niceeval 0.9.1 → 0.10.2

package/dist/report/index.js

@@ -3,28 +3,37 @@
 //
 // import 边界即运行时边界:计算函数(*Data)会经句柄触碰文件系统(懒加载 artifact),
 // 只能进服务端 / 脚本;组件的渲染面纯同步零 IO。text 宿主遍历渲染不需要 react-dom
-// (renderReportToText);web 宿主的 renderReportToStaticHtml 在 ./web.ts,只有那一侧
+// (renderReportToText);web 宿主的 renderReportToStaticHtml 在 ./runtime/web.ts,只有那一侧
 // import react-dom。写报告文件的项目要装 react(.tsx 编译产物 import react/jsx-runtime)。
 // 指标与维度读取器
-export { assistantTurns, costUSD, defineMetric, durationMs, endToEndPassRate, examScore, executionReliability, repeatedFailedCommands, taskPassRate, tokens, } from "./metrics.js";
-export { flag, numericFlag, numericRunConfig, runConfig } from "./flag.js";
+export { assistantTurns, costUSD, defineMetric, durationMs, endToEndPassRate, examScore, executionReliability, repeatedFailedCommands, taskPassRate, tokens, } from "./model/metrics.js";
+export { flag, label, numericFlag, numericLabel, numericRunConfig, runConfig } from "./model/flag.js";
 // 报告定义与组件基座
-export { buildReportMeta, defineReport, isReportDefinition, pickReportPage, renderReportToText, renderReportTreeToText, reportTitleText, resolveReportTitle, ReportPageNotFoundError, DEFAULT_PAGE_ID, } from "./report.js";
-export { defineComponent, createTextContext, renderNodeToText, resolveReportTree, validateReportTree, ResolveMemo } from "./tree.js";
-// 排版原语(八个内置双面组件)
-export { Col, Row, Section, Style, Tab, Table, Tabs, Text } from "./primitives.js";
+export { buildReportMeta, defineReport, isReportDefinition, resolveReportTitle, DEFAULT_PAGE_ID, } from "./definition/report.js";
+export { pickReportPage, renderReportToText, renderReportTreeToText, reportTitleText, ReportPageNotFoundError, ReportPageNeedsLocatorError, } from "./runtime/text.js";
+export { defineComponent, createTextContext, renderNodeToText, resolveReportTree, validateReportTree, ResolveMemo } from "./definition/tree.js";
+// 排版原语(十个内置双面组件)
+export { Col, Grid, Row, Section, Stat, Style, Tab, Table, Tabs, Text } from "./definition/primitives.js";
 // 文本排版工具箱:自定义组件的 text 面用的就是官方组件那把尺子。
 // 表格有 <Table> 承担,这里只给表以外的形态用 —— 尤其别拿 String.prototype.padEnd 对齐:
 // 它数 UTF-16 码元不数显示列宽,一带中文列就撕歪。renderAlignedRows 刻意不导出。
-export { stringWidth, padDisplay as padEnd, padStartDisplay as padStart, wrapDisplay as wrapText, indentBlock as indent, textBar as bar, joinColumns as columns, } from "./text/layout.js";
+export { stringWidth, padDisplay as padEnd, padStartDisplay as padStart, wrapDisplay as wrapText, indentBlock as indent, textBar as bar, joinColumns as columns, } from "./model/text-layout.js";
 // locale:官方组件 chrome 文案的语言(内置词典覆盖 en / zh-CN,其它 locale 走回退)
-export { DEFAULT_REPORT_LOCALE, localizedTextEquals, resolveLocalizedText, resolveMetricLabel } from "./locale.js";
+export { DEFAULT_REPORT_LOCALE, localizedTextEquals, resolveLocalizedText, resolveMetricLabel } from "./model/locale.js";
 // 官方双面组件(spec / data 双形态;配套 *Data 计算函数在下面成对导出)
 // 与站点组件(Hero / HeroCard / PoweredBy / ScopeWarnings / CopyFixPrompt / TraceWaterfall)
-export { AttemptList, CopyFixPrompt, DeltaTable, EvalList, ExperimentComparison, ExperimentList, FailureList, Hero, HeroCard, MetricBars, MetricLine, MetricMatrix, MetricScatter, MetricTable, PoweredBy, Scoreboard, ScopeSummary, ScopeWarnings, TraceWaterfall, } from "./components.js";
+export { ScopeSummary, ExperimentComparison } from "./components/summaries/index.js";
+export { AttemptList, EvalList, ExperimentList, FailureList } from "./components/entity-lists/index.js";
+export { DeltaTable, MetricBars, MetricLine, MetricMatrix, MetricScatter, MetricTable, Scoreboard, } from "./components/metric-views/index.js";
+export { CopyFixPrompt, Hero, HeroCard, PoweredBy, ScopeWarnings, TraceWaterfall, } from "./components/site-components/index.js";
+// Attempt 详情组件族(docs/feature/reports/library/attempt-detail.md):11 个叶子 + 2 个
+// 只装配叶子的组合组件(AttemptAssessment / AttemptDetail),都从 niceeval/report 导出。
+export { AttemptAssertions, AttemptAssessment, AttemptConversation, AttemptDetail, AttemptDiagnostics, AttemptDiff, AttemptError, AttemptFixPrompt, AttemptSource, AttemptSummary, AttemptTimeline, AttemptTrace, AttemptUsage, } from "./components/attempt-detail/index.js";
 // 计算函数(组件解析面的具名形式,与组件成对;spec 形态下由管线代调,data 形态与
 // 嵌入场景下由作者手工调)
-export { attemptListData, copyFixPromptData, deltaTableData, evalListData, experimentComparisonData, experimentListData, heroData, metricLineData, metricMatrixData, metricScatterData, metricTableData, pairsByFlag, scopeSummaryData, scopeWarningsData, scoreboardData, traceWaterfallData, } from "./compute.js";
-// experiment id 的组键推导(id 的目录前缀,如 `compare/bub-low` 的 `compare`)。
-// 重新导出,让自定义报告能按同一份口径把 experiment 分组,不必自己重写这两行逻辑。
-export { experimentGroupOf } from "../shared/aggregate.js";
+export { scopeSummaryData } from "./components/summaries/compute.js";
+export { attemptListData, evalListData, experimentListData } from "./components/entity-lists/compute.js";
+export { deltaTableData, metricLineData, metricMatrixData, metricScatterData, metricTableData, pairsByFlag, scoreboardData, } from "./components/metric-views/compute.js";
+export { copyFixPromptData, heroData, scopeWarningsData, traceWaterfallData, } from "./components/site-components/compute.js";
+// Attempt 详情组件族的计算函数:输入恒为单个 AttemptEvidence,同步纯派生(不读文件、不 fetch)。
+export { attemptAssertionsData, attemptConversationData, attemptDiagnosticsData, attemptDiffData, attemptErrorData, attemptFixPromptData, attemptSourceData, attemptSummaryData, attemptTimelineData, attemptTraceData, attemptUsageData, } from "./components/attempt-detail/compute.js";

package/dist/report/{aggregate.d.ts → model/aggregate.d.ts}

@@ -1,8 +1,8 @@
-import type { AttemptHandle, ScopeWarning, Snapshot } from "../results/types.ts";
-import { type AttemptLocator } from "../results/locator.ts";
-import type { Aggregator, DimensionInput, Metric, MetricCell, MetricColumn, NumericAxis, ReportInput } from "./types.ts";
+import type { AttemptHandle, ScopeWarning, Snapshot } from "../../results/types.ts";
+import { type AttemptLocator } from "../../results/locator.ts";
+import type { Aggregator, DimensionInput, Metric, MetricCell, MetricColumn, NumericAxis, ReportInput, SeriesInput } from "./types.ts";
 import { type LocalizedText } from "./locale.ts";
-import { evalPrefixPredicate } from "../shared/aggregate.ts";
+import { evalPrefixPredicate } from "../../shared/aggregate.ts";
 export declare function resolveInput(input: ReportInput): {
     snapshots: readonly Snapshot[];
     warnings: readonly ScopeWarning[];
@@ -42,6 +42,14 @@ export declare const MISSING_GROUP_KEY: string;
  */
 export declare function refDisplayKey(value: unknown): [display: string, canonical: string];
 export declare function dimensionKey(dimension: DimensionInput, item: Item): string;
+/**
+ * series 类选项(SeriesInput)的复合维度解析:数组形态的维度 name 依声明顺序以 ` × ` 连接,
+ * 每个 attempt 的维度值为各成员显示键以 ` · ` 连接(缺失成员沿用 `(missing)` 显示键参与
+ * 连接,与单维度缺失同一条规则)。单维度形态与 dimensionName / dimensionKey 严格等价。
+ */
+export declare function seriesMembers(series: SeriesInput): readonly DimensionInput[];
+export declare function seriesName(series: SeriesInput): string;
+export declare function seriesKey(series: SeriesInput, item: Item): string;
 /**
  * 按维度分组;维度 domain 按稳定 key 字典序(Unicode)排列,不让文件扫描顺序渗进报告。
  * flag / runConfig 维度做显示键冲突检测:不同原始值生成同一显示键时报错并要求改用

package/dist/report/{aggregate.js → model/aggregate.js}

@@ -5,12 +5,12 @@
 // 为什么是两级:earlyExit 默认开,失败的题天然比通过的题样本多;平铺求均值会把分数
 // 和重试策略纠缠在一起(eval A=[1]、eval B=[0,0,0] 平铺 = 0.25,两级宏平均 = 0.5)。
 // 自定义维度把同一道题的 attempt 分进不同组时,第一级折叠发生在各组内部。
-import { dedupeAttempts } from "../results/select.js";
-import { encodeAttemptLocator } from "../results/locator.js";
-import { flagValueOf, runConfigValueOf } from "./flag.js";
+import { dedupeAttempts } from "../../results/select.js";
+import { encodeAttemptLocator } from "../../results/locator.js";
+import { flagValueOf, labelValueOf, runConfigValueOf } from "./flag.js";
 import { formatMetricValue, localizedDisplay } from "./format.js";
 import { localeText } from "./locale.js";
-import { evalPrefixPredicate } from "../shared/aggregate.js";
+import { evalPrefixPredicate } from "../../shared/aggregate.js";
 // 复合键分隔符:NUL 不会出现在 eval id / experimentId / ISO 时间里,拼接键不会串味
 const KEY_SEP = "\u0000";
 export function resolveInput(input) {
@@ -84,7 +84,9 @@ export function evalGroupOf(id) {
 function isDimensionRef(dimension) {
     return (typeof dimension === "object" &&
         "kind" in dimension &&
-        (dimension.kind === "flag" || dimension.kind === "runConfig"));
+        (dimension.kind === "flag" ||
+            dimension.kind === "runConfig" ||
+            dimension.kind === "label"));
 }
 /** flag / runConfig 未声明时的显示键:内置文案,en / zh-CN 同形。 */
 export const MISSING_GROUP_KEY = localeText("en", "cell.missingValue");
@@ -110,9 +112,11 @@ export function refDisplayKey(value) {
     return [canonicalJson(value), `json:${canonicalJson(value)}`];
 }
 function refValueOf(ref, item) {
-    return ref.kind === "flag"
-        ? flagValueOf(item.attempt, ref.name)
-        : runConfigValueOf(item.attempt, ref.name);
+    if (ref.kind === "flag")
+        return flagValueOf(item.attempt, ref.name);
+    if (ref.kind === "label")
+        return labelValueOf(item.attempt, ref.name);
+    return runConfigValueOf(item.attempt, ref.name);
 }
 export function dimensionKey(dimension, item) {
     if (typeof dimension !== "string") {
@@ -141,6 +145,22 @@ export function dimensionKey(dimension, item) {
         }
     }
 }
+/**
+ * series 类选项(SeriesInput)的复合维度解析:数组形态的维度 name 依声明顺序以 ` × ` 连接,
+ * 每个 attempt 的维度值为各成员显示键以 ` · ` 连接(缺失成员沿用 `(missing)` 显示键参与
+ * 连接,与单维度缺失同一条规则)。单维度形态与 dimensionName / dimensionKey 严格等价。
+ */
+export function seriesMembers(series) {
+    return Array.isArray(series) ? series : [series];
+}
+export function seriesName(series) {
+    return seriesMembers(series).map(dimensionName).join(" × ");
+}
+export function seriesKey(series, item) {
+    return seriesMembers(series)
+        .map((member) => dimensionKey(member, item))
+        .join(" · ");
+}
 /**
  * 按维度分组;维度 domain 按稳定 key 字典序(Unicode)排列,不让文件扫描顺序渗进报告。
  * flag / runConfig 维度做显示键冲突检测:不同原始值生成同一显示键时报错并要求改用

package/dist/report/{flag.d.ts → model/flag.d.ts}

@@ -1,7 +1,9 @@
-import type { AttemptHandle } from "../results/types.ts";
+import type { AttemptHandle } from "../../results/types.ts";
 import type { DimensionOptions, DimensionRef, NumericAxis, NumericAxisOptions, NumericRunConfigAxisOptions, RunConfigKey } from "./types.ts";
 /** experiment 声明的 flags(快照级投影优先;第三方落盘只拼在条目上时回退 result.experiment)。 */
 export declare function flagValueOf(attempt: AttemptHandle, name: string): unknown;
+/** experiment 声明的报告归类标注 labels(同 flags 的读取回退链;值域 string | number)。 */
+export declare function labelValueOf(attempt: AttemptHandle, name: string): unknown;
 /**
  * runConfig() 的取值:读快照的 `ExperimentRunInfo` 投影,外加桥接到快照顶层权威字段的
  * `model` / `agent` 两个键(与 "model" / "agent" 内置维度同一读法,不另造第二套口径)。
@@ -14,6 +16,12 @@ export declare function runConfigValueOf(attempt: AttemptHandle, name: RunConfig
  * 缺失值显示内置文案 `(missing)`;显示键冲突时计算报错并要求改用 CustomDimension。
  */
 export declare function flag(name: string, options?: DimensionOptions): DimensionRef;
+/**
+ * 把 experiment 声明的一个报告标注(`ExperimentDef.labels` 的键)当分组维度,与 {@link flag}
+ * 同一套用法。labels 是纯报告侧的归类坐标:不透传运行时、不参与可比性配置;报告不从
+ * experiment id 字符串猜语义,归类只认声明(docs/feature/experiments/library.md「labels」)。
+ */
+export declare function label(name: string, options?: DimensionOptions): DimensionRef;
 /**
  * 把一项顶层运行配置当分组维度,与 {@link flag} 同一套用法。读快照的 `ExperimentRunInfo`
  * 投影;可用键由 RunConfigKey 在类型层穷尽(那张接口的字段全集,外加桥接到快照顶层权威
@@ -25,6 +33,11 @@ export declare function runConfig(name: RunConfigKey, options?: DimensionOptions
  * (折线不绘该点并报告缺失),不猜 `low < medium < high`。
  */
 export declare function numericFlag(name: string, options?: NumericAxisOptions): NumericAxis;
+/**
+ * MetricLine 的 x 轴:只接受 number 值的 label。labels 由作者亲手声明,要数值轴就直接
+ * 声明成 number,不设 map;字符串值返回 null(折线不绘该点并报告缺失)。
+ */
+export declare function numericLabel(name: string, options?: NumericAxisOptions): NumericAxis;
 /**
  * 顶层运行配置作数值轴:数值配置直接返回该值;字符串配置必须显式给 `map`
  * (如 { low: 1, medium: 2, high: 3 })。未声明、未投影、非数值或未命中 map 的值返回 null。

package/dist/report/{flag.js → model/flag.js}

@@ -8,6 +8,11 @@ export function flagValueOf(attempt, name) {
     const info = attempt.snapshot?.experiment ?? attempt.result.experiment;
     return info?.flags?.[name];
 }
+/** experiment 声明的报告归类标注 labels(同 flags 的读取回退链;值域 string | number)。 */
+export function labelValueOf(attempt, name) {
+    const info = attempt.snapshot?.experiment ?? attempt.result.experiment;
+    return info?.labels?.[name];
+}
 /**
  * runConfig() 的取值:读快照的 `ExperimentRunInfo` 投影,外加桥接到快照顶层权威字段的
  * `model` / `agent` 两个键(与 "model" / "agent" 内置维度同一读法,不另造第二套口径)。
@@ -41,6 +46,20 @@ export function flag(name, options) {
         ...(options?.unit !== undefined ? { unit: options.unit } : {}),
     };
 }
+/**
+ * 把 experiment 声明的一个报告标注(`ExperimentDef.labels` 的键)当分组维度,与 {@link flag}
+ * 同一套用法。labels 是纯报告侧的归类坐标:不透传运行时、不参与可比性配置;报告不从
+ * experiment id 字符串猜语义,归类只认声明(docs/feature/experiments/library.md「labels」)。
+ */
+export function label(name, options) {
+    assertName(name, "label", "the key declared in the experiment's labels");
+    return {
+        kind: "label",
+        name,
+        ...(options?.label !== undefined ? { label: options.label } : {}),
+        ...(options?.unit !== undefined ? { unit: options.unit } : {}),
+    };
+}
 /**
  * 把一项顶层运行配置当分组维度,与 {@link flag} 同一套用法。读快照的 `ExperimentRunInfo`
  * 投影;可用键由 RunConfigKey 在类型层穷尽(那张接口的字段全集,外加桥接到快照顶层权威
@@ -71,6 +90,22 @@ export function numericFlag(name, options) {
         },
     };
 }
+/**
+ * MetricLine 的 x 轴:只接受 number 值的 label。labels 由作者亲手声明,要数值轴就直接
+ * 声明成 number,不设 map;字符串值返回 null(折线不绘该点并报告缺失)。
+ */
+export function numericLabel(name, options) {
+    assertName(name, "numericLabel", "the key declared in the experiment's labels");
+    return {
+        name,
+        ...(options?.label !== undefined ? { label: options.label } : {}),
+        ...(options?.unit !== undefined ? { unit: options.unit } : {}),
+        of(attempt) {
+            const value = labelValueOf(attempt, name);
+            return typeof value === "number" && Number.isFinite(value) ? value : null;
+        },
+    };
+}
 /**
  * 顶层运行配置作数值轴:数值配置直接返回该值;字符串配置必须显式给 `map`
  * (如 { low: 1, medium: 2, high: 3 })。未声明、未投影、非数值或未命中 map 的值返回 null。

package/dist/report/{format.d.ts → model/format.d.ts}

@@ -1,13 +1,12 @@
-import type { Verdict } from "../types.ts";
+import type { Verdict } from "../../types.ts";
 import { type LocalizedText, type ReportLocale } from "./locale.ts";
 /**
- * experiment 行的显示名：给了父路径 `relativeTo` 且它确是前缀，就去掉 `relativeTo + "/"`，
- * 只留 id 末段——用在已经以组为标题的上下文（如默认 `ExperimentComparison` 的每组面板）里，
- * 避免每行重复文件夹名。组键就是 experiment id 的父目录，因此这里的末段与 `MetricScatter`
- * 点标签取的末段同源。不给 `relativeTo`、或它不是前缀（如根目录单例组）时原样返回完整 id。
- * 完整 id 仍是排序 / 着色 / 折叠的键，调用方不要拿这个显示名当身份用。
+ * 一组 id 的显示名：每个 id 缩成在这组里唯一的最短路径后缀，重名逐步加长到能区分为止
+ * （与 `MetricScatter` 点标签同一算法，两处共用本函数以保证同一份 experiment id 在散点和
+ * 列表里缩成同一个显示名）。单个 id、或所有 id 深度不同时也照常缩到各自的最短唯一后缀。
+ * 完整 id 不受影响，调用方仍用它做排序 / 过滤 / 折叠的身份键，这里只产出显示名。
  */
-export declare function experimentDisplayName(experimentId: string, relativeTo?: string): string;
+export declare function shortestUniqueLabels(ids: readonly string[]): Map<string, string>;
 export declare function formatMetricValue(value: number, unit?: string): string;
 /**
  * MetricCell.display 的生成:为官方生成面覆盖的每个 locale(DISPLAY_LOCALES)各生成一份;
@@ -25,6 +24,13 @@ export declare function formatDurationMs(ms: number): string;
 export declare function formatUSD(usd: number): string;
 /** 0..1 的比率 → 整数百分比。 */
 export declare function formatPercent(ratio: number): string;
+/** ISO 时间 → 本地化到分钟的短日期时间；不可解析时原样返回。 */
+export declare function formatReportDateTime(iso: string, locale: ReportLocale): string;
+/** 时间范围的两端；同日省略右侧日期，同年省略右侧年份，减少卡片中的重复噪音。 */
+export declare function formatReportDateTimeRange(fromIso: string, toIso: string, locale: ReportLocale): {
+    from: string;
+    to: string;
+};
 /** passed / failed / errored / skipped 的判定符。 */
 export declare function verdictMark(verdict: Verdict): string;
 /**

package/dist/report/{format.js → model/format.js}

@@ -3,17 +3,42 @@
 // metric.display 可整体覆盖;这里只负责默认。
 import { DISPLAY_LOCALES } from "./locale.js";
 /**
- * experiment 行的显示名：给了父路径 `relativeTo` 且它确是前缀，就去掉 `relativeTo + "/"`，
- * 只留 id 末段——用在已经以组为标题的上下文（如默认 `ExperimentComparison` 的每组面板）里，
- * 避免每行重复文件夹名。组键就是 experiment id 的父目录，因此这里的末段与 `MetricScatter`
- * 点标签取的末段同源。不给 `relativeTo`、或它不是前缀（如根目录单例组）时原样返回完整 id。
- * 完整 id 仍是排序 / 着色 / 折叠的键，调用方不要拿这个显示名当身份用。
+ * 一组 id 的显示名：每个 id 缩成在这组里唯一的最短路径后缀，重名逐步加长到能区分为止
+ * （与 `MetricScatter` 点标签同一算法，两处共用本函数以保证同一份 experiment id 在散点和
+ * 列表里缩成同一个显示名）。单个 id、或所有 id 深度不同时也照常缩到各自的最短唯一后缀。
+ * 完整 id 不受影响，调用方仍用它做排序 / 过滤 / 折叠的身份键，这里只产出显示名。
  */
-export function experimentDisplayName(experimentId, relativeTo) {
-    if (relativeTo && experimentId.startsWith(`${relativeTo}/`)) {
-        return experimentId.slice(relativeTo.length + 1);
+export function shortestUniqueLabels(ids) {
+    const segsOf = (id) => id.split("/").filter(Boolean);
+    const depth = new Map(ids.map((id) => [id, 1]));
+    for (;;) {
+        const byLabel = new Map();
+        for (const id of ids) {
+            const segs = segsOf(id);
+            const label = segs.slice(-Math.min(depth.get(id), segs.length)).join("/") || id;
+            byLabel.set(label, [...(byLabel.get(label) ?? []), id]);
+        }
+        let grew = false;
+        for (const group of byLabel.values()) {
+            if (group.length < 2)
+                continue;
+            for (const id of group) {
+                const segs = segsOf(id);
+                if (depth.get(id) < segs.length) {
+                    depth.set(id, depth.get(id) + 1);
+                    grew = true;
+                }
+            }
+        }
+        if (!grew) {
+            const out = new Map();
+            for (const id of ids) {
+                const segs = segsOf(id);
+                out.set(id, segs.slice(-Math.min(depth.get(id), segs.length)).join("/") || id);
+            }
+            return out;
+        }
     }
-    return experimentId;
 }
 /** 一位小数、去掉无意义的 ".0" 尾巴。 */
 function trimmed(n) {
@@ -103,6 +128,51 @@ export function formatUSD(usd) {
 export function formatPercent(ratio) {
     return `${Math.round(ratio * 100)}%`;
 }
+function formatReportDate(date, locale, options) {
+    try {
+        return new Intl.DateTimeFormat(locale, options).format(date);
+    }
+    catch {
+        return new Intl.DateTimeFormat("en", options).format(date);
+    }
+}
+const FULL_REPORT_DATE_TIME = {
+    year: "numeric",
+    month: "short",
+    day: "numeric",
+    hour: "2-digit",
+    minute: "2-digit",
+    hourCycle: "h23",
+};
+/** ISO 时间 → 本地化到分钟的短日期时间；不可解析时原样返回。 */
+export function formatReportDateTime(iso, locale) {
+    const date = new Date(iso);
+    return Number.isNaN(date.valueOf()) ? iso : formatReportDate(date, locale, FULL_REPORT_DATE_TIME);
+}
+/** 时间范围的两端；同日省略右侧日期，同年省略右侧年份，减少卡片中的重复噪音。 */
+export function formatReportDateTimeRange(fromIso, toIso, locale) {
+    const fromDate = new Date(fromIso);
+    const toDate = new Date(toIso);
+    if (Number.isNaN(fromDate.valueOf()) || Number.isNaN(toDate.valueOf())) {
+        return { from: formatReportDateTime(fromIso, locale), to: formatReportDateTime(toIso, locale) };
+    }
+    const sameYear = fromDate.getFullYear() === toDate.getFullYear();
+    const sameDay = sameYear && fromDate.getMonth() === toDate.getMonth() && fromDate.getDate() === toDate.getDate();
+    const toOptions = sameDay
+        ? { hour: "2-digit", minute: "2-digit", hourCycle: "h23" }
+        : {
+            month: "short",
+            day: "numeric",
+            hour: "2-digit",
+            minute: "2-digit",
+            hourCycle: "h23",
+            ...(!sameYear && { year: "numeric" }),
+        };
+    return {
+        from: formatReportDate(fromDate, locale, FULL_REPORT_DATE_TIME),
+        to: formatReportDate(toDate, locale, toOptions),
+    };
+}
 // ── 实体列表(ExperimentList / EvalList / AttemptList)共用的判定符 ──
 /** passed / failed / errored / skipped 的判定符。 */
 export function verdictMark(verdict) {

package/dist/report/{locale.d.ts → model/locale.d.ts}

@@ -1,4 +1,4 @@
-import type { LocalizedText } from "../types.ts";
+import type { LocalizedText } from "../../types.ts";
 export type { LocalizedText };
 /** 报告渲染的 locale(BCP 47 标签,开放);默认 "en"(`niceeval show` 缺省输出不变)。 */
 export type ReportLocale = string;
@@ -18,6 +18,8 @@ declare const en: {
     readonly "verdict.failed": "failed";
     readonly "verdict.errored": "errored";
     readonly "verdict.skipped": "skipped";
+    /** AttemptSummary 的重试序号(0 起的 identity.attempt 显示前 +1)。 */
+    readonly "attemptSummary.attempt": "attempt {n}";
     /** 全 null / 无样本的统一文案,绝不画 0。 */
     readonly "cell.missing": "no data";
     readonly "cell.measuredTitle": "{samples}/{total} attempts measured";
@@ -36,11 +38,12 @@ declare const en: {
     /** <Table> 的 locator 列表头(行带 locator 时自动追加)。 */
     readonly "table.attempt": "attempt";
     readonly "experimentList.experiment": "Experiment";
-    readonly "experimentList.avgDuration": "Avg duration";
-    readonly "experimentList.passRate": "End-to-end pass rate";
+    readonly "experimentList.avgDuration": "Avg. time";
+    readonly "experimentList.passRate": "Pass rate";
+    readonly "experimentList.passRateDescription": "End-to-end pass rate: passed = 1; failed and errored = 0";
     readonly "experimentList.tokens": "Tokens";
     readonly "experimentList.cost": "Cost";
-    readonly "experimentList.result": "Result";
+    readonly "experimentList.result": "Results";
     readonly "experimentList.status": "Status";
     readonly "experimentList.evalAttempt": "Eval / Attempt";
     readonly "experimentList.duration": "Duration";
@@ -56,30 +59,27 @@ declare const en: {
     /** ScopeSummary 的 KPI 标签。 */
     readonly "scopeSummary.experiments": "Experiments";
     readonly "scopeSummary.evals": "Evals";
-    readonly "scopeSummary.attempts": "attempts";
-    readonly "scopeSummary.passRate": "End-to-end pass rate";
+    readonly "scopeSummary.attempts": "Attempts";
+    readonly "scopeSummary.passRate": "Pass rate";
     readonly "scopeSummary.totalCost": "Total cost";
-    readonly "scopeSummary.range": "{from} – {to}";
-    readonly "scopeSummary.votesEval": "eval verdicts";
-    readonly "scopeSummary.votesAttempt": "attempt verdicts";
+    readonly "scopeSummary.lastRun": "Last run · {time}";
+    readonly "scopeSummary.runRange": "Run range · {from} – {to}";
+    readonly "scopeSummary.costCoverage": "Cost available for {samples}/{total} attempts";
+    readonly "scopeSummary.votesEval": "Eval results";
+    readonly "scopeSummary.votesAttempt": "Attempt results";
     readonly "overview.experiments.one": "{n} experiment";
     readonly "overview.experiments.other": "{n} experiments";
     readonly "overview.evalsCount": "{n} evals";
     readonly "overview.attemptsCount": "{n} attempts";
     readonly "overview.evals": "Evals";
-    readonly "overview.passRate": "End-to-end pass rate";
+    readonly "overview.passRate": "Pass rate";
     readonly "overview.totalCost": "Cost";
-    readonly "experimentComparison.groups": "Experiment groups";
-    readonly "experimentComparison.group": "Experiment group";
-    readonly "experimentComparison.results": "Eval results";
-    readonly "experimentComparison.lastRun": "Last run";
-    readonly "experimentComparison.selectGroup": "Select experiment group {group}";
-    readonly "experimentComparison.commandsHead": "Group details:";
-    readonly "experimentComparison.empty": "No experiment groups";
     /** 方向提示唯一文案:轴向已随 better 反正,「更好」恒指向右上(两轴都声明 better 时才显示)。 */
     readonly "scatter.betterUpperRight": "better → upper right";
     /** 0 个可画点:x/y 指标没有可用数据。 */
     readonly "scatter.noData": "No data to plot {x} × {y}";
+    /** 标题行尾的归类维度标注(series 维度存在时显示)。 */
+    readonly "scatter.groupedBy": "grouped by {dim}";
     readonly "pointsMissing.one": "{n} point missing data";
     readonly "pointsMissing.other": "{n} points missing data";
     readonly "attemptList.empty": "No attempts";

package/dist/report/{locale.js → model/locale.js}

@@ -43,6 +43,8 @@ const en = {
     "verdict.failed": "failed",
     "verdict.errored": "errored",
     "verdict.skipped": "skipped",
+    /** AttemptSummary 的重试序号(0 起的 identity.attempt 显示前 +1)。 */
+    "attemptSummary.attempt": "attempt {n}",
     /** 全 null / 无样本的统一文案,绝不画 0。 */
     "cell.missing": "no data",
     "cell.measuredTitle": "{samples}/{total} attempts measured",
@@ -61,11 +63,12 @@ const en = {
     /** <Table> 的 locator 列表头(行带 locator 时自动追加)。 */
     "table.attempt": "attempt",
     "experimentList.experiment": "Experiment",
-    "experimentList.avgDuration": "Avg duration",
-    "experimentList.passRate": "End-to-end pass rate",
+    "experimentList.avgDuration": "Avg. time",
+    "experimentList.passRate": "Pass rate",
+    "experimentList.passRateDescription": "End-to-end pass rate: passed = 1; failed and errored = 0",
     "experimentList.tokens": "Tokens",
     "experimentList.cost": "Cost",
-    "experimentList.result": "Result",
+    "experimentList.result": "Results",
     "experimentList.status": "Status",
     "experimentList.evalAttempt": "Eval / Attempt",
     "experimentList.duration": "Duration",
@@ -81,30 +84,27 @@ const en = {
     /** ScopeSummary 的 KPI 标签。 */
     "scopeSummary.experiments": "Experiments",
     "scopeSummary.evals": "Evals",
-    "scopeSummary.attempts": "attempts",
-    "scopeSummary.passRate": "End-to-end pass rate",
+    "scopeSummary.attempts": "Attempts",
+    "scopeSummary.passRate": "Pass rate",
     "scopeSummary.totalCost": "Total cost",
-    "scopeSummary.range": "{from} – {to}",
-    "scopeSummary.votesEval": "eval verdicts",
-    "scopeSummary.votesAttempt": "attempt verdicts",
+    "scopeSummary.lastRun": "Last run · {time}",
+    "scopeSummary.runRange": "Run range · {from} – {to}",
+    "scopeSummary.costCoverage": "Cost available for {samples}/{total} attempts",
+    "scopeSummary.votesEval": "Eval results",
+    "scopeSummary.votesAttempt": "Attempt results",
     "overview.experiments.one": "{n} experiment",
     "overview.experiments.other": "{n} experiments",
     "overview.evalsCount": "{n} evals",
     "overview.attemptsCount": "{n} attempts",
     "overview.evals": "Evals",
-    "overview.passRate": "End-to-end pass rate",
+    "overview.passRate": "Pass rate",
     "overview.totalCost": "Cost",
-    "experimentComparison.groups": "Experiment groups",
-    "experimentComparison.group": "Experiment group",
-    "experimentComparison.results": "Eval results",
-    "experimentComparison.lastRun": "Last run",
-    "experimentComparison.selectGroup": "Select experiment group {group}",
-    "experimentComparison.commandsHead": "Group details:",
-    "experimentComparison.empty": "No experiment groups",
     /** 方向提示唯一文案:轴向已随 better 反正,「更好」恒指向右上(两轴都声明 better 时才显示)。 */
     "scatter.betterUpperRight": "better → upper right",
     /** 0 个可画点:x/y 指标没有可用数据。 */
     "scatter.noData": "No data to plot {x} × {y}",
+    /** 标题行尾的归类维度标注(series 维度存在时显示)。 */
+    "scatter.groupedBy": "grouped by {dim}",
     "pointsMissing.one": "{n} point missing data",
     "pointsMissing.other": "{n} points missing data",
     "attemptList.empty": "No attempts",
@@ -174,6 +174,7 @@ const zhCN = {
     "verdict.failed": "失败",
     "verdict.errored": "错误",
     "verdict.skipped": "跳过",
+    "attemptSummary.attempt": "第 {n} 次",
     "cell.missing": "无数据",
     "cell.measuredTitle": "{samples}/{total} 次 attempt 测得",
     "cell.noneMeasurableTitle": "0/{total} 次 attempt 可测",
@@ -191,7 +192,8 @@ const zhCN = {
     "table.attempt": "Attempt",
     "experimentList.experiment": "实验",
     "experimentList.avgDuration": "平均耗时",
-    "experimentList.passRate": "端到端成功率",
+    "experimentList.passRate": "通过率",
+    "experimentList.passRateDescription": "端到端通过率：passed = 1；failed 和 errored = 0",
     "experimentList.tokens": "Tokens",
     "experimentList.cost": "成本",
     "experimentList.result": "结果",
@@ -206,30 +208,26 @@ const zhCN = {
     "entityList.moreFailures.other": "+{n} 条其它失败",
     "table.columnsHidden.one": "(还有 {n} 列未列出)",
     "table.columnsHidden.other": "(还有 {n} 列未列出)",
-    "scopeSummary.experiments": "实验数",
-    "scopeSummary.evals": "题目",
-    "scopeSummary.attempts": "次 attempt",
-    "scopeSummary.passRate": "端到端成功率",
+    "scopeSummary.experiments": "实验",
+    "scopeSummary.evals": "Eval",
+    "scopeSummary.attempts": "Attempt",
+    "scopeSummary.passRate": "通过率",
     "scopeSummary.totalCost": "总成本",
-    "scopeSummary.range": "{from} – {to}",
-    "scopeSummary.votesEval": "Eval 计票",
-    "scopeSummary.votesAttempt": "Attempt 计票",
+    "scopeSummary.lastRun": "最近运行 · {time}",
+    "scopeSummary.runRange": "运行范围 · {from} – {to}",
+    "scopeSummary.costCoverage": "{samples}/{total} 次有成本数据",
+    "scopeSummary.votesEval": "Eval 结果",
+    "scopeSummary.votesAttempt": "Attempt 结果",
     "overview.experiments.one": "{n} 个实验",
     "overview.experiments.other": "{n} 个实验",
-    "overview.evalsCount": "{n} 道题",
+    "overview.evalsCount": "{n} 个 Eval",
     "overview.attemptsCount": "{n} 次 attempt",
     "overview.evals": "Eval",
-    "overview.passRate": "端到端成功率",
+    "overview.passRate": "通过率",
     "overview.totalCost": "成本",
-    "experimentComparison.groups": "实验组",
-    "experimentComparison.group": "实验组",
-    "experimentComparison.results": "Eval 结果",
-    "experimentComparison.lastRun": "最后运行",
-    "experimentComparison.selectGroup": "选择实验组 {group}",
-    "experimentComparison.commandsHead": "查看组内详情:",
-    "experimentComparison.empty": "没有实验组",
     "scatter.betterUpperRight": "越靠右上越好",
     "scatter.noData": "{x} × {y} 没有可绘制的数据",
+    "scatter.groupedBy": "按 {dim} 归类",
     "pointsMissing.one": "{n} 个点缺数据",
     "pointsMissing.other": "{n} 个点缺数据",
     "attemptList.empty": "没有 attempt",

package/dist/report/{metrics.d.ts → model/metrics.d.ts}

@@ -1,4 +1,4 @@
-import type { EvalResult } from "../types.ts";
+import type { EvalResult } from "../../types.ts";
 import type { Metric } from "./types.ts";
 /**
  * 定义一个指标。内置指标与自定义指标是同一个类型,没有特权;
@@ -10,7 +10,7 @@ export declare function defineMetric<const Name extends string>(def: Metric<Name
 export declare function attemptCostUSD(result: EvalResult): number | null;
 /**
  * 条件答题质量:passed = 1,failed = 0,errored 记 null 不进分母。
- * 这是「已形成可信判定」条件下的诊断指标,不能简称默认成功率
+ * 这是「已形成可信判定」条件下的诊断指标,不能简称默认通过率
  * (docs/feature/reports/library.md「内置指标」)。
  */
 export declare const taskPassRate: Metric<"task-pass-rate">;

package/dist/report/{metrics.js → model/metrics.js}

@@ -40,7 +40,7 @@ export function attemptCostUSD(result) {
 }
 /**
  * 条件答题质量:passed = 1,failed = 0,errored 记 null 不进分母。
- * 这是「已形成可信判定」条件下的诊断指标,不能简称默认成功率
+ * 这是「已形成可信判定」条件下的诊断指标,不能简称默认通过率
  * (docs/feature/reports/library.md「内置指标」)。
  */
 export const taskPassRate = defineMetric({
@@ -86,7 +86,7 @@ export const executionReliability = defineMetric({
  */
 export const endToEndPassRate = defineMetric({
     name: "end-to-end-pass-rate",
-    label: { en: "End-to-end pass rate", "zh-CN": "端到端成功率" },
+    label: { en: "Pass rate", "zh-CN": "通过率" },
     description: "End-to-end composite: passed = 1, failed / errored = 0. Split blame with taskPassRate and executionReliability.",
     better: "higher",
     unit: "%",