niceeval 0.4.5 → 0.5.0

package/src/report/react/fixtures.ts

@@ -1,22 +1,23 @@
-// 手工 fixture:模拟计算层(niceeval/report 的 table()/matrix()/… )的产物。
-// 仅供 render.test.tsx 与 scripts/report-react-demo.tsx 使用,不从入口导出。
+// 手工 fixture:模拟计算函数(MetricTable.data / MetricMatrix.data / …)的产物。
+// 仅供渲染测试与 scripts/report-react-demo.tsx 使用,不从入口导出。
 // 数字刻意覆盖诚实细节:coverage 角标(samples<total)、全 null 格子、
 // 稀疏矩阵、缺数据的散点、delta 的 null 不硬算、truncated 计数。
 
 import type {
   CaseListData,
   DeltaData,
+  LineData,
   MatrixData,
   MetricColumn,
   OverviewData,
   ScatterData,
   ScoreboardData,
   TableData,
-} from "./data.ts";
+} from "../types.ts";
 
-export const passRateColumn: MetricColumn = { key: "pass-rate", label: "通过率", unit: "%", better: "higher" };
-export const codeLinesColumn: MetricColumn = { key: "code-lines", label: "代码行数", unit: "lines", better: "lower" };
-export const costColumn: MetricColumn = { key: "cost", label: "成本", unit: "$", better: "lower" };
+export const passRateColumn: MetricColumn = { key: "pass-rate", label: "pass rate", unit: "%", better: "higher" };
+export const codeLinesColumn: MetricColumn = { key: "code-lines", label: "code lines", unit: "lines", better: "lower" };
+export const costColumn: MetricColumn = { key: "cost", label: "cost", unit: "$", better: "lower" };
 
 export const overviewData: OverviewData = {
   snapshots: [
@@ -33,7 +34,15 @@ export const overviewData: OverviewData = {
     costUSD: null, // 全部 attempt 都没报成本:null,组件必须显示缺数据而不是 $0
     durationMs: 261_000,
   },
-  warnings: ["compare/bub 快照缺 3 个 eval 的结果"],
+  warnings: [
+    {
+      kind: "partial-coverage",
+      experimentId: "compare/bub",
+      covered: 9,
+      total: 12,
+      message: "snapshot covers 9 of 12 evals seen in history; re-run `niceeval exp compare/bub` for a full snapshot",
+    },
+  ],
 };
 
 export const overviewWithCost: OverviewData = {
@@ -42,7 +51,7 @@ export const overviewWithCost: OverviewData = {
   warnings: [],
 };
 
-export const tableData: TableData = {
+export const tableData: TableData<"pass-rate" | "code-lines"> = {
   dimension: "agent",
   // 行顺序故意不按 passRate 排:组件必须按传入顺序渲染,不重排
   columns: [passRateColumn, codeLinesColumn],
@@ -50,9 +59,9 @@ export const tableData: TableData = {
     {
       key: "codex",
       cells: {
-        "pass-rate": { value: 0.5, display: "50%", samples: 6, total: 6 },
+        "pass-rate": { value: 0.5, display: "50%", samples: 6, total: 6, refs: [] },
         // 全 null:一个有效样本都没有 → 缺数据文案,绝不画 0
-        "code-lines": { value: null, display: "", samples: 0, total: 6 },
+        "code-lines": { value: null, display: "—", samples: 0, total: 6, refs: [] },
       },
     },
     {
@@ -66,7 +75,7 @@ export const tableData: TableData = {
           refs: [{ run: "run-a", result: 0 }],
         },
         // samples < total:有 attempt 测不了 → 覆盖率角标 5/6
-        "code-lines": { value: 120, display: "120 lines", samples: 5, total: 6 },
+        "code-lines": { value: 120, display: "120 lines", samples: 5, total: 6, refs: [] },
       },
     },
   ],
@@ -81,23 +90,32 @@ export const matrixData: MatrixData = {
     {
       row: "algebra/quadratic",
       column: "bub",
-      cell: { value: 1, display: "100%", samples: 2, total: 2, refs: [{ run: "run-b", result: 3 }, { run: "run-b", result: 7 }] },
+      cell: {
+        value: 1,
+        display: "100%",
+        samples: 2,
+        total: 2,
+        refs: [
+          { run: "run-b", result: 3 },
+          { run: "run-b", result: 7 },
+        ],
+      },
     },
     {
       row: "algebra/quadratic",
       column: "codex",
-      cell: { value: 0, display: "0%", samples: 3, total: 3 },
+      cell: { value: 0, display: "0%", samples: 3, total: 3, refs: [] },
     },
     {
       row: "geometry/angles",
       column: "bub",
-      cell: { value: 0.5, display: "50%", samples: 2, total: 2 },
+      cell: { value: 0.5, display: "50%", samples: 2, total: 2, refs: [] },
     },
   ],
 };
 
 export const scoreboardData: ScoreboardData = {
-  of: "agent",
+  dimension: "agent",
   fullMarks: 100,
   weights: [{ prefix: "algebra/", weight: 2 }],
   rows: [
@@ -130,32 +148,76 @@ export const scatterData: ScatterData = {
     {
       key: "compare/bub-low",
       series: "bub",
-      x: { value: 5, display: "$5.00", samples: 6, total: 6 },
-      y: { value: 0.5, display: "50%", samples: 6, total: 6 },
+      x: { value: 5, display: "$5.00", samples: 6, total: 6, refs: [] },
+      y: { value: 0.5, display: "50%", samples: 6, total: 6, refs: [] },
     },
     {
       key: "compare/bub-high",
       series: "bub",
-      x: { value: 10, display: "$10.00", samples: 6, total: 6 },
-      y: { value: 0.9, display: "90%", samples: 6, total: 6 },
+      x: { value: 10, display: "$10.00", samples: 6, total: 6, refs: [] },
+      y: { value: 0.9, display: "90%", samples: 6, total: 6, refs: [] },
     },
     {
       key: "compare/codex-mid",
       series: "codex",
-      x: { value: 7, display: "$7.00", samples: 6, total: 6 },
-      y: { value: 0.6, display: "60%", samples: 6, total: 6 },
+      x: { value: 7, display: "$7.00", samples: 6, total: 6, refs: [] },
+      y: { value: 0.6, display: "60%", samples: 6, total: 6, refs: [] },
     },
     {
       // x 缺数据:这个点不画,注脚报 1 个点缺数据
       key: "compare/codex-broken",
       series: "codex",
-      x: { value: null, display: "", samples: 0, total: 6 },
-      y: { value: 0.7, display: "70%", samples: 6, total: 6 },
+      x: { value: null, display: "—", samples: 0, total: 6, refs: [] },
+      y: { value: 0.7, display: "70%", samples: 6, total: 6, refs: [] },
+    },
+  ],
+};
+
+export const lineData: LineData = {
+  x: { key: "latencyMs", label: "Simulated latency", unit: "ms" },
+  series: "agents",
+  y: passRateColumn,
+  rows: [
+    {
+      key: "ultra/agents-1-lat-100",
+      series: "1 agents",
+      x: 100,
+      xDisplay: "100ms",
+      y: { value: 0.4, display: "40%", samples: 6, total: 6, refs: [] },
+    },
+    {
+      key: "ultra/agents-1-lat-300",
+      series: "1 agents",
+      x: 300,
+      xDisplay: "300ms",
+      y: { value: 0.3, display: "30%", samples: 6, total: 6, refs: [] },
+    },
+    {
+      key: "ultra/agents-16-lat-100",
+      series: "16 agents",
+      x: 100,
+      xDisplay: "100ms",
+      y: { value: 0.8, display: "80%", samples: 6, total: 6, refs: [] },
+    },
+    {
+      key: "ultra/agents-16-lat-300",
+      series: "16 agents",
+      x: 300,
+      xDisplay: "300ms",
+      y: { value: 0.7, display: "70%", samples: 6, total: 6, refs: [] },
+    },
+    {
+      // 未声明 param 的 experiment:作轴不画点,注脚报数
+      key: "ultra/legacy",
+      series: "1 agents",
+      x: null,
+      xDisplay: "",
+      y: { value: 0.5, display: "50%", samples: 6, total: 6, refs: [] },
     },
   ],
 };
 
-export const deltaData: DeltaData = {
+export const deltaData: DeltaData<"pass-rate" | "cost"> = {
   columns: [passRateColumn, costColumn],
   rows: [
     {
@@ -165,15 +227,15 @@ export const deltaData: DeltaData = {
       cells: {
         // 通过率 +12pp:better higher → 好(绿)
         "pass-rate": {
-          a: { value: 0.5, display: "50%", samples: 6, total: 6 },
-          b: { value: 0.62, display: "62%", samples: 6, total: 6 },
+          a: { value: 0.5, display: "50%", samples: 6, total: 6, refs: [] },
+          b: { value: 0.62, display: "62%", samples: 6, total: 6, refs: [] },
           delta: 0.12,
           display: "+12pp",
         },
         // 成本 +$0.15:better lower → 坏(红)
         cost: {
-          a: { value: 0.2, display: "$0.20", samples: 6, total: 6 },
-          b: { value: 0.35, display: "$0.35", samples: 6, total: 6 },
+          a: { value: 0.2, display: "$0.20", samples: 6, total: 6, refs: [] },
+          b: { value: 0.35, display: "$0.35", samples: 6, total: 6, refs: [] },
           delta: 0.15,
           display: "+$0.15",
         },
@@ -185,17 +247,17 @@ export const deltaData: DeltaData = {
       b: { experimentId: "compare/codex--agents-md" },
       cells: {
         "pass-rate": {
-          a: { value: 0.4, display: "40%", samples: 6, total: 6 },
-          b: { value: 0.4, display: "40%", samples: 6, total: 6 },
+          a: { value: 0.4, display: "40%", samples: 6, total: 6, refs: [] },
+          b: { value: 0.4, display: "40%", samples: 6, total: 6, refs: [] },
           delta: 0,
-          display: "±0pp",
+          display: "±0",
         },
         // A 侧缺数据 → delta null:显示缺,不硬算
         cost: {
-          a: { value: null, display: "", samples: 0, total: 6 },
-          b: { value: 0.3, display: "$0.30", samples: 6, total: 6 },
+          a: { value: null, display: "—", samples: 0, total: 6, refs: [] },
+          b: { value: 0.3, display: "$0.30", samples: 6, total: 6, refs: [] },
           delta: null,
-          display: "",
+          display: "—",
         },
       },
     },
@@ -213,8 +275,8 @@ export const caseListData: CaseListData = {
         {
           name: "roots-correct",
           score: 0,
-          detail: "期望 x=2,得到 x=3",
-          evidence: "judge: 求根公式代入时符号写反",
+          detail: "expected x=2, got x=3",
+          evidence: "judge: sign flipped when substituting into the quadratic formula",
         },
       ],
       durationMs: 32_000,

package/src/report/react/format.ts

@@ -1,34 +1,9 @@
-// 展示层的小格式化。注意边界:MetricCell 一律自带 display(格式化发生在计算侧),
-// 组件不重算;这里只服务 OverviewData 这类携带裸数字的字段(耗时、成本、比率),
-// 以及 class 名拼接。缺数据的统一文案也钉在这里,保证各组件同词。
+// web 面的小工具:展示格式化统一住在计算侧的 ../format.ts(两个渲染面同一份),
+// 这里 re-export 并补 class 名拼接。MetricCell 一律自带 display,组件不重算。
 
-/** 全 null / 无样本的统一文案。绝不画 0(docs/reports.md「null ≠ 0」)。 */
-export const MISSING_TEXT = "no data";
+export { MISSING_TEXT, formatDurationMs, formatPercent, formatUSD } from "../format.ts";
 
 /** 拼 class 名:过滤空值,末尾接使用者透传的 className。 */
 export function cx(...parts: (string | undefined | false)[]): string {
   return parts.filter(Boolean).join(" ");
 }
-
-/** 毫秒 → 人读耗时("850ms" / "1.2s" / "4m 20s" / "1h 4m")。 */
-export function formatDurationMs(ms: number): string {
-  if (ms < 1000) return `${Math.round(ms)}ms`;
-  const s = ms / 1000;
-  if (s < 60) return `${s.toFixed(1)}s`;
-  const m = Math.floor(s / 60);
-  if (m < 60) return `${m}m ${Math.round(s % 60)}s`;
-  const h = Math.floor(m / 60);
-  return `${h}h ${m % 60}m`;
-}
-
-/** 美元金额;小额保留更多位数,不四舍五入成 $0.00 的假零。 */
-export function formatUSD(usd: number): string {
-  if (usd === 0) return "$0";
-  const digits = Math.abs(usd) >= 0.01 ? 2 : 4;
-  return `$${usd.toFixed(digits)}`;
-}
-
-/** 0..1 的比率 → 整数百分比。 */
-export function formatPercent(ratio: number): string {
-  return `${Math.round(ratio * 100)}%`;
-}

package/src/report/react/index.tsx

@@ -1,34 +1,51 @@
-// niceeval/report/react —— 纯渲染的报告积木(docs/reports.md 第一档,原型实现)。
+// niceeval/report/react —— 零件复用的导出点:把某一块指标表嵌进已有内部面板时从这里
+// import。导出的组件与 niceeval/report 是同一批双面组件(web 面即 React 渲染,零 IO、
+// 可进 "use client",不 hydrate 也完整),都带自己的 data 计算函数;那是零件的复用,
+// 不是另一套报告系统 —— 报告的家在官方宿主(--report)。
 //
 // 契约:
-//   - 组件只认「算好的可序列化数据」:零 IO、零 hooks、零数据操作,可进 "use client";
-//   - 不 hydrate 也完整:renderToStaticMarkup 的产物即成品——排序在数据侧预排,
-//     hover 退化为 title,下钻是普通 <a>,展开折叠是 <details>;
+//   - 组件只认「算好的可序列化数据」:零 hooks、零数据操作;
 //   - 样式随包发布:配套 ./styles.css(nre-* 稳定类名),使用者在其后加载覆盖即可;
 //   - 跨块配色一致:维度键 → 稳定散列 → 固定调色板下标(colors.ts)。
-//
-// 数据从哪来:niceeval/report 的计算函数(table()/matrix()/scoreboard()/scatter()/
-// overview()/delta()/cases())。计算层在并行实验中实现;类型暂由 ./data.ts 声明,
-// 合并时切换为从计算层 re-export。
 
-export { RunOverview } from "./RunOverview.tsx";
-export { MetricTable } from "./MetricTable.tsx";
-export { MetricMatrix } from "./MetricMatrix.tsx";
-export { Scoreboard } from "./Scoreboard.tsx";
-export { MetricScatter } from "./MetricScatter.tsx";
-export { DeltaTable } from "./DeltaTable.tsx";
-export { CaseList } from "./CaseList.tsx";
+export {
+  CaseList,
+  DeltaTable,
+  MetricBars,
+  MetricLine,
+  MetricMatrix,
+  MetricScatter,
+  MetricTable,
+  RunOverview,
+  Scoreboard,
+} from "../components.tsx";
+export type {
+  CaseListProps,
+  DeltaTableProps,
+  MetricLineProps,
+  MetricMatrixProps,
+  MetricScatterProps,
+  MetricTableProps,
+  RunOverviewProps,
+  ScoreboardProps,
+} from "../components.tsx";
 
-// 数据契约类型(临时家在 ./data.ts,见该文件头部说明)
+// 数据契约类型(家在 ../types.ts,「算」与「画」两侧共用同一份)
 export type {
   AttemptRef,
-  MetricColumn,
-  MetricCell,
-  TableData,
+  CaseListData,
+  DeltaData,
+  LineAxis,
+  LineData,
   MatrixData,
-  ScoreboardData,
-  ScatterData,
+  MetricCell,
+  MetricColumn,
   OverviewData,
-  DeltaData,
-  CaseListData,
-} from "./data.ts";
+  ScatterData,
+  ScoreboardData,
+  SelectionWarning,
+  TableData,
+} from "../types.ts";
+
+// 稳定配色(自定义组件想与官方组件同键同色时用)
+export { NRE_PALETTE, colorClassForKey, colorHexForKey, colorIndexForKey } from "./colors.ts";

package/src/report/react/render.test.tsx

@@ -54,7 +54,7 @@ describe("RunOverview", () => {
     expect(html).toContain("2 snapshots");
     expect(html).toContain("compare/bub");
     expect(html).toContain("2026-07-01T10:00:00Z");
-    expect(html).toContain("compare/bub 快照缺 3 个 eval 的结果");
+    expect(html).toContain("snapshot covers 9 of 12 evals seen in history");
     expect(html).toContain("nre-warnings");
   });
 });
@@ -69,9 +69,9 @@ describe("MetricTable", () => {
   });
 
   it("列头带 label、unit 与 better 方向", () => {
-    expect(html).toContain("通过率");
+    expect(html).toContain("pass rate");
     expect(html).toContain("(%)");
-    expect(html).toContain("代码行数");
+    expect(html).toContain("code lines");
     expect(html).toContain("↑");
     expect(html).toContain("↓");
   });
@@ -96,7 +96,7 @@ describe("MetricMatrix", () => {
   const html = renderToStaticMarkup(<MetricMatrix data={matrixData} attemptHref={attemptHref} />);
 
   it("caption 标出指标与行列维度", () => {
-    expect(html).toContain("通过率");
+    expect(html).toContain("pass rate");
     expect(html).toContain("eval × agent");
   });
 
@@ -158,8 +158,8 @@ describe("MetricScatter", () => {
 
   it("内联 SVG + 轴标签", () => {
     expect(html).toContain("<svg");
-    expect(html).toContain("成本($)");
-    expect(html).toContain("通过率(%)");
+    expect(html).toContain("cost($)");
+    expect(html).toContain("pass rate(%)");
   });
 
   it("better:lower 的 x 轴反向:便宜($5)在贵($10)右边,好的角落恒在右上", () => {
@@ -239,8 +239,8 @@ describe("CaseList", () => {
   it("逐条失败断言:name、score、detail、evidence", () => {
     expect(html).toContain("roots-correct");
     expect(html).toContain("score 0");
-    expect(html).toContain("期望 x=2,得到 x=3");
-    expect(html).toContain("judge: 求根公式代入时符号写反");
+    expect(html).toContain("expected x=2, got x=3");
+    expect(html).toContain("judge: sign flipped when substituting into the quadratic formula");
   });
 
   it("errored 的 error 摘要", () => {

package/src/report/react/styles.css

@@ -300,3 +300,80 @@ table.nre,
 .nre-case-list .nre-case-empty {
   color: #737373;
 }
+
+/* ---- MetricBars ---- */
+.nre-metric-bars {
+  margin: 1rem 0;
+}
+.nre-metric-bars .nre-bars-svg {
+  max-width: 640px;
+  width: 100%;
+  height: auto;
+  display: block;
+}
+.nre-metric-bars .nre-bars-legend {
+  color: #737373;
+  font-size: 0.85em;
+  margin-top: 0.25rem;
+}
+.nre-metric-bars .nre-bars-metric {
+  margin-right: 1em;
+  font-weight: 600;
+  color: #171717;
+}
+.nre-metric-bars .nre-legend-key {
+  margin-right: 0.75em;
+}
+
+/* ---- MetricLine ---- */
+.nre-metric-line {
+  margin: 1rem 0;
+}
+.nre-metric-line .nre-line-svg {
+  max-width: 640px;
+  width: 100%;
+  height: auto;
+  display: block;
+}
+.nre-metric-line .nre-line-missing {
+  color: #b45309;
+  font-size: 0.85em;
+  margin: 0.25rem 0 0;
+}
+.nre-metric-line .nre-line-point-link circle:hover {
+  stroke: #171717;
+  stroke-width: 1.5;
+}
+
+/* ---- 排版原语 Row / Col / Section / Text ---- */
+.nre-col {
+  display: flex;
+  flex-direction: column;
+  gap: 1rem;
+}
+.nre-row {
+  display: flex;
+  flex-wrap: wrap;
+  gap: 1.5rem;
+  align-items: flex-start;
+}
+.nre-row > * {
+  flex: 1 1 20rem;
+  min-width: 0;
+}
+.nre-section .nre-section-title {
+  font-size: 1.1em;
+  font-weight: 700;
+  margin: 0 0 0.5rem;
+}
+.nre-text {
+  color: #525252;
+  margin: 0.25rem 0;
+}
+
+/* ---- DefaultReport(官方水位整块)---- */
+.nre-default-report {
+  display: flex;
+  flex-direction: column;
+  gap: 1rem;
+}