurements, args.unansweredReason);
+    const beside = INSTRUMENT_BESIDE_JUDGE[row.id];
+    verdicts.push({
+      id: row.id,
+      severity: row.severity,
+      method: row.evaluation,
+      status: row.status,
+      source: s.source,
+      verdict: read.verdict,
+      evidence: read.evidence,
+      ...beside !== void 0 && row.evaluation === "judge" ? { instrument: besideRead(beside, args.measurements) } : {}
+    });
   }
   return {
     criteriaSetId: args.selected.criteriaSetId,
@@ -4733,7 +4801,15 @@ function resolveCriteriaBlock(args) {
     selectorVersion: CRITERIA_SELECTOR_VERSION,
     context: args.context,
     selection: args.selected.selection,
-    verdicts
+    verdicts,
+    ...args.call !== void 0 ? {
+      call: {
+        answer: args.call.answer,
+        // Kept answers whose id names no row of the bank: the judge answered something nobody asked.
+        unknownIds: args.answers.flat().filter((a) => !byId.has(a.id)).length,
+        ...args.call.finishReason !== void 0 ? { finishReason: args.call.finishReason } : {}
+      }
+    } : {}
   };
 }
 var CRITERIA_PROPS_MAX_CHARS = 2e3;
@@ -4753,7 +4829,7 @@ ${text}
 }
 function buildCriteriaJudgeBlock(bank, selected, inputs = {}) {
   const byId = new Map(bankRows(bank).map((r) => [r.id, r]));
-  const asked = selected.selection.map((s) => byId.get(s.id)).filter((r) => r !== void 0 && r.evaluation === "judge");
+  const asked = selected.selection.map((s) => byId.get(s.id)).filter((r) => r !== void 0 && r.evaluation === "judge").filter((r) => inputs.postActionCapture === true || !REQUIRES_POST_ACTION_CAPTURE.has(r.id));
   if (asked.length === 0) return "";
   const lines = asked.map((r) => `- ${r.id} (${r.severity}): ${r.text} Evidence must name: ${r.evidence}`);
   return "## Criteria \u2014 answer each one from the FRAME only\nRead these off the screenshot. Never infer them from the request, the styling profile or any direction text above.\n" + (inputs.frame !== void 0 ? `${criteriaFrameLine(inputs.frame)}
@@ -4938,6 +5014,356 @@ function judgeDesignIdentity(src) {
 function resetJudgeDesignIdentityCache() {
   identityCache.clear();
 }
+
+// src/evaluation/criteria/edge.ts
+var EDGE_TOLERANCE_PX = 2;
+var CARD = "card";
+var same = (a, b) => Math.abs(a - b) <= EDGE_TOLERANCE_PX;
+var overlaps = (a, b) => a.left < b.right && b.left < a.right;
+var axis = (b) => (b.left + b.right) / 2;
+function contentEdge(surface, blocks) {
+  const texts = blocks.filter((b) => b.within === surface.id && b.kind !== "surface" && b.role !== "centred");
+  if (texts.length === 0) return void 0;
+  const first = texts.reduce((a, b) => b.top < a.top ? b : a);
+  return first.textEdge ?? first.edge;
+}
+function columns(direct) {
+  const all = direct.filter((b) => b.kind !== "surface");
+  const bridges = (b) => {
+    const under = all.filter((x) => x !== b && overlaps(x, b));
+    return under.some((x, i) => under.slice(i + 1).some((y) => !overlaps(x, y)));
+  };
+  const flow = all.filter((b) => !bridges(b));
+  const parentPool = all.filter((b) => bridges(b));
+  const parent = flow.map((_, i) => i);
+  const find = (i) => parent[i] === i ? i : parent[i] = find(parent[i]);
+  for (let i = 0; i < flow.length; i += 1) for (let j = i + 1; j < flow.length; j += 1) if (overlaps(flow[i], flow[j])) parent[find(i)] = find(j);
+  const groups = /* @__PURE__ */ new Map();
+  flow.forEach((b, i) => groups.set(find(i), [...groups.get(find(i)) ?? [], b]));
+  const cols = [...groups.values()];
+  for (const s of direct.filter((b) => b.kind === "surface")) {
+    const hit = cols.filter((c) => c.some((b) => b.kind !== "surface" && overlaps(b, s)));
+    if (hit.length === 1) hit[0].push(s);
+    else parentPool.push(s);
+  }
+  const lp = parentPool.map((_, i) => i);
+  const lfind = (i) => lp[i] === i ? i : lp[i] = lfind(lp[i]);
+  for (let i = 0; i < parentPool.length; i += 1) for (let j = i + 1; j < parentPool.length; j += 1) if (overlaps(parentPool[i], parentPool[j])) lp[lfind(i)] = lfind(j);
+  const lgroups = /* @__PURE__ */ new Map();
+  parentPool.forEach((b, i) => lgroups.set(lfind(i), [...lgroups.get(lfind(i)) ?? [], b]));
+  cols.push(...lgroups.values());
+  return cols;
+}
+function checkColumn(container, col, blocks) {
+  const surfaces = col.filter((b) => b.kind === "surface");
+  const grounds = surfaces.flatMap((s) => {
+    const c = s.role === "table" ? void 0 : contentEdge(s, blocks);
+    return c === void 0 ? [] : [c];
+  });
+  const flow = col.filter((b) => b.kind !== "surface" && b.role !== "centred" && b.role !== "list-item");
+  const listItems = col.filter((b) => b.role === "list-item");
+  const candidates = [...flow.flatMap((b) => [b.edge, ...b.textEdge !== void 0 ? [b.textEdge] : []]), ...surfaces.flatMap((s) => [s.edge]), ...grounds];
+  const blockEdges = (b) => [b.edge, ...b.textEdge !== void 0 ? [b.textEdge] : []];
+  const endEdgesBesides = (b) => [...surfaces.map((s) => s.right), ...flow.filter((x) => x !== b && x.kind === "controls").map((x) => x.right)];
+  let best = candidates.length === 0 ? [] : null;
+  for (const E of candidates) {
+    const allowed = [E, ...grounds];
+    const off = [];
+    for (const b of flow) {
+      if (blockEdges(b).some((x) => allowed.some((a) => same(x, a)))) continue;
+      if (b.kind === "controls" && endEdgesBesides(b).some((r) => same(b.right, r))) continue;
+      const nearest = allowed.reduce((n, a) => Math.abs(a - b.edge) < Math.abs(n - b.edge) ? a : n);
+      off.push({ container, block: b.id, at: b.edge, nearest, offPx: Math.abs(b.edge - nearest) });
+    }
+    for (const s of surfaces) {
+      const c = s.role === "table" ? void 0 : contentEdge(s, blocks);
+      if (same(s.edge, E) || c !== void 0 && same(c, E)) continue;
+      const nearest = c !== void 0 && Math.abs(c - E) < Math.abs(s.edge - E) ? c : s.edge;
+      off.push({ container, block: s.id, at: nearest, nearest: E, offPx: Math.abs(nearest - E) });
+    }
+    if (best === null || off.length < best.length || off.length === best.length && sum(off) < sum(best)) best = off;
+    if (best.length === 0) break;
+  }
+  const offsets = best ?? [];
+  if (listItems.length > 1) {
+    const e0 = listItems[0].edge;
+    for (const li of listItems.slice(1)) if (!same(li.edge, e0)) offsets.push({ container, block: li.id, at: li.edge, nearest: e0, offPx: Math.abs(li.edge - e0) });
+  }
+  const centred = col.filter((b) => b.role === "centred");
+  const edgeAligned = col.filter((b) => b.role !== "centred");
+  const groupTop = centred.length > 0 ? Math.min(...centred.map((c) => c.top)) : 0;
+  const groupEndsColumn = centred.length > 1 && edgeAligned.every((b) => b.top < groupTop);
+  if (centred.length > 0 && edgeAligned.some((b) => b.kind !== "surface") && !groupEndsColumn) {
+    const E = flow[0]?.edge ?? edgeAligned[0].edge;
+    for (const c of centred) offsets.push({ container, block: c.id, at: c.edge, nearest: E, offPx: Math.abs(c.edge - E), reason: "a centred block in an edge-aligned column" });
+  } else if (centred.length > 1) {
+    const a0 = axis(centred[0]);
+    for (const c of centred.slice(1)) if (!same(axis(c), a0)) offsets.push({ container, block: c.id, at: axis(c), nearest: a0, offPx: Math.abs(axis(c) - a0) });
+  }
+  return offsets;
+}
+var sum = (o) => o.reduce((s, x) => s + x.offPx, 0);
+function judgeEdges(measured, opts = {}) {
+  if (opts.placeholdersOnly === true) return { verdict: "n/a", offsets: [], evidence: "only loading placeholders are shown; not evaluated" };
+  let blocks = measured;
+  const kept = blocks.filter((b) => b.role !== "placeholder");
+  const content = kept.filter((b) => b.kind !== "surface").length;
+  if (content < 2) return { verdict: "n/a", offsets: [], evidence: kept.length === 0 ? "no blocks were measured" : "fewer than two blocks once placeholders are set aside; not evaluated" };
+  blocks = kept;
+  const offsets = [];
+  const containers = [CARD, ...blocks.filter((b) => b.kind === "surface" && b.role !== "table").map((b) => b.id)];
+  for (const container of containers) {
+    const direct = blocks.filter((b) => (b.within ?? CARD) === container);
+    for (const col of columns(direct)) offsets.push(...checkColumn(container, col, blocks));
+  }
+  if (offsets.length === 0) return { verdict: "pass", offsets, evidence: "every column shares its edge, or sits on a stacked surface\u2019s content edge, within 2 px" };
+  const worst = offsets.reduce((a, b) => b.offPx > a.offPx ? b : a);
+  const lines = offsets.map((o) => `${o.block} at x${round(o.at)} is ${round(o.offPx)} px from x${round(o.nearest)} (in ${o.container}${o.reason !== void 0 ? `; ${o.reason}` : ""})`);
+  return { verdict: "fail", offsets, evidence: `${offsets.length} block(s) off the allowed edges, largest ${round(worst.offPx)} px: ${lines.join("; ")}` };
+}
+var round = (n) => (Math.round(n * 10) / 10).toString();
+
+// src/evaluation/criteria/edge-probe.ts
+var EDGE_BLOCKS_EXPRESSION = String.raw`(() => {
+  const root = document.getElementById('root');
+  if (!root) return null;
+  const rootBox = root.getBoundingClientRect();
+  const x0 = rootBox.left;
+  const px = (v) => Math.round((v - x0) * 2) / 2;
+  const CONTROL = 'button,input,select,textarea,[role="button"],[role="tab"],[role="switch"],[role="checkbox"],[role="radio"],[role="option"],[role="link"],a[href]';
+  const ALNUM = /[\p{L}\p{N}]/u;
+  const css = (el) => getComputedStyle(el);
+  const clear = (c) => !c || c === 'transparent' || /rgba\([^)]*,\s*0(\.0+)?\)$/.test(c);
+  // Visible through every ancestor (an ancestor at opacity 0 or visibility hidden hides it), and not degenerate.
+  const shown = (el) => {
+    const s = css(el);
+    if (s.display === 'none' || s.visibility === 'hidden' || Number(s.opacity) === 0) return false;
+    if (typeof el.checkVisibility === 'function' && !el.checkVisibility({ opacityProperty: true, visibilityProperty: true })) return false;
+    const r = el.getBoundingClientRect();
+    return r.width > 1 && r.height > 1;
+  };
+  const inline = (s) => /^inline/.test(s.display) && s.display !== 'inline-block' && s.display !== 'inline-flex' && s.display !== 'inline-grid';
+  const groundOf = (el) => {
+    for (let p = el.parentElement; p; p = p.parentElement) { const b = css(p).backgroundColor; if (!clear(b)) return b; }
+    return 'rgb(255, 255, 255)';
+  };
+  const sides = (s) => ['Top', 'Right', 'Bottom', 'Left'].filter((k) => parseFloat(s['border' + k + 'Width']) > 0 && s['border' + k + 'Style'] !== 'none' && !clear(s['border' + k + 'Color'])).length;
+  const bounded = (el) => { const s = css(el); return sides(s) >= 1 || (!clear(s.backgroundColor) && s.backgroundColor !== groundOf(el)); };
+  // A control: a native or ARIA one, or a bounded box the visitor clicks (a pointer cursor or a tab stop) that holds no
+  // control of its own and is not the whole card. Its contents are not blocks of the column (decided 2026-10-02).
+  const isControl = (el) => {
+    const s = css(el);
+    if (inline(s)) return false;
+    if (el.matches(CONTROL)) return true;
+    if (el === root) return false;
+    const clickable = s.cursor === 'pointer' || (el.hasAttribute('tabindex') && el.tabIndex >= 0);
+    if (!clickable || !bounded(el) || el.querySelector(CONTROL)) return false;
+    const r = el.getBoundingClientRect();
+    return r.width * r.height < 0.6 * rootBox.width * rootBox.height;
+  };
+  const isTable = (el) => el.matches('table,[role="table"],[role="grid"]');
+  const isSurface = (el) => {
+    if (el === root || isControl(el)) return false;
+    const s = css(el);
+    if (/^inline/.test(s.display) || s.display === 'contents') return false;
+    if (isTable(el)) return true;
+    return sides(s) >= 3 || (!clear(s.backgroundColor) && s.backgroundColor !== groundOf(el)) || (s.boxShadow && s.boxShadow !== 'none');
+  };
+  const path = (el) => {
+    const parts = [];
+    for (let e = el; e && e !== root && parts.length < 4; e = e.parentElement) {
+      const i = e.parentElement ? Array.prototype.indexOf.call(e.parentElement.children, e) + 1 : 1;
+      parts.unshift(e.tagName.toLowerCase() + ':' + i);
+    }
+    return parts.join('>');
+  };
+  const label = (el) => { const t = (el.innerText || '').replace(/\s+/g, ' ').trim(); return path(el) + (t ? ' "' + t.slice(0, 24) + '"' : ''); };
+  const surfaces = [];
+  const surfaceOf = (el, self) => { for (let p = self ? el : el.parentElement; p && p !== root; p = p.parentElement) if (surfaces.indexOf(p) >= 0) return p; return null; };
+  const inTable = (el) => { for (let p = el.parentElement; p && p !== root; p = p.parentElement) if (isTable(p)) return true; return false; };
+  const inControl = (el) => { for (let p = el; p && p !== root; p = p.parentElement) if (isControl(p)) return p; return null; };
+  const all = Array.prototype.slice.call(root.querySelectorAll('*')).filter(shown);
+  for (const el of all) if (isSurface(el) && !inTable(el) && !inControl(el)) surfaces.push(el);
+  const ids = new Map();
+  const idOf = (el) => { if (!ids.has(el)) ids.set(el, label(el)); return ids.get(el); };
+  const blocks = [];
+  // Per block: the element it stands for and its first line's box, for rows.
+  const owners = [];
+  const lines = [];
+  // Per block: its raw line boxes and left, so a row reads the line two items share, not a wrapped item's widest line.
+  const rects = [];
+  const rawLeft = [];
+  const words = [];
+  const within = (el, self) => { const s = surfaceOf(el, self); return s ? { within: idOf(s) } : {}; };
+  for (const s of surfaces) {
+    const r = s.getBoundingClientRect();
+    owners.push(s); lines.push({ top: r.top, bottom: r.bottom }); rects.push([r]); rawLeft.push(r.left); words.push('');
+    // A placeholder: a box with no text, no control and no image or icon inside it (a loading bar). Not a block.
+    const holds = Array.prototype.some.call(s.querySelectorAll('*'), (el) => shown(el) && (el.matches(CONTROL) || el.tagName === 'IMG' || el.tagName.toLowerCase() === 'svg')) || ALNUM.test(s.innerText || '');
+    blocks.push(Object.assign({ id: idOf(s), kind: 'surface', edge: px(r.left), left: px(r.left), right: px(r.right), top: Math.round(r.top) }, within(s, false), isTable(s) ? { role: 'table' } : holds ? {} : { role: 'placeholder' }));
+  }
+  const blockOf = (node) => { for (let p = node.parentElement; p && p !== root; p = p.parentElement) if (!inline(css(p))) return p; return root; };
+  const groups = new Map();
+  const walker = document.createTreeWalker(root, NodeFilter.SHOW_TEXT);
+  for (let n = walker.nextNode(); n; n = walker.nextNode()) {
+    if (!/\S/.test(n.data) || !n.parentElement || n.parentElement.closest('style,script,noscript,template') || !shown(n.parentElement) || inControl(n.parentElement) || inTable(n.parentElement)) continue;
+    const b = blockOf(n);
+    if (!groups.has(b)) groups.set(b, []);
+    groups.get(b).push(n);
+  }
+  // Ink: text whose own colour is not transparent. The edge is the first visible character in ink (decided 2026-10-02):
+  // leading invisible text, private-use icon-font glyphs and formatting characters are not characters of the block,
+  // while a currency sign or a quotation mark is.
+  const inked = (n) => !clear(css(n.parentElement).color) && ALNUM.test(n.data);
+  const textRects = (nodes) => {
+    const rs = [];
+    for (const n of nodes) { if (!inked(n)) continue; const range = document.createRange(); range.selectNodeContents(n); for (const r of range.getClientRects()) if (r.width > 0 && r.height > 0) rs.push(r); }
+    return rs;
+  };
+  const glyphAt = (n, i) => { const range = document.createRange(); range.setStart(n, i); range.setEnd(n, i + 1); const r = range.getClientRects()[0]; return r || null; };
+  const firstGlyph = (nodes) => {
+    for (const n of nodes) { if (!inked(n)) continue; const i = n.data.search(/[^\s\p{Co}\p{Cf}]/u); if (i < 0) continue; const r = glyphAt(n, i); if (r) return r; }
+    return null;
+  };
+  // A marker paints: a loaded image, an SVG with a filled or stroked shape, or an element with its own fill, border
+  // or visible text. An element that paints nothing is not a marker.
+  const painted = (m) => {
+    if (!shown(m)) return false;
+    if (m.tagName === 'IMG') return m.complete && m.naturalWidth > 0;
+    if (m.tagName.toLowerCase() === 'svg' || m.closest('svg')) {
+      const shapes = m.querySelectorAll('path,circle,rect,line,polyline,polygon,ellipse,use,text');
+      return Array.prototype.some.call(shapes, (sh) => { const st = css(sh); return (st.fill !== 'none' && !clear(st.fill)) || (st.stroke !== 'none' && !clear(st.stroke)); });
+    }
+    const s = css(m);
+    if (!clear(s.backgroundColor) || sides(s) >= 1 || ((m.textContent || '').trim().length > 0 && !clear(s.color))) return true;
+    // A wrapper that paints nothing itself is a marker when an image or icon inside it paints.
+    return Array.prototype.some.call(m.querySelectorAll('svg,img'), (k) => painted(k));
+  };
+  const markerBefore = (b, textLeft, lineTop, lineBottom) => {
+    const cands = [];
+    const consider = (m) => { if (!painted(m)) return; const r = m.getBoundingClientRect(); if (r.left < textLeft - 2 && r.bottom > lineTop && r.top < lineBottom) cands.push(r.left); };
+    for (const m of b.querySelectorAll('svg,img,[aria-hidden="true"],i')) if ((m.textContent || '').trim().length <= 2) consider(m);
+    const p = b.parentElement;
+    if (p && p !== root && /flex/.test(css(p).display) && !/column/.test(css(p).flexDirection)) for (const sib of p.children) { if (sib === b) break; if (sib.matches('svg,img,i,[aria-hidden="true"]') || !(sib.textContent || '').trim()) consider(sib); }
+    return cands.length ? Math.min.apply(null, cands) : null;
+  };
+  let texts = 0;
+  let ornaments = 0;
+  for (const [b, nodes] of groups) {
+    if (!nodes.some((n) => ALNUM.test(n.data))) { ornaments += 1; continue; }
+    const rs = textRects(nodes);
+    const g = firstGlyph(nodes);
+    if (!rs.length || g === null) continue;
+    texts += 1;
+    const first = g.left;
+    const left = Math.min.apply(null, rs.map((r) => r.left));
+    const right = Math.max.apply(null, rs.map((r) => r.right));
+    const top = Math.min.apply(null, rs.map((r) => r.top));
+    const marker = markerBefore(b, first, g.top, g.bottom);
+    const s = css(b);
+    const box = b.getBoundingClientRect();
+    const parent = b.parentElement ? css(b.parentElement) : null;
+    const centredText = s.textAlign === 'center' && right - left < box.width * 0.9;
+    const centredBox = parent && /flex/.test(parent.display) && /column/.test(parent.flexDirection) && parent.alignItems === 'center';
+    const list = b.matches('li') || s.display === 'list-item' ? b.closest('ul,ol') : null;
+    const inProse = list && list.previousElementSibling && list.previousElementSibling.matches('p');
+    const block = { id: idOf(b), kind: 'text', edge: px(marker !== null ? marker : first), left: px(Math.min(left, marker !== null ? marker : left)), right: px(right), top: Math.round(top) };
+    if (marker !== null) block.textEdge = px(first);
+    if (centredText || centredBox) block.role = 'centred';
+    else if (inProse) block.role = 'list-item';
+    owners.push(b); lines.push({ top: g.top, bottom: g.bottom }); rects.push(rs); rawLeft.push(Math.min(left, marker !== null ? marker : left)); words.push(nodes.map((n) => n.data).join('').trim());
+    blocks.push(Object.assign(block, within(b, true)));
+  }
+  const ctlGroups = new Map();
+  for (const el of all) { if (!isControl(el) || inTable(el) || (el.parentElement && inControl(el.parentElement))) continue; const p = el.parentElement || root; if (!ctlGroups.has(p)) ctlGroups.set(p, []); ctlGroups.get(p).push(el); }
+  let controls = 0;
+  for (const [p, els] of ctlGroups) {
+    const rs = els.map((e) => e.getBoundingClientRect());
+    const topRow = Math.min.apply(null, rs.map((r) => r.top));
+    let firstEl = null; let firstR = null;
+    els.forEach((e, i) => { const r = rs[i]; if (r.top - topRow < r.height / 2 && (!firstR || r.left < firstR.left)) { firstEl = e; firstR = r; } });
+    let edge = firstR.left;
+    if (!bounded(firstEl)) { const tw = document.createTreeWalker(firstEl, NodeFilter.SHOW_TEXT); const ns = []; for (let n = tw.nextNode(); n; n = tw.nextNode()) ns.push(n); const g = firstGlyph(ns); if (g) edge = g.left; }
+    controls += 1;
+    const ps = css(p);
+    const centred = ps.justifyContent === 'center' || (ps.textAlign === 'center' && !/flex|grid/.test(ps.display));
+    const block = { id: idOf(p) + ' [controls]', kind: 'controls', edge: px(edge), left: px(Math.min.apply(null, rs.map((r) => r.left))), right: px(Math.max.apply(null, rs.map((r) => r.right))), top: Math.round(topRow) };
+    if (centred) block.role = 'centred';
+    owners.push(firstEl); lines.push({ top: firstR.top, bottom: firstR.bottom }); rects.push(rs); rawLeft.push(Math.min.apply(null, rs.map((r) => r.left))); words.push('');
+    blocks.push(Object.assign(block, within(firstEl, true)));
+  }
+  // Rows (decided 2026-10-01, read by geometry 2026-10-02): items set on one line, however the CSS sets them. An item
+  // after another measured item on its line that holds one block (a text, a control, or a box holding at most one
+  // text) belongs to the row and is not measured; an item with a stack inside it is a column.
+  const dropped = new Set();
+  const sameLine = (a, b) => Math.min(a.bottom, b.bottom) - Math.max(a.top, b.top) > Math.min(a.bottom - a.top, b.bottom - b.top) / 2;
+  const contains = (o, el) => o === el || o.contains(el);
+  const innerTexts = (i) => owners.map((el, k) => k).filter((k) => k !== i && blocks[k].kind !== 'surface' && contains(owners[i], owners[k]));
+  const isPill = (i) => blocks[i].kind === 'surface' && innerTexts(i).length <= 1 && !owners.some((el, k) => k !== i && blocks[k].kind === 'surface' && contains(owners[i], el));
+  const itemOf = (el, other) => { let e = el; while (e.parentElement && e.parentElement !== root && !e.parentElement.contains(other)) e = e.parentElement; return e; };
+  for (let i = 0; i < blocks.length; i += 1) {
+    if (blocks[i].kind === 'surface' && !isPill(i)) continue;
+    const before = blocks.findIndex((bj, j) => j !== i && bj.within === blocks[i].within && !(bj.kind === 'surface' && !isPill(j)) && !contains(owners[j], owners[i]) && !contains(owners[i], owners[j]) && rects[j].some((r) => sameLine(r, lines[i]) && r.right <= rawLeft[i] + 1));
+    if (before < 0) continue;
+    // A number or bullet in front of a text on its line is a hanging marker, not a row: the pair is one block with two
+    // edges, the marker's and the text's.
+    if (blocks[i].kind === 'text' && blocks[before].kind === 'text' && /^[\p{N}\p{P}\p{S}]{1,3}$/u.test(words[before])) {
+      blocks[i] = Object.assign({}, blocks[i], { edge: blocks[before].edge, left: Math.min(blocks[i].left, blocks[before].left), textEdge: blocks[i].edge });
+      delete blocks[i].role;
+      dropped.add(before);
+      continue;
+    }
+    const item = itemOf(owners[i], owners[before]);
+    const inside = owners.map((el, k) => k).filter((k) => contains(item, owners[k]) && blocks[k].kind !== 'surface');
+    const boxes = owners.map((el, k) => k).filter((k) => contains(item, owners[k]) && blocks[k].kind === 'surface' && k !== i);
+    if (blocks[i].kind === 'surface') { if (boxes.length === 0) { dropped.add(i); for (const k of innerTexts(i)) dropped.add(k); } continue; }
+    if (inside.length === 1 && boxes.length === 0) dropped.add(i);
+  }
+  const kept = blocks.filter((_, i) => !dropped.has(i));
+  // Placeholders: a frame with no text in ink shows nothing to evaluate, whatever controls or boxes it draws.
+  return { blocks: kept, placeholdersOnly: texts === 0, skipped: { rowItems: dropped.size, ornaments } };
+})()`;
+var KINDS = ["text", "controls", "surface"];
+var ROLES = ["list-item", "table", "centred", "placeholder"];
+var isKind = (v) => KINDS.some((k) => k === v);
+var isRole = (v) => ROLES.some((r) => r === v);
+var num = (v) => typeof v === "number" && Number.isFinite(v);
+var field = (o, key) => Reflect.get(o, key);
+function parseEdgeProbe(value) {
+  if (typeof value !== "object" || value === null) return null;
+  const raw = field(value, "blocks");
+  const placeholdersOnly = field(value, "placeholdersOnly");
+  if (!Array.isArray(raw) || typeof placeholdersOnly !== "boolean") return null;
+  const skippedRaw = field(value, "skipped");
+  if (typeof skippedRaw !== "object" || skippedRaw === null) return null;
+  const rowItems = field(skippedRaw, "rowItems");
+  const ornaments = field(skippedRaw, "ornaments");
+  if (!num(rowItems) || !num(ornaments)) return null;
+  const blocks = [];
+  for (const entry of raw) {
+    if (typeof entry !== "object" || entry === null) return null;
+    const [id, kind, edge, left, right, top] = ["id", "kind", "edge", "left", "right", "top"].map((k) => field(entry, k));
+    const [within, textEdge, role] = ["within", "textEdge", "role"].map((k) => field(entry, k));
+    if (typeof id !== "string" || !isKind(kind) || !num(edge) || !num(left) || !num(right) || !num(top)) return null;
+    if (within !== void 0 && typeof within !== "string") return null;
+    if (textEdge !== void 0 && !num(textEdge)) return null;
+    if (role !== void 0 && !isRole(role)) return null;
+    blocks.push({
+      id,
+      kind,
+      edge,
+      left,
+      right,
+      top,
+      ...within !== void 0 ? { within } : {},
+      ...textEdge !== void 0 ? { textEdge } : {},
+      ...role !== void 0 ? { role } : {}
+    });
+  }
+  return { blocks, placeholdersOnly, skipped: { rowItems, ornaments } };
+}
+
+// src/evaluation/visual-evaluator.ts
 function canvasFitPolicy(canvas) {
   switch (canvas) {
     case "xs-chat-card":
@@ -5021,6 +5447,7 @@ function canvasOverflowIssue(canvas, viewport, contentHeight, verdict) {
   const box = isDeclaredViewport(canvas, viewport) ? `declared ${viewport.width}\xD7${viewport.height}; class box ${classBox.width}\xD7${classBox.height}` : `${viewport.width}\xD7${viewport.height}`;
   return {
     dimension: "canvas-overflow",
+    origin: "instrument",
     severity: verdict === "fail" ? "critical" : "major",
     description: `Rendered content is ${contentHeight}px tall on the ${canvas} canvas (${box}) \u2014 ${hidden}px ${verdict === "fail" ? "is cut off: the inline card does not scroll" : "sits below the first screen"}.`,
     fix: "Fit the composition to the canvas: fewer and shorter sections, one compact row of chips, no hero taller than a third of the box, no fixed min-heights or tall paddings \u2014 measure against the canvas, not the page."
@@ -5029,6 +5456,7 @@ function canvasOverflowIssue(canvas, viewport, contentHeight, verdict) {
 function canvasOverflowXIssue(canvas, viewport, hiddenPx, verdict) {
   return {
     dimension: "canvas-overflow-x",
+    origin: "instrument",
     severity: verdict === "fail" ? "critical" : "major",
     description: `Content runs ${hiddenPx}px past the ${canvas} card's ${viewport.width}px width \u2014 the card's frame clips it, so a visitor sees a cut-off row or a sideways scroll.`,
     fix: "Keep every row within the card: wrap long text (overflow-wrap: anywhere), let a wide table or code block scroll inside its own container, and give no child a fixed width wider than the card."
@@ -5039,6 +5467,7 @@ function canvasBlankIssue(canvas, viewport) {
   const box = isDeclaredViewport(canvas, viewport) ? `declared ${viewport.width}\xD7${viewport.height}; class box ${classBox.width}\xD7${classBox.height}` : `${viewport.width}\xD7${viewport.height}`;
   return {
     dimension: "canvas-blank",
+    origin: "instrument",
     severity: "critical",
     description: `Nothing painted on the ${canvas} canvas (${box}): the capture is one flat colour \u2014 the component mounted and rendered no pixels.`,
     fix: "Return and paint the component tree: a root that renders nothing is a blank, not an empty state. Check that the default export returns its JSX, that no early return yields null on the sample props, and that the first paint shows visible content."
@@ -5260,6 +5689,19 @@ async function measureContentHeight(page, expression = CONTENT_HEIGHT_EXPRESSION
     return null;
   }
 }
+async function measureEdges(page) {
+  try {
+    const probe = parseEdgeProbe(await page.evaluate(EDGE_BLOCKS_EXPRESSION));
+    if (probe === null) {
+      console.warn("[visual-eval] edge probe unavailable: the page returned no readable blocks");
+      return null;
+    }
+    return judgeEdges(probe.blocks, { placeholdersOnly: probe.placeholdersOnly });
+  } catch (e) {
+    console.warn(`[visual-eval] edge probe unavailable: ${e instanceof Error ? e.message : String(e)}`);
+    return null;
+  }
+}
 async function captureScreenshotDetailed(html, viewport = DEFAULT_SCREENSHOT_VIEWPORT, deps = {}, capture = "full-page", natural = { padPx: JUDGE_INLINE_PAD_PX }) {
   try {
     const launchOptions = await resolveLaunchOptions(viewport, deps);
@@ -5281,12 +5723,14 @@ async function captureScreenshotDetailed(html, viewport = DEFAULT_SCREENSHOT_VIE
         }
         const framed = cardHeightPx === null ? viewport.height - 2 * natural.padPx : floorPx !== void 0 ? Math.max(cardHeightPx, floorPx) : cardHeightPx;
         const clip = naturalClip(viewport, framed, natural.padPx);
+        const edge2 = await measureEdges(page);
         const screenshot2 = await page.screenshot({ type: "png", fullPage: false, clip });
-        return { png: Buffer.from(screenshot2), contentHeight: cardHeightPx, overflowX };
+        return { png: Buffer.from(screenshot2), contentHeight: cardHeightPx, overflowX, edge: edge2 };
       }
       const contentHeight = await measureContentHeight(page);
+      const edge = await measureEdges(page);
       const screenshot = await page.screenshot({ type: "png", fullPage: capture === "full-page" });
-      return { png: Buffer.from(screenshot), contentHeight };
+      return { png: Buffer.from(screenshot), contentHeight, edge };
     } finally {
       await browser.close();
     }
@@ -5415,6 +5859,43 @@ function overflowXVerdict(canvas, frame) {
   if (px === void 0 || px === null || px <= OVERFLOW_X_ROUNDING_PX || frame.policy.overflow === "none") return null;
   return canvasOverflowXIssue(canvas, frame.viewport, px, frame.policy.overflow);
 }
+function judgeSalvageOf(error, text, finishReason) {
+  const cause = finishReason === void 0 ? "unknown" : finishReason === "length" ? "output-cap" : finishReason === "stop" ? "malformed" : "stopped";
+  const message = error instanceof Error ? error.message : String(error);
+  const position = /position (\d+)/.exec(message);
+  const start = text.indexOf("{");
+  const offset = position !== null ? (start > 0 ? start : 0) + Number(position[1]) : void 0;
+  const parseError = message.startsWith("Visual evaluator did not return valid JSON") ? "no complete JSON object in the answer" : message.replace(/"(?:[^"\\]|\\.)*"/g, '"\u2026"').slice(0, 160);
+  return { cause, ...finishReason !== void 0 ? { finishReason } : {}, parseError, ...offset !== void 0 ? { offset } : {}, chars: text.length };
+}
+function judgeRecordOf(k, parsed, samples, criteria, requestedModel) {
+  const salvaged = parsed.flatMap((a) => a.salvage !== void 0 ? [a.salvage.cause] : []);
+  const answered = [...parsed, ...criteria?.kind === "ok" ? [criteria] : []];
+  const served = [...new Set(answered.flatMap((a) => a.servedModel !== void 0 ? [a.servedModel] : []))];
+  const criteriaSalvaged = criteria?.kind === "ok" ? criteria.salvage?.cause : void 0;
+  return {
+    k,
+    rule: "median",
+    samples,
+    sigma: populationSigma(samples),
+    notes: parsed.map((a) => a.result.critique ?? ""),
+    ...salvaged.length > 0 ? { salvaged } : {},
+    ...criteriaSalvaged !== void 0 ? { criteriaSalvaged } : {},
+    models: { requested: requestedModel, served }
+  };
+}
+function callFacts(response) {
+  return {
+    ...response.finishReason !== void 0 ? { finishReason: response.finishReason } : {},
+    ...response.servedModel !== void 0 ? { servedModel: response.servedModel } : {}
+  };
+}
+var NO_CRITERIA_ARRAY = { array: false, received: 0, kept: 0, dropped: { notObject: 0, badId: 0, badVerdict: 0 } };
+function criteriaCallOf(criteria) {
+  if (criteria?.kind !== "ok") return {};
+  const answer = criteria.result.criteriaCounts ?? NO_CRITERIA_ARRAY;
+  return { call: { answer, ...criteria.finishReason !== void 0 ? { finishReason: criteria.finishReason } : {} } };
+}
 async function judgeFrame(judge, config, model, png, originalPrompt, profileBlock, canvas, k, criteriaBlock) {
   const [scoring, criteria] = await Promise.all([
     Promise.all(Array.from({ length: k }, () => judgeAndParse(judge, config, model, png, originalPrompt, profileBlock, canvas))),
@@ -5426,7 +5907,10 @@ function criteriaCallTokens(criteria) {
   return criteria?.kind === "ok" ? { inputTokens: criteria.inputTokens, outputTokens: criteria.outputTokens } : void 0;
 }
 function criteriaReads(criteria) {
-  return criteria?.kind === "ok" ? [criteria.result.criteriaAnswers ?? []] : [];
+  return criteria?.kind === "ok" && criteria.salvage === void 0 ? [criteria.result.criteriaAnswers ?? []] : [];
+}
+function criteriaUnansweredReason(criteria) {
+  return criteria?.kind === "ok" && criteria.salvage !== void 0 ? { unansweredReason: `not answered: the criteria answer was salvaged (${criteria.salvage.cause})` } : {};
 }
 function judgeCountFor(config, canvas) {
   const k = Math.max(1, Math.floor(config.judgeK ?? 1));
@@ -5449,12 +5933,16 @@ async function judgeAndParse(judge, config, model, screenshot, originalPrompt, p
     const response = await judge(config, model, VISUAL_EVAL_PROMPT, screenshot, originalPrompt, profileBlock, criteriaBlock);
     try {
       const result = parseVisualResponse(response.text, config.passThreshold);
-      return { kind: "ok", result, inputTokens: response.inputTokens, outputTokens: response.outputTokens };
+      return { kind: "ok", result, inputTokens: response.inputTokens, outputTokens: response.outputTokens, ...callFacts(response) };
     } catch (e) {
       const salvaged = salvageTruncatedVisualAnswer(response.text, config.passThreshold);
       if (salvaged !== null) {
-        console.warn(`[visual-eval] visual_judge_truncated${canvas ? ` canvas=${canvas}` : ""}: answer cut after ${response.text.length} chars \u2014 score recovered from the closed prefix, issues list partial`);
-        return { kind: "ok", result: salvaged, inputTokens: response.inputTokens, outputTokens: response.outputTokens };
+        const salvage = judgeSalvageOf(e, response.text, response.finishReason);
+        const token = criteriaBlock.length > 0 ? "visual_criteria_salvaged" : "visual_judge_salvaged";
+        console.warn(
+          `[visual-eval] ${token}${canvas ? ` canvas=${canvas}` : ""} cause=${salvage.cause} finish=${salvage.finishReason ?? "unreported"} at=${salvage.offset ?? "?"}/${salvage.chars} (${salvage.parseError}): score recovered from the closed prefix, issues list partial`
+        );
+        return { kind: "ok", result: salvaged, inputTokens: response.inputTokens, outputTokens: response.outputTokens, salvage, ...callFacts(response) };
       }
       const reason = `judge answer unparsable: ${e instanceof Error ? e.message : String(e)}`;
       if (attempt === 1) {
@@ -5533,17 +6021,11 @@ async function runVisualEvaluationDetailed(context, config, deps = {}) {
       const result2 = representative.result;
       result2.finalScore = median;
       result2.passed = median >= config.passThreshold;
-      result2.inputTokens = parsed.reduce((sum, a) => sum + a.inputTokens, 0);
-      result2.outputTokens = parsed.reduce((sum, a) => sum + a.outputTokens, 0);
+      result2.inputTokens = parsed.reduce((sum2, a) => sum2 + a.inputTokens, 0);
+      result2.outputTokens = parsed.reduce((sum2, a) => sum2 + a.outputTokens, 0);
       const response2 = { inputTokens: result2.inputTokens, outputTokens: result2.outputTokens };
       const criteriaTokens = criteriaCallTokens(criteriaAnswer);
-      const judgeRecord = {
-        k,
-        rule: "median",
-        samples,
-        sigma: populationSigma(samples),
-        notes: parsed.map((a) => a.result.critique ?? "")
-      };
+      const judgeRecord = judgeRecordOf(k, parsed, samples, criteriaAnswer, model);
       const contentHeight = attempt2.contentHeight;
       const overflow = contentHeight !== null && contentHeight > viewport.height;
       const fitIssue = fitVerdict(canvas, frame);
@@ -5567,7 +6049,9 @@ async function runVisualEvaluationDetailed(context, config, deps = {}) {
         context: criteriaContext,
         selected: criteriaSelected,
         answers: criteriaReads(criteriaAnswer),
-        measurements: { overflow, contentHeight, viewportHeight: viewport.height, inkRatio }
+        measurements: { overflow, contentHeight, viewportHeight: viewport.height, inkRatio, edge: attempt2.edge ?? null },
+        ...criteriaUnansweredReason(criteriaAnswer),
+        ...criteriaCallOf(criteriaAnswer)
       }) : void 0;
       perCanvasResults.push(result2);
       perCanvas.push({
@@ -5629,7 +6113,7 @@ function canvasScopedIssue(canvas, issue) {
 }
 function aggregateCanvasResults(perCanvas, results) {
   const n = results.length;
-  const mean = (pick) => Math.round(results.reduce((sum, r) => sum + pick(r), 0) / n);
+  const mean = (pick) => Math.round(results.reduce((sum2, r) => sum2 + pick(r), 0) / n);
   const dimensions = {
     completeness: mean((r) => r.dimensions.completeness),
     visualPolish: mean((r) => r.dimensions.visualPolish),
@@ -5649,7 +6133,18 @@ function aggregateCanvasResults(perCanvas, results) {
     outputTokens += r.outputTokens ?? 0;
   });
   const criteriaSpent = perCanvas.flatMap((c) => c.criteriaTokens !== void 0 ? [c.criteriaTokens] : []);
+  const judgeSalvage = {
+    scoring: {
+      salvaged: perCanvas.reduce((sum2, c) => sum2 + (c.judge.salvaged?.length ?? 0), 0),
+      answers: perCanvas.reduce((sum2, c) => sum2 + c.judge.samples.length, 0)
+    },
+    criteria: {
+      salvaged: perCanvas.filter((c) => c.judge.criteriaSalvaged !== void 0).length,
+      answered: perCanvas.filter((c) => c.criteriaTokens !== void 0).length
+    }
+  };
   return {
+    judgeSalvage,
     passed: perCanvas.every((c) => c.passed),
     finalScore: mean((r) => r.finalScore),
     dimensions,
@@ -5659,8 +6154,8 @@ function aggregateCanvasResults(perCanvas, results) {
     outputTokens,
     ...criteriaSpent.length > 0 ? {
       criteriaTokens: {
-        inputTokens: criteriaSpent.reduce((sum, t) => sum + t.inputTokens, 0),
-        outputTokens: criteriaSpent.reduce((sum, t) => sum + t.outputTokens, 0)
+        inputTokens: criteriaSpent.reduce((sum2, t) => sum2 + t.inputTokens, 0),
+        outputTokens: criteriaSpent.reduce((sum2, t) => sum2 + t.outputTokens, 0)
       }
     } : {},
     canvases: [...perCanvas]
@@ -5746,7 +6241,9 @@ async function judgeStoredCapture(input, config, deps = {}) {
     context: criteriaContext,
     selected: criteriaSelected,
     answers: criteriaReads(criteriaAnswer),
-    measurements: { overflow, contentHeight, viewportHeight: viewport.height, inkRatio }
+    measurements: { overflow, contentHeight, viewportHeight: viewport.height, inkRatio },
+    ...criteriaUnansweredReason(criteriaAnswer),
+    ...criteriaCallOf(criteriaAnswer)
   }) : void 0;
   const criteriaTokens = criteriaCallTokens(criteriaAnswer);
   return {
@@ -5757,13 +6254,13 @@ async function judgeStoredCapture(input, config, deps = {}) {
       score: median,
       passed,
       issues,
-      judge: { k, rule: "median", samples, sigma: populationSigma(samples), notes: parsed.map((a) => a.result.critique ?? "") },
+      judge: judgeRecordOf(k, parsed, samples, criteriaAnswer, model),
       contentHeight,
       overflow,
       inkRatio,
       ...criteria !== void 0 ? { criteria } : {},
-      inputTokens: parsed.reduce((sum, a) => sum + a.inputTokens, 0),
-      outputTokens: parsed.reduce((sum, a) => sum + a.outputTokens, 0),
+      inputTokens: parsed.reduce((sum2, a) => sum2 + a.inputTokens, 0),
+      outputTokens: parsed.reduce((sum2, a) => sum2 + a.outputTokens, 0),
       ...criteriaTokens !== void 0 ? { criteriaTokens } : {}
     }
   };
@@ -5802,7 +6299,7 @@ function salvageTruncatedVisualAnswer(text, passThreshold) {
     hierarchy,
     aesthetics,
     issues,
-    critique: `[truncated judge answer \u2014 score recovered from the closed prefix; ${issues.length} issue(s) kept, the rest and the critique were cut]`
+    critique: `[salvaged judge answer \u2014 it did not parse; score recovered from the closed prefix; ${issues.length} issue(s) kept, the rest and the critique lost]`
   });
   return parseVisualResponse(rebuilt, passThreshold);
 }
@@ -5831,31 +6328,43 @@ function parseVisualResponse(text, passThreshold) {
   };
   const issues = (raw.issues || []).map((i) => ({
     dimension: i.dimension || "visual",
+    origin: "judge",
     severity: i.severity || "minor",
     description: i.description || "",
     fix: i.fix || ""
   }));
-  const criteriaAnswers = parseCriteriaAnswers(raw.criteria);
+  const criteria = parseCriteriaAnswers(raw.criteria);
   return {
     passed: finalScore >= passThreshold,
     finalScore,
     dimensions,
     issues,
     critique: raw.critique,
-    ...criteriaAnswers !== void 0 ? { criteriaAnswers } : {}
+    ...criteria.answers !== void 0 ? { criteriaAnswers: criteria.answers } : {},
+    criteriaCounts: criteria.counts
   };
 }
 function parseCriteriaAnswers(raw) {
-  if (!Array.isArray(raw)) return void 0;
-  const out = [];
+  if (!Array.isArray(raw)) return { counts: NO_CRITERIA_ARRAY };
+  const answers = [];
+  const dropped = { notObject: 0, badId: 0, badVerdict: 0 };
   for (const entry of raw) {
-    if (typeof entry !== "object" || entry === null) continue;
+    if (typeof entry !== "object" || entry === null) {
+      dropped.notObject += 1;
+      continue;
+    }
     const o = entry;
-    if (typeof o.id !== "string") continue;
-    if (o.verdict !== "pass" && o.verdict !== "fail" && o.verdict !== "n/a") continue;
-    out.push({ id: o.id, verdict: o.verdict, evidence: typeof o.evidence === "string" ? o.evidence : "" });
+    if (typeof o.id !== "string") {
+      dropped.badId += 1;
+      continue;
+    }
+    if (o.verdict !== "pass" && o.verdict !== "fail" && o.verdict !== "n/a") {
+      dropped.badVerdict += 1;
+      continue;
+    }
+    answers.push({ id: o.id, verdict: o.verdict, evidence: typeof o.evidence === "string" ? o.evidence : "" });
   }
-  return out;
+  return { answers, counts: { array: true, received: raw.length, kept: answers.length, dropped } };
 }
 function getDefaultVisualModel(provider) {
   switch (provider) {
@@ -5875,7 +6384,12 @@ async function runVisualEval(context, config, deps = {}) {
   const issues = (result.issues || []).map(toEvalIssue);
   const summary = summarizeVisualResult(result);
   const coverage = { status: "ran" };
-  return summary === void 0 ? { issues, coverage } : { issues, summary, coverage };
+  const tokens = result.inputTokens !== void 0 && result.outputTokens !== void 0 ? {
+    inputTokens: result.inputTokens,
+    outputTokens: result.outputTokens,
+    ...result.criteriaTokens !== void 0 ? { criteria: result.criteriaTokens } : {}
+  } : void 0;
+  return { issues, ...summary !== void 0 ? { summary } : {}, coverage, ...tokens !== void 0 ? { tokens } : {} };
 }
 function toEvalIssue(issue) {
   return {
@@ -5885,7 +6399,8 @@ function toEvalIssue(issue) {
     subcategory: issue.dimension,
     severity: issue.severity === "critical" ? "critical" : "major",
     description: issue.description,
-    fix: issue.fix || ""
+    fix: issue.fix || "",
+    ...issue.origin !== void 0 ? { origin: issue.origin } : {}
   };
 }
 function cssTokensForAppTheme(theme, mode = "light", themeId) {
@@ -5925,6 +6440,14 @@ function criteriaContextFor(src) {
   };
 }
 
-export { CANVAS_BREAKPOINTS, CANVAS_CLASSES, CANVAS_DESCRIPTORS, CANVAS_VIEWPORTS, CARD_HEIGHT_EXPRESSION, CONTENT_HEIGHT_EXPRESSION, CRITERIA, CRITERIA_SELECTOR_VERSION, DEFAULT_EVAL_MAX_ROUNDS, DEFAULT_QUALITY_CONFIG, HOST_COLOUR_PATTERN, HOST_FRAME_BOUNDS, INK_TOLERANCE, INSTRUMENT_BY_ID, JUDGE_GROUND_MARGIN_PX, JUDGE_INK_INSET_PX, LLM_EVAL_STATIC_CRITERIA, MAX_EVAL_ROUNDS_HARD_LIMIT, VISUAL_EVAL_PROMPT, VISUAL_JUDGE_MAX_OUTPUT_TOKENS, VISUAL_JUDGE_PROMPT_DIGEST, VISUAL_JUDGE_PROMPT_VERSION, bankRows, buildCodingCriteriaSummary, buildCriteriaJudgeBlock, buildFixPrompt, callMultimodalLLM, canonicalCriteriaContext, canvasBlankIssue, canvasDisplayModeForShell, canvasFitPolicy, canvasForRendering, canvasForViewportWidth, canvasOverflowIssue, captureScreenshot, captureScreenshotDetailed, chromaOfCssTokens, computeEvaluationScore, createEvaluationToolsServer, criteriaBankSchema, criteriaContextFor, criteriaFor, criteriaSetIdFor, cssTokensForAppTheme, decodePng, describeCanvas, designTreeSha256, extractCompiledCode, extractCompiledCodeFromMessage, extractEvalResult, extractSourceCode, extractSourceCodeFromMessage, extractToolResultTexts, getActionableIssues, getCriteriaByPriority, getCriterionById, getEvaluatorSystemPrompt, getLLMCriteria, isBlocked, judgeDesignIdentity, judgeStoredCapture, matches, notApplicableCoverage, parseCriteriaBank, priorityForIssue, readInkExtent, resetJudgeDesignIdentityCache, resolveCriteriaBlock, resolveDesignPackageDir, resolveLaunchOptions, runAxisChecks, runEvaluation, runEvaluationLoop, runVisualEval, runVisualEvaluation, runVisualEvaluationDetailed, selectCriteria, summarizeVisualResult };
+// src/evaluation/judge-spend.ts
+function judgedFrameTokens(result) {
+  return {
+    inputTokens: (result.inputTokens ?? 0) + (result.criteriaTokens?.inputTokens ?? 0),
+    outputTokens: (result.outputTokens ?? 0) + (result.criteriaTokens?.outputTokens ?? 0)
+  };
+}
+
+export { CANVAS_BREAKPOINTS, CANVAS_CLASSES, CANVAS_DESCRIPTORS, CANVAS_VIEWPORTS, CARD_HEIGHT_EXPRESSION, CONTENT_HEIGHT_EXPRESSION, CRITERIA, CRITERIA_SELECTOR_VERSION, DEFAULT_EVAL_MAX_ROUNDS, DEFAULT_QUALITY_CONFIG, EDGE_BLOCKS_EXPRESSION, EDGE_TOLERANCE_PX, HOST_COLOUR_PATTERN, HOST_FRAME_BOUNDS, INK_TOLERANCE, INSTRUMENT_BESIDE_JUDGE, INSTRUMENT_BY_ID, JUDGE_GROUND_MARGIN_PX, JUDGE_INK_INSET_PX, LLM_EVAL_STATIC_CRITERIA, MAX_EVAL_ROUNDS_HARD_LIMIT, VISUAL_EVAL_PROMPT, VISUAL_JUDGE_MAX_OUTPUT_TOKENS, VISUAL_JUDGE_PROMPT_DIGEST, VISUAL_JUDGE_PROMPT_VERSION, bankRows, buildCodingCriteriaSummary, buildCriteriaJudgeBlock, buildFixPrompt, callMultimodalLLM, canonicalCriteriaContext, canvasBlankIssue, canvasDisplayModeForShell, canvasFitPolicy, canvasForRendering, canvasForViewportWidth, canvasOverflowIssue, captureScreenshot, captureScreenshotDetailed, chromaOfCssTokens, computeEvaluationScore, createEvaluationToolsServer, criteriaBankSchema, criteriaContextFor, criteriaFor, criteriaSetIdFor, cssTokensForAppTheme, decodePng, describeCanvas, designTreeSha256, extractCompiledCode, extractCompiledCodeFromMessage, extractEvalResult, extractSourceCode, extractSourceCodeFromMessage, extractToolResultTexts, getActionableIssues, getCriteriaByPriority, getCriterionById, getEvaluatorSystemPrompt, getLLMCriteria, isBlocked, judgeDesignIdentity, judgeEdges, judgeStoredCapture, judgedFrameTokens, matches, notApplicableCoverage, parseCriteriaBank, parseEdgeProbe, priorityForIssue, readInkExtent, resetJudgeDesignIdentityCache, resolveCriteriaBlock, resolveDesignPackageDir, resolveLaunchOptions, runAxisChecks, runEvaluation, runEvaluationLoop, runVisualEval, runVisualEvaluation, runVisualEvaluationDetailed, selectCriteria, summarizeVisualResult };
 //# sourceMappingURL=index.js.map
 //# sourceMappingURL=index.js.map