niceeval 0.9.1 → 0.10.0

package/src/report/{compute.ts → components/metric-views/compute.ts}

@@ -1,7 +1,7 @@
 // 计算函数(*Data):ReportInput → 一份组件数据。跑在 Node 侧,产物是算好的、可序列化的
-// 普通 JSON(终值 + 渲染提示,不含公式);渲染面(web/text)只做展示。
-// 它们是双面组件解析面的具名形式(MetricTable / metricTableData),与组件成对导出,
-// 只住在 niceeval/report(docs/feature/reports/library.md「数据计算与缓存边界」)。
+// 普通 JSON(终值 + 渲染提示,不含公式);渲染面(web/text)只做展示。指标图形族
+// (MetricTable / MetricMatrix / MetricBars / Scoreboard / MetricScatter / MetricLine /
+// DeltaTable)的 *Data 与配套 Options 都住在这里(docs/feature/reports/library/metric-views.md)。
 //
 // 共同约定(docs/feature/reports/architecture.md「指标聚合不变量」):
 // - 第一参收 ReportInput = Scope | readonly Snapshot[];warnings 不进组件数据(宿主统一显示);
@@ -12,19 +12,10 @@
 // - core 中立:只认 Metric / Dimension 接口,不出现具体 agent 名的分支。
 
 import type {
-  AttemptListItem,
-  AttemptLocator,
-  CopyFixPromptData,
   DeltaData,
   DeltaPair,
   DimensionInput,
-  EvalListItem,
-  ExperimentComparisonData,
-  ExperimentComparisonGroupData,
-  ExperimentListEvalRow,
-  ExperimentListItem,
   FlagPairs,
-  HeroData,
   LineData,
   MatrixData,
   Metric,
@@ -32,19 +23,14 @@ import type {
   NumericAxis,
   ReportInput,
   ScatterData,
-  ScopeSummaryData,
-  ScopeWarning,
   ScoreboardData,
+  SeriesInput,
   TableData,
-  TraceSpanSummary,
-  TraceWaterfallRow,
-  VerdictTally,
-} from "./types.ts";
-import type { EvalResult, JsonValue, TraceSpan } from "../types.ts";
-import type { Snapshot } from "../results/types.ts";
-import { comparabilityConfigOf, deepEqualJson } from "../results/select.ts";
-import { evalLevelStats, foldEvalVerdict } from "../shared/verdict.ts";
-import { experimentGroupOf } from "../shared/aggregate.ts";
+} from "../../model/types.ts";
+import type { JsonValue } from "../../../types.ts";
+import type { AttemptLocator } from "../../../results/locator.ts";
+import type { Snapshot } from "../../../results/types.ts";
+import { comparabilityConfigOf, deepEqualJson } from "../../../results/select.ts";
 import {
   assertUniqueMetricNames,
   axisValueOf,
@@ -52,10 +38,8 @@ import {
   computeCell,
   dimensionKey,
   dimensionName,
-  displayValue,
   evalGroupOf,
   evalIdOf,
-  evaluateMetric,
   experimentIdOf,
   filterItems,
   fullEvalKey,
@@ -63,15 +47,15 @@ import {
   locatorOf,
   refDisplayKey,
   resolveInput,
-  snapshotKeyOf,
+  seriesKey,
+  seriesName,
   toColumn,
   type Item,
-} from "./aggregate.ts";
-import { attemptCostUSD, costUSD, durationMs, endToEndPassRate, examScore, tokens } from "./metrics.ts";
-import { formatMetricValue, formatPlainNumber, localizedDisplay } from "./format.ts";
-import { compactAssertionSummary, primaryAssertionSummary, summaryText } from "../scoring/display.ts";
-import { defineMetric } from "./metrics.ts";
-import type { LocalizedText } from "./locale.ts";
+} from "../../model/aggregate.ts";
+import { examScore } from "../../model/metrics.ts";
+import { formatMetricValue, formatPlainNumber, localizedDisplay } from "../../model/format.ts";
+import type { LocalizedText } from "../../model/locale.ts";
+import { selectedEvalsOnly } from "../shared-compute.ts";
 
 // ───────────────────────── metricTableData ─────────────────────────
 
@@ -173,290 +157,6 @@ export async function metricMatrixData(input: ReportInput, options: MetricMatrix
   };
 }
 
-// ───────────────────────── 实体列表(experimentListData / evalListData / attemptListData)─────────────────────────
-
-/**
- * 一次 attempt 的单行结果摘要(Scoring display 契约):failed 取主失败断言摘要(不含
- * "+N more",N 单独进 moreFailures),errored 取结构化 error 的一层摘要
- * (phase · code · message),passed / skipped 为 null。
- */
-function failureSummaryOf(result: EvalResult): { summary: string | null; more: number } {
-  if (result.verdict === "errored" && result.error !== undefined) {
-    const parts = [result.error.phase, result.error.code, result.error.message].filter(
-      (part): part is string => typeof part === "string" && part.length > 0,
-    );
-    return { summary: summaryText(parts.join(" · ")), more: 0 };
-  }
-  if (result.verdict === "failed" || result.verdict === "errored") {
-    const primary = primaryAssertionSummary(result.assertions, result.verdict);
-    if (primary !== undefined) {
-      return {
-        summary: compactAssertionSummary({ ...primary, additionalFailures: 0 }),
-        more: primary.additionalFailures,
-      };
-    }
-    if (result.verdict === "errored" && result.skipReason !== undefined) {
-      return { summary: summaryText(result.skipReason), more: 0 };
-    }
-    return { summary: null, more: 0 };
-  }
-  return { summary: null, more: 0 };
-}
-
-/** AttemptList / ExperimentList / EvalList 共用的叶子构造:一个 Item → 一个 AttemptListItem。 */
-async function attemptListItemOf(item: Item): Promise<AttemptListItem> {
-  const result = item.attempt.result;
-  const { summary, more } = failureSummaryOf(result);
-  return {
-    experimentId: experimentIdOf(item),
-    evalId: evalIdOf(item),
-    attempt: result.attempt,
-    agent: result.agent,
-    verdict: result.verdict,
-    failureSummary: summary,
-    moreFailures: more,
-    examScore: await computeCell(examScore, [item]),
-    durationMs: result.durationMs,
-    costUSD: attemptCostUSD(result),
-    locator: locatorOf(item),
-  };
-}
-
-/** `attemptListData(input)`:每个 Attempt 一项,顺序取自 Scope 展平顺序(不重排)。 */
-export async function attemptListData(input: ReportInput): Promise<AttemptListItem[]> {
-  const { snapshots } = resolveInput(input);
-  const items = collectItems(snapshots);
-  return Promise.all(items.map((item) => attemptListItemOf(item)));
-}
-
-/** `evalListData(input)`:每个 `experimentId + evalId` 一项,按 evalId 再按 experimentId 升序。 */
-export async function evalListData(input: ReportInput): Promise<EvalListItem[]> {
-  const { snapshots } = resolveInput(input);
-  const items = collectItems(snapshots);
-  const groups = new Map<string, Item[]>();
-  for (const item of items) {
-    const key = fullEvalKey(item);
-    const list = groups.get(key);
-    if (list) list.push(item);
-    else groups.set(key, [item]);
-  }
-  const out: EvalListItem[] = [];
-  for (const group of groups.values()) {
-    const sorted = [...group].sort((a, b) => a.attempt.result.attempt - b.attempt.result.attempt);
-    const verdict = foldEvalVerdict(sorted.map((item) => item.attempt.result));
-    const attempts = await Promise.all(sorted.map((item) => attemptListItemOf(item)));
-    out.push({
-      experimentId: experimentIdOf(sorted[0]!),
-      evalId: evalIdOf(sorted[0]!),
-      verdict,
-      examScore: await computeCell(examScore, sorted),
-      durationMs: await computeCell(durationMs, sorted),
-      costUSD: await computeCell(costUSD, sorted),
-      attempts,
-    });
-  }
-  out.sort((a, b) => a.evalId.localeCompare(b.evalId) || a.experimentId.localeCompare(b.experimentId));
-  return out;
-}
-
-/**
- * `experimentListData(input)`:每个 experiment 一项,展开到每道 Eval;初始按端到端成功率
- * 从高到低(缺数据沉底,同分按 id)。一行只有一套 agent / model / flags 是输入约束:
- * 宿主注入的 current() Scope 保证每个 experiment 只由可比性配置一致的快照拼成;作者自选
- * Snapshot[] 时若同一 experiment 混入不一致的可比性配置,按完整用户反馈失败并指引——
- * 看跨配置演化用 snapshot 维度或 MetricLine,不把两套配置拼成一行冒充单一配置。
- */
-export async function experimentListData(input: ReportInput): Promise<ExperimentListItem[]> {
-  const { snapshots } = resolveInput(input);
-
-  // 可比性配置单义检查:同一 experiment 的输入快照必须共享一套可比性配置。
-  const configByExperiment = new Map<string, { snapshot: Snapshot; config: unknown }>();
-  for (const snapshot of snapshots) {
-    const config = comparabilityConfigOf(snapshot);
-    const existing = configByExperiment.get(snapshot.experimentId);
-    if (existing === undefined) {
-      configByExperiment.set(snapshot.experimentId, { snapshot, config });
-    } else if (!deepEqualJson(existing.config, config)) {
-      throw new Error(
-        `experimentListData got inconsistent comparability configs for experiment "${snapshot.experimentId}" ` +
-          `(snapshots ${existing.snapshot.startedAt} and ${snapshot.startedAt} differ in agent/model/reasoningEffort/flags/budget/timeoutMs/sandbox). ` +
-          "One row shows one configuration — it cannot honestly merge two. To chart evolution across configs, " +
-          'use the "snapshot" dimension or MetricLine; to show the current level, pass results.current() which selects a single config per experiment.',
-      );
-    }
-  }
-
-  const items = collectItems(snapshots);
-  const groups = groupItems(items, "experiment");
-  const out: ExperimentListItem[] = [];
-  for (const [experimentId, group] of groups) {
-    const stats = summarizeItems(group);
-    const newest = [...group].sort((a, b) => b.snapshot.startedAt.localeCompare(a.snapshot.startedAt))[0]!;
-    const evalGroups = groupItems(group, "eval");
-    const evalRows: ExperimentListEvalRow[] = [];
-    for (const [evalId, evalItems] of evalGroups) {
-      const sorted = [...evalItems].sort((a, b) => a.attempt.result.attempt - b.attempt.result.attempt);
-      const verdict = foldEvalVerdict(sorted.map((item) => item.attempt.result));
-      const attempts = await Promise.all(sorted.map((item) => attemptListItemOf(item)));
-      evalRows.push({
-        evalId,
-        verdict,
-        durationMs: await computeCell(durationMs, sorted),
-        costUSD: await computeCell(costUSD, sorted),
-        attempts,
-      });
-    }
-    const experiment = newest.snapshot.experiment ?? newest.attempt.result.experiment;
-    const model = newest.attempt.result.model ?? newest.snapshot.model;
-    out.push({
-      experimentId,
-      agent: newest.snapshot.agent || newest.attempt.result.agent,
-      ...(model !== undefined ? { model } : {}),
-      ...(experiment?.flags ? { flags: experiment.flags } : {}),
-      evalVerdicts: stats.verdicts,
-      endToEndPassRate: await computeCell(endToEndPassRate, group),
-      costUSD: await computeCell(costUSD, group),
-      durationMs: await computeCell(durationMs, group),
-      tokens: await computeCell(tokens, group),
-      evals: stats.evals,
-      attempts: stats.attempts,
-      lastRunAt: stats.lastRunAt!,
-      evalRows,
-    });
-  }
-  // 初始态按端到端成功率(endToEndPassRate)从高到低,缺数据沉底;同分按 experiment id 稳定排序。
-  out.sort((a, b) => {
-    const va = a.endToEndPassRate.value;
-    const vb = b.endToEndPassRate.value;
-    if (va === null && vb === null) return a.experimentId.localeCompare(b.experimentId);
-    if (va === null) return 1;
-    if (vb === null) return -1;
-    return vb - va || a.experimentId.localeCompare(b.experimentId);
-  });
-  return out;
-}
-
-// ───────────────────────── scopeSummaryData ─────────────────────────
-
-/** costUSD 的求和投影:两级都 sum(题内多轮求和 + 跨题求和 = 全量求和),display 走 $。 */
-const totalCostMetric = defineMetric({
-  name: "total-cost",
-  label: costUSD.label,
-  unit: "$",
-  value: costUSD.value,
-  aggregate: { perEval: "sum", acrossEvals: "sum" },
-});
-
-function tallyOf(): VerdictTally {
-  return { passed: 0, failed: 0, errored: 0, skipped: 0 };
-}
-
-/** 一批 Item 的组级统计(experimentListData / scopeSummaryData 共用)。 */
-function summarizeItems(items: Item[]): {
-  experiments: number;
-  evals: number;
-  attempts: number;
-  verdicts: VerdictTally;
-  lastRunAt: string | undefined;
-} {
-  const experimentIds = new Set<string>();
-  for (const item of items) experimentIds.add(experimentIdOf(item));
-  const stats = evalLevelStats(
-    items.map((item) => ({ verdict: item.attempt.result.verdict, key: fullEvalKey(item) })),
-    (r) => r.key,
-  );
-  let lastRunAt: string | undefined;
-  for (const item of items) {
-    const startedAt = item.snapshot.startedAt;
-    if (lastRunAt === undefined || startedAt > lastRunAt) lastRunAt = startedAt;
-  }
-  return {
-    experiments: experimentIds.size,
-    evals: stats.evals,
-    attempts: items.length,
-    verdicts: { passed: stats.passed, failed: stats.failed, errored: stats.errored, skipped: stats.skipped },
-    lastRunAt,
-  };
-}
-
-/**
- * `scopeSummaryData(input)`:范围摘要——快照时间窗、experiment / eval / attempt 数、
- * 两级判定计票、端到端成功率与总成本(docs/feature/reports/library/summaries.md)。
- * data 恒携带两级计票;成功率来自官方两级指标引擎,不从任一计票重算。
- */
-export async function scopeSummaryData(input: ReportInput): Promise<ScopeSummaryData> {
-  const { snapshots } = resolveInput(input);
-  const items = collectItems(snapshots);
-
-  let earliest: string | null = null;
-  let latest: string | null = null;
-  for (const snapshot of snapshots) {
-    if (earliest === null || snapshot.startedAt < earliest) earliest = snapshot.startedAt;
-    if (latest === null || snapshot.startedAt > latest) latest = snapshot.startedAt;
-  }
-
-  const stats = summarizeItems(items);
-  const attemptVerdicts = tallyOf();
-  for (const item of items) attemptVerdicts[item.attempt.result.verdict] += 1;
-
-  return {
-    range: { earliestStartedAt: earliest, latestStartedAt: latest },
-    experiments: stats.experiments,
-    evals: stats.evals,
-    attempts: stats.attempts,
-    evalVerdicts: stats.verdicts,
-    attemptVerdicts,
-    endToEndPassRate: await computeCell(endToEndPassRate, items),
-    totalCostUSD: await computeCell(totalCostMetric, items),
-  };
-}
-
-// ───────────────────────── experimentComparisonData ─────────────────────────
-
-/** 完整父路径是组键;没有父路径的 experiment 不能互相比,自己形成单例组。 */
-export function experimentComparisonGroupKey(experimentId: string): string {
-  return experimentGroupOf(experimentId) ?? experimentId;
-}
-
-/** 每组散点的唯一口径:默认 definition 与公开计算共用,不各写一份。 */
-const COMPARISON_SCATTER_OPTIONS: MetricScatterOptions = {
-  points: "experiment",
-  series: "agent",
-  x: costUSD,
-  y: endToEndPassRate,
-};
-
-/**
- * `experimentComparisonData(input)`:先把 input 按可比组分区(experiment id 的完整父路径),
- * 再为每组分别计算 ScopeSummary、成本 × 端到端成功率散点和 ExperimentList——分区发生在任何
- * 指标计算之前,组外 attempt 不可能污染该组的坐标尺度、series、成功率、成本、排序或缺数据计数。
- */
-export async function experimentComparisonData(input: ReportInput): Promise<ExperimentComparisonData> {
-  const { snapshots } = resolveInput(input);
-  const snapshotsByGroup = new Map<string, Snapshot[]>();
-  for (const snapshot of snapshots) {
-    const key = experimentComparisonGroupKey(snapshot.experimentId);
-    const group = snapshotsByGroup.get(key);
-    if (group) group.push(snapshot);
-    else snapshotsByGroup.set(key, [snapshot]);
-  }
-  const groups = await Promise.all(
-    [...snapshotsByGroup.entries()]
-      .sort(([a], [b]) => (a < b ? -1 : a > b ? 1 : 0))
-      .map(async ([key, groupSnapshots]): Promise<ExperimentComparisonGroupData> => {
-        const [summary, scatter, experiments] = await Promise.all([
-          scopeSummaryData(groupSnapshots),
-          metricScatterData(groupSnapshots, COMPARISON_SCATTER_OPTIONS),
-          experimentListData(groupSnapshots),
-        ]);
-        return { key, summary, scatter, experiments };
-      }),
-  );
-  return { groups };
-}
-
-// ───────────────────────── scoreboardData ─────────────────────────
-
 export interface ScoreboardOptions {
   rows: DimensionInput;
   /** 固定题集;eval id 必须唯一。元素引用运行时数据,类型放宽为普通数组,空数组在计算时报错。 */
@@ -644,13 +344,11 @@ function subjectDisplay(earned: number, possible: number): LocalizedText {
   return `${formatPlainNumber(earned)}/${formatPlainNumber(possible)} (${formatMetricValue(ratio, "%")})`;
 }
 
-// ───────────────────────── metricScatterData ─────────────────────────
-
 export interface MetricScatterOptions {
   /** 点维度:每个点 = 该组 attempt 的聚合。 */
   points: DimensionInput;
-  /** 可选:只决定颜色和分组,默认不连线。 */
-  series?: DimensionInput;
+  /** 决定颜色和图例归类,默认不连线(连线是呈现 prop `connect`);数组形态解析为复合维度。 */
+  series?: SeriesInput;
   x: Metric;
   y: Metric;
   /** eval id 前缀过滤,同 CLI 位置参数语义。 */
@@ -658,7 +356,7 @@ export interface MetricScatterOptions {
 }
 
 export async function metricScatterData(input: ReportInput, options: MetricScatterOptions): Promise<ScatterData> {
-  const { snapshots } = resolveInput(input);
+  const snapshots = selectedEvalsOnly(resolveInput(input).snapshots);
   const items = filterItems(collectItems(snapshots), options.evals);
   const groups = groupItems(items, options.points);
   const rows: ScatterData["rows"] = [];
@@ -666,26 +364,25 @@ export async function metricScatterData(input: ReportInput, options: MetricScatt
     rows.push({
       key,
       // 组内取第一条解析系列:点维度细于系列维度时(experiment ⊂ agent)天然一致
-      ...(options.series ? { series: dimensionKey(options.series, group[0]!) } : {}),
+      ...(options.series ? { series: seriesKey(options.series, group[0]!) } : {}),
       x: await computeCell(options.x, group),
       y: await computeCell(options.y, group), // 任一轴 null 的点留在 rows 里:组件不画,但注脚要报的数就从这里数
     });
   }
   return {
     pointDimension: dimensionName(options.points),
-    ...(options.series ? { seriesDimension: dimensionName(options.series) } : {}),
+    ...(options.series ? { seriesDimension: seriesName(options.series) } : {}),
     x: toColumn(options.x),
     y: toColumn(options.y),
     rows,
   };
 }
 
-// ───────────────────────── metricLineData ─────────────────────────
-
 export interface MetricLineOptions {
-  /** x 轴:NumericAxis(numericFlag() / numericRunConfig() 或自定义 of),不解析 experiment 命名。 */
+  /** x 轴:NumericAxis(numericFlag() / numericLabel() / numericRunConfig() 或自定义 of),不解析 experiment 命名。 */
   x: NumericAxis;
-  series?: DimensionInput;
+  /** 数组形态解析为复合维度。 */
+  series?: SeriesInput;
   y: Metric;
   /** eval id 前缀过滤,同 CLI 位置参数语义。 */
   evals?: string | readonly string[];
@@ -719,7 +416,7 @@ export async function metricLineData(input: ReportInput, options: MetricLineOpti
           "Fix of() to read experiment-level configuration (numericFlag()/numericRunConfig() do this by construction).",
       );
     }
-    const series = options.series ? dimensionKey(options.series, item) : undefined;
+    const series = options.series ? seriesKey(options.series, item) : undefined;
     const bucketKey = `${series ?? ""}${x === null ? "null" : String(x)}`;
     const bucket = buckets.get(bucketKey);
     if (bucket) bucket.items.push(item);
@@ -752,7 +449,7 @@ export async function metricLineData(input: ReportInput, options: MetricLineOpti
       label: options.x.label ?? options.x.name,
       ...(options.x.unit !== undefined ? { unit: options.x.unit } : {}),
     },
-    ...(options.series ? { seriesDimension: dimensionName(options.series) } : {}),
+    ...(options.series ? { seriesDimension: seriesName(options.series) } : {}),
     y: toColumn(options.y),
     rows,
   };
@@ -762,8 +459,10 @@ export async function metricLineData(input: ReportInput, options: MetricLineOpti
 
 /**
  * 按 flag 派生 A/B 对(docs/feature/reports/library/metric-views.md「DeltaTable」):
- * 配对域 = 同可比组 + 删除该 flag 后可比性配置深相等;a 取 baseline(缺省 = 未声明该 flag),
- * b 侧该 flag 的每个其它取值各成一对;label 自动 `<a 末段> · <flag>=<显示键>`。
+ * 配对域 = input Scope 内删除该 flag 后可比性配置深相等(不额外按 experiment id 的目录前缀分组
+ * ——architecture.md「Scope 是默认报告的比较边界」同一条契约);a 取 baseline(缺省 = 未声明该
+ * flag),b 侧该 flag 的每个其它取值各成一对;label 自动 `<完整 a experiment id> · <flag>=<显示键>`,
+ * 排序仍按 a 的末段、再按 flag 显示键(不受完整 id 的字符串差异影响)。
  */
 export function pairsByFlag(name: string, options?: { baseline?: JsonValue }): FlagPairs {
   if (typeof name !== "string" || name.length === 0) {
@@ -790,13 +489,13 @@ function isFlagPairs(pairs: DeltaTableOptions["pairs"]): pairs is FlagPairs {
   return typeof pairs === "object" && pairs !== null && !Array.isArray(pairs) && (pairs as FlagPairs).kind === "flagPairs";
 }
 
-/** experiment id 相对可比组的末段。 */
+/** experiment id 的末段(最后一个 `/` 之后;无 `/` 时是完整 id)。 */
 function experimentTail(experimentId: string): string {
   const slash = experimentId.lastIndexOf("/");
   return slash === -1 ? experimentId : experimentId.slice(slash + 1);
 }
 
-/** 派生配对:同可比组 + 删除该 flag 后可比性配置深相等。返回 pair 列表与配对域实验数。 */
+/** 派生配对:input Scope 内删除该 flag 后可比性配置深相等。返回 pair 列表与配对域实验数。 */
 function derivePairsByFlag(
   snapshots: readonly Snapshot[],
   spec: FlagPairs,
@@ -820,8 +519,9 @@ function derivePairsByFlag(
     const flagValue = config.flags?.[spec.flag];
     const reduced = { ...config, flags: { ...config.flags } };
     delete reduced.flags[spec.flag];
-    const group = experimentGroupOf(id) ?? id;
-    entries.push({ id, flagValue, bucket: `${group}${JSON.stringify(sortedJson(reduced))}` });
+    // 配对域只是 input Scope + 删除该 flag 后的可比性配置深相等;不额外按 experiment id 的
+    // 目录前缀分组——组件不从路径猜比较边界(architecture.md「Scope 是默认报告的比较边界」)。
+    entries.push({ id, flagValue, bucket: JSON.stringify(sortedJson(reduced)) });
   }
 
   const baseline = spec.baseline; // undefined = 未声明该 flag 的实验作 a
@@ -832,28 +532,34 @@ function derivePairsByFlag(
     else buckets.set(entry.bucket, [entry]);
   }
 
-  const pairs: DeltaPair[] = [];
+  interface DerivedPair {
+    a: string;
+    b: string;
+    label: string;
+    flagKey: string;
+  }
+  const derived: DerivedPair[] = [];
   for (const bucket of buckets.values()) {
     const aSide = bucket.filter((e) => deepEqualJson(e.flagValue, baseline));
     const bSide = bucket.filter((e) => !deepEqualJson(e.flagValue, baseline));
     for (const a of aSide) {
       for (const b of bSide) {
-        pairs.push({
-          a: a.id,
-          b: b.id,
-          label: `${experimentTail(a.id)} · ${spec.flag}=${refDisplayKey(b.flagValue)[0]}`,
-        });
+        const flagKey = refDisplayKey(b.flagValue)[0]!;
+        // label 用完整 a experiment id 自动命名,不截断成末段——目录前缀不同的两个实验
+        // 配成一对时,末段可能撞名,完整 id 才能唯一标识这一对来自哪个实验。
+        derived.push({ a: a.id, b: b.id, label: `${a.id} · ${spec.flag}=${flagKey}`, flagKey });
       }
     }
   }
-  pairs.sort((p, q) => {
+  // 排序只看 a 的末段与 flag 显示键,不看完整 label 字符串——完整 a id 只服务显示,
+  // 不该让目录前缀的字符串差异打乱本该相邻的 a 末段分组。
+  derived.sort((p, q) => {
     const ta = experimentTail(p.a);
     const tb = experimentTail(q.a);
     if (ta !== tb) return ta < tb ? -1 : 1;
-    const la = p.label as string;
-    const lb = q.label as string;
-    return la < lb ? -1 : la > lb ? 1 : 0;
+    return p.flagKey < q.flagKey ? -1 : p.flagKey > q.flagKey ? 1 : 0;
   });
+  const pairs: DeltaPair[] = derived.map(({ a, b, label }) => ({ a, b, label }));
   return { pairs, experiments: byExperiment.size };
 }
 
@@ -974,135 +680,3 @@ function deltaOutcome(metric: Metric, delta: number | null): "improved" | "regre
   const better = metric.better ?? "higher";
   return (delta > 0) === (better === "higher") ? "improved" : "regressed";
 }
-
-// ───────────────────────── 站点组件的计算函数(hero / warnings / fix prompt / trace)─────────────────────────
-
-/**
- * `heroData(input)`:站点标题区的运行 meta——`latestStartedAt` 取范围内最新快照的开始时间
- * (空范围为 null,不编造当前时间),`snapshots` 计贡献当前水位的快照数
- * (docs/feature/reports/library/site-components.md「HeroCard」)。
- */
-export async function heroData(input: ReportInput): Promise<HeroData> {
-  const { snapshots } = resolveInput(input);
-  let latest: string | null = null;
-  for (const snapshot of snapshots) {
-    if (latest === null || snapshot.startedAt > latest) latest = snapshot.startedAt;
-  }
-  return { latestStartedAt: latest, snapshots: snapshots.length };
-}
-
-/**
- * `scopeWarningsData(input)`:Scope 携带的挑选警告原样透出;`input` 是裸 `Snapshot[]` 时
- * 没有挑选过程、没有警告,返回空数组,也如实
- * (docs/feature/reports/library/site-components.md「ScopeWarnings」)。
- */
-export async function scopeWarningsData(input: ReportInput): Promise<readonly ScopeWarning[]> {
-  return resolveInput(input).warnings;
-}
-
-/**
- * `copyFixPromptData(input)`:把范围内全部失败(verdict 为 failed / errored 的 attempt)
- * 整理成一段可交给 coding agent 的修复 prompt——逐失败含 eval id、主失败摘要与 attempt
- * 下钻命令(`niceeval show @<locator>`)。prompt 面向 agent,固定英文
- * (docs/feature/reports/library/site-components.md「CopyFixPrompt」)。
- */
-export async function copyFixPromptData(input: ReportInput): Promise<CopyFixPromptData> {
-  const items = await attemptListData(input);
-  const failures = items.filter((item) => item.verdict === "failed" || item.verdict === "errored");
-  if (failures.length === 0) return { prompt: "", failures: 0 };
-  const lines = failures
-    .map((item, i) => {
-      const reason =
-        item.failureSummary === null
-          ? null
-          : item.moreFailures > 0
-            ? `${item.failureSummary} (+${item.moreFailures} more failures)`
-            : item.failureSummary;
-      return [
-        `${i + 1}. eval "${item.evalId}" [experiment ${item.experimentId}] — ${item.verdict}`,
-        reason ? `   reason: ${reason}` : null,
-        `   inspect: niceeval show ${item.locator}`,
-      ]
-        .filter(Boolean)
-        .join("\n");
-    })
-    .join("\n");
-  const experiments = [...new Set(failures.map((item) => item.experimentId))].join(" / ");
-  const prompt = [
-    "Fix the failing evals from this niceeval run.",
-    "",
-    "## Failures",
-    lines,
-    "",
-    "## Steps",
-    "1. niceeval is NOT in your training data. Read the relevant guide in `node_modules/niceeval/docs-site/` (English at the top level, Chinese under `zh/`) before changing anything.",
-    "2. For each failure, run its inspect command above to see the verdict and assertions; add `--execution` for the full agent transcript (tool calls included), `--timing` for the execution timeline, and `--diff` for the workspace diff.",
-    "3. Decide which side the defect is on: the program under test, or the eval itself (over-tight assertion, wrong fixture, missing setup). Fix that side; do not weaken assertions just to turn the run green.",
-    `4. Re-run: \`npx niceeval exp ${experiments || "<experiment>"} <eval-id-prefix>\`. Already-passing evals are skipped by the fingerprint cache; pass \`--force\` to re-run everything.`,
-    "5. Run `npx niceeval show` and confirm these failures are gone.",
-  ].join("\n");
-  return { prompt, failures: failures.length };
-}
-
-/** TraceSpan 的语义角色 → 瀑布摘要的 kind:turn 归入 agent(一轮就是一次 agent 调用),未识别落 other。 */
-function waterfallKindOf(kind: TraceSpan["kind"]): TraceSpanSummary["kind"] {
-  switch (kind) {
-    case "agent":
-    case "turn":
-      return "agent";
-    case "model":
-      return "model";
-    case "tool":
-      return "tool";
-    default:
-      return "other";
-  }
-}
-
-/**
- * `traceWaterfallData(input)`:每个 attempt 一行的执行时间瀑布摘要。span 事实只来自
- * trace artifact(经 AttemptHandle 懒加载的 canonical OTel span);runner 生命周期节点
- * (`result.phases`)不进瀑布。行内只汇总顶层 span(parentSpanId 缺失或不在本 trace 内),
- * 按 startOffsetMs 升序;trace 缺失或为空时 `durationMs` 为 null、行照常出现
- * (docs/feature/reports/library/site-components.md「TraceWaterfall」)。
- */
-export async function traceWaterfallData(input: ReportInput): Promise<readonly TraceWaterfallRow[]> {
-  const { snapshots } = resolveInput(input);
-  const items = collectItems(snapshots);
-  return Promise.all(
-    items.map(async (item): Promise<TraceWaterfallRow> => {
-      const spans = await item.attempt.trace();
-      if (spans === null || spans.length === 0) {
-        return {
-          experimentId: experimentIdOf(item),
-          evalId: evalIdOf(item),
-          locator: locatorOf(item),
-          durationMs: null,
-          spans: [],
-        };
-      }
-      const t0 = Math.min(...spans.map((s) => s.startMs));
-      const t1 = Math.max(...spans.map((s) => s.endMs));
-      const ids = new Set(spans.map((s) => s.spanId));
-      const topLevel = spans.filter((s) => s.parentSpanId === undefined || !ids.has(s.parentSpanId));
-      const summaries = topLevel
-        .map(
-          (s): TraceSpanSummary => ({
-            name: s.name,
-            kind: waterfallKindOf(s.kind),
-            startOffsetMs: s.startMs - t0,
-            durationMs: s.endMs - s.startMs,
-            failed: s.status === "error",
-          }),
-        )
-        .sort((a, b) => a.startOffsetMs - b.startOffsetMs);
-      return {
-        experimentId: experimentIdOf(item),
-        evalId: evalIdOf(item),
-        locator: locatorOf(item),
-        durationMs: Math.max(0, t1 - t0),
-        spans: summaries,
-      };
-    }),
-  );
-}