niceeval 0.11.4-canary.20 → 0.12.0

package/dist/report/model/format.js

@@ -1,7 +1,6 @@
 // unit 驱动的内置格式化(docs/feature/reports/library/presentation.md):
 //   "%" → 87%    "ms" → 1.2s    "$" → $0.31    其余 → 1.2k 缩写(带 unit 后缀)
 // MetricValue.format 可覆盖；公开入口是 formatMetricValue / formatAxisTick。
-import { gapParts } from "../../sample/index.js";
 import { DEFAULT_REPORT_LOCALE, DISPLAY_LOCALES, localeText, } from "./locale.js";
 /**
  * 一组 id 的显示名：每个 id 缩成在这组里唯一的最短路径后缀，重名逐步加长到能区分为止
@@ -145,6 +144,7 @@ const MISSING_CODE_KEYS = {
     noSamples: "cell.missing",
     notRun: "cell.notRun",
     unscorable: "cell.unscorable",
+    noCurrentResult: "cell.noCurrentResult",
 };
 /**
  * `missing` 格的本地化原因。code 是结构化代码,不是显示文本
@@ -260,16 +260,37 @@ export function formatReportDateTimeRange(fromIso, toIso, locale) {
         to: formatReportDate(toDate, locale, toOptions),
     };
 }
-// ── 实体列表(ExperimentList / EvalList / AttemptList)共用的时效标注 ──
+// ── 实体列表(ExperimentList / EvalList / AttemptList)共用的相对时距 ──
+/** 时距词表按区间取的单位;`{n}` 插值,en 是紧凑后缀,zh-CN 带空格接量词。 */
+const TIME_DISTANCE_KEYS = {
+    minute: "timeDistance.minute",
+    hour: "timeDistance.hour",
+    day: "timeDistance.day",
+    month: "timeDistance.month",
+};
+const MS_PER_MINUTE = 60_000;
+const MS_PER_HOUR = 3_600_000;
+const MS_PER_DAY = 86_400_000;
+const MS_PER_MONTH = MS_PER_DAY * 30;
 /**
- * 历史执行的紧凑时距("3d" / "2h" / "5m" / "10s"):自 `startedAt` 起算,渲染时刻由调用方
- * 传入(`nowIso` 缺省当前时刻)——粒度阈值复用 `gapParts`(与曾经的 stale-run message
- * 同一套单源,见 `results/select.ts`),只是这里的呈现是紧凑单字母,不是完整单词
- * (docs/feature/reports/library.md「时效标注」)。
+ * 一段时长(毫秒) → 当前 locale 的紧凑相对时距("12d" / "12 天")。按区间选粒度最大的单位
+ * 取整,结果恒不小于一个单位(docs/feature/reports/library/presentation.md「相对时距是数据,
+ * 不是文案」):不足 1 小时→分钟,不足 1 天→小时,不足 30 天→天,30 天及以上→月。
+ * `locator` 格与占位行参考的 `staleSinceMs` 都经这一个入口,不各写各的措辞;不加箭头、回环
+ * 之类的装饰记号——记号不携带时长,读者还要先学会它。
  */
-export function formatHistoricalGap(startedAtIso, nowIso = new Date().toISOString()) {
-    const { n, unit } = gapParts(startedAtIso, nowIso);
-    return `${n}${unit[0]}`;
+export function formatTimeDistance(ms, locale = DEFAULT_REPORT_LOCALE) {
+    const abs = Math.max(0, ms);
+    if (abs < MS_PER_HOUR) {
+        return localeText(locale, TIME_DISTANCE_KEYS.minute, { n: Math.max(1, Math.round(abs / MS_PER_MINUTE)) });
+    }
+    if (abs < MS_PER_DAY) {
+        return localeText(locale, TIME_DISTANCE_KEYS.hour, { n: Math.max(1, Math.round(abs / MS_PER_HOUR)) });
+    }
+    if (abs < MS_PER_MONTH) {
+        return localeText(locale, TIME_DISTANCE_KEYS.day, { n: Math.max(1, Math.round(abs / MS_PER_DAY)) });
+    }
+    return localeText(locale, TIME_DISTANCE_KEYS.month, { n: Math.max(1, Math.round(abs / MS_PER_MONTH)) });
 }
 // ── 实体列表(ExperimentList / EvalList / AttemptList)共用的判定符 ──
 /** passed / failed / errored / skipped 的判定符。 */

package/dist/report/model/locale.d.ts

@@ -20,12 +20,19 @@ declare const en: {
     readonly "verdict.skipped": "skipped";
     /** AttemptSummary 的重试序号(0 起的 identity.attempt 显示前 +1)。 */
     readonly "attemptSummary.attempt": "attempt {n}";
+    /** formatTimeDistance 的四区间读法(presentation.md「相对时距是数据,不是文案」)。 */
+    readonly "timeDistance.minute": "{n}m";
+    readonly "timeDistance.hour": "{n}h";
+    readonly "timeDistance.day": "{n}d";
+    readonly "timeDistance.month": "{n}mo";
     /** 全 null / 无样本的统一文案,绝不画 0。 */
     readonly "cell.missing": "no data";
     /** missing 格 code=`notRun`:固定题集里这道题没有 attempt。 */
     readonly "cell.notRun": "not run";
     /** missing 格 code=`unscorable`:有 attempt,但读数测不出。 */
     readonly "cell.unscorable": "unscorable";
+    /** missing 格 code=`noCurrentResult`:当前配置下这道题没有结果(覆盖缺口占位行)。 */
+    readonly "cell.noCurrentResult": "no result for current config";
     readonly "cell.measuredTitle": "{samples}/{total} attempts measured";
     readonly "cell.noneMeasurableTitle": "0/{total} attempts measurable";
     readonly "cell.coverageTitle": "coverage {samples}/{total}: this metric is null for the remaining attempts";
@@ -136,10 +143,19 @@ declare const en: {
     readonly "issues.details.one": "{n} warning";
     readonly "issues.details.other": "{n} issues";
     readonly "issues.badge.unfinishedSnapshot": "unfinished";
-    /** 覆盖缺口(scope.coverage)与时效标注(见 entity-lists.md「时效标注」「ExperimentList」)。 */
+    /** 覆盖缺口(scope.coverage)。 */
     readonly "overview.evalsCountPartial": "{covered}/{total} evals";
-    readonly "experimentList.historicalAttempts": "↩ {n}/{m} attempts";
-    readonly "experimentList.noResultsForConfig": "no results under the current configuration";
+    /** 覆盖构成四段的段名(docs/feature/reports/components/summaries/experiment-table.md「覆盖构成」)。 */
+    readonly "coverage.fresh": "fresh";
+    readonly "coverage.historical": "historical";
+    readonly "coverage.stale": "stale";
+    readonly "coverage.notRun": "not run";
+    /** locator 格历史执行时距的 hover 说明(三条通道见「时效不写字」)。 */
+    readonly "experimentList.historicalTooltip": "Historical execution — carried over or pulled in from an earlier run; still counts toward every total.";
+    /** 覆盖缺口占位行参考的 hover 说明。 */
+    readonly "experimentList.staleReferenceTooltip": "Reference only — not comparable with the current configuration, does not count toward any total.";
+    /** 「只看新执行」开关的 web 面文案(experiment-table.md「只看新执行」)。 */
+    readonly "experimentTable.freshOnlyToggle": "Fresh executions only";
     /** Hero / HeroCard 的运行 meta(hero.noRuns 是 latestStartedAt 为 null 时的内置文案)。 */
     readonly "hero.lastRun": "Last run {time}";
     readonly "hero.noRuns": "No runs yet";

package/dist/report/model/locale.js

@@ -46,12 +46,19 @@ const en = {
     "verdict.skipped": "skipped",
     /** AttemptSummary 的重试序号(0 起的 identity.attempt 显示前 +1)。 */
     "attemptSummary.attempt": "attempt {n}",
+    /** formatTimeDistance 的四区间读法(presentation.md「相对时距是数据,不是文案」)。 */
+    "timeDistance.minute": "{n}m",
+    "timeDistance.hour": "{n}h",
+    "timeDistance.day": "{n}d",
+    "timeDistance.month": "{n}mo",
     /** 全 null / 无样本的统一文案,绝不画 0。 */
     "cell.missing": "no data",
     /** missing 格 code=`notRun`:固定题集里这道题没有 attempt。 */
     "cell.notRun": "not run",
     /** missing 格 code=`unscorable`:有 attempt,但读数测不出。 */
     "cell.unscorable": "unscorable",
+    /** missing 格 code=`noCurrentResult`:当前配置下这道题没有结果(覆盖缺口占位行)。 */
+    "cell.noCurrentResult": "no result for current config",
     "cell.measuredTitle": "{samples}/{total} attempts measured",
     "cell.noneMeasurableTitle": "0/{total} attempts measurable",
     "cell.coverageTitle": "coverage {samples}/{total}: this metric is null for the remaining attempts",
@@ -162,10 +169,19 @@ const en = {
     "issues.details.one": "{n} warning",
     "issues.details.other": "{n} issues",
     "issues.badge.unfinishedSnapshot": "unfinished",
-    /** 覆盖缺口(scope.coverage)与时效标注(见 entity-lists.md「时效标注」「ExperimentList」)。 */
+    /** 覆盖缺口(scope.coverage)。 */
     "overview.evalsCountPartial": "{covered}/{total} evals",
-    "experimentList.historicalAttempts": "↩ {n}/{m} attempts",
-    "experimentList.noResultsForConfig": "no results under the current configuration",
+    /** 覆盖构成四段的段名(docs/feature/reports/components/summaries/experiment-table.md「覆盖构成」)。 */
+    "coverage.fresh": "fresh",
+    "coverage.historical": "historical",
+    "coverage.stale": "stale",
+    "coverage.notRun": "not run",
+    /** locator 格历史执行时距的 hover 说明(三条通道见「时效不写字」)。 */
+    "experimentList.historicalTooltip": "Historical execution — carried over or pulled in from an earlier run; still counts toward every total.",
+    /** 覆盖缺口占位行参考的 hover 说明。 */
+    "experimentList.staleReferenceTooltip": "Reference only — not comparable with the current configuration, does not count toward any total.",
+    /** 「只看新执行」开关的 web 面文案(experiment-table.md「只看新执行」)。 */
+    "experimentTable.freshOnlyToggle": "Fresh executions only",
     /** Hero / HeroCard 的运行 meta(hero.noRuns 是 latestStartedAt 为 null 时的内置文案)。 */
     "hero.lastRun": "Last run {time}",
     "hero.noRuns": "No runs yet",
@@ -218,9 +234,14 @@ const zhCN = {
     "verdict.errored": "错误",
     "verdict.skipped": "跳过",
     "attemptSummary.attempt": "第 {n} 次",
+    "timeDistance.minute": "{n} 分钟",
+    "timeDistance.hour": "{n} 小时",
+    "timeDistance.day": "{n} 天",
+    "timeDistance.month": "{n} 个月",
     "cell.missing": "无数据",
     "cell.notRun": "未跑到",
     "cell.unscorable": "测不出",
+    "cell.noCurrentResult": "当前配置下无结果",
     "cell.measuredTitle": "{samples}/{total} 次 attempt 测得",
     "cell.noneMeasurableTitle": "0/{total} 次 attempt 可测",
     "cell.coverageTitle": "覆盖率 {samples}/{total}:其余 attempt 测不了这个指标",
@@ -321,8 +342,13 @@ const zhCN = {
     "issues.details.other": "{n} 条原始警告",
     "issues.badge.unfinishedSnapshot": "未收尾",
     "overview.evalsCountPartial": "{covered}/{total} 个 Eval",
-    "experimentList.historicalAttempts": "↩ {n}/{m} attempts",
-    "experimentList.noResultsForConfig": "当前配置下无结果",
+    "coverage.fresh": "新执行",
+    "coverage.historical": "历史执行",
+    "coverage.stale": "过期结论",
+    "coverage.notRun": "未跑到",
+    "experimentList.historicalTooltip": "历史执行——携带自上一轮或跨 Run 拼入的结果；仍计入所有汇总。",
+    "experimentList.staleReferenceTooltip": "仅作参考——与当前配置不可比，不进任何计数。",
+    "experimentTable.freshOnlyToggle": "只看新执行",
     "hero.lastRun": "最后运行 {time}",
     "hero.noRuns": "暂无运行",
     "hero.composedRuns": "由 {n} 次运行合成",

package/dist/report/model/types.d.ts

@@ -240,8 +240,10 @@ export interface DeltaCell {
     attempts: readonly AttemptLocator[];
     totalTokens?: number;
     totalCostUSD?: number;
-    /** true 时该格来自跨快照携带的历史执行,渲染为 ↩ 时效标注。 */
+    /** true 时该格来自跨快照携带的历史执行(时效标注见 experiment-table.md「时效不写字」)。 */
     historical: boolean;
+    /** `historical` 为 true 时,距今的毫秒数(渲染时刻起算);新执行时省略,不伪造 0。 */
+    staleSinceMs?: number;
 }
 export interface DeltaData {
     byDimension: string;
@@ -262,6 +264,12 @@ export interface DeltaData {
             tokens?: number;
             costUSD?: number;
         }>;
+        /**
+         * 键是条件值,只在该条件缺席这道题(`cells` 无键)且记录里存在不可比历史判定时才有条目
+         * (docs/feature/reports/show/compare.md「— ✓ 12d」);它不进 `totals`、`delta` 与配对覆盖的
+         * 任何一个数,只提供 locator 下钻。
+         */
+        references?: globalThis.Record<string, StaleConclusionReference>;
     }>;
     /** 各条件自身覆盖面的描述,分母是该条件有结果的 eval 数;不用于跨条件直接归因。 */
     totals: globalThis.Record<string, {
@@ -463,8 +471,20 @@ export interface AttemptListItem {
     startedAt: string;
     /** 历史执行:携带条目,或来自该实验在 Sample 中最新快照之外的快照;false = 最新一次运行实测。 */
     historical: boolean;
+    /** `historical` 为 true 时,距今的毫秒数(渲染时刻起算);新执行时省略,不伪造 0。 */
+    staleSinceMs?: number;
     locator: AttemptLocator;
 }
+/**
+ * 覆盖缺口占位行的参考:记录里存在、但与当前基准 configHash 不可比的最近一条判定
+ * (docs/feature/reports/components/summaries/experiment-table.md「覆盖缺口的两档占位行」)。
+ * 它不进任何计数,只提供下钻入口。
+ */
+export interface StaleConclusionReference {
+    readonly locator: AttemptLocator;
+    readonly verdict: Verdict | "skipped";
+    readonly staleSinceMs: number;
+}
 /**
  * `EvalList` 一项 = 一个 `experimentId + evalId`(同一个 Eval 跑在两个 experiment 上是
  * 两条不同结果,不合并)。失败原因只存在于各 AttemptListItem,不在 Eval 父项重复一份。
@@ -523,10 +543,16 @@ export interface ExperimentListItem {
     evals: number;
     /** 这个 experiment 覆盖的 attempt 总数(原始计数,含多轮重试)。 */
     attempts: number;
-    /** 历史执行的 attempt 数(分母是 attempts);时效标注「↩ n/m attempts」的数据源。 */
+    /** 历史执行的 attempt 数(分母是 attempts)。 */
     historicalAttempts: number;
     /** 已知 eval 并集里、当前口径下没有任何 attempt 的题(来自 `scope.coverage`);渲染为占位行。 */
     missingEvalIds: string[];
+    /**
+     * `missingEvalIds` 里、记录中存在不可比历史判定的那些题的参考(evalId → 参考);
+     * 其余缺口题没有条目,占位行只给「当前配置下无结果」。`sample.fresh` 为 true 时整份为空——
+     * 读者已声明只看新执行,占位行就不再把被排除的历史结论请回来。
+     */
+    staleReferences: Readonly<globalThis.Record<string, StaleConclusionReference>>;
     /** 所含快照中最近的 startedAt。 */
     lastRunAt: string;
     evalRows: ExperimentListEvalRow[];
@@ -685,8 +711,8 @@ export interface AttemptConversationData {
 /** 独立的生命周期命令证据项;不属于 Conversation 的轮次或消息。 */
 export interface AttemptCommandEvidence extends CommandExitEvidence {
     key: string;
-    /** 消费层由 checked + 非零 exitCode 推导,不落盘。 */
-    classification: "observed" | "failed";
+    /** 消费层由 checked + exitCode 推导,不落盘:exitCode 为 0 是 "succeeded";非零时 checked 才是 "failed",否则是 "observed"。 */
+    classification: "succeeded" | "observed" | "failed";
     /** 从关联 timing node 派生,缺少 timing 时省略。 */
     durationMs?: number;
 }
@@ -729,6 +755,17 @@ export interface UsageTableData {
     usage?: Usage;
     estimatedCostUSD?: number;
 }
+/**
+ * `AttemptFacts` 的 data:attempt 作用域 `ctx.fact()` 上报的运行事实完整键值表
+ * (见 docs/feature/record/architecture.md#facts运行事实)，按落盘的 key 插入顺序排列。
+ * `AttemptRecord.facts` 缺失或为空对象时整个 data 为 null,不渲染空表。
+ */
+export interface AttemptFactsData {
+    facts: {
+        key: string;
+        value: string | number | boolean;
+    }[];
+}
 /** `AttemptTrace` 的 data:不与 runner 节点合并的原始 OTel span 列表;没有 trace 时 null。 */
 export interface AttemptTraceData {
     /** text 面拼 `niceeval show <locator> --timing` 下钻命令用;web 面不需要。 */

package/dist/report/presentation.d.ts

@@ -18,21 +18,68 @@ export type DimensionDeclarations = Readonly<Record<string, DimensionDeclaration
 export type DimensionPins = Readonly<Record<string, Readonly<Record<string, number>>>>;
 export declare const VISUAL_SLOT_COUNT = 24;
 export type SeriesVariant = 1 | 2 | 3 | 4;
-/** 槽 1..24 → (色板下标 1..6, 形状变体 1..4)，与 README 槽序表逐项对应。 */
+/**
+ * 槽 1..24 → (色板下标 1..6, 形状变体 1..4)。
+ * 与 components/README「视觉编码容量」槽序表一致:1–6 第一变体、7–12 第二变体,依此类推。
+ * `colorIndex = (slot-1) % 6 + 1`,`variant = floor((slot-1)/6) + 1`。
+ */
 export declare function seriesChannelsOf(slot: number): {
     colorIndex: number;
     variant: SeriesVariant;
 };
-export interface PresentedValue {
+export interface PresentationIdentity {
+    /** 完整维度值，作为排序、筛选、React key 与证据身份。 */
     readonly value: string;
+    /** 当前页完整 label keyset 内生成的显示名。 */
     readonly label: string;
-    readonly seriesSlot?: number;
-    readonly colorIndex?: number;
-    readonly variant?: SeriesVariant;
 }
+export interface LabelPresentation extends PresentationIdentity {
+    readonly kind: "label";
+}
+export interface ColorPresentation extends PresentationIdentity {
+    readonly kind: "color";
+    /** `var(--niceeval-color-series-N)`。 */
+    readonly color: string;
+}
+export interface LineSeriesPresentation extends PresentationIdentity {
+    readonly kind: "series";
+    readonly mark: "line";
+    readonly stroke: string;
+    readonly strokeDasharray: string;
+    readonly marker: {
+        readonly path: string;
+        readonly viewBox: string;
+        readonly fill: string;
+        readonly stroke: string;
+    };
+}
+export interface ScatterSeriesPresentation extends PresentationIdentity {
+    readonly kind: "series";
+    readonly mark: "scatter";
+    readonly marker: LineSeriesPresentation["marker"];
+}
+export interface FillSeriesPresentation extends PresentationIdentity {
+    readonly kind: "series";
+    readonly mark: "bar" | "area";
+    /** 颜色，或可直接使用的 `url(#pattern-id)`。 */
+    readonly fill: string;
+    readonly stroke: string;
+    readonly strokeDasharray: string;
+}
+export type DimensionPresentation = LabelPresentation | ColorPresentation | LineSeriesPresentation | ScatterSeriesPresentation | FillSeriesPresentation;
+export type PresentationFor<E extends DimensionEncoding> = E extends {
+    kind: "label";
+} ? LabelPresentation : E extends {
+    kind: "color";
+} ? ColorPresentation : LineSeriesPresentation | ScatterSeriesPresentation | FillSeriesPresentation;
+/**
+ * 句柄上的呈现面。`at(i)` 返回可直接交给 SVG/CSS 的值;text 面恒为 label 支。
+ * `labels` 是本维度整页 label keyset 的最短唯一后缀表,便于探针与调试。
+ */
 export interface PresentedDimension {
+    readonly length: number;
     readonly labels: ReadonlyMap<string, string>;
-    at(index: number): PresentedValue;
+    at(index: number): DimensionPresentation;
 }
 export declare class UndeclaredDimensionValueError extends Error {
     readonly handle: string;

package/dist/report/presentation.js

@@ -1,6 +1,11 @@
 import { shortestUniqueLabels as labels } from "./model/format.js";
+import { seriesColorVar, seriesFill, seriesMarker, seriesStrokeDasharray, } from "./assets/series-encoding.js";
 export const VISUAL_SLOT_COUNT = 24;
-/** 槽 1..24 → (色板下标 1..6, 形状变体 1..4)，与 README 槽序表逐项对应。 */
+/**
+ * 槽 1..24 → (色板下标 1..6, 形状变体 1..4)。
+ * 与 components/README「视觉编码容量」槽序表一致:1–6 第一变体、7–12 第二变体,依此类推。
+ * `colorIndex = (slot-1) % 6 + 1`,`variant = floor((slot-1)/6) + 1`。
+ */
 export function seriesChannelsOf(slot) {
     if (!Number.isInteger(slot) || slot < 1 || slot > VISUAL_SLOT_COUNT) {
         throw new Error(`seriesSlot must be an integer in [1, ${VISUAL_SLOT_COUNT}], got ${slot}.`);
@@ -8,7 +13,7 @@ export function seriesChannelsOf(slot) {
     const zero = slot - 1;
     return {
         colorIndex: (zero % 6) + 1,
-        variant: ((((zero % 4) + Math.floor(zero / 12)) % 4) + 1),
+        variant: (Math.floor(zero / 6) + 1),
     };
 }
 export class UndeclaredDimensionValueError extends Error {
@@ -33,6 +38,35 @@ export function presentDimension(declaration) {
     return presented;
 }
 export const shortestUniqueLabels = labels;
+function presentChannels(value, label, seriesSlot, encoding) {
+    const { colorIndex, variant } = seriesChannelsOf(seriesSlot);
+    if (encoding.kind === "color") {
+        return { kind: "color", value, label, color: seriesColorVar(colorIndex) };
+    }
+    if (encoding.kind !== "series") {
+        return { kind: "label", value, label };
+    }
+    const stroke = seriesColorVar(colorIndex);
+    const strokeDasharray = seriesStrokeDasharray(variant);
+    const marker = seriesMarker(colorIndex, variant);
+    switch (encoding.mark) {
+        case "line":
+            return { kind: "series", mark: "line", value, label, stroke, strokeDasharray, marker };
+        case "scatter":
+            return { kind: "series", mark: "scatter", value, label, marker };
+        case "bar":
+        case "area":
+            return {
+                kind: "series",
+                mark: encoding.mark,
+                value,
+                label,
+                fill: seriesFill(colorIndex, variant),
+                stroke,
+                strokeDasharray,
+            };
+    }
+}
 /**
  * 页级呈现分配：收集整页 dimensions() 声明，产出 label / visual 两套 keyset 上的映射。
  * pins 原样占位；未钉键以稳定散列为起点、撞槽按显示键字典序线性探测。
@@ -79,6 +113,7 @@ export function allocatePageDimensions(handles, pins = {}, options = { face: "we
         const values = declaration.values;
         const encoding = declaration.encoding;
         byHandle.set(handle, {
+            length: values.length,
             labels: dimLabels,
             at(index) {
                 const value = values[index];
@@ -86,15 +121,15 @@ export function allocatePageDimensions(handles, pins = {}, options = { face: "we
                     throw new UndeclaredDimensionValueError(`dimension handle "${handle}" has no value at index ${index} (declared length ${values.length}).`, handle, index);
                 }
                 const label = dimLabels.get(value) ?? value;
+                // text 面恒返回 label 支:不上 ANSI 色,拿不到颜色、线型或 pattern。
                 if (options.face === "text" || encoding.kind === "label") {
-                    return { value, label };
+                    return { kind: "label", value, label };
                 }
                 const seriesSlot = slots?.get(value);
                 if (seriesSlot === undefined) {
                     throw new UndeclaredDimensionValueError(`dimension handle "${handle}" value ${JSON.stringify(value)} has no seriesSlot on this page.`, handle, index);
                 }
-                const channels = seriesChannelsOf(seriesSlot);
-                return { value, label, seriesSlot, colorIndex: channels.colorIndex, variant: channels.variant };
+                return presentChannels(value, label, seriesSlot, encoding);
             },
         });
     }

package/dist/report/react/index.d.ts

@@ -2,12 +2,12 @@ export { Callouts, Chart, Col, Conversation, CopyBlock, DiffView, Grid, Row, Sec
 export { HeroCard } from "../components/site-components/HeroCard.tsx";
 export type { HeroLink, HeroLogo } from "../components/site-components/hero-types.ts";
 export { PoweredBy } from "../components/site-components/PoweredBy.tsx";
-export type { AttemptAssertionsData, AttemptConversationData, AttemptConversationReply, AttemptConversationRound, AttemptDiagnosticsData, AttemptDiffData, AttemptErrorData, AttemptFixPromptData, AttemptListItem, AttemptLocator, AttemptSummaryData, AttemptTimelineData, AttemptTraceData, CopyFixPromptData, Dataset, DatasetField, DatasetRow, EvalListItem, ExperimentListEvalRow, ExperimentListItem, HeroData, MetricColumn, MetricValue, SampleSummaryContent, SampleIssue, SnapshotDiagnosticsData, SnapshotDiagnosticsItem, TraceSpanSummary, TraceWaterfallRow, UsageTableData, VerdictTally, } from "../model/types.ts";
+export type { AttemptAssertionsData, AttemptConversationData, AttemptConversationReply, AttemptConversationRound, AttemptDiagnosticsData, AttemptDiffData, AttemptErrorData, AttemptFixPromptData, AttemptListItem, AttemptLocator, AttemptSummaryData, AttemptTimelineData, AttemptTraceData, CopyFixPromptData, Dataset, DatasetField, DatasetRow, EvalListItem, ExperimentListEvalRow, ExperimentListItem, HeroData, MetricColumn, MetricValue, SampleSummaryContent, SampleIssue, SnapshotDiagnosticsData, SnapshotDiagnosticsItem, StaleConclusionReference, TraceSpanSummary, TraceWaterfallRow, UsageTableData, VerdictTally, } from "../model/types.ts";
 export type { Cell, VerdictCounts } from "../definition/cell.ts";
 export { formatCellText } from "../definition/cell.ts";
 export type { AttemptEvidence, AttemptEvidenceCapabilities } from "../../record/attempt-evidence.ts";
-export { formatAxisTick, formatInstant, formatMetricValue, missingText } from "../model/format.ts";
+export { formatAxisTick, formatInstant, formatMetricValue, formatTimeDistance, missingText } from "../model/format.ts";
 export { presentDimension, shortestUniqueLabels } from "../presentation.ts";
-export type { DimensionDeclaration, DimensionEncoding, PresentedDimension } from "../presentation.ts";
+export type { ColorPresentation, DimensionDeclaration, DimensionEncoding, DimensionPresentation, FillSeriesPresentation, LabelPresentation, LineSeriesPresentation, PresentedDimension, ScatterSeriesPresentation, } from "../presentation.ts";
 export { DEFAULT_REPORT_LOCALE, localizedTextEquals, resolveLocalizedText, resolveMetricLabel, } from "../model/locale.ts";
 export type { LocalizedText, ReportLocale } from "../model/locale.ts";

package/dist/report/react/index.js

@@ -12,7 +12,7 @@ export { HeroCard } from "../components/site-components/HeroCard.js";
 export { PoweredBy } from "../components/site-components/PoweredBy.js";
 export { formatCellText } from "../definition/cell.js";
 // 格式化与呈现工具箱(docs/feature/reports/library/presentation.md):自有 React 页面与报告面同实现。
-export { formatAxisTick, formatInstant, formatMetricValue, missingText } from "../model/format.js";
+export { formatAxisTick, formatInstant, formatMetricValue, formatTimeDistance, missingText } from "../model/format.js";
 export { presentDimension, shortestUniqueLabels } from "../presentation.js";
 // locale(官方组件 chrome 文案的语言;LocalizedText 的按 locale 解析也用它)
 export { DEFAULT_REPORT_LOCALE, localizedTextEquals, resolveLocalizedText, resolveMetricLabel, } from "../model/locale.js";

package/dist/report/slices/compute.js

@@ -12,7 +12,7 @@
 // - core 中立:只认 Metric / Dimension 接口,不出现具体 agent 名的分支。
 import { comparabilityConfigOf, deepEqualJson } from "../../sample/index.js";
 import { foldEvalVerdict } from "../../shared/verdict.js";
-import { assertUniqueMetricNames, axisValueOf, collectItems, computeCell, dimensionKey, dimensionName, evalGroupOf, evalIdOf, evaluateMetric, experimentIdOf, filterItems, fullEvalKey, groupItems, historicalOf, locatorOf, refDisplayKey, resolveInput, seriesKey, seriesName, toColumn, } from "../model/aggregate.js";
+import { assertUniqueMetricNames, axisValueOf, collectItems, computeCell, dimensionKey, dimensionName, evalGroupOf, evalIdOf, evaluateMetric, experimentIdOf, filterItems, fullEvalKey, groupItems, historicalOf, locatorOf, msSince, refDisplayKey, resolveInput, seriesKey, seriesName, staleReferenceOf, toColumn, } from "../model/aggregate.js";
 import { costUSD as costUSDMetric, examScore, tokens as tokensMetric, totalScore as totalScoreMetric } from "../model/metrics.js";
 import { formatMetricValue, formatPlainNumber } from "../model/format.js";
 import { metricFieldOf } from "../model/dataset.js";
@@ -420,6 +420,10 @@ async function buildDeltaCell(items) {
     const verdict = foldEvalVerdict(items.map((item) => ({ verdict: item.attempt.result.verdict })));
     const refs = new Set();
     let historical = false;
+    // 距今最久的历史执行决定这一格的时距——多个 attempt 折进同一格时,时距说的是「这批结论里
+    // 最旧的证据有多旧」,不是最新那条(与 experiment-table 的 attempt 行时距同一份数据来源,
+    // 只是这里先跨 attempt 折叠成一格)。
+    let staleStartedAt;
     let scoreSum = 0;
     let scoreCount = 0;
     let tokensSum = 0;
@@ -428,8 +432,12 @@ async function buildDeltaCell(items) {
     let costCount = 0;
     for (const item of items) {
         refs.add(locatorOf(item));
-        if (historicalOf(item))
+        if (historicalOf(item)) {
             historical = true;
+            const startedAt = item.attempt.result.startedAt ?? item.run.startedAt;
+            if (staleStartedAt === undefined || startedAt < staleStartedAt)
+                staleStartedAt = startedAt;
+        }
         if (evaluationKind === "points") {
             const value = await evaluateMetric(totalScoreMetric, item.attempt);
             if (value !== null) {
@@ -458,10 +466,21 @@ async function buildDeltaCell(items) {
         ...(tokensCount > 0 ? { totalTokens: tokensSum } : {}),
         ...(costCount > 0 ? { totalCostUSD: costSum } : {}),
         historical,
+        ...(staleStartedAt !== undefined ? { staleSinceMs: msSince(staleStartedAt) } : {}),
     };
 }
+/** 每个 experiment 当前基准的 configHash:该 experiment 在 `runs` 里最新快照的 configHash。 */
+function latestConfigHashByExperiment(runs) {
+    const latest = new Map();
+    for (const run of runs) {
+        const existing = latest.get(run.experimentId);
+        if (existing === undefined || run.startedAt > existing.startedAt)
+            latest.set(run.experimentId, run);
+    }
+    return new Map([...latest].map(([id, run]) => [id, run.configHash]));
+}
 export async function deltaTableData(input, options) {
-    const { runs, attempts } = resolveInput(input);
+    const { runs, attempts, historyAttempts, fresh } = resolveInput(input);
     const items = filterItems(collectItems(runs, attempts), options.evals);
     let conditions;
     let experiments;
@@ -526,6 +545,13 @@ export async function deltaTableData(input, options) {
         for (const evalId of byEval.keys())
             evalIdsSet.add(evalId);
     const knownEvalIds = [...evalIdsSet].sort();
+    // 每个条件背后的 experiment id 集合(by: "experiment" 时条件字符串本身通常就是唯一一个 id);
+    // 过期结论参考按各自 experiment 的当前基准 configHash 判定。
+    const experimentIdsByCondition = new Map();
+    for (const [condition, conditionItems] of itemsByCondition) {
+        experimentIdsByCondition.set(condition, new Set(conditionItems.map((item) => experimentIdOf(item))));
+    }
+    const anchorConfigHashByExperiment = latestConfigHashByExperiment(runs);
     const baseline = conditions[0];
     const rows = [];
     for (const evalId of knownEvalIds) {
@@ -558,7 +584,29 @@ export async function deltaTableData(input, options) {
                 }
             }
         }
-        rows.push({ key: evalId, flipped, cells, ...(delta ? { delta } : {}) });
+        // 过期结论参考(show/compare.md「— ✓ 12d」):只在该条件缺席这道题、且记录里存在与该
+        // experiment 当前基准不可比的历史判定时才给,口径与 experiment-table 两档占位行一致——
+        // `fresh` 为 true 时不给参考,读者已声明只看新执行。它只写进 references,不进 cells,
+        // 因此天然不参与 totals / delta / 配对覆盖的任何一个数。
+        let references;
+        if (!fresh) {
+            for (const condition of conditions) {
+                if (cells[condition] !== undefined)
+                    continue;
+                const experimentIds = experimentIdsByCondition.get(condition);
+                if (!experimentIds || experimentIds.size === 0)
+                    continue;
+                const candidates = historyAttempts.filter((a) => experimentIds.has(a.experimentId) &&
+                    a.evalId === evalId &&
+                    a.run.configHash !== anchorConfigHashByExperiment.get(a.experimentId));
+                const reference = staleReferenceOf(candidates);
+                if (reference) {
+                    references ??= {};
+                    references[condition] = reference;
+                }
+            }
+        }
+        rows.push({ key: evalId, flipped, cells, ...(delta ? { delta } : {}), ...(references ? { references } : {}) });
     }
     // 各条件自身覆盖面:分母是该条件有结果的 eval 数,不看其它条件是否也覆盖了这道题。
     const totals = {};

package/dist/report/slices/content.d.ts

@@ -1,6 +1,7 @@
 import type { TableContent, TableContentRow } from "../definition/cell.ts";
 import type { DeltaData, StabilityMatrixCell, StabilityMatrixData } from "../model/types.ts";
-export declare function deltaTableContent(data: DeltaData): TableContent;
+import { type ReportLocale } from "../model/locale.ts";
+export declare function deltaTableContent(data: DeltaData, locale?: ReportLocale): TableContent;
 export interface StabilityContentRow extends TableContentRow {
     readonly evalId: string;
     /** 全部条件历史执行中通过次数为 0 且执行数 > 0。 */