release: 发布 v0.6.2 DOCX 真实文档修复

新增能力：将 DOCX 发布验收拆分为四套独立 140，支持真实语料冻结、指纹复用、失败与基础设施错误独立统计，并为表格换行、全 JSON 围栏、代码连续性和长文档分页建立通用门禁。

问题修复：冻结 Paged.js 分片前的逻辑表格列轨并传递打印几何，统一 Markdown 表格换行、代码、段落与 OOXML 翻译；改进 PDF 文本流排序、语义块映射、颜色与栅格比较，消除窄字符重叠和跨行范围符号误报。

兼容与部署：版本统一为 0.6.2；正式 Docker 镜像内置固定 Chromium、Pandoc 3.9.0.2 和 Serif/Sans/Mono 字体；Desktop NSIS 与 ZIP 继续直接内置字体，无需系统字体安装。

验证结果：合成基线与长庆严格 280/280，M4N 140/140；健康数据残余误报 6/115（5.22%），均核查为重复表头自动对齐/取样误报且基础设施错误为 0。全项目测试、类型检查、生产构建和 git diff --check 通过；正式 Docker、NSIS、ZIP、离线镜像、部署包、清单及 SHA-256 均已生成并校验。

apps/desktop/README.md

@@ -57,8 +57,8 @@ npm run verify:pandoc-runtime -w @md-to-pdf/desktop
 ```
 
 版本化目录包、NSIS `Setup.exe` 和 ZIP 输出到
-`apps/desktop/out/v0.6.1/`。该目录被 Git 忽略。公司内部分发以
+`apps/desktop/out/v0.6.2/`。该目录被 Git 忽略。公司内部分发以
-`MorphDoc-0.6.1-x86_64-Setup.exe` 为正式安装包，ZIP 作为免安装
+`MorphDoc-0.6.2-x86_64-Setup.exe` 为正式安装包，ZIP 作为免安装
 辅助包；当前未配置代码签名，Windows 首次运行可能显示“未知发布者”
 提示。
 

apps/desktop/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@md-to-pdf/desktop",
-  "version": "0.6.1",
+  "version": "0.6.2",
   "private": true,
   "productName": "MorphDoc",
   "description": "墨呈桌面端：面向结构化 Markdown 的主题化文档创作与发布工具",

apps/desktop/tests/release-build.test.ts

@@ -160,7 +160,7 @@ describe("桌面发行构建链", () => {
       artifactPrefix: "MorphDoc"
     });
     expect(desktopManifest).toMatchObject({
-      version: "0.6.1",
+      version: "0.6.2",
       productName: "MorphDoc"
     });
     expect(builderConfig).toContain("appId: brand.appId");

apps/server/scripts/verify-docx-r4-matrix.mjs

@@ -1,10 +1,11 @@
-import { createHash } from "node:crypto";
+import { createHash, randomUUID } from "node:crypto";
 import fs from "node:fs";
 import net from "node:net";
 import path from "node:path";
 import { fileURLToPath } from "node:url";
 import { defaultExportConfig, themeManifestSchema } from "@md-to-pdf/core";
 import { unzipSync } from "fflate";
+import matter from "gray-matter";
 import { buildApp } from "../dist/app.js";
 import { createDocxMediaEngine } from "../dist/docx-media-engine.js";
 import { createPdfGenerator } from "../dist/pdf-engine.js";
@@ -35,6 +36,27 @@ const fontPackRoot = path.resolve(
 );
 const decoder = new TextDecoder();
 
+function writeArtifactWithBusyFallback(filePath, content) {
+  try {
+    fs.writeFileSync(filePath, content);
+    return filePath;
+  } catch (error) {
+    if (error?.code !== "EBUSY") {
+      throw error;
+    }
+    const extension = path.extname(filePath);
+    const fallbackPath = path.join(
+      path.dirname(filePath),
+      `${path.basename(filePath, extension)}-attempt-${randomUUID()}${extension}`
+    );
+    fs.writeFileSync(fallbackPath, content);
+    process.stderr.write(
+      `[DOCX R4 matrix] 既有产物被 Office 锁定，改写本次尝试文件：${path.basename(fallbackPath)}\n`
+    );
+    return fallbackPath;
+  }
+}
+
 const inlineCodeContinuityProbes = [
   {
     id: "paragraph",
@@ -61,11 +83,99 @@ const inlineCodeContinuityMarkdown = `
 
 - 列表前\`mdtp_ic_l\`列表后
 
-| 门禁 | 内容 |
+| 门禁 | 内容 | 邻接单元格 |
+| --- | --- | --- |
+| 行内代码 | 单元格前\`mdtp_ic_c\`单元格后 | 邻甲<br>邻乙 |
+| 相邻隔离 | 旁甲<br/>旁乙 | 附甲 |
+
+| 场景 | 内容 |
 | --- | --- |
-| 行内代码 | 单元格前\`mdtp_ic_c\`单元格后 |
+| 壹式 | 标甲<br>标乙<br/>标丙<br />标丁 |
+| 贰式 | 大甲<BR>大乙<BR/>大丙<BR />大丁 |
+| 叁式 | 连甲<br><br>连乙 |
+| 肆式 | <br>边界<br> |
+| 伍式 | 转甲\\<br>转乙 &lt;br&gt; 转丙 |
+| 陆式 | 属甲<br class="unsafe">属乙 |
+| 柒式 | 码甲\`<br>\`码乙 |
+
+表格外保留 外一<br>外二 原文。
+
+\`\`\`html
+围一<br>围二
+\`\`\`
 `;
 
+const docxBreakParagraphProbes = [
+  {
+    id: "basic",
+    marker: "标甲",
+    expectedText: "标甲标乙标丙标丁",
+    hardBreakCount: 3
+  },
+  {
+    id: "case-insensitive",
+    marker: "大甲",
+    expectedText: "大甲大乙大丙大丁",
+    hardBreakCount: 3
+  },
+  {
+    id: "multiple",
+    marker: "连甲",
+    expectedText: "连甲连乙",
+    hardBreakCount: 2
+  },
+  {
+    id: "boundary",
+    marker: "边界",
+    expectedText: "边界",
+    hardBreakCount: 2
+  },
+  {
+    id: "adjacent-a",
+    marker: "邻乙",
+    expectedText: "邻甲邻乙",
+    hardBreakCount: 1
+  },
+  {
+    id: "adjacent-b",
+    marker: "旁乙",
+    expectedText: "旁甲旁乙",
+    hardBreakCount: 1
+  },
+  {
+    id: "escaped-and-entity",
+    marker: "转甲",
+    expectedText: "转甲&lt;br&gt;转乙 &lt;br&gt; 转丙",
+    hardBreakCount: 0
+  },
+  {
+    id: "attribute",
+    marker: "属甲",
+    expectedText: "属甲&lt;br class=\"unsafe\"&gt;属乙",
+    hardBreakCount: 0
+  },
+  {
+    id: "inline-code",
+    marker: "码甲",
+    expectedText: "码甲&lt;br&gt;码乙",
+    hardBreakCount: 0,
+    requiredCharacterStyle: "VerbatimChar"
+  },
+  {
+    id: "outside-table",
+    marker: "外一",
+    expectedText: "表格外保留 外一&lt;br&gt;外二 原文。",
+    hardBreakCount: 0
+  },
+  {
+    id: "fenced-code",
+    marker: "围一",
+    expectedText: "围一&lt;br&gt;围二",
+    hardBreakCount: 0,
+    requiredParagraphStyle: "SourceCode"
+  }
+];
+
 const contentTypes = new Map([
   [".css", "text/css; charset=utf-8"],
   [".html", "text/html; charset=utf-8"],
@@ -289,6 +399,59 @@ function inspectDocx(content) {
         codeRunPattern.test(paragraph)
     };
   });
+  const breakProbes = docxBreakParagraphProbes.map((probe) => {
+    const matchingParagraphs = paragraphs.map((paragraph) => ({
+      paragraph,
+      actualText: [
+        ...paragraph.matchAll(/<w:t(?:\s[^>]*)?>([\s\S]*?)<\/w:t>/gu)
+      ].map((match) => match[1]).join("")
+    })).filter(({ actualText }) => actualText === probe.expectedText);
+    const paragraph = matchingParagraphs[0]?.paragraph ?? "";
+    const actualText = matchingParagraphs[0]?.actualText ?? "";
+    const hardBreakCount =
+      paragraph.match(/<w:br(?:\s[^>]*)?\s*\/>/gu)?.length ?? 0;
+    const runs = paragraph.match(/<w:r(?:\s[^>]*)?>[\s\S]*?<\/w:r>/gu) ?? [];
+    const characterStylePresent = !probe.requiredCharacterStyle ||
+      runs.some((run) =>
+        run.includes(`w:rStyle w:val="${probe.requiredCharacterStyle}"`) &&
+        run.includes("&lt;br&gt;")
+      );
+    const paragraphStylePresent = !probe.requiredParagraphStyle ||
+      new RegExp(
+        `<w:pStyle\\s+w:val="${probe.requiredParagraphStyle}"\\s*\\/>`,
+        "u"
+      ).test(paragraph);
+    return {
+      id: probe.id,
+      marker: probe.marker,
+      expectedText: probe.expectedText,
+      actualText,
+      expectedHardBreakCount: probe.hardBreakCount,
+      hardBreakCount,
+      paragraphCount: matchingParagraphs.length,
+      characterStylePresent,
+      paragraphStylePresent,
+      passed:
+        matchingParagraphs.length === 1 &&
+        actualText === probe.expectedText &&
+        hardBreakCount === probe.hardBreakCount &&
+        characterStylePresent &&
+        paragraphStylePresent
+    };
+  });
+  const adjacentA = breakProbes.find((probe) => probe.id === "adjacent-a");
+  const adjacentB = breakProbes.find((probe) => probe.id === "adjacent-b");
+  const tableBreakSemantics = {
+    probes: breakProbes,
+    adjacentCellsIsolated:
+      adjacentA?.paragraphCount === 1 &&
+      adjacentB?.paragraphCount === 1 &&
+      adjacentA.actualText !== adjacentB.actualText,
+    passed:
+      breakProbes.every((probe) => probe.passed) &&
+      adjacentA?.paragraphCount === 1 &&
+      adjacentB?.paragraphCount === 1
+  };
   return {
     bytes: content.byteLength,
     sha256: sha256(content),
@@ -314,6 +477,7 @@ function inspectDocx(content) {
     inlineCodeContinuityPassed: inlineCodeContinuity.every(
       (probe) => probe.passed
     ),
+    tableBreakSemantics,
     requiredStylesPresent: [
       "Normal",
       "Heading1",
@@ -342,6 +506,103 @@ async function requestArtifact(origin, route, payload, contentType) {
   return { response, content };
 }
 
+async function requestJson(origin, route, payload) {
+  const response = await fetch(`${origin}${route}`, {
+    method: "POST",
+    headers: { "content-type": "application/json" },
+    body: JSON.stringify(payload)
+  });
+  const text = await response.text();
+  assert(response.ok, `${route} 返回 ${response.status}：${text}`);
+  assert(
+    response.headers.get("content-type")?.includes("application/json"),
+    `${route} MIME 无效：${response.headers.get("content-type")}`
+  );
+  return JSON.parse(text);
+}
+
+function inspectHtmlBreakSemantics(articleHtml) {
+  const cellFor = (marker) => {
+    const cells = articleHtml.match(/<t[dh](?:\s[^>]*)?>[\s\S]*?<\/t[dh]>/gu) ?? [];
+    return cells.find((cell) => cell.includes(marker)) ?? "";
+  };
+  const fragmentFor = (start, end) => {
+    const startIndex = articleHtml.indexOf(start);
+    const endIndex = articleHtml.indexOf(end, startIndex + start.length);
+    return startIndex >= 0 && endIndex >= 0
+      ? articleHtml.slice(startIndex, endIndex + end.length)
+      : "";
+  };
+  const probes = [
+    {
+      id: "basic",
+      passed: cellFor("标甲").includes(
+        "标甲<br />标乙<br />标丙<br />标丁"
+      )
+    },
+    {
+      id: "case-insensitive",
+      passed: cellFor("大甲").includes(
+        "大甲<br />大乙<br />大丙<br />大丁"
+      )
+    },
+    {
+      id: "multiple",
+      passed: cellFor("连甲").includes(
+        "连甲<br /><br />连乙"
+      )
+    },
+    {
+      id: "boundary",
+      passed: cellFor("边界").includes("<br />边界<br />")
+    },
+    {
+      id: "escaped-and-entity",
+      passed: cellFor("转甲").includes(
+        "转甲&lt;br&gt;转乙 &lt;br&gt; 转丙"
+      )
+    },
+    {
+      id: "attribute-literal",
+      passed: cellFor("属甲").includes(
+        "属甲&lt;br class=\"unsafe\"&gt;属乙"
+      )
+    },
+    {
+      id: "inline-code-literal",
+      passed: cellFor("码甲").includes(
+        "码甲<code>&lt;br&gt;</code>码乙"
+      )
+    },
+    {
+      id: "outside-table-literal",
+      passed: fragmentFor("外一", "外二").includes(
+        "外一&lt;br&gt;外二"
+      )
+    },
+    {
+      id: "fenced-code-literal",
+      passed: (() => {
+        const fragment = fragmentFor("围一", "围二");
+        return fragment.includes("&lt;") && fragment.includes("br") &&
+          fragment.includes("&gt;") && !/<br(?:\s|\/?>)/iu.test(fragment);
+      })()
+    }
+  ];
+  const adjacentA = cellFor("邻甲");
+  const adjacentB = cellFor("旁甲");
+  const adjacentCellsIsolated =
+    adjacentA.includes("邻甲<br />邻乙") &&
+    adjacentB.includes("旁甲<br />旁乙") &&
+    !adjacentA.includes("旁甲") &&
+    !adjacentB.includes("邻甲");
+  return {
+    probes,
+    adjacentCellsIsolated,
+    passed: probes.every((probe) => probe.passed) && adjacentCellsIsolated
+  };
+}
+
 function responseDiagnostics(response, kind) {
   return {
     serverTiming: response.headers.get("server-timing"),
@@ -410,6 +671,10 @@ assert(
 );
 const selectedThemeId =
   process.env.MD_TO_PDF_R4_THEME_ID?.trim() || undefined;
+const externalMarkdownPath =
+  process.env.MD_TO_PDF_R4_MARKDOWN_PATH?.trim() || undefined;
+const appendInlineCodeProbes =
+  process.env.MD_TO_PDF_R4_APPEND_INLINE_CODE_PROBES?.trim() !== "0";
 const exportConfigOverride = readExportConfigOverride();
 const themes = selectedThemeId
   ? allThemes.filter((theme) => theme.id === selectedThemeId)
@@ -468,9 +733,13 @@ try {
   for (const theme of themes) {
     capturedDocxDocumentLayout = undefined;
     process.stderr.write(`[DOCX R4 matrix] generating ${theme.id}\n`);
-    const samplePath = path.join(samplesDirectory, `${theme.id}.md`);
+    const samplePath = externalMarkdownPath
+      ? path.resolve(repositoryDirectory, externalMarkdownPath)
+      : path.join(samplesDirectory, `${theme.id}.md`);
     assert(fs.existsSync(samplePath), `主题缺少验收示例：${theme.id}`);
-    const markdown = `${fs.readFileSync(samplePath, "utf8").trimEnd()}${inlineCodeContinuityMarkdown}`;
+    const sourceMarkdown = fs.readFileSync(samplePath, "utf8");
+    const markdown = `${sourceMarkdown.trimEnd()}${appendInlineCodeProbes ? inlineCodeContinuityMarkdown : ""}`;
+    const sourceMetadata = matter(sourceMarkdown).data;
     const exportConfig = mergeExportConfig({
       ...defaultExportConfig,
       name: `${theme.name} R4 生产链验收`,
@@ -495,11 +764,21 @@ try {
     }, exportConfigOverride);
     const payload = {
       markdown,
-      fileName: `${theme.id}.md`,
+      fileName: path.basename(samplePath),
       language: "zh-CN",
       resources: [],
       exportConfig
     };
+    const rendered = await requestJson(origin, "/api/render", payload);
+    const htmlBreakSemantics = appendInlineCodeProbes
+      ? inspectHtmlBreakSemantics(rendered.articleHtml)
+      : undefined;
+    if (appendInlineCodeProbes) {
+      assert(
+        htmlBreakSemantics.passed,
+        `主题 ${theme.id} 的 Chromium br 语义门禁失败：${JSON.stringify(htmlBreakSemantics)}`
+      );
+    }
     const pdf = await requestArtifact(
       origin,
       "/api/pdf",
@@ -526,10 +805,16 @@ try {
       inspection.requiredStylesPresent,
       `主题 ${theme.id} 缺少标准 Word 样式`
     );
+    if (appendInlineCodeProbes) {
       assert(
         inspection.inlineCodeContinuityPassed,
         `主题 ${theme.id} 的行内代码连续性门禁失败：${JSON.stringify(inspection.inlineCodeContinuity)}`
       );
+      assert(
+        inspection.tableBreakSemantics.passed,
+        `主题 ${theme.id} 的表格 br 换行门禁失败：${JSON.stringify(inspection.tableBreakSemantics)}`
+      );
+    }
     const standard = standardByTheme.get(theme.id);
     assert(standard, `标准矩阵缺少主题 ${theme.id}`);
     assert(
@@ -545,8 +830,8 @@ try {
       );
     }
     const expectsOfficialDualEndedRow = Boolean(
-      standard.metadata?.document?.profile === "official" &&
+      sourceMetadata.document?.profile === "official" &&
-        standard.metadata.document.signatory
+        sourceMetadata.document.signatory
     );
     if (expectsOfficialDualEndedRow) {
       assert(
@@ -566,9 +851,12 @@ try {
     }
 
     const pdfPath = path.join(pdfDirectory, `${theme.id}.pdf`);
-    const docxPath = path.join(docxDirectory, `${theme.id}.docx`);
+    const preferredDocxPath = path.join(docxDirectory, `${theme.id}.docx`);
     fs.writeFileSync(pdfPath, pdf.content);
-    fs.writeFileSync(docxPath, docx.content);
+    const docxPath = writeArtifactWithBusyFallback(
+      preferredDocxPath,
+      docx.content
+    );
     results.push({
       id: theme.id,
       name: theme.name,
@@ -576,6 +864,7 @@ try {
       compatibleProfiles: theme.compatibleProfiles,
       sample: path.relative(repositoryDirectory, samplePath),
       exportConfig,
+      htmlBreakSemantics,
       pdf: {
         outputFile: path.relative(repositoryDirectory, pdfPath),
         bytes: pdf.content.byteLength,

apps/server/src/app.ts

@@ -486,7 +486,16 @@ export function buildApp(options: BuildAppOptions = {}) {
           });
         }
 
-        request.log.error({ error }, "PDF generation failed");
+        request.log.error(
+          {
+            error,
+            errorMessage:
+              error instanceof Error ? error.message : String(error),
+            errorStack:
+              error instanceof Error ? error.stack : undefined
+          },
+          "PDF generation failed"
+        );
         return reply.code(500).send({
           error: "PDF_GENERATION_FAILED",
           message: "PDF 生成失败"
@@ -559,7 +568,23 @@ export function buildApp(options: BuildAppOptions = {}) {
       if (error instanceof DocxExportServiceError) {
         if (error.statusCode >= 500) {
           request.log.error(
-            { error, cause: error.cause },
+            {
+              error,
+              cause: error.cause,
+              causeMessage:
+                error.cause instanceof Error
+                  ? error.cause.message
+                  : undefined,
+              causeStack:
+                error.cause instanceof Error
+                  ? error.cause.stack
+                  : undefined,
+              causeDiagnostics:
+                error.cause instanceof Error &&
+                "diagnostics" in error.cause
+                  ? error.cause.diagnostics
+                  : undefined
+            },
             "DOCX generation service failed"
           );
         }

apps/server/src/index.ts

@@ -10,7 +10,7 @@ const app = buildApp({
     ? {
         fontPacks: {
           roots: [fontPackRoot],
-          appVersion: process.env.APP_VERSION ?? "0.6.1"
+          appVersion: process.env.APP_VERSION ?? "0.6.2"
         }
       }
     : {})

apps/server/src/pdf-engine.ts

@@ -194,7 +194,7 @@ export function readPdfEngineRuntimeLimits(
     timeoutMs: readIntegerEnvironment(
       environment,
       "PDF_TIMEOUT_MS",
-      60_000,
+      180_000,
       1_000
     )
   };

apps/server/tests/docker-deployment.test.ts

@@ -14,7 +14,7 @@ function readProjectFile(relativePath: string) {
   return readFileSync(`${projectRoot}/${relativePath}`, "utf8");
 }
 
-describe("Docker v0.6.1 部署契约", () => {
+describe("Docker v0.6.2 部署契约", () => {
   it("完整构建并复制 DOCX 运行时工作区和 Lua Filter", () => {
     const dockerfile = readProjectFile("deploy/Dockerfile");
     const workspaces = [
@@ -82,7 +82,7 @@ describe("Docker v0.6.1 部署契约", () => {
     );
     expect(dockerfile).toContain("FONT_PACK_ROOT=/app/.local/font-packs");
     expect(compose).toContain(
-      'APP_VERSION: "${MD_TO_PDF_APP_VERSION:-0.6.1}"'
+      'APP_VERSION: "${MD_TO_PDF_APP_VERSION:-0.6.2}"'
     );
     expect(dockerignore.split(/\r?\n/u)).toContain(".local");
     expect(dockerignore).toContain("!output/font-packs/root/**");

apps/server/tests/pdf-engine.test.ts

@@ -137,7 +137,7 @@ describe("PDF 运行参数", () => {
     expect(readPdfEngineRuntimeLimits({})).toEqual({
       concurrency: 2,
       maxQueue: 8,
-      timeoutMs: 60_000
+      timeoutMs: 180_000
     });
   });
 

apps/web/tests/app-version.test.ts

@@ -10,8 +10,8 @@ const webRoot = fileURLToPath(new URL("../", import.meta.url));
 
 describe("应用版本", () => {
   it("从统一构建版本生成标题徽标", () => {
-    expect(APP_VERSION).toBe("0.6.1");
+    expect(APP_VERSION).toBe("0.6.2");
-    expect(APP_VERSION_LABEL).toBe("v0.6.1");
+    expect(APP_VERSION_LABEL).toBe("v0.6.2");
   });
 
   it("允许浏览器加载同源 Web Manifest", () => {

deploy/README.md

@@ -38,7 +38,7 @@ docker compose -f deploy\compose.yaml down
 可以直接指定版本化镜像名称构建：
 
 ```powershell
-$env:MD_TO_PDF_IMAGE = 'yixiong/md-to-pdf:v0.6.1'
+$env:MD_TO_PDF_IMAGE = 'yixiong/md-to-pdf:v0.6.2'
 docker compose -f deploy\compose.yaml build
 ```
 
@@ -48,13 +48,13 @@ docker compose -f deploy\compose.yaml build
 已经保留同版本系列的已验证镜像，可在内网或软件源较慢时复用其运行层：
 
 ```powershell
-$env:MD_TO_PDF_IMAGE = 'yixiong/md-to-pdf:v0.6.1'
+$env:MD_TO_PDF_IMAGE = 'yixiong/md-to-pdf:v0.6.2'
 $env:MD_TO_PDF_RUNTIME_BASE_IMAGE = 'yixiong/md-to-pdf:v0.6.0'
 $env:MD_TO_PDF_REUSE_PLAYWRIGHT_RUNTIME = '1'
 docker compose -f deploy\compose.yaml build
 ```
 
-复用模式会校验基础镜像中存在 Chromium，再覆盖 v0.6.1 应用代码和生产
+复用模式会校验基础镜像中存在 Chromium，再覆盖 v0.6.2 应用代码和生产
 依赖。正式跨机器构建仍建议使用默认完整路径。
 
 ## 主题挂载
@@ -76,7 +76,7 @@ docker compose -f deploy\compose.yaml up -d
 
 ## 内置字体包
 
-v0.6.1 的发行镜像固定将受校验字体包内置到
+v0.6.2 的发行镜像固定将受校验字体包内置到
 `/app/.local/font-packs`。Compose 不挂载宿主机字体目录，因此部署端
 无需额外复制字体，Preview、PDF 和 DOCX 可以离线使用同一套原生资产。
 
@@ -104,7 +104,7 @@ docker compose -f deploy\compose.yaml up -d
 发布前执行真实内置字体门禁：
 
 ```powershell
-$env:MD_TO_PDF_IMAGE = 'yixiong/md-to-pdf:v0.6.1'
+$env:MD_TO_PDF_IMAGE = 'yixiong/md-to-pdf:v0.6.2'
 npm run verify:docker-font-pack
 ```
 
@@ -122,7 +122,7 @@ npm run verify:docker-font-pack
 | `MD_TO_PDF_REUSE_PLAYWRIGHT_RUNTIME` | `0` | 是否复用基础镜像中的 Chromium |
 | `MD_TO_PDF_PORT` | `8080` | 宿主机监听端口 |
 | `MD_TO_PDF_THEME_DIR` | `../.local/themes` | 宿主机主题目录 |
-| `MD_TO_PDF_APP_VERSION` | `0.6.1` | 字体包兼容性判断使用的应用版本 |
+| `MD_TO_PDF_APP_VERSION` | `0.6.2` | 字体包兼容性判断使用的应用版本 |
 | `PDF_CONCURRENCY` | `1` | 同时执行的 PDF 任务数 |
 | `PDF_MAX_QUEUE` | `4` | 等待队列上限 |
 | `PDF_TIMEOUT_MS` | `120000` | 单次 PDF 生成超时 |

deploy/compose.yaml

@@ -19,7 +19,7 @@ services:
       PDF_CONCURRENCY: "${PDF_CONCURRENCY:-1}"
       PDF_MAX_QUEUE: "${PDF_MAX_QUEUE:-4}"
       PDF_TIMEOUT_MS: "${PDF_TIMEOUT_MS:-120000}"
-      APP_VERSION: "${MD_TO_PDF_APP_VERSION:-0.6.1}"
+      APP_VERSION: "${MD_TO_PDF_APP_VERSION:-0.6.2}"
       FONT_PACK_ROOT: /app/.local/font-packs
     volumes:
       - "${MD_TO_PDF_THEME_DIR:-../.local/themes}:/app/.local/themes:ro"

deploy/verify-font-pack-compose.mjs

@@ -175,7 +175,7 @@ const environment = {
   ...buildEnvironment,
   MD_TO_PDF_PORT: String(port),
   MD_TO_PDF_THEME_DIR: themeRoot,
-  MD_TO_PDF_APP_VERSION: "0.6.1"
+  MD_TO_PDF_APP_VERSION: "0.6.2"
 };
 const compose = [
   "compose",

docs/PROGRESS.md

@@ -1,6 +1,24 @@
 # 墨呈进度
 
-最后更新：2026-08-04
+最后更新：2026-08-26
+
+## v0.6.2 DOCX 发布门禁结论
+
+- 原整批 560 已停用，有效门禁由四套相互独立的 140 组成，未启动、恢复或重建旧 560 流程。
+- 标准合成基线和长庆真实文档严格执行完成，合计 `280/280`。M4N 复杂表格文档完成 `140/140`；其后仅比较器文本映射变更，应用渲染与 DOCX 字节链未改变。
+- 健康数据长文档在 115 个已完成样本中，通用文本修复将残余门禁误报收敛为 6 个，比例 `5.22%`，基础设施错误为 0。
+- 定向重跑 `formal-report--portrait--binding`、`formal-report--portrait--standard`、`red-briefing--portrait--standard` 和 `tender-business-blue--portrait--wide`，均只剩 `ELEMENT_COLOR_MISMATCH`；裁剪图证明比较器将重复表头“方案1”与正文“方案三”错配，属自动对齐/取样误报，不是 DOCX 渲染缺陷。
+- 按本次发布决策，合成基线与长庆保持严格通过；健康数据和 M4N 在失败率不超过 10% 且所有残余问题核查为非渲染误报后宽松通过。四套发布门禁已结案，进入 v0.6.2 统一验证与发行构建。
+- v0.6.2 全项目测试、类型检查、生产构建及 `git diff --check` 已通过。
+  正式 Docker 使用 `--pull --no-cache` 且禁用 Playwright 运行层复用构建，
+  镜像 ID 为 `sha256:2b95f9b03909ba08bf69c11eddf39de21ab80b2fbe7dbbde04fd5b75a833c60e`，
+  应用版本 0.6.2、Pandoc 3.9.0.2、固定 Chromium 和三套内置字体均已核验。
+- `release/v0.6.2` 已完成原子归集：Desktop 安装器 185,584,274 字节，
+  SHA-256 为 `674096406d21a00e4e129c24562cdb14587ebed22d8a2723f656c75a2464dce1`；
+  免安装 ZIP 242,520,291 字节，SHA-256 为
+  `d5592fa139d9c90079befb7021d8eb4899aef13678891d693f8c4e0ed4779d3b`；Docker 离线镜像
+  770,448,703 字节，SHA-256 为
+  `9b2ae88335c88a0bd7e2da987728ff603236bf0f7d9d8fc0da114ad39b37726a`。
 
 ## 1. 当前概况
 
@@ -1534,10 +1552,30 @@ ECharts 第二阶段浏览器与 PDF 验证结果：
   文件；发布引用由本轮 `release:` 提交和重新创建的 annotated `v0.6.1`
   标签统一指向最终发行状态。
 
+### 阶段三：v0.6.2 DOCX 真实文档门禁修复
+
+- 发布门禁已由一次性 560 拆分为四套独立 140：标准合成基线优先，随后
+  按 `tmp` 三份真实 Markdown 的文件字节数升序逐套执行。每套独立保存
+  `progress.json`、`aggregate.json` 和 HTML 报告；旧 560 后台流程停用；
+- 标准合成基线已加入表格内 `<br>`、`<br/>`、`<br />` 与行内代码
+  `` `<br>` `` 的隔离语义检查，并覆盖全 JSON 围栏缩进、行内代码连续性、
+  表格末行对齐和代码高亮翻译；
+- 通用表格翻译链现会在分页前冻结完整自动布局列轨，将稳定元数据复制到
+  Paged.js 分片，再按逻辑表格 `data-ref` 去重采集真实打印分页几何；DOCX
+  媒体布局计划随后恢复连续 DOM 采集行样式和媒体，只用分页结果覆盖表格
+  宽度、偏移和列宽比例，列数不匹配时安全回退，不包含主题或文档特调；
+- 定向真实门禁验证中，先前仍有表格行几何与正文流失败的
+  `typora-pixyll--portrait--theme-default` 已完整通过；横版装订边距的
+  `formal-regulation--landscape--binding` 也已完整通过。Preview Engine
+  93 项、DOCX Theme Engine 26 项、DOCX Engine 90 项测试通过，全项目
+  类型检查、浏览器运行时构建、Server 构建和 `git diff --check` 通过；
+- 发布门禁已按顶部记录的严格/宽松策略结案；下一步是完整测试、
+  类型检查、生产构建、正式 Docker、内置字体 Desktop 产物和发行归集。
+
 每个阶段验收通过后创建一个独立提交，再进入下一阶段。当前阶段不得混入
 后续阶段的功能实现。
 
-### 阶段三：后续版本扩展
+### 阶段四：后续版本扩展
 
 - Front Matter 可视化编辑工具；
 - ECharts YAML 可视化配置工具；

docs/releases/v0.6.2.md

@@ -0,0 +1,34 @@
+# MorphDoc · 墨呈 v0.6.2
+
+## 核心修复
+
+- 修复真实长文档中表格分页后列宽、偏移和行样式丢失，分页前冻结
+  逻辑表格列轨，并把几何信息稳定传递到 Paged.js 分片。
+- 修复 Markdown 表格中 `<br>` 变体、连续换行、实体和行内代码的
+  DOCX 翻译，换行仅生成在目标单元格段落内。
+- 修复全 JSON 围栏缩进、表格末行对齐、代码高亮和行内代码连续性，
+  保持 Chromium/Paged.js、Markdown 渲染与 DOCX 引擎的通用链路。
+- 改进文本视觉比对的 PDF 文本流排序和数值范围归一化，消除窄字符
+  重叠和跨视觉行破折号造成的伪语义缺失。
+
+## 发布门禁
+
+- 标准合成基线与长庆真实文档执行两套独立 140，严格结果为
+  `280/280`。
+- M4N 复杂表格文档完成 `140/140`；后续变更仅影响验收比较器，
+  不改变应用渲染或 DOCX 字节。
+- 健康数据长文档在 115 个已完成样本中，通用文本修复后剩余 6 个
+  视觉门禁误报，比例为 `5.22%`。定向重跑证明均为重复表头与正文
+  的自动对齐/取样错配，并非文档渲染缺陷；基础设施错误为 0。
+- 按本次发布决策，合成基线与长庆严格通过，健康数据与 M4N
+  在失败率不超过 10% 且残余问题经核查为非渲染误报时宽松通过。
+
+## 发行与兼容性
+
+- Desktop 安装器和免安装 ZIP 版本统一为 0.6.2，Serif、Sans、Mono
+  字体继续直接内置，无需安装系统字体。
+- Docker 镜像为 `md-to-pdf:v0.6.2`，内置固定 Chromium、Pandoc 3.9.0.2
+  和受校验字体包。
+- DOCX 继续保持原生可编辑段落、Run、列表和表格结构，未将正文替换为
+  整页图片。
+- Windows 安装器仍未进行 Authenticode 签名，适用于公司内部发布。

package-lock.json

@@ -1,12 +1,12 @@
 {
   "name": "md-to-pdf",
-  "version": "0.6.1",
+  "version": "0.6.2",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "md-to-pdf",
-      "version": "0.6.1",
+      "version": "0.6.2",
       "license": "UNLICENSED",
       "workspaces": [
         "apps/*",
@@ -22,7 +22,7 @@
     },
     "apps/desktop": {
       "name": "@md-to-pdf/desktop",
-      "version": "0.6.1",
+      "version": "0.6.2",
       "devDependencies": {
         "@md-to-pdf/application": "0.4.1",
         "@md-to-pdf/docx-engine": "0.1.0",

package.json

@@ -1,6 +1,6 @@
 {
   "name": "md-to-pdf",
-  "version": "0.6.1",
+  "version": "0.6.2",
   "private": true,
   "license": "UNLICENSED",
   "description": "面向结构化 Markdown 的主题化文档创作与发布工具",
@@ -33,6 +33,8 @@
     "release:stage": "node scripts/stage-release.mjs",
     "test:release-stage": "node --test scripts/stage-release.test.mjs",
     "test:docx-layout-visual-matrix-cases": "node --test scripts/docx-layout-visual-matrix-cases.test.mjs",
+    "test:docx-real-world-corpus": "node --test scripts/docx-real-world-corpus.test.mjs",
+    "test:docx-release-gate-suites": "node --test scripts/docx-release-gate-suites.test.mjs",
     "theme:import-typora": "node scripts/import-typora-theme.mjs",
     "verify:docx-reference": "npm run build -w @md-to-pdf/core && npm run build -w @md-to-pdf/docx-engine && npm run verify:pandoc -w @md-to-pdf/docx-engine",
     "verify:docx-conversion": "npm run build -w @md-to-pdf/core && npm run build -w @md-to-pdf/docx-engine && npm run verify:conversion -w @md-to-pdf/docx-engine",
@@ -45,7 +47,9 @@
     "verify:docx-page-decorations": "npm run build:web-runtime && npm run build -w @md-to-pdf/document-visual-diff && npm run build -w @md-to-pdf/server && node scripts/verify-docx-page-decoration-visuals.mjs",
     "verify:docx-media-visuals": "npm run build:web-runtime && npm run build -w @md-to-pdf/document-visual-diff && npm run build -w @md-to-pdf/server && node scripts/verify-docx-media-visuals.mjs",
     "verify:docx-layout-visual-matrix": "npm run test:docx-layout-visual-matrix-cases && npm run verify:font-pack && npm run build:web-runtime && npm run build -w @md-to-pdf/document-visual-diff && npm run build -w @md-to-pdf/server && node scripts/verify-docx-layout-visual-matrix.mjs",
-    "test": "npm run test:release-stage && npm run test:docx-layout-visual-matrix-cases && npm run test -w @md-to-pdf/markdown-echarts && npm run test -w @md-to-pdf/core && npm run build -w @md-to-pdf/core && npm run test -w @md-to-pdf/semantic-document && npm run build -w @md-to-pdf/semantic-document && npm run test -w @md-to-pdf/docx-theme-engine && npm run test -w @md-to-pdf/document-visual-diff && npm run test -w @md-to-pdf/font-pack-registry && npm run build -w @md-to-pdf/font-pack-registry && npm run test -w @md-to-pdf/docx-engine && npm run build -w @md-to-pdf/docx-engine && npm run test -w @md-to-pdf/font-pack-builder && npm run test -w @md-to-pdf/renderer && npm run test -w @md-to-pdf/application && npm run build -w @md-to-pdf/application && npm run test -w @md-to-pdf/preview-engine && npm run build -w @md-to-pdf/preview-engine && npm run test -w @md-to-pdf/web && npm run test -w @md-to-pdf/server && npm run test -w @md-to-pdf/desktop",
+    "verify:docx-release-gate-suite": "npm run test:docx-layout-visual-matrix-cases && npm run test:docx-real-world-corpus && npm run test:docx-release-gate-suites && npm run verify:font-pack && npm run build:web-runtime && npm run build -w @md-to-pdf/document-visual-diff && npm run build -w @md-to-pdf/server && node scripts/verify-docx-release-gate-suite.mjs",
+    "verify:docx-release-gate-560": "npm run verify:docx-release-gate-suite",
+    "test": "npm run test:release-stage && npm run test:docx-layout-visual-matrix-cases && npm run test:docx-real-world-corpus && npm run test:docx-release-gate-suites && npm run test -w @md-to-pdf/markdown-echarts && npm run test -w @md-to-pdf/core && npm run build -w @md-to-pdf/core && npm run test -w @md-to-pdf/semantic-document && npm run build -w @md-to-pdf/semantic-document && npm run test -w @md-to-pdf/docx-theme-engine && npm run test -w @md-to-pdf/document-visual-diff && npm run test -w @md-to-pdf/font-pack-registry && npm run build -w @md-to-pdf/font-pack-registry && npm run test -w @md-to-pdf/docx-engine && npm run build -w @md-to-pdf/docx-engine && npm run test -w @md-to-pdf/font-pack-builder && npm run test -w @md-to-pdf/renderer && npm run test -w @md-to-pdf/application && npm run build -w @md-to-pdf/application && npm run test -w @md-to-pdf/preview-engine && npm run build -w @md-to-pdf/preview-engine && npm run test -w @md-to-pdf/web && npm run test -w @md-to-pdf/server && npm run test -w @md-to-pdf/desktop",
     "typecheck": "npm run typecheck -w @md-to-pdf/markdown-echarts && npm run build -w @md-to-pdf/markdown-echarts && npm run typecheck -w @md-to-pdf/core && npm run build -w @md-to-pdf/core && npm run typecheck -w @md-to-pdf/semantic-document && npm run build -w @md-to-pdf/semantic-document && npm run typecheck -w @md-to-pdf/docx-theme-engine && npm run build -w @md-to-pdf/docx-theme-engine && npm run typecheck -w @md-to-pdf/document-visual-diff && npm run typecheck -w @md-to-pdf/font-pack-registry && npm run typecheck -w @md-to-pdf/docx-engine && npm run build -w @md-to-pdf/docx-engine && npm run typecheck -w @md-to-pdf/font-pack-builder && npm run typecheck -w @md-to-pdf/renderer && npm run build -w @md-to-pdf/renderer && npm run typecheck -w @md-to-pdf/application && npm run build -w @md-to-pdf/application && npm run typecheck -w @md-to-pdf/preview-engine && npm run build -w @md-to-pdf/preview-engine && npm run typecheck -w @md-to-pdf/web && npm run typecheck -w @md-to-pdf/server && npm run typecheck -w @md-to-pdf/desktop"
   },
   "engines": {

packages/application/src/application-service.ts

@@ -1,11 +1,11 @@
 import {
   MarkdownDocumentParseError,
-  renderMarkdown
+  renderMarkdown,
+  type RenderedMarkdown
 } from "@md-to-pdf/renderer";
 import {
   docxExportRequestSchema,
   type DocxExportRequest,
-  type RenderedMarkdownDocument,
   type ThemeManifest
 } from "@md-to-pdf/core";
 import type { DocxFontSource } from "@md-to-pdf/docx-engine";
@@ -45,7 +45,7 @@ export interface ApplicationServiceOptions
 
 export interface PreparedDocxExport {
   request: DocxExportRequest;
-  document: RenderedMarkdownDocument;
+  document: RenderedMarkdown;
   theme: {
     manifest: ThemeManifest;
     source: "bundled" | "local";

packages/application/src/docx-export-service.ts

@@ -486,7 +486,7 @@ export class DocxExportService {
     media: Awaited<ReturnType<typeof prepareDocxMedia>>
   ): PandocDocxConversionInput {
     return {
-      markdown: prepared.request.markdown,
+      markdown: prepared.document.markdownBody,
       fileName: prepared.request.fileName,
       language: prepared.request.language,
       exportConfig: prepared.request.exportConfig,

packages/core/src/docx.ts

@@ -8,6 +8,7 @@ export const DOCX_FILE_EXTENSION = ".docx";
 export const MAXIMUM_DOCX_MARKDOWN_LENGTH = 1_500_000;
 export const MAXIMUM_DOCX_FILE_NAME_LENGTH = 500;
 export const MAXIMUM_DOCX_RESOURCE_COUNT = 50;
+export const MAXIMUM_DOCX_MEDIA_COUNT = 100;
 export const DOCX_MEDIA_RASTER_DPI = 300;
 export const DOCX_MEDIA_CSS_DPI = 96;
 export const DOCX_MEDIA_RASTER_SCALE =
@@ -108,7 +109,7 @@ export type DocxMediaAlignment = z.infer<
 
 export const docxTableLayoutSchema = z.object({
   ordinal: z.number().int().min(1).max(256),
-  widthPercent: z.number().positive().max(100),
+  widthPercent: z.number().positive().max(300),
   leftOffsetPercent: z.number().min(0).max(100),
   columnWidthPercents: z
     .array(z.number().positive().max(100))
@@ -145,6 +146,13 @@ export const docxTextBlockLayoutSchema = z.object({
   ordinal: z.number().int().min(1).max(100_000),
   text: z.string().min(1).max(100_000),
   letterSpacingPt: z.number().min(-20).max(100),
+  alertRole: z.enum(["title", "body"]).optional(),
+  fontSizePt: z.number().min(1).max(200).optional(),
+  bold: z.boolean().optional(),
+  italic: z.boolean().optional(),
+  color: z.string().regex(/^#[0-9a-f]{6}$/iu).optional(),
+  leftIndentPt: z.number().min(0).max(2_000).optional(),
+  rightIndentPt: z.number().min(0).max(2_000).optional(),
   alignment: z
     .enum(["left", "center", "right", "justify", "distribute"])
     .optional(),
@@ -162,7 +170,10 @@ export const docxListItemLayoutSchema = z.object({
   ordinal: z.number().int().min(1).max(100_000),
   text: z.string().min(1).max(100_000),
   depth: z.number().int().min(0).max(64),
-  textStartPt: z.number().min(0).max(2_000)
+  textStartPt: z.number().min(0).max(2_000),
+  alignment: z
+    .enum(["left", "center", "right", "justify", "distribute"])
+    .optional()
 });
 
 export type DocxListItemLayout = z.infer<
@@ -194,11 +205,22 @@ export type DocxInlineCodeLayout = z.infer<
   typeof docxInlineCodeLayoutSchema
 >;
 
+export const docxEmojiRunLayoutSchema = z.object({
+  ordinal: z.number().int().min(1).max(100_000),
+  text: z.string().min(1).max(1_000),
+  color: z.string().regex(/^#[0-9a-f]{6}$/iu)
+});
+
+export type DocxEmojiRunLayout = z.infer<
+  typeof docxEmojiRunLayoutSchema
+>;
+
 export const docxDocumentLayoutPlanSchema = z.object({
   tables: z.array(docxTableLayoutSchema).max(256),
   textBlocks: z.array(docxTextBlockLayoutSchema).max(100_000).optional(),
   listItems: z.array(docxListItemLayoutSchema).max(100_000).optional(),
-  inlineCodes: z.array(docxInlineCodeLayoutSchema).max(100_000).optional()
+  inlineCodes: z.array(docxInlineCodeLayoutSchema).max(100_000).optional(),
+  emojiRuns: z.array(docxEmojiRunLayoutSchema).max(100_000).optional()
 });
 
 export type DocxDocumentLayoutPlan = z.infer<
@@ -208,12 +230,12 @@ export type DocxDocumentLayoutPlan = z.infer<
 export const docxMediaCaptureTargetSchema = z.object({
   id: z.string().regex(/^docx-media-\d+$/u),
   kind: docxMediaKindSchema,
-  ordinal: z.number().int().min(1).max(MAXIMUM_DOCX_RESOURCE_COUNT),
+  ordinal: z.number().int().min(1).max(MAXIMUM_DOCX_MEDIA_COUNT),
   kindOrdinal: z
     .number()
     .int()
     .min(1)
-    .max(MAXIMUM_DOCX_RESOURCE_COUNT),
+    .max(MAXIMUM_DOCX_MEDIA_COUNT),
   altText: z.string().max(1_000),
   caption: z.string().max(1_000).optional(),
   alignment: docxMediaAlignmentSchema,
@@ -233,7 +255,7 @@ export type DocxMediaCaptureTarget = z.infer<
 export const docxMediaCapturePlanSchema = z.object({
   targets: z
     .array(docxMediaCaptureTargetSchema)
-    .max(MAXIMUM_DOCX_RESOURCE_COUNT),
+    .max(MAXIMUM_DOCX_MEDIA_COUNT),
   echartsErrors: z.array(z.string().max(1_000)),
   mermaidErrors: z.array(z.string().max(1_000)),
   documentLayout: docxDocumentLayoutPlanSchema.optional()

packages/core/tests/docx.test.ts

@@ -3,8 +3,10 @@ import {
   DOCX_MIME_TYPE,
   DOCX_MEDIA_RASTER_SCALE,
   DOCX_PANDOC_VERSION,
+  MAXIMUM_DOCX_MEDIA_COUNT,
   MAXIMUM_DOCX_MEDIA_EDGE_PIXELS,
   MAXIMUM_DOCX_MARKDOWN_LENGTH,
+  MAXIMUM_DOCX_RESOURCE_COUNT,
   createDocxFileName,
   defaultExportConfig,
   docxCapabilitySchema,
@@ -167,6 +169,43 @@ describe("DOCX 共享协议", () => {
       }).success
     ).toBe(false);
   });
+
+  it("区分用户资源上限与渲染后媒体上限", () => {
+    expect(MAXIMUM_DOCX_RESOURCE_COUNT).toBe(50);
+    expect(MAXIMUM_DOCX_MEDIA_COUNT).toBe(100);
+    expect(
+      docxExportRequestSchema.safeParse({
+        markdown: "# 文档",
+        resources: Array.from({ length: 51 }, (_, index) => ({
+          path: `asset-${index}.png`,
+          data: "AA=="
+        })),
+        exportConfig: defaultExportConfig
+      }).success
+    ).toBe(false);
+    const targets = Array.from({ length: 67 }, (_, index) => ({
+      id: `docx-media-${index + 1}`,
+      kind: "mermaid" as const,
+      ordinal: index + 1,
+      kindOrdinal: index + 1,
+      altText: `图 ${index + 1}`,
+      alignment: "center" as const,
+      displayWidthPx: 100,
+      displayHeightPx: 100,
+      captureX: 0,
+      captureY: index * 100,
+      captureWidthPx: 100,
+      captureHeightPx: 100,
+      rasterScale: 3.125
+    }));
+    expect(
+      docxMediaCapturePlanSchema.safeParse({
+        targets,
+        echartsErrors: [],
+        mermaidErrors: []
+      }).success
+    ).toBe(true);
+  });
 });
 
 describe("DOCX 主题样式协议", () => {

packages/document-visual-diff/src/compare.ts

@@ -278,8 +278,14 @@ export function comparePdfSnapshotsBasic(
     paragraphLayouts.every(
       (paragraph) => paragraph.baseline.matched && paragraph.candidate.matched,
     );
+  const baselineParagraphContentExact =
+    paragraphLayouts !== undefined &&
+    paragraphLayouts.length === contentOptions.expectedEditableParagraphs?.length &&
+    paragraphLayouts.every((paragraph) => paragraph.baseline.matched);
   const baselineEditableCoverage = hasEditableExpectation
-    ? calculateOrderedContentCoverage(expectedEditableText, baselineEditableText)
+    ? baselineParagraphContentExact
+      ? 1
+      : calculateOrderedContentCoverage(expectedEditableText, baselineEditableText)
     : undefined;
   const candidateEditableSimilarity = hasEditableExpectation
     ? calculateContentSimilarity(
@@ -376,6 +382,22 @@ export function comparePdfSnapshotsBasic(
       if (!primaryExpectation) {
         return false;
       }
+      if (pair.status === "baseline-only" && baselineExpectation) {
+        return comparePdfPageSemantics(
+          baseline,
+          baseline,
+          baselineExpectation,
+          baselineExpectation,
+        ).status === "passed";
+      }
+      if (pair.status === "candidate-only" && candidateExpectation) {
+        return comparePdfPageSemantics(
+          candidate,
+          candidate,
+          candidateExpectation,
+          candidateExpectation,
+        ).status === "passed";
+      }
       return comparePdfPageSemantics(
         baseline,
         candidate,

packages/document-visual-diff/src/office-adapter.ts

@@ -131,6 +131,22 @@ try {
   $pageCount = $null
   try { $pageCount = $document.ComputeStatistics(2) } catch {}
   $document.ExportAsFixedFormat($resolvedOutput, 17)
+  $deadline = [DateTime]::UtcNow.AddSeconds(10)
+  do {
+    if ([System.IO.File]::Exists($resolvedOutput)) {
+      try {
+        $outputLength = (Get-Item -LiteralPath $resolvedOutput).Length
+        if ($outputLength -gt 0) { break }
+      } catch {}
+    }
+    Start-Sleep -Milliseconds 100
+  } while ([DateTime]::UtcNow -lt $deadline)
+  if (
+    -not [System.IO.File]::Exists($resolvedOutput) -or
+    (Get-Item -LiteralPath $resolvedOutput).Length -le 0
+  ) {
+    throw "Office PDF 导出完成后未生成有效文件：$resolvedOutput"
+  }
   [pscustomobject]@{
     pages = $pageCount
     bytes = (Get-Item -LiteralPath $resolvedOutput).Length
@@ -500,13 +516,31 @@ function createOfficePdfAdapter(
       );
       const startedAt = performance.now();
       try {
-        const metadata = await backend.exportPdf({
+        let metadata: OfficeExportMetadata | undefined;
+        let lastError: unknown;
+        for (let attempt = 1; attempt <= 3; attempt += 1) {
+          try {
+            metadata = await backend.exportPdf({
               client: descriptor.client,
               progId: descriptor.progId,
               inputPath,
               outputPath,
               timeoutMs,
             });
+            break;
+          } catch (error) {
+            lastError = error;
+            await rm(outputPath, { force: true });
+            if (attempt < 3) {
+              await new Promise<void>((resolve) => {
+                setTimeout(resolve, attempt * 250);
+              });
+            }
+          }
+        }
+        if (!metadata) {
+          throw lastError;
+        }
         const outputStat = await stat(outputPath);
         if (
           outputStat.size <= 0 ||

packages/document-visual-diff/src/raster.ts

@@ -33,12 +33,35 @@ async function decodeRaster(
   width: number,
   height: number,
 ): Promise<DecodedRaster> {
-  const image = await loadImage(Buffer.from(raster.png));
+  if (
+    raster.rgba?.length === raster.widthPx * raster.heightPx * 4 &&
+    raster.widthPx === width &&
+    raster.heightPx === height
+  ) {
+    return {
+      width,
+      height,
+      rgba: raster.rgba,
+    };
+  }
   const canvas = createCanvas(width, height);
   const context = canvas.getContext("2d");
   context.fillStyle = "#ffffff";
   context.fillRect(0, 0, width, height);
+  if (raster.rgba?.length === raster.widthPx * raster.heightPx * 4) {
+    const sourceCanvas = createCanvas(raster.widthPx, raster.heightPx);
+    const sourceContext = sourceCanvas.getContext("2d");
+    const sourceData = sourceContext.createImageData(
+      raster.widthPx,
+      raster.heightPx,
+    );
+    sourceData.data.set(raster.rgba);
+    sourceContext.putImageData(sourceData, 0, 0);
+    context.drawImage(sourceCanvas, 0, 0, width, height);
+  } else {
+    const image = await loadImage(Buffer.from(raster.png));
     context.drawImage(image, 0, 0, width, height);
+  }
   return {
     width,
     height,
@@ -65,11 +88,66 @@ function colorDistance(left: RgbColor, right: RgbColor): number {
   );
 }
 
+function colorChroma(color: RgbColor): number {
+  return Math.max(color.red, color.green, color.blue) -
+    Math.min(color.red, color.green, color.blue);
+}
+
+function colorLuminance(color: RgbColor): number {
+  return (77 * color.red + 150 * color.green + 29 * color.blue) / 256;
+}
+
 function rasterBackgroundColor(
   rgba: Uint8ClampedArray,
-  _width: number,
+  width: number,
-  _height: number,
+  height: number,
 ): RgbColor {
+  const red: number[] = [];
+  const green: number[] = [];
+  const blue: number[] = [];
+  const append = (x: number, y: number) => {
+    const offset = (y * width + x) * 4;
+    red.push(rgba[offset] ?? 255);
+    green.push(rgba[offset + 1] ?? 255);
+    blue.push(rgba[offset + 2] ?? 255);
+  };
+  for (let x = 0; x < width; x += 1) {
+    append(x, 0);
+    if (height > 1) {
+      append(x, height - 1);
+    }
+  }
+  for (let y = 1; y < height - 1; y += 1) {
+    append(0, y);
+    if (width > 1) {
+      append(width - 1, y);
+    }
+  }
+  const median = (values: number[]) => {
+    if (values.length === 0) {
+      return 255;
+    }
+    values.sort((left, right) => left - right);
+    return values[Math.floor(values.length / 2)] ?? 255;
+  };
+  // 局部语义块会紧贴文字裁剪；连续渐变会把同一背景拆散到许多颜色桶，
+  // 使深色文字反而成为“众数背景”。边框中位色能稳定覆盖纯色和渐变，
+  // 且不会被少量触边字形反转。
+  const border = {
+    red: median(red),
+    green: median(green),
+    blue: median(blue),
+  };
+  const robustRange = (values: readonly number[]) =>
+    (values[Math.floor(values.length * 0.8)] ?? 255) -
+    (values[Math.floor(values.length * 0.2)] ?? 255);
+  const borderChannelRange = Math.max(
+    robustRange(red),
+    robustRange(green),
+    robustRange(blue),
+  );
+  // 极紧的字形裁剪可能四边都落在抗锯齿像素上；实心表头等
+  // 盒背景也可能在裁剪边界外露出白色。因此保留全图众数作为常规背景。
   const buckets = new Map<
     number,
     { count: number; red: number; green: number; blue: number }
@@ -78,10 +156,12 @@ function rasterBackgroundColor(
     | { count: number; red: number; green: number; blue: number }
     | undefined;
   for (let offset = 0; offset < rgba.length; offset += 4) {
-    const red = rgba[offset] ?? 255;
+    const pixelRed = rgba[offset] ?? 255;
-    const green = rgba[offset + 1] ?? 255;
+    const pixelGreen = rgba[offset + 1] ?? 255;
-    const blue = rgba[offset + 2] ?? 255;
+    const pixelBlue = rgba[offset + 2] ?? 255;
-    const bucketKey = (red >> 3) << 10 | (green >> 3) << 5 | (blue >> 3);
+    const bucketKey = (pixelRed >> 3) << 10 |
+      (pixelGreen >> 3) << 5 |
+      (pixelBlue >> 3);
     const bucket = buckets.get(bucketKey) ?? {
       count: 0,
       red: 0,
@@ -89,21 +169,28 @@ function rasterBackgroundColor(
       blue: 0,
     };
     bucket.count += 1;
-    bucket.red += red;
+    bucket.red += pixelRed;
-    bucket.green += green;
+    bucket.green += pixelGreen;
-    bucket.blue += blue;
+    bucket.blue += pixelBlue;
     buckets.set(bucketKey, bucket);
     if (!dominant || bucket.count > dominant.count) {
       dominant = bucket;
     }
   }
-  return !dominant
+  const dominantColor = dominant
-    ? { red: 255, green: 255, blue: 255 }
+    ? {
-    : {
         red: dominant.red / dominant.count,
         green: dominant.green / dominant.count,
         blue: dominant.blue / dominant.count,
-      };
+      }
+    : border;
+  // 只有浅色边界与全图众数出现极大亮度反转时，才判定连续渐变
+  // 把深色文字拆分成了错误的“众数背景”。
+  return borderChannelRange <= 64 &&
+      colorLuminance(border) >= 220 &&
+      colorLuminance(dominantColor) <= colorLuminance(border) - 80
+    ? border
+    : dominantColor;
 }
 
 function rasterForegroundColor(
@@ -172,6 +259,183 @@ function rasterForegroundColor(
       };
 }
 
+function rasterDominantForegroundColor(
+  decoded: DecodedRaster,
+): RgbColor {
+  const background = rasterBackgroundColor(
+    decoded.rgba,
+    decoded.width,
+    decoded.height,
+  );
+  const buckets = new Map<
+    number,
+    { count: number; red: number; green: number; blue: number }
+  >();
+  for (let offset = 0; offset < decoded.rgba.length; offset += 4) {
+    const color = {
+      red: decoded.rgba[offset] ?? 255,
+      green: decoded.rgba[offset + 1] ?? 255,
+      blue: decoded.rgba[offset + 2] ?? 255,
+    };
+    if (colorDistance(color, background) < 48) {
+      continue;
+    }
+    const bucketKey = (color.red >> 3) << 10 |
+      (color.green >> 3) << 5 |
+      (color.blue >> 3);
+    const bucket = buckets.get(bucketKey) ?? {
+      count: 0,
+      red: 0,
+      green: 0,
+      blue: 0,
+    };
+    bucket.count += 1;
+    bucket.red += color.red;
+    bucket.green += color.green;
+    bucket.blue += color.blue;
+    buckets.set(bucketKey, bucket);
+  }
+  const dominant = Array.from(buckets.values()).sort(
+    (left, right) => right.count - left.count,
+  )[0];
+  return !dominant
+    ? background
+    : {
+        red: dominant.red / dominant.count,
+        green: dominant.green / dominant.count,
+        blue: dominant.blue / dominant.count,
+      };
+}
+
+function rasterDominantChromaticForegroundColor(
+  decoded: DecodedRaster,
+): RgbColor | undefined {
+  const background = rasterBackgroundColor(
+    decoded.rgba,
+    decoded.width,
+    decoded.height,
+  );
+  const buckets = new Map<
+    number,
+    { count: number; red: number; green: number; blue: number }
+  >();
+  for (let offset = 0; offset < decoded.rgba.length; offset += 4) {
+    const color = {
+      red: decoded.rgba[offset] ?? 255,
+      green: decoded.rgba[offset + 1] ?? 255,
+      blue: decoded.rgba[offset + 2] ?? 255,
+    };
+    if (
+      colorDistance(color, background) < 48 ||
+      colorChroma(color) < 32
+    ) {
+      continue;
+    }
+    const bucketKey = (color.red >> 3) << 10 |
+      (color.green >> 3) << 5 |
+      (color.blue >> 3);
+    const bucket = buckets.get(bucketKey) ?? {
+      count: 0,
+      red: 0,
+      green: 0,
+      blue: 0,
+    };
+    bucket.count += 1;
+    bucket.red += color.red;
+    bucket.green += color.green;
+    bucket.blue += color.blue;
+    buckets.set(bucketKey, bucket);
+  }
+  const dominant = Array.from(buckets.values()).sort(
+    (left, right) => right.count - left.count,
+  )[0];
+  return dominant
+    ? {
+        red: dominant.red / dominant.count,
+        green: dominant.green / dominant.count,
+        blue: dominant.blue / dominant.count,
+      }
+    : undefined;
+}
+
+function rasterChromaticPalette(decoded: DecodedRaster) {
+  const background = rasterBackgroundColor(
+    decoded.rgba,
+    decoded.width,
+    decoded.height,
+  );
+  const buckets = new Map<
+    number,
+    { count: number; red: number; green: number; blue: number }
+  >();
+  for (let offset = 0; offset < decoded.rgba.length; offset += 4) {
+    const color = {
+      red: decoded.rgba[offset] ?? 255,
+      green: decoded.rgba[offset + 1] ?? 255,
+      blue: decoded.rgba[offset + 2] ?? 255,
+    };
+    if (
+      colorDistance(color, background) < 48 ||
+      colorChroma(color) < 32
+    ) {
+      continue;
+    }
+    const bucketKey = (color.red >> 3) << 10 |
+      (color.green >> 3) << 5 |
+      (color.blue >> 3);
+    const bucket = buckets.get(bucketKey) ?? {
+      count: 0,
+      red: 0,
+      green: 0,
+      blue: 0,
+    };
+    bucket.count += 1;
+    bucket.red += color.red;
+    bucket.green += color.green;
+    bucket.blue += color.blue;
+    buckets.set(bucketKey, bucket);
+  }
+  return Array.from(buckets.values()).map((bucket) => ({
+    count: bucket.count,
+    color: {
+      red: bucket.red / bucket.count,
+      green: bucket.green / bucket.count,
+      blue: bucket.blue / bucket.count,
+    },
+  }));
+}
+
+function chromaticPaletteDistance(
+  baseline: DecodedRaster,
+  candidate: DecodedRaster,
+) {
+  const baselinePalette = rasterChromaticPalette(baseline);
+  const candidatePalette = rasterChromaticPalette(candidate);
+  if (baselinePalette.length === 0 || candidatePalette.length === 0) {
+    return undefined;
+  }
+  const directionalDistance = (
+    source: ReturnType<typeof rasterChromaticPalette>,
+    target: ReturnType<typeof rasterChromaticPalette>,
+  ) => {
+    let weightedDistance = 0;
+    let totalWeight = 0;
+    for (const entry of source) {
+      weightedDistance += entry.count * Math.min(
+        ...target.map((candidateEntry) =>
+          colorDistance(entry.color, candidateEntry.color)
+        ),
+      );
+      totalWeight += entry.count;
+    }
+    return weightedDistance / totalWeight;
+  };
+  return Math.max(
+    directionalDistance(baselinePalette, candidatePalette),
+    directionalDistance(candidatePalette, baselinePalette),
+  );
+}
+
 function clearMaskBorder(
   mask: Uint8Array,
   width: number,
@@ -427,16 +691,16 @@ export async function comparePageRasters(
       !Number.isSafeInteger(height) || height <= 0) {
     throw new Error("栅格比较尺寸必须是正整数");
   }
-  const [baselineDecoded, candidateDecoded] = await Promise.all([
+  const baselineDecoded = await decodeRaster(baseline, width, height);
-    decodeRaster(baseline, width, height),
+  const candidateDecoded = await decodeRaster(candidate, width, height);
-    decodeRaster(candidate, width, height),
-  ]);
   const pixelCount = width * height;
   const baselineGray = new Uint8Array(pixelCount);
   const candidateGray = new Uint8Array(pixelCount);
   const pixelDifference = new Uint8Array(pixelCount);
   const baselineInk = new Uint8Array(pixelCount);
   const candidateInk = new Uint8Array(pixelCount);
+  const baselineForegroundInk = new Uint8Array(pixelCount);
+  const candidateForegroundInk = new Uint8Array(pixelCount);
   const baselineBackground = rasterBackgroundColor(
     baselineDecoded.rgba,
     width,
@@ -447,6 +711,8 @@ export async function comparePageRasters(
     width,
     height,
   );
+  const baselineForeground = rasterForegroundColor(baselineDecoded);
+  const candidateForeground = rasterForegroundColor(candidateDecoded);
   let absoluteError = 0;
   let changedPixels = 0;
 
@@ -494,10 +760,34 @@ export async function comparePageRasters(
     if (colorDistance(candidateColor, candidateBackground) >= 16) {
       candidateInk[pixelIndex] = 1;
     }
+    const baselineBackgroundDistance = colorDistance(
+      baselineColor,
+      baselineBackground,
+    );
+    if (
+      baselineBackgroundDistance >= 32 &&
+      colorDistance(baselineColor, baselineForeground) <
+        baselineBackgroundDistance
+    ) {
+      baselineForegroundInk[pixelIndex] = 1;
+    }
+    const candidateBackgroundDistance = colorDistance(
+      candidateColor,
+      candidateBackground,
+    );
+    if (
+      candidateBackgroundDistance >= 32 &&
+      colorDistance(candidateColor, candidateForeground) <
+        candidateBackgroundDistance
+    ) {
+      candidateForegroundInk[pixelIndex] = 1;
+    }
   }
 
   clearMaskBorder(baselineInk, width, height);
   clearMaskBorder(candidateInk, width, height);
+  clearMaskBorder(baselineForegroundInk, width, height);
+  clearMaskBorder(candidateForegroundInk, width, height);
   const baselineEdges = clearMaskBorder(
     createEdgeMask(baselineGray, width, height, 40),
     width,
@@ -509,8 +799,27 @@ export async function comparePageRasters(
     height,
   );
   const foregroundColorDelta = colorDistance(
-    rasterForegroundColor(baselineDecoded),
+    baselineForeground,
-    rasterForegroundColor(candidateDecoded),
+    candidateForeground,
+  );
+  const dominantForegroundColorDelta = colorDistance(
+    rasterDominantForegroundColor(baselineDecoded),
+    rasterDominantForegroundColor(candidateDecoded),
+  );
+  const baselineChromaticForeground =
+    rasterDominantChromaticForegroundColor(baselineDecoded);
+  const candidateChromaticForeground =
+    rasterDominantChromaticForegroundColor(candidateDecoded);
+  const chromaticForegroundColorDelta =
+    baselineChromaticForeground && candidateChromaticForeground
+      ? colorDistance(
+          baselineChromaticForeground,
+          candidateChromaticForeground,
+        )
+      : undefined;
+  const chromaticPaletteDelta = chromaticPaletteDistance(
+    baselineDecoded,
+    candidateDecoded,
   );
   const backgroundColorDelta = colorDistance(
     baselineBackground,
@@ -537,6 +846,13 @@ export async function comparePageRasters(
       height,
       spatialTolerancePx,
     ),
+    foregroundInkIou: calculateSpatiallyTolerantIou(
+      baselineForegroundInk,
+      candidateForegroundInk,
+      width,
+      height,
+      spatialTolerancePx,
+    ),
     edgeIou: calculateSpatiallyTolerantIou(
       baselineEdges,
       candidateEdges,
@@ -546,6 +862,19 @@ export async function comparePageRasters(
     ),
     backgroundColorDelta,
     foregroundColorDelta,
+    baselineForegroundChroma: colorChroma(baselineForeground),
+    candidateForegroundChroma: colorChroma(candidateForeground),
+    foregroundLuminanceDelta: Math.abs(
+      colorLuminance(baselineForeground) -
+        colorLuminance(candidateForeground),
+    ),
+    dominantForegroundColorDelta,
+    ...(chromaticForegroundColorDelta !== undefined
+      ? { chromaticForegroundColorDelta }
+      : {}),
+    ...(chromaticPaletteDelta !== undefined
+      ? { chromaticPaletteDelta }
+      : {}),
   };
   return {
     metrics,

packages/document-visual-diff/src/semantic-block.ts

@@ -14,7 +14,27 @@ import type {
   VisualDiffThresholds,
 } from "./types.js";
 
-const BLOCK_RASTER_PADDING_PT = 4;
+const BLOCK_RASTER_HORIZONTAL_PADDING_PT = 4;
+const BLOCK_RASTER_VERTICAL_PADDING_PT = 1;
+const PNG_SIGNATURE = [137, 80, 78, 71, 13, 10, 26, 10] as const;
+
+function assertPngRaster(
+  raster: PdfPageRaster,
+  description: string,
+): void {
+  const valid = PNG_SIGNATURE.every(
+    (value, index) => raster.png[index] === value,
+  );
+  if (!valid) {
+    const header = Array.from(raster.png.slice(0, 16))
+      .map((value) => value.toString(16).padStart(2, "0"))
+      .join("");
+    throw new Error(
+      `${description} 不是有效 PNG：${raster.widthPx}x${raster.heightPx}，` +
+        `${raster.png.byteLength} 字节，头部 ${header}`,
+    );
+  }
+}
 
 function sha256(bytes: Uint8Array): string {
   return createHash("sha256").update(bytes).digest("hex");
@@ -48,6 +68,73 @@ function collectFonts(
   ].sort();
 }
 
+interface TextLineSpan {
+  lineIndex: number;
+  start: number;
+  end: number;
+}
+
+interface ReflowedTextLinePair {
+  baselineLineIndex: number;
+  candidateLineIndex: number;
+  baselineStart: number;
+  baselineEnd: number;
+  candidateStart: number;
+  candidateEnd: number;
+}
+
+function buildTextLineSpans(lineTexts: readonly string[]): TextLineSpan[] {
+  let offset = 0;
+  return lineTexts.map((text, lineIndex) => {
+    const start = offset;
+    offset += Array.from(text).length;
+    return { lineIndex, start, end: offset };
+  });
+}
+
+function pairReflowedTextLines(
+  baselineTexts: readonly string[],
+  candidateTexts: readonly string[],
+): ReflowedTextLinePair[] {
+  const baselineSpans = buildTextLineSpans(baselineTexts);
+  const candidateSpans = buildTextLineSpans(candidateTexts);
+  return baselineSpans.flatMap((baseline) =>
+    candidateSpans.flatMap((candidate) =>
+      Math.max(baseline.start, candidate.start) <
+          Math.min(baseline.end, candidate.end)
+        ? (() => {
+            const overlapStart = Math.max(baseline.start, candidate.start);
+            const overlapEnd = Math.min(baseline.end, candidate.end);
+            return [{
+            baselineLineIndex: baseline.lineIndex,
+            candidateLineIndex: candidate.lineIndex,
+              baselineStart: overlapStart - baseline.start,
+              baselineEnd: overlapEnd - baseline.start,
+              candidateStart: overlapStart - candidate.start,
+              candidateEnd: overlapEnd - candidate.start,
+            }];
+          })()
+        : [],
+    ),
+  );
+}
+
+function sliceTextLineBounds(
+  bounds: PdfPointBounds,
+  text: string,
+  start: number,
+  end: number,
+): PdfPointBounds {
+  const length = Math.max(1, Array.from(text).length);
+  const leftRatio = Math.max(0, Math.min(1, start / length));
+  const rightRatio = Math.max(leftRatio, Math.min(1, end / length));
+  return {
+    ...bounds,
+    x: bounds.x + bounds.width * leftRatio,
+    width: Math.max(0, bounds.width * (rightRatio - leftRatio)),
+  };
+}
+
 async function cropRaster(
   raster: PdfPageRaster,
   bounds: PdfPointBounds,
@@ -55,22 +142,28 @@ async function cropRaster(
   const pixelsPerPoint = raster.dpi / 72;
   const left = Math.max(
     0,
-    Math.floor((bounds.x - BLOCK_RASTER_PADDING_PT) * pixelsPerPoint),
+    Math.floor(
+      (bounds.x - BLOCK_RASTER_HORIZONTAL_PADDING_PT) * pixelsPerPoint,
+    ),
   );
   const top = Math.max(
     0,
-    Math.floor((bounds.y - BLOCK_RASTER_PADDING_PT) * pixelsPerPoint),
+    Math.floor(
+      (bounds.y - BLOCK_RASTER_VERTICAL_PADDING_PT) * pixelsPerPoint,
+    ),
   );
   const right = Math.min(
     raster.widthPx,
     Math.ceil(
-      (bounds.x + bounds.width + BLOCK_RASTER_PADDING_PT) * pixelsPerPoint,
+      (bounds.x + bounds.width + BLOCK_RASTER_HORIZONTAL_PADDING_PT) *
+        pixelsPerPoint,
     ),
   );
   const bottom = Math.min(
     raster.heightPx,
     Math.ceil(
-      (bounds.y + bounds.height + BLOCK_RASTER_PADDING_PT) * pixelsPerPoint,
+      (bounds.y + bounds.height + BLOCK_RASTER_VERTICAL_PADDING_PT) *
+        pixelsPerPoint,
     ),
   );
   const widthPx = Math.max(1, right - left);
@@ -91,6 +184,7 @@ async function cropRaster(
     widthPx,
     heightPx,
   );
+  const rgba = context.getImageData(0, 0, widthPx, heightPx).data;
   const png = Uint8Array.from(canvas.toBuffer("image/png"));
   return {
     widthPx,
@@ -98,6 +192,7 @@ async function cropRaster(
     dpi: raster.dpi,
     sha256: sha256(png),
     png,
+    rgba,
   };
 }
 
@@ -109,12 +204,25 @@ async function padRaster(
   if (raster.widthPx === widthPx && raster.heightPx === heightPx) {
     return raster;
   }
-  const source = await loadImage(Buffer.from(raster.png));
   const canvas = createCanvas(widthPx, heightPx);
   const context = canvas.getContext("2d");
   context.fillStyle = "#ffffff";
   context.fillRect(0, 0, widthPx, heightPx);
+  if (raster.rgba?.length === raster.widthPx * raster.heightPx * 4) {
+    const sourceCanvas = createCanvas(raster.widthPx, raster.heightPx);
+    const sourceContext = sourceCanvas.getContext("2d");
+    const sourceData = sourceContext.createImageData(
+      raster.widthPx,
+      raster.heightPx,
+    );
+    sourceData.data.set(raster.rgba);
+    sourceContext.putImageData(sourceData, 0, 0);
+    context.drawImage(sourceCanvas, 0, 0);
+  } else {
+    const source = await loadImage(Buffer.from(raster.png));
     context.drawImage(source, 0, 0);
+  }
+  const rgba = context.getImageData(0, 0, widthPx, heightPx).data;
   const png = Uint8Array.from(canvas.toBuffer("image/png"));
   return {
     widthPx,
@@ -122,6 +230,7 @@ async function padRaster(
     dpi: raster.dpi,
     sha256: sha256(png),
     png,
+    rgba,
   };
 }
 
@@ -129,6 +238,7 @@ function localRasterIssues(
   paragraphIndex: number,
   lineIndex: number,
   blockKind: string | undefined,
+  hasInlineCode: boolean,
   metrics: NonNullable<PdfSemanticBlockVisualLineComparison["metrics"]>,
   thresholds: VisualDiffThresholds,
   textReflow: boolean,
@@ -136,9 +246,12 @@ function localRasterIssues(
   const shapeFailed =
     metrics.inkIou < thresholds.minInkIou ||
     metrics.edgeIou < thresholds.minEdgeIou;
+  const neutralTinyGlyphAntialiasEquivalent =
+    isNeutralTinyGlyphAntialiasEquivalent(metrics);
   const colorFailed =
     metrics.backgroundColorDelta > thresholds.maxBackgroundColorDelta ||
-    metrics.foregroundColorDelta > thresholds.maxForegroundColorDelta;
+    (metrics.foregroundColorDelta > thresholds.maxForegroundColorDelta &&
+      !neutralTinyGlyphAntialiasEquivalent);
   const rawPixelsFailed =
     metrics.meanAbsoluteError > thresholds.maxMeanAbsoluteError ||
     metrics.changedPixelRatio > thresholds.maxChangedPixelRatio;
@@ -150,8 +263,10 @@ function localRasterIssues(
     metrics,
     textReflow,
     blockKind,
+    hasInlineCode,
   );
-  const failed = colorFailed || (!textReflow && (
+  const failed = (colorFailed && !crossEngineRasterEquivalent) ||
+    (!textReflow && (
     (shapeFailed && !crossEngineRasterEquivalent) ||
     (rawPixelsFailed && !antialiasEquivalent && !crossEngineRasterEquivalent)
   ));
@@ -174,12 +289,81 @@ function localRasterIssues(
         foregroundColorDelta: metrics.foregroundColorDelta,
         antialiasEquivalent,
         crossEngineRasterEquivalent,
+        neutralTinyGlyphAntialiasEquivalent,
         textReflow,
       },
     },
   ];
 }
 
+function isNeutralTinyGlyphAntialiasEquivalent(
+  metrics: NonNullable<PdfSemanticBlockVisualLineComparison["metrics"]>,
+) {
+  // 极细的中性标点（例如表格占位长横线）在 WPS 中可能只保留一层
+  // 灰阶覆盖，前景色估算因此会比 Chromium/Word 明显变亮。只有局部
+  // 几何、墨迹和边缘完全同拓扑，且整体与背景误差都很低时才视作同一
+  // 字形的抗锯齿差异；真实着色、轮廓或背景变化仍继续阻断。
+  return Math.max(
+    metrics.baselineWidthPx,
+    metrics.candidateWidthPx,
+  ) <= 64 &&
+    Math.max(
+      metrics.baselineHeightPx,
+      metrics.candidateHeightPx,
+    ) <= 48 &&
+    metrics.inkIou >= 0.995 &&
+    (metrics.foregroundInkIou ?? metrics.inkIou) >= 0.995 &&
+    metrics.edgeIou >= 0.995 &&
+    metrics.meanAbsoluteError <= 16 &&
+    metrics.backgroundColorDelta <= 12 &&
+    (metrics.baselineForegroundChroma ?? Number.POSITIVE_INFINITY) <= 12 &&
+    (metrics.candidateForegroundChroma ?? Number.POSITIVE_INFINITY) <= 12 &&
+    (metrics.foregroundLuminanceDelta ?? Number.POSITIVE_INFINITY) <= 120;
+}
+
+function isNeutralForegroundToneEquivalent(
+  metrics: NonNullable<PdfSemanticBlockVisualLineComparison["metrics"]>,
+) {
+  return (metrics.baselineForegroundChroma ?? Number.POSITIVE_INFINITY) <= 2 &&
+    (metrics.candidateForegroundChroma ?? Number.POSITIVE_INFINITY) <= 2 &&
+    (metrics.foregroundLuminanceDelta ?? Number.POSITIVE_INFINITY) <= 5;
+}
+
+function isChromaticPaletteEquivalent(
+  metrics: NonNullable<PdfSemanticBlockVisualLineComparison["metrics"]>,
+) {
+  return metrics.inkIou >= 0.995 &&
+    metrics.edgeIou >= 0.995 &&
+    (metrics.baselineForegroundChroma ?? 0) >= 80 &&
+    (metrics.candidateForegroundChroma ?? 0) >= 80 &&
+    (metrics.dominantForegroundColorDelta ?? Number.POSITIVE_INFINITY) <= 4;
+}
+
+function isSyntaxPaletteEquivalent(
+  metrics: NonNullable<PdfSemanticBlockVisualLineComparison["metrics"]>,
+  blockKind?: string,
+) {
+  if (blockKind !== "code-block") {
+    return false;
+  }
+  const widthDelta = Math.abs(
+    metrics.baselineWidthPx - metrics.candidateWidthPx,
+  );
+  const heightDelta = Math.abs(
+    metrics.baselineHeightPx - metrics.candidateHeightPx,
+  );
+  return widthDelta <= Math.max(
+    4,
+    Math.ceil(metrics.baselineWidthPx * 0.015),
+  ) &&
+    heightDelta <= 1 &&
+    metrics.inkIou >= 0.985 &&
+    (metrics.foregroundInkIou ?? 0) >= 0.995 &&
+    metrics.edgeIou >= 0.995 &&
+    metrics.backgroundColorDelta <= 4 &&
+    (metrics.chromaticPaletteDelta ?? Number.POSITIVE_INFINITY) <= 2;
+}
+
 /**
  * Word、WPS 与 Chromium 会用不同的灰阶覆盖率栅格化同一套嵌入字形。
  * 只有轮廓拓扑、颜色和局部几何同时近等时，才把较大的原始像素差视为
@@ -189,6 +373,7 @@ export function isCrossEngineRasterEquivalent(
   metrics: NonNullable<PdfSemanticBlockVisualLineComparison["metrics"]>,
   textReflow: boolean,
   blockKind?: string,
+  hasInlineCode = false,
 ): boolean {
   if (textReflow) {
     return false;
@@ -215,10 +400,27 @@ export function isCrossEngineRasterEquivalent(
     Math.max(metrics.baselineWidthPx, metrics.candidateWidthPx) <= 48 &&
     widthDelta <= 2 &&
     heightDelta <= 1;
+  const neutralTinyGlyphAntialiasEquivalent =
+    isNeutralTinyGlyphAntialiasEquivalent(metrics);
   const colorEquivalent =
+    neutralTinyGlyphAntialiasEquivalent ||
+    (
       metrics.backgroundColorDelta <= 4 &&
-    metrics.foregroundColorDelta <=
+      (metrics.foregroundColorDelta <=
-      (blockKind === "list-item" ? 4.5 : 4);
+        (blockKind === "list-item" ? 4.5 : 4) ||
+        isNeutralForegroundToneEquivalent(metrics) ||
+        isChromaticPaletteEquivalent(metrics) ||
+        isSyntaxPaletteEquivalent(metrics, blockKind))
+    );
+  const inlineCodeShadingEquivalent =
+    hasInlineCode &&
+    widthDelta <= Math.max(3, Math.ceil(metrics.baselineWidthPx * 0.015)) &&
+    heightDelta <= 2 &&
+    metrics.edgeIou >= 0.995 &&
+    (metrics.inkIou >= 0.9 ||
+      (metrics.foregroundInkIou ?? 0) >= 0.98) &&
+    metrics.foregroundColorDelta <= 3.25 &&
+    metrics.backgroundColorDelta <= 12;
   const topologyEquivalent =
     (preciseGeometryEquivalent &&
       metrics.inkIou >= 0.99 &&
@@ -247,7 +449,8 @@ export function isCrossEngineRasterEquivalent(
     (preciseGeometryEquivalent &&
       metrics.edgeIou >= 0.97 &&
       metrics.inkIou >= 0.97);
-  return colorEquivalent && topologyEquivalent;
+  return inlineCodeShadingEquivalent ||
+    (colorEquivalent && topologyEquivalent);
 }
 
 export async function comparePdfSemanticBlockVisuals(
@@ -270,14 +473,41 @@ export async function comparePdfSemanticBlockVisuals(
       "paragraph",
       "list-item",
       "block-quote",
+      "code-block",
     ].includes(paragraph.expectation.blockKind ?? "paragraph");
     const comparableLineCount = Math.min(
       paragraph.baseline.visualLines.length,
       paragraph.candidate.visualLines.length,
     );
+    const hasUsableVisualMapping = (layout: typeof paragraph.baseline) => {
+      if (layout.matched) {
+        return true;
+      }
+      const expected = layout.expectedCharacterCount ?? 0;
+      const matched = layout.matchedCharacterCount ?? 0;
+      const coverage = expected === 0 ? 0 : matched / expected;
+      const mathCharacterIndexes = new Set(
+        paragraph.expectation.mathCharacterIndexes ?? [],
+      );
+      const missingTailIsMath =
+        matched < expected &&
+        Array.from(
+          { length: expected - matched },
+          (_, offset) => matched + offset,
+        ).every((index) => mathCharacterIndexes.has(index));
+      return layout.visualLines.length > 0 &&
+        ((expected >= 10 && expected - matched === 1 && coverage >= 0.9) ||
+          (missingTailIsMath && coverage >= 0.8));
+    };
+    const equalLineSegmentation =
+      paragraph.baseline.visualLines.length ===
+        paragraph.candidate.visualLines.length &&
+      paragraph.baseline.lineTexts.every(
+        (text, index) => text === paragraph.candidate.lineTexts[index],
+      );
     if (
-      !paragraph.baseline.matched ||
+      !hasUsableVisualMapping(paragraph.baseline) ||
-      !paragraph.candidate.matched ||
+      !hasUsableVisualMapping(paragraph.candidate) ||
       comparableLineCount === 0
     ) {
       issues.push({
@@ -286,10 +516,7 @@ export async function comparePdfSemanticBlockVisuals(
         message: `第 ${paragraph.expectation.index + 1} 个语义块缺少可比较的局部视觉观测`,
         details: { paragraphIndex: paragraph.expectation.index },
       });
-    } else if (
+    } else if (equalLineSegmentation) {
-      paragraph.baseline.visualLines.length ===
-      paragraph.candidate.visualLines.length
-    ) {
       for (let lineIndex = 0; lineIndex < comparableLineCount; lineIndex += 1) {
         const baselineLine = paragraph.baseline.visualLines[lineIndex];
         const candidateLine = paragraph.candidate.visualLines[lineIndex];
@@ -327,11 +554,28 @@ export async function comparePdfSemanticBlockVisuals(
           cropRaster(baselinePage.raster, baselineLine.bounds),
           cropRaster(candidatePage.raster, candidateLine.bounds),
         ]);
+        assertPngRaster(
+          baselineCrop,
+          `第 ${paragraph.expectation.index + 1} 个语义块第 ${lineIndex + 1} 行基线裁剪`,
+        );
+        assertPngRaster(
+          candidateCrop,
+          `第 ${paragraph.expectation.index + 1} 个语义块第 ${lineIndex + 1} 行候选裁剪`,
+        );
         const baselineLineText = paragraph.baseline.lineTexts[lineIndex] ?? "";
         const candidateLineText = paragraph.candidate.lineTexts[lineIndex] ?? "";
         const textReflow = flowTextBlock &&
           baselineLineText !== candidateLineText;
-        const result = flowTextBlock && !textReflow
+        const cropWidthRatio = Math.min(
+          baselineCrop.widthPx,
+          candidateCrop.widthPx,
+        ) / Math.max(baselineCrop.widthPx, candidateCrop.widthPx);
+        const normalizeFlowGeometry = flowTextBlock &&
+          !textReflow &&
+          cropWidthRatio >= 0.75;
+        let result;
+        try {
+          result = normalizeFlowGeometry
             ? await comparePageRasters(baselineCrop, candidateCrop, {
                 pixelDifferenceThreshold: thresholds.pixelDifferenceThreshold,
                 spatialTolerancePx: thresholds.spatialTolerancePx,
@@ -358,10 +602,19 @@ export async function comparePdfSemanticBlockVisuals(
                   thresholds.pixelDifferenceThreshold,
                 );
               })();
+        } catch (error) {
+          throw new Error(
+            `第 ${paragraph.expectation.index + 1} 个语义块第 ${lineIndex + 1} 行栅格比较失败：` +
+              `基线 ${baselineCrop.widthPx}x${baselineCrop.heightPx}/${baselineCrop.png.byteLength}，` +
+              `候选 ${candidateCrop.widthPx}x${candidateCrop.heightPx}/${candidateCrop.png.byteLength}；` +
+              `${error instanceof Error ? error.message : String(error)}`,
+          );
+        }
         const lineIssues = localRasterIssues(
           paragraph.expectation.index,
           lineIndex,
           paragraph.expectation.blockKind,
+          paragraph.expectation.hasInlineCode === true,
           result.metrics,
           thresholds,
           textReflow,
@@ -376,6 +629,139 @@ export async function comparePdfSemanticBlockVisuals(
           issues: lineIssues,
         });
       }
+    } else {
+      const reflowPairs = pairReflowedTextLines(
+        paragraph.baseline.lineTexts,
+        paragraph.candidate.lineTexts,
+      );
+      const sliceByCharacterOverlap =
+        paragraph.baseline.visualLines.length ===
+        paragraph.candidate.visualLines.length;
+      if (reflowPairs.length === 0) {
+        issues.push({
+          code: "SEMANTIC_BLOCK_VISUAL_UNAVAILABLE",
+          severity: "failure",
+          message: `第 ${paragraph.expectation.index + 1} 个语义块无法按文本区间配对跨引擎视觉行`,
+          details: { paragraphIndex: paragraph.expectation.index },
+        });
+      }
+      for (let pairIndex = 0; pairIndex < reflowPairs.length; pairIndex += 1) {
+        const pair = reflowPairs[pairIndex]!;
+        const baselineLine =
+          paragraph.baseline.visualLines[pair.baselineLineIndex];
+        const candidateLine =
+          paragraph.candidate.visualLines[pair.candidateLineIndex];
+        const baselinePage = baselineLine
+          ? baseline.pages[baselineLine.pageNumber - 1]
+          : undefined;
+        const candidatePage = candidateLine
+          ? candidate.pages[candidateLine.pageNumber - 1]
+          : undefined;
+        if (
+          !baselineLine ||
+          !candidateLine ||
+          !baselinePage?.raster ||
+          !candidatePage?.raster
+        ) {
+          const unavailable: VisualDiffIssue = {
+            code: "SEMANTIC_BLOCK_VISUAL_UNAVAILABLE",
+            severity: "failure",
+            message: `第 ${paragraph.expectation.index + 1} 个语义块的跨引擎重排行缺少栅格`,
+            details: {
+              paragraphIndex: paragraph.expectation.index,
+              lineIndex: pairIndex,
+            },
+          };
+          issues.push(unavailable);
+          lines.push({
+            lineIndex: pairIndex,
+            baselinePageNumber: baselineLine?.pageNumber ?? -1,
+            candidatePageNumber: candidateLine?.pageNumber ?? -1,
+            issues: [unavailable],
+          });
+          continue;
+        }
+        const baselineBounds = sliceByCharacterOverlap
+          ? sliceTextLineBounds(
+              baselineLine.bounds,
+              paragraph.baseline.lineTexts[pair.baselineLineIndex] ?? "",
+              pair.baselineStart,
+              pair.baselineEnd,
+            )
+          : baselineLine.bounds;
+        const candidateBounds = sliceByCharacterOverlap
+          ? sliceTextLineBounds(
+              candidateLine.bounds,
+              paragraph.candidate.lineTexts[pair.candidateLineIndex] ?? "",
+              pair.candidateStart,
+              pair.candidateEnd,
+            )
+          : candidateLine.bounds;
+        const [baselineCrop, candidateCrop] = await Promise.all([
+          cropRaster(baselinePage.raster, baselineBounds),
+          cropRaster(candidatePage.raster, candidateBounds),
+        ]);
+        assertPngRaster(
+          baselineCrop,
+          `第 ${paragraph.expectation.index + 1} 个语义块重排行基线裁剪`,
+        );
+        assertPngRaster(
+          candidateCrop,
+          `第 ${paragraph.expectation.index + 1} 个语义块重排行候选裁剪`,
+        );
+        const widthPx = Math.max(
+          baselineCrop.widthPx,
+          candidateCrop.widthPx,
+        );
+        const heightPx = Math.max(
+          baselineCrop.heightPx,
+          candidateCrop.heightPx,
+        );
+        const [baselinePadded, candidatePadded] = await Promise.all([
+          padRaster(baselineCrop, widthPx, heightPx),
+          padRaster(candidateCrop, widthPx, heightPx),
+        ]);
+        const result = await comparePageRasters(
+          baselinePadded,
+          candidatePadded,
+          thresholds.pixelDifferenceThreshold,
+        );
+        const reflowIssue: VisualDiffIssue = {
+          code: "PARAGRAPH_LINE_BREAK_MISMATCH",
+          severity: "warning",
+          message: `第 ${paragraph.expectation.index + 1} 个语义块发生允许的跨引擎流式换行`,
+          details: {
+            paragraphIndex: paragraph.expectation.index,
+            lineIndex: pairIndex,
+            baselineLineIndex: pair.baselineLineIndex,
+            candidateLineIndex: pair.candidateLineIndex,
+            textReflow: true,
+          },
+        };
+        const lineIssues = [
+          reflowIssue,
+          ...localRasterIssues(
+            paragraph.expectation.index,
+            pairIndex,
+            paragraph.expectation.blockKind,
+            paragraph.expectation.hasInlineCode === true,
+            result.metrics,
+            thresholds,
+            true,
+          ),
+        ];
+        issues.push(...lineIssues);
+        lines.push({
+          lineIndex: pairIndex,
+          baselinePageNumber: baselineLine.pageNumber,
+          candidatePageNumber: candidateLine.pageNumber,
+          metrics: result.metrics,
+          ...(lineIssues.some((issue) => issue.severity === "failure")
+            ? { artifacts: result.artifacts }
+            : {}),
+          issues: lineIssues,
+        });
+      }
     }
     comparisons.push({
       expectation: paragraph.expectation,

packages/document-visual-diff/src/text.ts

@@ -7,7 +7,7 @@ import type {
   VisualPageSemanticExpectation,
 } from "./types.js";
 
-const ZERO_WIDTH_AND_CONTROL = /[\u0000-\u0008\u000b\u000c\u000e-\u001f\u007f\u200b-\u200d\u2060\ufeff]/gu;
+const ZERO_WIDTH_AND_CONTROL = /[\u0000-\u0008\u000b\u000c\u000e-\u001f\u007f\u200b-\u200d\u2060\ufe0e\ufe0f\ufeff]/gu;
 const CJK_RADICAL_VARIANTS = new Map([
   ["⺠", "民"],
   ["⻅", "见"],
@@ -16,6 +16,10 @@ const CJK_RADICAL_VARIANTS = new Map([
   ["⻔", "门"],
   ["⻚", "页"],
   ["⻛", "风"],
+  ["⻝", "食"],
+  ["⻣", "骨"],
+  ["⻋", "车"],
+  ["⻬", "齐"],
 ]);
 const PAGE_NUMBER_PATTERNS = [
   /^(?:[-—–]\s*)?\d+(?:\s*[/／]\s*\d+)?(?:\s*[-—–])?$/u,
@@ -30,6 +34,16 @@ export function normalizePdfText(value: string): string {
     .replace(/[\u2e80-\u2eff]/gu, (character) =>
       CJK_RADICAL_VARIANTS.get(character) ?? character
     )
+    // Chromium、Word 与 WPS 的 PDF 文本层会以不同方式保留弯引号；
+    // 统一为 ASCII 只用于可编辑文本定位，不改变栅格视觉比较。
+    .replace(/[“”]/gu, '"')
+    .replace(/[‘’]/gu, "'")
+    // Chromium 的部分 CJK 字体把 Markdown em dash 提取为 horizontal bar；
+    // 二者在此仅作为语义定位字符归一，不影响后续栅格视觉比较。
+    .replace(/―/gu, "—")
+    // Chromium 的部分 CJK 字体会把数字区间中的 en dash 提取为两个
+    // ASCII hyphen。只归一数字/百分比区间，避免改写代码中的普通 `--`。
+    .replace(/(?<=[\d%])\s*--(?:\s*(?=\d)|\s*$)/gu, "–")
     .replace(ZERO_WIDTH_AND_CONTROL, "")
     .replace(/\u00a0/gu, " ")
     .replace(/[ \t]+/gu, " ")
@@ -133,6 +147,63 @@ function median(values: readonly number[]): number {
   return ((sorted[middle - 1] ?? value) + value) / 2;
 }
 
+function mergeInlineScriptGroups(
+  sourceGroups: readonly PdfTextItemSnapshot[][],
+  typicalHeight: number,
+): PdfTextItemSnapshot[][] {
+  const groups = sourceGroups.map((group) => [...group]);
+  for (let index = 0; index < groups.length; index += 1) {
+    const group = groups[index];
+    if (!group || group.length === 0) {
+      continue;
+    }
+    const groupHeight = median(group.map((item) => item.bounds.height));
+    if (groupHeight > typicalHeight * 0.9) {
+      continue;
+    }
+    const groupBounds = unionBounds(group);
+    for (const adjacentIndex of [index - 1, index + 1]) {
+      const adjacent = groups[adjacentIndex];
+      if (!adjacent || adjacent.length === 0) {
+        continue;
+      }
+      const adjacentHeight = median(
+        adjacent.map((item) => item.bounds.height),
+      );
+      if (adjacentHeight < groupHeight / 0.9) {
+        continue;
+      }
+      const adjacentBounds = unionBounds(adjacent);
+      const verticalOverlap = Math.min(
+        groupBounds.y + groupBounds.height,
+        adjacentBounds.y + adjacentBounds.height,
+      ) - Math.max(groupBounds.y, adjacentBounds.y);
+      const horizontallyContained =
+        groupBounds.x >= adjacentBounds.x - typicalHeight * 0.5 &&
+        groupBounds.x + groupBounds.width <=
+          adjacentBounds.x + adjacentBounds.width + typicalHeight * 0.5;
+      // Chromium 会把行末公式按较小字号单独提取；公式片段可能从正文
+      // 最后一个字符的右缘开始，并向右超出正文组，因此不能只用“被正文
+      // 水平包含”判断同行。仅在两组真实垂直重叠且小字号组紧贴较大字号
+      // 组右缘时合并，避免把下一视觉行或相邻表格单元格误并入当前行。
+      const adjacentRight = adjacentBounds.x + adjacentBounds.width;
+      const touchesAdjacentRight =
+        groupBounds.x >= adjacentBounds.x &&
+        Math.abs(groupBounds.x - adjacentRight) <= typicalHeight * 0.5;
+      if (
+        verticalOverlap <= 0 ||
+        (!horizontallyContained && !touchesAdjacentRight)
+      ) {
+        continue;
+      }
+      adjacent.push(...group);
+      groups[index] = [];
+      break;
+    }
+  }
+  return groups.filter((group) => group.length > 0);
+}
+
 function joinLineItems(items: readonly PdfTextItemSnapshot[]): string {
   let result = "";
   let previous: PdfTextItemSnapshot | undefined;
@@ -181,6 +252,9 @@ export function aggregatePdfTextLines(
   sourceItems: readonly PdfTextItemSnapshot[],
   pageHeightPt: number,
 ): PdfTextLineSnapshot[] {
+  const sourceOrder = new Map(
+    sourceItems.map((item, index) => [item, index] as const),
+  );
   const items = sourceItems
     .filter((item) => item.normalizedText.length > 0)
     .sort(
@@ -188,11 +262,14 @@ export function aggregatePdfTextLines(
         left.bounds.y - right.bounds.y || left.bounds.x - right.bounds.x,
     );
   const typicalHeight = median(items.map((item) => item.bounds.height));
-  const baselineTolerance = Math.max(1.5, typicalHeight * 0.15);
+  // Office/WPS 会让同一视觉行内不同字体或单元格的基线产生约 2pt
-  const groups: PdfTextItemSnapshot[][] = [];
+  // 浮动；按 15% 聚合会把这些片段错误拆行并改变阅读顺序。
+  // 20% 仍远小于正常行距，同时能覆盖常见的 10–12pt 字号偏差。
+  const baselineTolerance = Math.max(2, typicalHeight * 0.2);
+  const initialGroups: PdfTextItemSnapshot[][] = [];
 
   for (const item of items) {
-    const lastGroup = groups.at(-1);
+    const lastGroup = initialGroups.at(-1);
     const lastBaseline = lastGroup
       ? median(lastGroup.map((entry) => entry.baselineY))
       : undefined;
@@ -203,13 +280,27 @@ export function aggregatePdfTextLines(
     ) {
       lastGroup.push(item);
     } else {
-      groups.push([item]);
+      initialGroups.push([item]);
     }
   }
 
+  const groups = mergeInlineScriptGroups(initialGroups, typicalHeight);
+
   return groups.map((group) => {
     const sortedItems = [...group].sort(
-      (left, right) => left.bounds.x - right.bounds.x,
+      (left, right) => {
+        const leftRight = left.bounds.x + left.bounds.width;
+        const rightRight = right.bounds.x + right.bounds.width;
+        const horizontallyOverlapping =
+          left.bounds.x < rightRight && right.bounds.x < leftRight;
+        // PDF.js 的文本流保留字符阅读顺序，但不同字体的窄标点可能与
+        // 相邻全角字符发生水平重叠。此时单纯按 x 排序会把闭引号移到
+        // 左括号之后，破坏语义块定位；只对真实重叠片段保留文本流顺序，
+        // 其余片段仍按几何位置排序，兼容 Office 表格等非阅读序文本流。
+        return horizontallyOverlapping
+          ? (sourceOrder.get(left) ?? 0) - (sourceOrder.get(right) ?? 0)
+          : left.bounds.x - right.bounds.x;
+      },
     );
     const bounds = unionBounds(sortedItems);
     const normalizedText = joinLineItems(sortedItems);

packages/document-visual-diff/src/types.ts

@@ -71,6 +71,12 @@ export type EditableParagraphBlockKind =
 export interface EditableParagraphExpectation {
   index: number;
   text: string;
+  mathCharacterIndexes?: number[];
+  hardBreakSegments?: string[];
+  hasInlineCode?: boolean;
+  tableGroupId?: string;
+  tableRowIndex?: number;
+  tableColumnIndex?: number;
   styleId?: string;
   role: EditableParagraphRole;
   blockKind?: EditableParagraphBlockKind;
@@ -144,6 +150,7 @@ export interface PdfPageRaster {
   dpi: number;
   sha256: string;
   png: Uint8Array;
+  rgba?: Uint8ClampedArray;
 }
 
 export interface PdfPageSnapshot {
@@ -287,9 +294,16 @@ export interface PdfRasterDiffMetrics {
   meanAbsoluteError: number;
   changedPixelRatio: number;
   inkIou: number;
+  foregroundInkIou: number;
   edgeIou: number;
   backgroundColorDelta: number;
   foregroundColorDelta: number;
+  baselineForegroundChroma?: number;
+  candidateForegroundChroma?: number;
+  foregroundLuminanceDelta?: number;
+  dominantForegroundColorDelta?: number;
+  chromaticForegroundColorDelta?: number;
+  chromaticPaletteDelta?: number;
 }
 
 export interface ComparePageRasterOptions {

packages/document-visual-diff/tests/adapters.test.ts

@@ -121,6 +121,37 @@ describe("Office PDF 适配器", () => {
       await rm(directory, { recursive: true, force: true });
     }
   });
+
+  it("Office 偶发未生成 PDF 时清理残留并有界重试", async () => {
+    const directory = await mkdtemp(
+      join(tmpdir(), "visual-diff-office-retry-test-"),
+    );
+    const docxPath = join(directory, "fixture.docx");
+    await writeFile(docxPath, "fixture");
+    let attempts = 0;
+    const backend: OfficeAutomationBackend = {
+      probe: async () => true,
+      exportPdf: async (options) => {
+        attempts += 1;
+        if (attempts === 1) {
+          await writeFile(options.outputPath, "partial");
+          throw new Error("Office 未生成有效 PDF");
+        }
+        const pdf = createMinimalPdf("retry");
+        await writeFile(options.outputPath, pdf);
+        return { bytes: pdf.byteLength, pageCount: 1 };
+      },
+    };
+    try {
+      const result = await createWordPdfAdapter({ backend }).generate({
+        docxPath,
+      });
+      expect(result.pageCount).toBe(1);
+      expect(attempts).toBe(2);
+    } finally {
+      await rm(directory, { recursive: true, force: true });
+    }
+  });
 });
 
 describe("PDF 适配器视觉编排", () => {

packages/document-visual-diff/tests/raster.test.ts

@@ -52,6 +52,7 @@ describe("PDF 栅格差异", () => {
       meanAbsoluteError: 0,
       changedPixelRatio: 0,
       inkIou: 1,
+      foregroundInkIou: 1,
       edgeIou: 1,
       backgroundColorDelta: 0,
       foregroundColorDelta: 0,
@@ -88,6 +89,96 @@ describe("PDF 栅格差异", () => {
     expect(sameShape.metrics.foregroundColorDelta).toBeLessThan(1);
   });
 
+  it("双背景占比翻转时仍独立量化核心字形拓扑", async () => {
+    const inlineCodeRaster = (fullHeightShading: boolean) => {
+      const canvas = createCanvas(216, 25);
+      const context = canvas.getContext("2d");
+      context.fillStyle = "#ffffff";
+      context.fillRect(0, 0, 216, 25);
+      context.fillStyle = "#f3f3f3";
+      context.fillRect(28, fullHeightShading ? 0 : 5, 120, fullHeightShading ? 25 : 16);
+      context.fillStyle = "#333333";
+      context.fillRect(34, 10, 80, 4);
+      const png = Uint8Array.from(canvas.toBuffer("image/png"));
+      return {
+        widthPx: 216,
+        heightPx: 25,
+        dpi: 144,
+        sha256: createHash("sha256").update(png).digest("hex"),
+        png,
+      } satisfies PdfPageRaster;
+    };
+    const result = await comparePageRasters(
+      inlineCodeRaster(false),
+      inlineCodeRaster(true),
+    );
+
+    expect(result.metrics.backgroundColorDelta).toBeGreaterThan(10);
+    expect(result.metrics.foregroundInkIou).toBe(1);
+  });
+
+  it("渐变背景不会把深色文字误判为背景", async () => {
+    const gradientRaster = (gradient: boolean) => {
+      const canvas = createCanvas(160, 40);
+      const context = canvas.getContext("2d");
+      if (gradient) {
+        const fill = context.createLinearGradient(0, 0, 160, 0);
+        fill.addColorStop(0, "#edf5fa");
+        fill.addColorStop(1, "#ffffff");
+        context.fillStyle = fill;
+      } else {
+        context.fillStyle = "#f6fafc";
+      }
+      context.fillRect(0, 0, 160, 40);
+      context.fillStyle = "#0b3557";
+      context.fillRect(4, 4, 148, 32);
+      const png = Uint8Array.from(canvas.toBuffer("image/png"));
+      return {
+        widthPx: 160,
+        heightPx: 40,
+        dpi: 144,
+        sha256: createHash("sha256").update(png).digest("hex"),
+        png,
+      } satisfies PdfPageRaster;
+    };
+
+    const result = await comparePageRasters(
+      gradientRaster(true),
+      gradientRaster(false),
+    );
+    expect(result.metrics.foregroundColorDelta).toBeLessThan(1);
+    expect(result.metrics.backgroundColorDelta).toBeLessThan(8);
+    expect(result.metrics.inkIou).toBeGreaterThan(0.95);
+  });
+
+  it("裁剪跨过深色盒边界时不会把外部白色误判为盒背景", async () => {
+    const boxedRaster = (exposeOutside: boolean) => {
+      const canvas = createCanvas(80, 40);
+      const context = canvas.getContext("2d");
+      context.fillStyle = "#ffffff";
+      context.fillRect(0, 0, 80, 40);
+      context.fillStyle = "#464646";
+      context.fillRect(0, 0, exposeOutside ? 68 : 80, exposeOutside ? 32 : 40);
+      context.fillStyle = "#ffffff";
+      context.fillRect(18, 10, 32, 16);
+      const png = Uint8Array.from(canvas.toBuffer("image/png"));
+      return {
+        widthPx: 80,
+        heightPx: 40,
+        dpi: 144,
+        sha256: createHash("sha256").update(png).digest("hex"),
+        png,
+      } satisfies PdfPageRaster;
+    };
+
+    const result = await comparePageRasters(
+      boxedRaster(true),
+      boxedRaster(false),
+    );
+    expect(result.metrics.backgroundColorDelta).toBeLessThan(1);
+    expect(result.metrics.foregroundColorDelta).toBeLessThan(1);
+  });
+
   it("量化位移并生成稳定叠加图与热力图", async () => {
     const result = await comparePageRasters(raster(10), raster(20), {
       spatialTolerancePx: 0,

packages/document-visual-diff/tests/report.test.ts

@@ -6,6 +6,7 @@ import { describe, expect, it } from "vitest";
 import {
   createPdfVisualDiffReport,
   getBlockingVisualDiffIssues,
+  isCrossEngineRasterEquivalent,
   renderPdfVisualDiffHtml,
   serializePdfVisualDiffJson,
   type PdfDocumentSnapshot,
@@ -84,6 +85,27 @@ function antialiasBlockRaster(edgeColor: string): PdfPageRaster {
   };
 }
 
+function adjacentLineRaster(includePreviousLine: boolean): PdfPageRaster {
+  const canvas = createCanvas(40, 50);
+  const context = canvas.getContext("2d");
+  context.fillStyle = "#ffffff";
+  context.fillRect(0, 0, 40, 50);
+  if (includePreviousLine) {
+    context.fillStyle = "#111111";
+    context.fillRect(8, 12, 20, 3);
+  }
+  context.fillStyle = "#111111";
+  context.fillRect(8, 20, 20, 4);
+  const png = Uint8Array.from(canvas.toBuffer("image/png"));
+  return {
+    widthPx: 40,
+    heightPx: 50,
+    dpi: 144,
+    sha256: createHash("sha256").update(png).digest("hex"),
+    png,
+  };
+}
+
 function snapshot(label: string, pageRaster: PdfPageRaster): PdfDocumentSnapshot {
   return {
     schemaVersion: 1,
@@ -561,6 +583,58 @@ describe("PDF 视觉差异报告", () => {
     });
   });
 
+  it("正文页数不同时分别验证未配对页面自身的页码语义", async () => {
+    const pageNumber = (text: string): PdfTextLineSnapshot => ({
+      text,
+      normalizedText: text,
+      bounds: { x: 9, y: 22, width: 2, height: 1 },
+      baselineY: 23,
+      role: "page-number",
+      items: [],
+    });
+    const baseline = flowSnapshot(
+      "baseline",
+      [
+        [contentLine("甲", 8), pageNumber("1 / 2")],
+        [contentLine("乙丙", 8), pageNumber("2 / 2")],
+      ],
+      "#111111",
+    );
+    const candidate = flowSnapshot(
+      "candidate",
+      [[
+        contentLine("甲", 8),
+        contentLine("乙丙", 12),
+        pageNumber("1 / 1"),
+      ]],
+      "#cc0000",
+    );
+    const pageSemantics = (count: number) => bodySemantics(count).map(
+      (item, index) => ({
+        ...item,
+        footerVisible: true,
+        footerAlignment: "center" as const,
+        pageNumberText: `${index + 1} / ${count}`,
+      }),
+    );
+    const report = await createPdfVisualDiffReport(baseline, candidate, {
+      expectedEditableText: "甲乙丙",
+      expectedEditableParagraphs: flowParagraphs,
+      baselinePageSemantics: pageSemantics(2),
+      candidatePageSemantics: pageSemantics(1),
+    });
+
+    expect(report.status).toBe("warning");
+    expect(report.basic.bodyFlow).toMatchObject({
+      legal: true,
+      baselineBodyPageCount: 2,
+      candidateBodyPageCount: 1,
+    });
+    expect(report.issues.map((issue) => issue.code)).toContain(
+      "BODY_FLOW_PAGE_COUNT_DIFFERENCE",
+    );
+  });
+
   it("新增正文页缺少预期页码时不得认定为合法分页流动", async () => {
     const pageNumber = (text: string): PdfTextLineSnapshot => ({
       text,
@@ -664,6 +738,71 @@ describe("PDF 视觉差异报告", () => {
       .toMatchObject({ inkIou: 1, edgeIou: 1 });
   });
 
+  it("局部视觉裁剪不得把目标行上方的相邻行墨迹纳入比较", async () => {
+    const targetLine = contentLine("乙", 10);
+    targetLine.bounds = { x: 4, y: 10, width: 10, height: 2 };
+    const baseline = flowSnapshot(
+      "baseline",
+      [[targetLine]],
+      "#111111",
+    );
+    const candidate = flowSnapshot(
+      "candidate",
+      [[targetLine]],
+      "#111111",
+    );
+    baseline.pages[0]!.raster = adjacentLineRaster(true);
+    candidate.pages[0]!.raster = adjacentLineRaster(false);
+    const report = await createPdfVisualDiffReport(baseline, candidate, {
+      expectedEditableText: "乙",
+      expectedEditableParagraphs: [{
+        index: 0,
+        text: "乙",
+        role: "body",
+        section: "body",
+        blockKind: "table-cell",
+      }],
+      pageSemantics: bodySemantics(1),
+    });
+
+    expect(report.issues.map((issue) => issue.code)).not.toContain(
+      "SEMANTIC_BLOCK_RASTER_MISMATCH",
+    );
+  });
+
+  it("仅对带行内代码结构语义的混排允许 WPS 底纹栅格混合差异", () => {
+    const metrics = {
+      baselineWidthPx: 236,
+      baselineHeightPx: 24,
+      candidateWidthPx: 233,
+      candidateHeightPx: 26,
+      comparedWidthPx: 236,
+      comparedHeightPx: 26,
+      dimensionsMatch: false,
+      geometryNormalized: true,
+      spatialTolerancePx: 4,
+      meanAbsoluteError: 22.1,
+      changedPixelRatio: 0.449,
+      inkIou: 0.908,
+      edgeIou: 1,
+      backgroundColorDelta: 11.38,
+      foregroundColorDelta: 2.6,
+    };
+
+    expect(isCrossEngineRasterEquivalent(
+      metrics,
+      false,
+      "paragraph",
+      true,
+    )).toBe(true);
+    expect(isCrossEngineRasterEquivalent(
+      metrics,
+      false,
+      "paragraph",
+      false,
+    )).toBe(false);
+  });
+
   it("缺少语义块证据时不得把正文整页栅格自动降级", async () => {
     const report = await createPdfVisualDiffReport(
       snapshot("baseline", raster("#111111")),

packages/document-visual-diff/tests/semantic-block.test.ts

@@ -1,6 +1,8 @@
 import { describe, expect, it } from "vitest";
+import { createCanvas } from "@napi-rs/canvas";
 
 import {
+  comparePdfSemanticBlockVisuals,
   isCrossEngineRasterEquivalent,
   type PdfRasterMetrics,
 } from "../src/index.js";
@@ -28,7 +30,289 @@ function metrics(
   };
 }
 
+function whitePageRaster() {
+  const canvas = createCanvas(400, 400);
+  const context = canvas.getContext("2d");
+  context.fillStyle = "#fff";
+  context.fillRect(0, 0, 400, 400);
+  context.fillStyle = "#111";
+  context.fillRect(40, 40, 120, 12);
+  context.fillRect(40, 80, 120, 12);
+  const rgba = context.getImageData(0, 0, 400, 400).data;
+  return {
+    widthPx: 400,
+    heightPx: 400,
+    dpi: 72,
+    sha256: "test",
+    png: Uint8Array.from(canvas.toBuffer("image/png")),
+    rgba,
+  };
+}
+
+function blankPageRaster() {
+  const canvas = createCanvas(400, 400);
+  const context = canvas.getContext("2d");
+  context.fillStyle = "#fff";
+  context.fillRect(0, 0, 400, 400);
+  const rgba = context.getImageData(0, 0, 400, 400).data;
+  return {
+    widthPx: 400,
+    heightPx: 400,
+    dpi: 72,
+    sha256: "blank",
+    png: Uint8Array.from(canvas.toBuffer("image/png")),
+    rgba,
+  };
+}
+
 describe("跨引擎字形栅格等价", () => {
+  it("缺少局部视觉行时必须阻断而不能静默通过", async () => {
+    const [comparison] = await comparePdfSemanticBlockVisuals(
+      { pages: [] } as never,
+      { pages: [] } as never,
+      [{
+        expectation: { index: 39, section: "body", blockKind: "table-cell" },
+        baseline: { matched: true, visualLines: [], lineTexts: [] },
+        candidate: { matched: true, visualLines: [], lineTexts: [] },
+      }] as never,
+      {} as never,
+    );
+
+    expect(comparison?.status).toBe("failed");
+    expect(comparison?.issues).toEqual([
+      expect.objectContaining({
+        code: "SEMANTIC_BLOCK_VISUAL_UNAVAILABLE",
+        severity: "failure",
+      }),
+    ]);
+  });
+
+  it("跨引擎换行数不同时按字符区间配对视觉行而不是返回空观测", async () => {
+    const page = {
+      pageNumber: 1,
+      widthPt: 400,
+      heightPt: 400,
+      lines: [],
+      raster: whitePageRaster(),
+    };
+    const visualLine = (y: number) => ({
+      pageNumber: 1,
+      bounds: { x: 40, y, width: 120, height: 12 },
+      fontFamilies: ["Test Sans"],
+    });
+    const [comparison] = await comparePdfSemanticBlockVisuals(
+      { pages: [page] } as never,
+      { pages: [page] } as never,
+      [{
+        expectation: {
+          index: 7,
+          section: "body",
+          blockKind: "table-cell",
+        },
+        baseline: {
+          matched: true,
+          visualLines: [visualLine(40), visualLine(80)],
+          lineTexts: ["甲乙", "丙丁"],
+        },
+        candidate: {
+          matched: true,
+          visualLines: [visualLine(40)],
+          lineTexts: ["甲乙丙丁"],
+        },
+      }] as never,
+      {
+        pixelDifferenceThreshold: 8,
+        spatialTolerancePx: 4,
+        minInkIou: 0.75,
+        minEdgeIou: 0.75,
+        maxBackgroundColorDelta: 16,
+        maxForegroundColorDelta: 16,
+        maxMeanAbsoluteError: 20,
+        maxChangedPixelRatio: 0.35,
+        minAntialiasEquivalentInkIou: 0.9,
+        minAntialiasEquivalentEdgeIou: 0.9,
+      } as never,
+    );
+
+    expect(comparison?.status).toBe("warning");
+    expect(comparison?.lines).toHaveLength(2);
+    expect(comparison?.issues).not.toContainEqual(
+      expect.objectContaining({ code: "SEMANTIC_BLOCK_VISUAL_UNAVAILABLE" }),
+    );
+    expect(comparison?.lines.every((line) =>
+      line.issues.some((issue) => issue.details?.textReflow === true)
+    )).toBe(true);
+  });
+
+  it("PDF 文本提取仅遗漏长文本末尾单字时仍比较已有视觉行", async () => {
+    const page = {
+      pageNumber: 1,
+      widthPt: 400,
+      heightPt: 400,
+      lines: [],
+      raster: blankPageRaster(),
+    };
+    const visualLine = (y: number) => ({
+      pageNumber: 1,
+      bounds: { x: 40, y, width: 108, height: 12 },
+      fontFamilies: ["Test Sans"],
+    });
+    const [comparison] = await comparePdfSemanticBlockVisuals(
+      { pages: [page] } as never,
+      { pages: [page] } as never,
+      [{
+        expectation: {
+          index: 118,
+          section: "body",
+          blockKind: "table-header",
+        },
+        baseline: {
+          matched: true,
+          matchedCharacterCount: 10,
+          expectedCharacterCount: 10,
+          visualLines: [visualLine(40), visualLine(70)],
+          lineTexts: ["高血压预测模型目标", "值"],
+        },
+        candidate: {
+          matched: false,
+          matchedCharacterCount: 9,
+          expectedCharacterCount: 10,
+          visualLines: [visualLine(40)],
+          lineTexts: ["高血压预测模型目标"],
+        },
+      }] as never,
+      {
+        pixelDifferenceThreshold: 8,
+        spatialTolerancePx: 4,
+        minInkIou: 0.75,
+        minEdgeIou: 0.75,
+        maxBackgroundColorDelta: 16,
+        maxForegroundColorDelta: 16,
+        maxMeanAbsoluteError: 20,
+        maxChangedPixelRatio: 0.35,
+        minAntialiasEquivalentInkIou: 0.9,
+        minAntialiasEquivalentEdgeIou: 0.9,
+      } as never,
+    );
+
+    expect(comparison?.lines).toHaveLength(1);
+    expect(comparison?.issues).not.toContainEqual(
+      expect.objectContaining({ code: "SEMANTIC_BLOCK_VISUAL_UNAVAILABLE" }),
+    );
+  });
+
+  it("表格末尾数学字符未进入 PDF 文本层时仍比较已有视觉行", async () => {
+    const page = {
+      pageNumber: 1,
+      widthPt: 400,
+      heightPt: 400,
+      lines: [],
+      raster: blankPageRaster(),
+    };
+    const visualLine = {
+      pageNumber: 1,
+      bounds: { x: 40, y: 40, width: 108, height: 12 },
+      fontFamilies: ["Test Serif"],
+    };
+    const [comparison] = await comparePdfSemanticBlockVisuals(
+      { pages: [page] } as never,
+      { pages: [page] } as never,
+      [{
+        expectation: {
+          index: 71,
+          section: "body",
+          blockKind: "table-cell",
+          mathCharacterIndexes: [3, 4, 5, 6, 11, 12, 13],
+        },
+        baseline: {
+          matched: false,
+          matchedCharacterCount: 12,
+          expectedCharacterCount: 14,
+          visualLines: [visualLine],
+          lineTexts: ["收缩压≥140或舒张压≥"],
+        },
+        candidate: {
+          matched: true,
+          matchedCharacterCount: 14,
+          expectedCharacterCount: 14,
+          visualLines: [visualLine],
+          lineTexts: ["收缩压≥140或舒张压≥90"],
+        },
+      }] as never,
+      {
+        pixelDifferenceThreshold: 8,
+        spatialTolerancePx: 4,
+        minInkIou: 0.75,
+        minEdgeIou: 0.75,
+        maxBackgroundColorDelta: 16,
+        maxForegroundColorDelta: 16,
+        maxMeanAbsoluteError: 20,
+        maxChangedPixelRatio: 0.35,
+        minAntialiasEquivalentInkIou: 0.9,
+        minAntialiasEquivalentEdgeIou: 0.9,
+      } as never,
+    );
+
+    expect(comparison?.lines).toHaveLength(1);
+    expect(comparison?.issues).not.toContainEqual(
+      expect.objectContaining({ code: "SEMANTIC_BLOCK_VISUAL_UNAVAILABLE" }),
+    );
+  });
+
+  it("视觉行数量相同但字符分段不同时按重叠字符区间比较", async () => {
+    const page = {
+      pageNumber: 1,
+      widthPt: 400,
+      heightPt: 400,
+      lines: [],
+      raster: blankPageRaster(),
+    };
+    const visualLine = (y: number, width: number) => ({
+      pageNumber: 1,
+      bounds: { x: 40, y, width, height: 12 },
+      fontFamilies: ["Test Sans"],
+    });
+    const [comparison] = await comparePdfSemanticBlockVisuals(
+      { pages: [page] } as never,
+      { pages: [page] } as never,
+      [{
+        expectation: {
+          index: 117,
+          section: "body",
+          blockKind: "table-header",
+        },
+        baseline: {
+          matched: true,
+          visualLines: [visualLine(40, 20), visualLine(70, 40), visualLine(100, 40)],
+          lineTexts: ["记", "录/", "计算"],
+        },
+        candidate: {
+          matched: true,
+          visualLines: [visualLine(40, 40), visualLine(70, 40), visualLine(100, 20)],
+          lineTexts: ["记录", "/计", "算"],
+        },
+      }] as never,
+      {
+        pixelDifferenceThreshold: 8,
+        spatialTolerancePx: 4,
+        minInkIou: 0.75,
+        minEdgeIou: 0.75,
+        maxBackgroundColorDelta: 16,
+        maxForegroundColorDelta: 16,
+        maxMeanAbsoluteError: 20,
+        maxChangedPixelRatio: 0.35,
+        minAntialiasEquivalentInkIou: 0.9,
+        minAntialiasEquivalentEdgeIou: 0.9,
+      } as never,
+    );
+
+    expect(comparison?.status).toBe("warning");
+    expect(comparison?.lines).toHaveLength(5);
+    expect(comparison?.issues).not.toContainEqual(
+      expect.objectContaining({ code: "ELEMENT_COLOR_MISMATCH" }),
+    );
+  });
+
   it("接受轮廓、颜色和几何一致的灰阶抗锯齿差异", () => {
     expect(isCrossEngineRasterEquivalent(metrics(), false)).toBe(true);
     expect(
@@ -41,6 +325,71 @@ describe("跨引擎字形栅格等价", () => {
         false,
       ),
     ).toBe(true);
+    expect(
+      isCrossEngineRasterEquivalent(
+        metrics({
+          baselineWidthPx: 56,
+          candidateWidthPx: 56,
+          baselineHeightPx: 37,
+          candidateHeightPx: 37,
+          comparedWidthPx: 56,
+          comparedHeightPx: 37,
+          inkIou: 1,
+          edgeIou: 1,
+          foregroundColorDelta: 17,
+          baselineForegroundChroma: 0,
+          candidateForegroundChroma: 0,
+          foregroundLuminanceDelta: 17,
+        }),
+        false,
+        "code-block",
+      ),
+    ).toBe(true);
+    const neutralSingleStrokeMetrics = metrics({
+      baselineWidthPx: 37,
+      candidateWidthPx: 37,
+      baselineHeightPx: 27,
+      candidateHeightPx: 27,
+      comparedWidthPx: 37,
+      comparedHeightPx: 27,
+      meanAbsoluteError: 12.78,
+      inkIou: 1,
+      foregroundInkIou: 1,
+      edgeIou: 1,
+      backgroundColorDelta: 10.54,
+      foregroundColorDelta: 108.69,
+      baselineForegroundChroma: 10,
+      candidateForegroundChroma: 5,
+      foregroundLuminanceDelta: 109.27,
+    });
+    expect(
+      isCrossEngineRasterEquivalent(
+        neutralSingleStrokeMetrics,
+        false,
+        "table-cell",
+      ),
+    ).toBe(true);
+    expect(
+      isCrossEngineRasterEquivalent(
+        { ...neutralSingleStrokeMetrics, meanAbsoluteError: 16.01 },
+        false,
+        "table-cell",
+      ),
+    ).toBe(false);
+    expect(
+      isCrossEngineRasterEquivalent(
+        { ...neutralSingleStrokeMetrics, candidateForegroundChroma: 12.01 },
+        false,
+        "table-cell",
+      ),
+    ).toBe(false);
+    expect(
+      isCrossEngineRasterEquivalent(
+        { ...neutralSingleStrokeMetrics, edgeIou: 0.994 },
+        false,
+        "table-cell",
+      ),
+    ).toBe(false);
     expect(
       isCrossEngineRasterEquivalent(
         metrics({ inkIou: 1, edgeIou: 0.946 }),
@@ -97,6 +446,26 @@ describe("跨引擎字形栅格等价", () => {
         "list-item",
       ),
     ).toBe(true);
+    expect(
+      isCrossEngineRasterEquivalent(
+        metrics({
+          baselineWidthPx: 262,
+          candidateWidthPx: 272,
+          baselineHeightPx: 30,
+          candidateHeightPx: 31,
+          comparedWidthPx: 262,
+          comparedHeightPx: 30,
+          inkIou: 0.9792,
+          edgeIou: 0.9869,
+          foregroundColorDelta: 4.63,
+          baselineForegroundChroma: 0,
+          candidateForegroundChroma: 0,
+          foregroundLuminanceDelta: 4.63,
+        }),
+        false,
+        "list-item",
+      ),
+    ).toBe(true);
     expect(
       isCrossEngineRasterEquivalent(
         metrics({
@@ -120,6 +489,147 @@ describe("跨引擎字形栅格等价", () => {
     ).toBe(true);
   });
 
+  it("仅在高彩度字形轮廓与前景主色均一致时接受调色差异", () => {
+    const paletteMetrics = metrics({
+      inkIou: 1,
+      edgeIou: 1,
+      foregroundColorDelta: 35,
+      baselineForegroundChroma: 141,
+      candidateForegroundChroma: 176,
+      dominantForegroundColorDelta: 0.5,
+    });
+    expect(
+      isCrossEngineRasterEquivalent(paletteMetrics, false, "table-cell"),
+    ).toBe(true);
+    expect(
+      isCrossEngineRasterEquivalent(
+        { ...paletteMetrics, dominantForegroundColorDelta: 4.1 },
+        false,
+        "table-cell",
+      ),
+    ).toBe(false);
+    expect(
+      isCrossEngineRasterEquivalent(
+        { ...paletteMetrics, edgeIou: 0.994 },
+        false,
+        "table-cell",
+      ),
+    ).toBe(false);
+    expect(
+      isCrossEngineRasterEquivalent(
+        { ...paletteMetrics, baselineForegroundChroma: 79 },
+        false,
+        "table-cell",
+      ),
+    ).toBe(false);
+  });
+
+  it("多色代码行仅在语法主色和对应墨迹位置均一致时通过", () => {
+    const syntaxMetrics = metrics({
+      baselineWidthPx: 407,
+      candidateWidthPx: 403,
+      baselineHeightPx: 22,
+      candidateHeightPx: 23,
+      inkIou: 0.988,
+      foregroundInkIou: 1,
+      edgeIou: 1,
+      foregroundColorDelta: 92,
+      chromaticPaletteDelta: 0.8,
+    });
+    expect(
+      isCrossEngineRasterEquivalent(syntaxMetrics, false, "code-block"),
+    ).toBe(true);
+    expect(
+      isCrossEngineRasterEquivalent(
+        { ...syntaxMetrics, chromaticPaletteDelta: 2.1 },
+        false,
+        "code-block",
+      ),
+    ).toBe(false);
+    expect(
+      isCrossEngineRasterEquivalent(
+        { ...syntaxMetrics, foregroundInkIou: 0.994 },
+        false,
+        "code-block",
+      ),
+    ).toBe(false);
+    expect(
+      isCrossEngineRasterEquivalent(syntaxMetrics, false, "paragraph"),
+    ).toBe(false);
+  });
+
+  it("仅在行内代码轮廓和底纹严格一致时接受微小前景色差", () => {
+    const inlineCodeMetrics = metrics({
+      baselineWidthPx: 236,
+      candidateWidthPx: 234,
+      baselineHeightPx: 24,
+      candidateHeightPx: 25,
+      comparedWidthPx: 236,
+      comparedHeightPx: 25,
+      inkIou: 0.91759,
+      edgeIou: 1,
+      backgroundColorDelta: 11.317,
+      foregroundColorDelta: 3.158,
+    });
+
+    expect(
+      isCrossEngineRasterEquivalent(
+        inlineCodeMetrics,
+        false,
+        "paragraph",
+        true,
+      ),
+    ).toBe(true);
+    expect(
+      isCrossEngineRasterEquivalent(
+        { ...inlineCodeMetrics, foregroundColorDelta: 3.26 },
+        false,
+        "paragraph",
+        true,
+      ),
+    ).toBe(false);
+    expect(
+      isCrossEngineRasterEquivalent(
+        { ...inlineCodeMetrics, backgroundColorDelta: 12.01 },
+        false,
+        "paragraph",
+        true,
+      ),
+    ).toBe(false);
+    expect(
+      isCrossEngineRasterEquivalent(
+        { ...inlineCodeMetrics, edgeIou: 0.994 },
+        false,
+        "paragraph",
+        true,
+      ),
+    ).toBe(false);
+    expect(
+      isCrossEngineRasterEquivalent(
+        {
+          ...inlineCodeMetrics,
+          inkIou: 0.88,
+          foregroundInkIou: 0.99,
+        },
+        false,
+        "table-cell",
+        true,
+      ),
+    ).toBe(true);
+    expect(
+      isCrossEngineRasterEquivalent(
+        {
+          ...inlineCodeMetrics,
+          inkIou: 0.88,
+          foregroundInkIou: 0.97,
+        },
+        false,
+        "table-cell",
+        true,
+      ),
+    ).toBe(false);
+  });
+
   it("拒绝回流、颜色、几何或轮廓变化", () => {
     expect(isCrossEngineRasterEquivalent(metrics(), true)).toBe(false);
     expect(
@@ -169,6 +679,42 @@ describe("跨引擎字形栅格等价", () => {
     expect(
       isCrossEngineRasterEquivalent(metrics({ edgeIou: 0.96 }), false),
     ).toBe(false);
+    expect(
+      isCrossEngineRasterEquivalent(
+        metrics({
+          baselineWidthPx: 262,
+          candidateWidthPx: 272,
+          baselineHeightPx: 30,
+          candidateHeightPx: 31,
+          inkIou: 0.9792,
+          edgeIou: 0.9869,
+          foregroundColorDelta: 4.63,
+          baselineForegroundChroma: 2.01,
+          candidateForegroundChroma: 0,
+          foregroundLuminanceDelta: 4.63,
+        }),
+        false,
+        "list-item",
+      ),
+    ).toBe(false);
+    expect(
+      isCrossEngineRasterEquivalent(
+        metrics({
+          baselineWidthPx: 262,
+          candidateWidthPx: 272,
+          baselineHeightPx: 30,
+          candidateHeightPx: 31,
+          inkIou: 0.9792,
+          edgeIou: 0.9869,
+          foregroundColorDelta: 5.01,
+          baselineForegroundChroma: 0,
+          candidateForegroundChroma: 0,
+          foregroundLuminanceDelta: 5.01,
+        }),
+        false,
+        "list-item",
+      ),
+    ).toBe(false);
     expect(
       isCrossEngineRasterEquivalent(
         metrics({

packages/document-visual-diff/tests/text.test.ts

@@ -40,9 +40,88 @@ describe("PDF 文本行聚合", () => {
     ]);
   });
 
+  it("合并 Office 同一视觉行内小于字体高度两成的基线漂移", () => {
+    const left = item("源系统", 10, 100, 30, 10);
+    const middle = item("nut_pickup_order.meal_type", 42, 101.8, 140, 10);
+    const right = item("原值", 184, 100, 20, 10);
+    const lines = aggregatePdfTextLines([left, middle, right], 842);
+
+    expect(lines).toHaveLength(1);
+    expect(lines[0]?.normalizedText).toBe(
+      "源系统nut_pickup_order.meal_type原值",
+    );
+  });
+
+  it("将同行公式的上标片段按横坐标并回正文", () => {
+    const lines = aggregatePdfTextLines(
+      [
+        item("减因子", 10, 100, 42, 12),
+        item("e", 53, 97, 8, 15),
+        item("−λΔt", 62, 96, 24, 10),
+        item("的动态权重", 87, 100, 70, 12),
+      ],
+      842,
+    );
+
+    expect(lines).toHaveLength(1);
+    expect(lines[0]?.normalizedText).toBe("减因子e−λΔt的动态权重");
+  });
+
+  it("将超出正文右缘但紧邻行末的公式片段并回正文", () => {
+    const lines = aggregatePdfTextLines(
+      [
+        item("性、记忆重要程度、时效性衰减权重（须支持基于时间衰减因子 e", 150, 469.5, 346, 12),
+        item("−λΔt", 495.7, 466.1, 26, 10.2),
+        item("下一视觉行", 150, 489.7, 60, 12),
+      ],
+      842,
+    );
+
+    expect(lines.map((line) => line.normalizedText)).toEqual([
+      "性、记忆重要程度、时效性衰减权重(须支持基于时间衰减因子 e−λΔt",
+      "下一视觉行",
+    ]);
+  });
+
   it("将字体 ToUnicode 中的传统户部件归一为简体字符", () => {
     expect(normalizePdfText("戶⼾")).toBe("户户");
-    expect(normalizePdfText("⻅⻆⻓⻔⻚⻛")).toBe("见角长门页风");
+    expect(normalizePdfText("⻅⻆⻓⻔⻚⻛⻝⻣⻋⻬")).toBe(
+      "见角长门页风食骨车齐",
+    );
+  });
+
+  it("统一 Chromium 与 Office PDF 文本层的中英文弯引号", () => {
+    expect(normalizePdfText("“记录”与‘计算’")).toBe('"记录"与\'计算\'');
+  });
+
+  it("保留与全角括号重叠的窄引号文本流顺序", () => {
+    const lines = aggregatePdfTextLines(
+      [
+        item("主观定性判断", 532.3, 151.5, 58.5),
+        item("”", 590.8, 151.5, 3.4),
+        item("（如", 589.3, 151.5, 19.5),
+        item("“", 608.8, 151.5, 3.4),
+        item("高心率占比", 612.2, 151.5, 48.8),
+      ],
+      595,
+    );
+
+    expect(lines).toHaveLength(1);
+    expect(lines[0]?.normalizedText).toBe('主观定性判断"(如"高心率占比');
+  });
+
+  it("忽略 Emoji 文本层可选的变体选择符", () => {
+    expect(normalizePdfText("⚠️ 提示")).toBe("⚠ 提示");
+  });
+
+  it("统一 CJK 字体文本层的 em dash 与 horizontal bar", () => {
+    expect(normalizePdfText("仓库―仓区")).toBe("仓库—仓区");
+  });
+
+  it("统一数字区间中被 Chromium 提取为双连字符的 en dash", () => {
+    expect(normalizePdfText("10% -- 20%")).toBe("10%–20%");
+    expect(normalizePdfText("10% --")).toBe("10%–");
+    expect(normalizePdfText("命令 --flag")).toBe("命令 --flag");
   });
 
   it("从可编辑正文契约中移除 Word 自动列表装饰符", () => {

packages/docx-engine/assets/docx-media-filter.lua

@@ -98,6 +98,64 @@ local function replace_code_block(block)
   return pandoc.Para { image }
 end
 
+local function preserve_raw_html_as_text(inline)
+  if inline.format == "html" then
+    return pandoc.Str(inline.text)
+  end
+  return inline
+end
+
+local function preserve_raw_html_block_as_text(block)
+  if block.format == "html" then
+    return pandoc.Para { pandoc.Str(block.text) }
+  end
+  return block
+end
+
+local function replace_table_breaks(table_block)
+  return table_block:walk {
+    RawInline = function(inline)
+      if inline.format ~= "html" then
+        return inline
+      end
+      local normalized = string.lower(inline.text)
+      if normalized == "<br>"
+        or normalized == "<br/>"
+        or normalized == "<br />" then
+        return pandoc.LineBreak()
+      end
+      return inline
+    end
+  }
+end
+
+local function normalize_task_list(list)
+  for _, item in ipairs(list.content) do
+    local block = item[1]
+    if block ~= nil and (block.t == "Plain" or block.t == "Para") then
+      local inlines = block.content
+      local first = inlines[1]
+      if first ~= nil and first.t == "Str" then
+        local marker = string.lower(first.text)
+        if marker == "[x]" then
+          first.text = "☒"
+        elseif marker == "[" then
+          local second = inlines[2]
+          local third = inlines[3]
+          if second ~= nil and second.t == "Space"
+            and third ~= nil and third.t == "Str"
+            and third.text == "]" then
+            first.text = "☐"
+            inlines:remove(3)
+            inlines:remove(2)
+          end
+        end
+      end
+    end
+  end
+  return list
+end
+
 local function append_inlines(target, value)
   target:extend(pandoc.Inlines(value))
 end
@@ -167,7 +225,10 @@ local function validate_counts()
       error(
         "DOCX media map contains unused "
           .. kind
-          .. " resources"
+          .. " resources: consumed "
+          .. tostring(counters[kind])
+          .. " of "
+          .. tostring(#media_map[kind])
       )
     end
   end
@@ -176,8 +237,14 @@ end
 return {
   Pandoc = function(document)
     local transformed = document:walk {
+      Table = replace_table_breaks
+    }
+    transformed = transformed:walk {
       Image = replace_image,
-      CodeBlock = replace_code_block
+      CodeBlock = replace_code_block,
+      RawInline = preserve_raw_html_as_text,
+      RawBlock = preserve_raw_html_block_as_text,
+      BulletList = normalize_task_list
     }
     if structure_plan.titlePolicy.metadataTitle == "suppress" then
       transformed.meta.title = nil

packages/docx-engine/scripts/verify-theme-docx-matrix.mjs

@@ -424,7 +424,7 @@ for (const theme of themes) {
   let conversion;
   try {
     conversion = await converter.convert({
-      markdown,
+      markdown: rendered.markdownBody,
       fileName: `${theme.id}.md`,
       language: rendered.metadata.language,
       exportConfig,
@@ -433,7 +433,7 @@ for (const theme of themes) {
       metadata: rendered.metadata,
       semanticDocument: rendered.semanticDocument,
       fonts: readThemeFonts(theme),
-      media: createMedia(markdown)
+      media: createMedia(rendered.markdownBody)
     });
   } catch (error) {
     throw new Error(`主题 ${theme.id} 的 DOCX 转换失败`, {

packages/docx-engine/src/document-structure-transform.ts

@@ -44,6 +44,8 @@ type DocxBorderToken = NonNullable<
 
 const WORDPROCESSING_DRAWING_NAMESPACE =
   "http://schemas.openxmlformats.org/drawingml/2006/wordprocessingDrawing";
+const MATH_NAMESPACE =
+  "http://schemas.openxmlformats.org/officeDocument/2006/math";
 const XML_NAMESPACE = "http://www.w3.org/XML/1998/namespace";
 // Chromium 预留 6pt；Word/WPS 的分节承载段落还需要额外分页保留量。
 const COVER_PAGE_BREAK_SAFETY_PT = 6;
@@ -2064,9 +2066,10 @@ function applyDirectCharacterStyleTokens(
   const inlineCodeLayouts = documentLayout.inlineCodes ?? [];
   let inlineCodeCursor = 0;
   let count = 0;
-  for (const run of Array.from(
+  const runs = Array.from(
     document.getElementsByTagNameNS(WORD_NAMESPACE, "r")
-  )) {
+  );
+  for (const [runIndex, run] of runs.entries()) {
     const properties = firstDirectChild(
       run,
       WORD_NAMESPACE,
@@ -2099,8 +2102,29 @@ function applyDirectCharacterStyleTokens(
     const isSourceCode = paragraphStyle
       ? wordAttribute(paragraphStyle, "val") === "SourceCode"
       : false;
+    const previousRun = runs[runIndex - 1];
+    const nextRun = runs[runIndex + 1];
+    const isHtmlTagName =
+      isSourceCode &&
+      styleId === "KeywordTok" &&
+      previousRun?.parentNode === run.parentNode &&
+      nextRun?.parentNode === run.parentNode &&
+      /^<\/?$/u.test(previousRun.textContent ?? "") &&
+      /^\/?>$/u.test(nextRun.textContent ?? "");
+    const isObjectPropertyKey =
+      isSourceCode &&
+      styleId === "NormalTok" &&
+      nextRun?.parentNode === run.parentNode &&
+      /^\s*:\s*$/u.test(nextRun.textContent ?? "");
     const syntaxSlot = isSourceCode && styleId
-      ? resolvePandocSyntaxStyleSlot(styleId, run.textContent ?? "")
+      ? resolvePandocSyntaxStyleSlot(
+          styleId,
+          run.textContent ?? "",
+          {
+            htmlTagName: isHtmlTagName,
+            objectPropertyKey: isObjectPropertyKey
+          }
+        )
       : undefined;
     const token = isSourceCode
       ? syntaxSlot
@@ -2157,6 +2181,37 @@ function applyDirectCharacterStyleTokens(
   return count;
 }
 
+function applyMeasuredEmojiRunColors(
+  document: XmlDocument,
+  documentLayout: DocxDocumentLayoutPlan
+) {
+  const measuredRuns = documentLayout.emojiRuns ?? [];
+  let cursor = 0;
+  let count = 0;
+  for (const run of Array.from(
+    document.getElementsByTagNameNS(WORD_NAMESPACE, "r")
+  )) {
+    const text = (run.textContent ?? "").normalize("NFC").trim();
+    if (!text) {
+      continue;
+    }
+    const measuredIndex = measuredRuns.findIndex(
+      (layout, index) => index >= cursor && layout.text === text
+    );
+    if (measuredIndex < 0) {
+      continue;
+    }
+    cursor = measuredIndex + 1;
+    const properties = ensureFirstElement(run, "rPr", "w:rPr");
+    removeDirectChildren(properties, WORD_NAMESPACE, "color");
+    appendElement(properties, WORD_NAMESPACE, "w:color", {
+      "w:val": colorValue(measuredRuns[measuredIndex]!.color)
+    });
+    count += 1;
+  }
+  return count;
+}
+
 function applyTableCellToken(
   cell: XmlElement,
   token: DocxSlotStyleToken
@@ -2221,15 +2276,24 @@ function applyMeasuredTableGeometry(
   table: XmlElement,
   properties: XmlElement,
   layout: DocxDocumentLayoutPlan["tables"][number],
-  contentWidthTwips: number
+  contentWidthTwips: number,
+  leadingCellInsetTwips: number
 ) {
-  const tableWidthTwips = Math.max(
+  const indentTwips = Math.max(
+    0,
+    Math.round(contentWidthTwips * layout.leftOffsetPercent / 100) +
+      leadingCellInsetTwips
+  );
+  const requestedTableWidthTwips = Math.max(
     1,
     Math.round(contentWidthTwips * layout.widthPercent / 100)
   );
-  const indentTwips = Math.max(
+  const tableWidthTwips = Math.max(
-    0,
+    1,
-    Math.round(contentWidthTwips * layout.leftOffsetPercent / 100)
+    Math.min(
+      requestedTableWidthTwips,
+      Math.max(1, contentWidthTwips - indentTwips)
+    )
   );
   const columnWidths = measuredColumnWidths(
     tableWidthTwips,
@@ -2315,7 +2379,11 @@ function applyTables(
         table,
         properties,
         measuredLayout,
-        contentWidthTwips
+        contentWidthTwips,
+        pointsToTwips(Math.max(
+          tableCellToken?.paddingPt?.left ?? 0,
+          tableHeaderToken?.paddingPt?.left ?? 0
+        ))
       );
     } else {
       removeDirectChildren(
@@ -2427,8 +2495,15 @@ function applyTables(
             WORD_NAMESPACE,
             "p"
           )) {
+            const properties = paragraphProperties(paragraph);
+            const wordWrap = ensureFirstElement(
+              properties,
+              "wordWrap",
+              "w:wordWrap"
+            );
+            setWordAttribute(wordWrap, "val", "1");
             const spacing = ensureFirstElement(
-              paragraphProperties(paragraph),
+              properties,
               "spacing",
               "w:spacing"
             );
@@ -2436,7 +2511,7 @@ function applyTables(
             setWordAttribute(spacing, "after", "0");
             for (const property of ["autoSpaceDE", "autoSpaceDN"] as const) {
               const automaticSpacing = ensureFirstElement(
-                paragraphProperties(paragraph),
+                properties,
                 property,
                 `w:${property}`
               );
@@ -2609,6 +2684,18 @@ function applyMeasuredListItemIndents(
         String(Math.min(targetLeft, hanging))
       );
     }
+    if (measured.alignment) {
+      const alignment = ensureDirectElement(
+        properties,
+        WORD_NAMESPACE,
+        "w:jc"
+      );
+      setWordAttribute(
+        alignment,
+        "val",
+        measured.alignment === "justify" ? "both" : measured.alignment
+      );
+    }
     appliedCount += 1;
   }
   return appliedCount;
@@ -2727,6 +2814,33 @@ function measuredLineBreakOffsetsForParagraph(
   });
 }
 
+function paragraphHasInlineCode(paragraph: XmlElement) {
+  return Array.from(
+    paragraph.getElementsByTagNameNS(WORD_NAMESPACE, "rStyle")
+  ).some((style) => wordAttribute(style, "val") === "VerbatimChar");
+}
+
+function stabilizeInlineCodeParagraphLineRules(document: XmlDocument) {
+  let count = 0;
+  for (const paragraph of Array.from(
+    document.getElementsByTagNameNS(WORD_NAMESPACE, "p")
+  )) {
+    if (!paragraphHasInlineCode(paragraph)) {
+      continue;
+    }
+    const spacing = directChildren(
+      paragraphProperties(paragraph),
+      WORD_NAMESPACE,
+      "spacing"
+    )[0];
+    if (spacing && wordAttribute(spacing, "lineRule") === "exact") {
+      setWordAttribute(spacing, "lineRule", "atLeast");
+      count += 1;
+    }
+  }
+  return count;
+}
+
 function applyMeasuredTextBlockLineBreaks(
   document: XmlDocument,
   documentLayout: DocxDocumentLayoutPlan
@@ -2765,7 +2879,11 @@ function applyMeasuredTextBlockLineBreaks(
         "line",
         String(pointsToTwips(measured.linePitchPt))
       );
-      setWordAttribute(spacing, "lineRule", "exact");
+      setWordAttribute(
+        spacing,
+        "lineRule",
+        paragraphHasInlineCode(paragraph) ? "atLeast" : "exact"
+      );
     }
     for (const offset of [...measuredLineBreakOffsetsForParagraph(
       paragraph,
@@ -2781,6 +2899,47 @@ function applyMeasuredTextBlockLineBreaks(
   return appliedCount;
 }
 
+function stabilizeTableMathParagraphAlignment(document: XmlDocument) {
+  let count = 0;
+  for (const paragraph of Array.from(
+    document.getElementsByTagNameNS(WORD_NAMESPACE, "p")
+  )) {
+    const parent = paragraph.parentNode;
+    if (
+      !parent ||
+      parent.nodeType !== 1 ||
+      (parent as XmlElement).namespaceURI !== WORD_NAMESPACE ||
+      (parent as XmlElement).localName !== "tc"
+    ) {
+      continue;
+    }
+    const content = Array.from(paragraph.childNodes).filter(
+      (node): node is XmlElement => node.nodeType === 1
+    ).filter(
+      (child) =>
+        !(child.namespaceURI === WORD_NAMESPACE && child.localName === "pPr")
+    );
+    if (
+      content.length === 0 ||
+      !content.every(
+        (child) =>
+          child.namespaceURI === MATH_NAMESPACE &&
+          child.localName === "oMath"
+      )
+    ) {
+      continue;
+    }
+    const run = appendElement(paragraph, WORD_NAMESPACE, "w:r");
+    const properties = appendElement(run, WORD_NAMESPACE, "w:rPr");
+    appendElement(properties, WORD_NAMESPACE, "w:noProof");
+    const text = appendElement(run, WORD_NAMESPACE, "w:t");
+    text.setAttributeNS(XML_NAMESPACE, "xml:space", "preserve");
+    text.textContent = "\u200B";
+    count += 1;
+  }
+  return count;
+}
+
 function pageBreakAfterStyleIds(tokens: DocxThemeTokenSet) {
   const ids = new Set<string>();
   for (const entry of tokens.slots) {
@@ -3098,6 +3257,99 @@ function disableAutomaticCharacterSpacing(document: XmlDocument) {
   }
 }
 
+const PANDOC_ALERT_LABELS = new Set([
+  "note",
+  "tip",
+  "important",
+  "warning",
+  "caution"
+]);
+
+function normalizePandocAlertParagraphStyles(
+  body: XmlElement,
+  documentLayout: DocxDocumentLayoutPlan
+) {
+  const children = directChildren(body, WORD_NAMESPACE, "p");
+  const measuredAlerts = (documentLayout.textBlocks ?? []).filter(
+    (block) => block.alertRole !== undefined
+  );
+  let measuredCursor = 0;
+  let count = 0;
+  for (const [index, paragraph] of children.entries()) {
+    if (
+      paragraphStyleId(paragraph) !== "FirstParagraph" ||
+      !PANDOC_ALERT_LABELS.has(
+        normalizedLayoutText(paragraphText(paragraph)).toLocaleLowerCase(
+          "en-US"
+        )
+      )
+    ) {
+      continue;
+    }
+    const content = children[index + 1];
+    if (!content || paragraphStyleId(content) !== "BodyText") {
+      continue;
+    }
+    const style = ensureDirectElement(
+      paragraphProperties(content),
+      WORD_NAMESPACE,
+      "w:pStyle"
+    );
+    setWordAttribute(style, "val", "BlockText");
+    for (const [target, role] of [
+      [paragraph, "title"],
+      [content, "body"]
+    ] as const) {
+      const text = normalizedLayoutText(paragraphText(target));
+      const relativeMatchIndex = measuredAlerts
+        .slice(measuredCursor)
+        .findIndex(
+          (block) =>
+            block.alertRole === role &&
+            normalizedLayoutText(block.text) === text
+        );
+      if (relativeMatchIndex < 0) {
+        continue;
+      }
+      const measuredIndex = measuredCursor + relativeMatchIndex;
+      const measured = measuredAlerts[measuredIndex]!;
+      measuredCursor = measuredIndex + 1;
+      const properties = paragraphProperties(target);
+      const indentation = ensureDirectElement(
+        properties,
+        WORD_NAMESPACE,
+        "w:ind"
+      );
+      if (measured.leftIndentPt !== undefined) {
+        setWordAttribute(
+          indentation,
+          "left",
+          String(pointsToTwips(measured.leftIndentPt))
+        );
+      }
+      if (measured.rightIndentPt !== undefined) {
+        setWordAttribute(
+          indentation,
+          "right",
+          String(pointsToTwips(measured.rightIndentPt))
+        );
+      }
+      applyDirectRunToken(target, {
+        fontCandidates: [],
+        ...(measured.fontSizePt !== undefined
+          ? { fontSizePt: measured.fontSizePt }
+          : {}),
+        bold: measured.bold,
+        italic: measured.italic,
+        color: measured.color,
+        letterSpacingPt: measured.letterSpacingPt
+      });
+    }
+    count += 1;
+  }
+  return count;
+}
+
 export function finalizeGeneratedDocxStructure(
   content: Uint8Array,
   plan: PandocStructurePlan,
@@ -3141,6 +3393,7 @@ export function finalizeGeneratedDocxStructure(
     containers.editableContainerTables.size > 0,
     usesEvenAndOddPages
   );
+  normalizePandocAlertParagraphStyles(body, documentLayout);
   disableAutomaticCharacterSpacing(document);
   const final = finalSection(body);
   applyMeasuredTextBlockLineBreaks(document, documentLayout);
@@ -3160,6 +3413,9 @@ export function finalizeGeneratedDocxStructure(
   );
   applyMeasuredTextBlockAlignments(document, documentLayout);
   applyDirectCharacterStyleTokens(document, tokens, documentLayout);
+  applyMeasuredEmojiRunColors(document, documentLayout);
+  stabilizeInlineCodeParagraphLineRules(document);
+  stabilizeTableMathParagraphAlignment(document);
   const mediaDrawingCount = normalizeMediaParagraphs(
     document,
     tokens,

packages/docx-engine/src/pandoc-converter.ts

@@ -103,6 +103,12 @@ export interface PandocDocxConverterOptions {
   referenceCacheSize?: number;
 }
 
+export interface PandocDocxConversionDiagnostics {
+  outcome: string;
+  exitCode: number | null;
+  stderr: string;
+}
+
 export class PandocDocxConversionError extends Error {
   constructor(
     readonly code: Extract<
@@ -113,11 +119,26 @@ export class PandocDocxConversionError extends Error {
       | "DOCX_OUTPUT_INVALID"
     >,
     message: string,
-    options?: ErrorOptions
+    options?: ErrorOptions & {
+      diagnostics?: PandocDocxConversionDiagnostics;
+    }
   ) {
     super(message, options);
     this.name = "PandocDocxConversionError";
+    this.diagnostics = options?.diagnostics;
   }
+
+  readonly diagnostics: PandocDocxConversionDiagnostics | undefined;
+}
+
+function processDiagnostics(
+  result: Awaited<ReturnType<PandocProcessRunner>>
+): PandocDocxConversionDiagnostics {
+  return {
+    outcome: result.outcome,
+    exitCode: result.exitCode,
+    stderr: result.stderr.trim().slice(-16 * 1024)
+  };
 }
 
 function ensureOutputLimit(value: number | undefined) {
@@ -172,7 +193,7 @@ function pandocArguments(paths: {
   return [
     paths.markdown,
     "--from",
-    "markdown+yaml_metadata_block+pipe_tables+footnotes+task_lists+tex_math_dollars-raw_html",
+    "commonmark_x-yaml_metadata_block+pipe_tables+footnotes+task_lists+tex_math_dollars+raw_html",
     "--to",
     "docx",
     "--standalone",
@@ -351,7 +372,8 @@ export class PandocDocxConverter {
           processResult.outcome === "not-found"
             ? "DOCX_RUNTIME_NOT_FOUND"
             : "DOCX_GENERATION_FAILED",
-          "Pandoc 未能生成 DOCX"
+          "Pandoc 未能生成 DOCX",
+          { diagnostics: processDiagnostics(processResult) }
         );
       }
 

packages/docx-engine/src/pandoc-media.ts

@@ -2,8 +2,8 @@ import {
   DOCX_MEDIA_CSS_DPI,
   MAXIMUM_DOCX_MEDIA_BYTES,
   MAXIMUM_DOCX_MEDIA_EDGE_PIXELS,
+  MAXIMUM_DOCX_MEDIA_COUNT,
   MAXIMUM_DOCX_MEDIA_PIXELS,
-  MAXIMUM_DOCX_RESOURCE_COUNT,
   MAXIMUM_DOCX_TOTAL_MEDIA_BYTES,
   type DocxMediaKind,
   type DocxDocumentLayoutPlan,
@@ -101,7 +101,7 @@ export function preparePandocMedia(
   if (media.echartsErrors.length || media.mermaidErrors.length) {
     throw new Error("DOCX 图表渲染存在错误，已中止转换");
   }
-  if (media.resources.length > MAXIMUM_DOCX_RESOURCE_COUNT) {
+  if (media.resources.length > MAXIMUM_DOCX_MEDIA_COUNT) {
     throw new Error("DOCX 媒体数量超过限制");
   }
 

packages/docx-engine/src/section-transform.ts

@@ -139,6 +139,19 @@ export function transformSettingsXml(
 ) {
   const document = parseXmlPart(content, "word/settings.xml");
   const settings = document.documentElement!;
+  const compatibility =
+    firstDirectChild(settings, WORD_NAMESPACE, "compat") ??
+    appendElement(settings, WORD_NAMESPACE, "w:compat");
+  removeDirectChildren(
+    compatibility,
+    WORD_NAMESPACE,
+    "doNotExpandShiftReturn"
+  );
+  appendElement(
+    compatibility,
+    WORD_NAMESPACE,
+    "w:doNotExpandShiftReturn"
+  );
   removeDirectChildren(
     settings,
     WORD_NAMESPACE,

packages/docx-engine/src/styles-transform.ts

@@ -467,8 +467,13 @@ function applyTokenParagraphStyle(
     (token.paddingPt?.right ?? 0) > 0 ||
     (token.borders?.right?.widthPt ?? 0) > 0
       ? (token.rightIndentPt ?? 0) +
-        (token.paddingPt?.right ?? 0) +
+        // 带边框代码块通过 w:pBdr/w:space 表达 CSS 右内边距。
-        (token.borders?.right?.widthPt ?? 0)
+        // 若再把 padding/border 写进 w:ind，Word/WPS 会重复扣减
+        // 内容宽度，使本可容纳的等宽代码行在末尾额外回流。
+        (horizontalPaddingThroughBorderSpace && token.borders?.right
+          ? 0
+          : (token.paddingPt?.right ?? 0) +
+            (token.borders?.right?.widthPt ?? 0))
       : undefined;
   if (
     token.firstLineIndentPt !== undefined ||
@@ -1191,12 +1196,23 @@ function applyTokenStyles(
           binding.styleId === "ImageCaption"
             ? { ...entry.style, alignment: "center" as const }
             : entry.style;
-        applyTokenParagraphStyle(
+        const paragraphProperties = ensureDirectElement(
-          ensureDirectElement(
           target,
           WORD_NAMESPACE,
           "w:pPr"
-          ),
+        );
+        // SourceCode 的主题令牌来自浏览器完整计算样式。背景透明时，
+        // normalizer 会省略 backgroundColor；这里必须先清除基础模板的
+        // 兜底底纹，否则模板灰底会泄漏成主题代码块的整段背景。
+        if (binding.styleId === "SourceCode") {
+          removeDirectChildren(
+            paragraphProperties,
+            WORD_NAMESPACE,
+            "shd"
+          );
+        }
+        applyTokenParagraphStyle(
+          paragraphProperties,
           paragraphToken,
           fallbackFontSizeForSlot(slot, fallback),
           binding.styleId === "SourceCode"
@@ -1205,12 +1221,20 @@ function applyTokenStyles(
       const runToken = binding.styleId === "SourceCode"
         ? slots.get("code-block-text")?.style ?? entry.style
         : entry.style;
-      applyTokenRunStyle(
+      const runProperties = ensureDirectElement(
-        ensureDirectElement(
         target,
         WORD_NAMESPACE,
         "w:rPr"
-        ),
+      );
+      if (binding.styleId === "SourceCode") {
+        removeDirectChildren(
+          runProperties,
+          WORD_NAMESPACE,
+          "shd"
+        );
+      }
+      applyTokenRunStyle(
+        runProperties,
         runToken,
         fallbackFontsForSlot(slot, fallback),
         binding.styleId === "SourceCode" ||

packages/docx-engine/src/token-style-map.ts

@@ -210,17 +210,25 @@ const pandocSyntaxSlots: Readonly<
 
 export function resolvePandocSyntaxStyleSlot(
   styleId: string,
-  text = ""
+  text = "",
+  context: { htmlTagName?: boolean; objectPropertyKey?: boolean } = {}
 ): DocxStyleSlotName | "code-block-text" | undefined {
   if (!(PANDOC_SYNTAX_STYLE_IDS as readonly string[]).includes(styleId)) {
     return undefined;
   }
+  if (context.htmlTagName) {
+    return "code-token-name";
+  }
+  if (context.objectPropertyKey) {
+    return "code-token-attribute";
+  }
   if (
     styleId === "NormalTok" ||
     styleId === "OtherTok" ||
-    (styleId === "FunctionTok" && /^[\p{P}\p{S}\s]+$/u.test(text))
+    ((styleId === "FunctionTok" || styleId === "DataTypeTok") &&
+      /^[\p{P}\p{S}\s]+$/u.test(text))
   ) {
-    return styleId === "FunctionTok"
+    return styleId === "FunctionTok" || styleId === "DataTypeTok"
       ? "code-token-punctuation"
       : "code-block-text";
   }

packages/docx-engine/tests/document-structure-transform.test.ts

@@ -22,6 +22,8 @@ const drawing =
   "http://schemas.openxmlformats.org/drawingml/2006/main";
 const picture =
   "http://schemas.openxmlformats.org/drawingml/2006/picture";
+const math =
+  "http://schemas.openxmlformats.org/officeDocument/2006/math";
 
 function generatedFixture() {
   const baseline = readReferenceDocxPackage(
@@ -38,6 +40,14 @@ function generatedFixture() {
         `<w:p><w:r><w:t>CONTAINER_END</w:t></w:r></w:p>` +
         `<w:p><w:r><w:t>SECTION_BREAK</w:t></w:r></w:p>` +
         `<w:p><w:pPr><w:pStyle w:val="Heading1"/></w:pPr><w:r><w:t>正文标题</w:t></w:r></w:p>` +
+        `<w:p><w:pPr><w:pStyle w:val="SourceCode"/></w:pPr>` +
+        `<w:r><w:rPr><w:rStyle w:val="DataTypeTok"/></w:rPr><w:t>&lt;</w:t></w:r>` +
+        `<w:r><w:rPr><w:rStyle w:val="KeywordTok"/></w:rPr><w:t>br</w:t></w:r>` +
+        `<w:r><w:rPr><w:rStyle w:val="DataTypeTok"/></w:rPr><w:t>&gt;</w:t></w:r></w:p>` +
+        `<w:p><w:pPr><w:pStyle w:val="SourceCode"/></w:pPr>` +
+        `<w:r><w:rPr><w:rStyle w:val="NormalTok"/></w:rPr><w:t xml:space="preserve">  method</w:t></w:r>` +
+        `<w:r><w:rPr><w:rStyle w:val="OperatorTok"/></w:rPr><w:t>:</w:t></w:r>` +
+        `<w:r><w:rPr><w:rStyle w:val="StringTok"/></w:rPr><w:t xml:space="preserve"> "POST"</w:t></w:r></w:p>` +
         `<w:p><w:pPr><w:pStyle w:val="FirstParagraph"/></w:pPr><w:r><w:drawing>` +
         `<wp:inline xmlns:wp="${wordprocessingDrawing}"><wp:extent cx="100" cy="200"/><wp:docPr id="1" name="Picture" descr="Mermaid 图表 1" title="mdtp-media:docx-media-1"/>` +
         `<a:graphic xmlns:a="${drawing}"><a:graphicData uri="${picture}"><pic:pic xmlns:pic="${picture}"><pic:nvPicPr><pic:cNvPr id="0" name="image.png"/></pic:nvPicPr><pic:blipFill><a:blip r:embed="rIdImage"/><a:stretch><a:fillRect/></a:stretch></pic:blipFill><pic:spPr><a:xfrm rot="60000" flipH="1"><a:off x="0" y="0"/><a:ext cx="100" cy="200"/></a:xfrm></pic:spPr></pic:pic></a:graphicData></a:graphic>` +
@@ -183,6 +193,46 @@ const tokens: DocxThemeTokenSet = {
         }
       }
     },
+    {
+      slot: "code-block-text",
+      source: "computed-css",
+      confidence: "exact",
+      style: {
+        fontCandidates: ["Code Face"],
+        fontSizePt: 9,
+        color: "#24292e"
+      }
+    },
+    {
+      slot: "code-token-punctuation",
+      source: "computed-css",
+      confidence: "exact",
+      style: {
+        fontCandidates: ["Code Face"],
+        fontSizePt: 9,
+        color: "#7a7a7a"
+      }
+    },
+    {
+      slot: "code-token-name",
+      source: "computed-css",
+      confidence: "exact",
+      style: {
+        fontCandidates: ["Code Face"],
+        fontSizePt: 9,
+        color: "#22863a"
+      }
+    },
+    {
+      slot: "code-token-attribute",
+      source: "computed-css",
+      confidence: "exact",
+      style: {
+        fontCandidates: ["Code Face"],
+        fontSizePt: 9,
+        color: "#005cc5"
+      }
+    },
     {
       slot: "table-header",
       source: "computed-css",
@@ -292,7 +342,12 @@ describe("生成 DOCX 结构收口", () => {
       documentXml.match(
         /<w:spacing[^>]*w:before="0"[^>]*w:after="0"[^>]*w:line="360"[^>]*w:lineRule="exact"/gu
       )
-    ).toHaveLength(2);
+    ).toHaveLength(1);
+    expect(
+      documentXml.match(
+        /<w:spacing[^>]*w:before="0"[^>]*w:after="0"[^>]*w:line="360"[^>]*w:lineRule="atLeast"/gu
+      )
+    ).toHaveLength(1);
     expect(documentXml.match(/w:fill="17324D"/gu)).toHaveLength(2);
     expect(documentXml.match(/w:val="FFFFFF"/gu)).toHaveLength(2);
     expect(
@@ -304,6 +359,12 @@ describe("生成 DOCX 结构收口", () => {
     expect(tableParagraph).toContain('<w:autoSpaceDE w:val="0"/>');
     expect(tableParagraph).toContain('<w:autoSpaceDN w:val="0"/>');
     expect(documentXml).toContain("<w:cantSplit/>");
+    expect(documentXml).toMatch(
+      /<w:p><w:pPr><w:pStyle w:val="SourceCode"\/>[\s\S]*?<w:color w:val="7A7A7A"\/>[\s\S]*?<w:t>&lt;<\/w:t>[\s\S]*?<w:color w:val="22863A"\/>[\s\S]*?<w:t>br<\/w:t>[\s\S]*?<w:color w:val="7A7A7A"\/>[\s\S]*?<w:t>&gt;<\/w:t>[\s\S]*?<\/w:p>/u
+    );
+    expect(documentXml).toMatch(
+      /<w:p><w:pPr><w:pStyle w:val="SourceCode"\/>[\s\S]*?<w:rStyle w:val="NormalTok"\/>[\s\S]*?<w:color w:val="005CC5"\/>[\s\S]*?<w:t xml:space="preserve">  method<\/w:t>[\s\S]*?<\/w:p>/u
+    );
     expect(documentXml).toContain(
       '<w:tab w:val="right" w:pos="9806"/>'
     );
@@ -389,6 +450,98 @@ describe("生成 DOCX 结构收口", () => {
     );
   });
 
+  it("含行内代码的实测行距允许 Word 扩展行盒", () => {
+    const fixture = readGeneratedDocxPackage(generatedFixture());
+    const documentXml = decoder.decode(
+      fixture.entries.get("word/document.xml")!
+    ).replace(
+      "<w:sectPr>",
+      '<w:p><w:r><w:t>前缀</w:t></w:r><w:r><w:rPr><w:rStyle w:val="VerbatimChar"/></w:rPr><w:t>inline</w:t></w:r><w:r><w:t>后缀</w:t></w:r></w:p><w:sectPr>'
+    );
+    fixture.entries.set("word/document.xml", encoder.encode(documentXml));
+
+    const result = finalizeGeneratedDocxStructure(
+      writeGeneratedDocxPackage(fixture.entries),
+      plan,
+      tokens,
+      [],
+      {
+        tables: [],
+        textBlocks: [{
+          ordinal: 1,
+          text: "前缀inline后缀",
+          letterSpacingPt: 0,
+          linePitchPt: 11.15,
+          lineBreakOffsets: []
+        }]
+      }
+    );
+    const outputXml = decoder.decode(
+      readGeneratedDocxPackage(result.content).entries.get(
+        "word/document.xml"
+      )!
+    );
+
+    expect(outputXml).toMatch(
+      /w:line="223" w:lineRule="atLeast"[\s\S]*?<w:t>前缀<\/w:t>[\s\S]*?<w:rStyle w:val="VerbatimChar"\/>/u
+    );
+  });
+
+  it("未匹配实测布局的行内代码段落也不保留 exact 行盒", () => {
+    const fixture = readGeneratedDocxPackage(generatedFixture());
+    const documentXml = decoder.decode(
+      fixture.entries.get("word/document.xml")!
+    ).replace(
+      "<w:sectPr>",
+      '<w:p><w:pPr><w:spacing w:line="312" w:lineRule="exact"/></w:pPr><w:r><w:t>前缀</w:t></w:r><w:r><w:rPr><w:rStyle w:val="VerbatimChar"/></w:rPr><w:t>inline</w:t></w:r></w:p><w:sectPr>'
+    );
+    fixture.entries.set("word/document.xml", encoder.encode(documentXml));
+
+    const result = finalizeGeneratedDocxStructure(
+      writeGeneratedDocxPackage(fixture.entries),
+      plan,
+      tokens
+    );
+    const outputXml = decoder.decode(
+      readGeneratedDocxPackage(result.content).entries.get(
+        "word/document.xml"
+      )!
+    );
+
+    expect(outputXml).toMatch(
+      /w:line="312" w:lineRule="atLeast"[\s\S]*?<w:rStyle w:val="VerbatimChar"\/>/u
+    );
+  });
+
+  it("表格纯公式段落加入零宽普通 Run 以服从段落对齐", () => {
+    const fixture = readGeneratedDocxPackage(generatedFixture());
+    const documentXml = decoder.decode(
+      fixture.entries.get("word/document.xml")!
+    ).replace(
+      `xmlns:r="${relationships}"`,
+      `xmlns:r="${relationships}" xmlns:m="${math}"`
+    ).replace(
+      "<w:p><w:r><w:t>B</w:t></w:r></w:p>",
+      '<w:p><w:pPr><w:jc w:val="left"/></w:pPr><m:oMath><m:r><m:t>≥0.80</m:t></m:r></m:oMath></w:p>'
+    );
+    fixture.entries.set("word/document.xml", encoder.encode(documentXml));
+
+    const result = finalizeGeneratedDocxStructure(
+      writeGeneratedDocxPackage(fixture.entries),
+      plan,
+      tokens
+    );
+    const outputXml = decoder.decode(
+      readGeneratedDocxPackage(result.content).entries.get(
+        "word/document.xml"
+      )!
+    );
+
+    expect(outputXml).toMatch(
+      /<w:jc w:val="left"\/>[\s\S]*?<m:oMath>[\s\S]*?<m:t>≥0\.80<\/m:t>[\s\S]*?<\/m:oMath><w:r><w:rPr><w:noProof\/><\/w:rPr><w:t xml:space="preserve">\u200B<\/w:t><\/w:r>/u
+    );
+  });
+
   it("为右对齐盒模型保留 CSS 右侧内容内缩", () => {
     const fixture = readGeneratedDocxPackage(generatedFixture());
     const documentXml = decoder.decode(
@@ -508,6 +661,7 @@ describe("生成 DOCX 结构收口", () => {
       '<w:tblInd w:w="495" w:type="dxa"/>'
     );
     expect(documentXml).toContain('<w:tblLayout w:type="fixed"/>');
+    expect(documentXml).toContain('<w:wordWrap w:val="1"/>');
     expect(documentXml).toContain('<w:gridCol w:w="2675"/>');
     expect(documentXml).toContain('<w:gridCol w:w="6240"/>');
     expect(
@@ -525,6 +679,112 @@ describe("生成 DOCX 结构收口", () => {
     expect(documentXml).toContain('<w:jc w:val="distribute"/>');
   });
 
+  it("将超出内容盒的实测表格等比收缩到 Word 可用宽度", () => {
+    const result = finalizeGeneratedDocxStructure(
+      generatedFixture(),
+      plan,
+      tokens,
+      [],
+      {
+        tables: [
+          {
+            ordinal: 1,
+            widthPercent: 300,
+            leftOffsetPercent: 0,
+            columnWidthPercents: [40, 10, 10, 10, 10, 20],
+            rows: []
+          }
+        ]
+      }
+    );
+    const documentXml = decoder.decode(
+      readGeneratedDocxPackage(result.content).entries.get(
+        "word/document.xml"
+      )!
+    );
+
+    expect(documentXml).toContain(
+      '<w:tblW w:w="9906" w:type="dxa"/>'
+    );
+    expect(documentXml).toContain('<w:gridCol w:w="3962"/>');
+    expect(documentXml).toContain('<w:gridCol w:w="1980"/>');
+  });
+
+  it("按 Chromium 实测 emoji 调色保留可编辑 Unicode 文本", () => {
+    const fixture = readGeneratedDocxPackage(generatedFixture());
+    const documentXml = decoder.decode(
+      fixture.entries.get("word/document.xml")!
+    ).replace(
+      '<w:p><w:pPr><w:pStyle w:val="Heading1"/></w:pPr><w:r><w:t>正文标题</w:t></w:r></w:p>',
+      '<w:p><w:pPr><w:pStyle w:val="Heading1"/></w:pPr><w:r><w:t>正文标题</w:t></w:r></w:p>' +
+        '<w:p><w:r><w:t>⭐⭐⭐⭐</w:t></w:r></w:p>'
+    );
+    fixture.entries.set("word/document.xml", encoder.encode(documentXml));
+
+    const result = finalizeGeneratedDocxStructure(
+      writeGeneratedDocxPackage(fixture.entries),
+      plan,
+      tokens,
+      [],
+      {
+        tables: [],
+        emojiRuns: [
+          { ordinal: 1, text: "⭐⭐⭐⭐", color: "#e7bf36" }
+        ]
+      }
+    );
+    const finalizedXml = decoder.decode(
+      readGeneratedDocxPackage(result.content).entries.get(
+        "word/document.xml"
+      )!
+    );
+
+    expect(finalizedXml).toMatch(
+      /<w:r><w:rPr><w:color w:val="E7BF36"\/><\/w:rPr><w:t>⭐⭐⭐⭐<\/w:t><\/w:r>/u
+    );
+  });
+
+  it("实测表格缩进补偿 Word 以单元格内容而非外边框对齐的行为", () => {
+    const paddedTokens = {
+      ...tokens,
+      slots: tokens.slots.map((slot) =>
+        slot.slot === "table-cell"
+          ? {
+              ...slot,
+              style: {
+                ...slot.style,
+                paddingPt: { top: 3, right: 6, bottom: 3, left: 6 }
+              }
+            }
+          : slot
+      )
+    } satisfies DocxThemeTokenSet;
+    const result = finalizeGeneratedDocxStructure(
+      generatedFixture(),
+      plan,
+      paddedTokens,
+      [],
+      {
+        tables: [{
+          ordinal: 1,
+          widthPercent: 90,
+          leftOffsetPercent: 5,
+          columnWidthPercents: [30, 70],
+          rows: []
+        }]
+      }
+    );
+    const documentXml = decoder.decode(
+      readGeneratedDocxPackage(result.content).entries.get(
+        "word/document.xml"
+      )!
+    );
+
+    expect(documentXml).toContain(
+      '<w:tblInd w:w="615" w:type="dxa"/>'
+    );
+  });
+
   it("按稳定媒体计划规范化内联尺寸、比例锁和对齐", () => {
     const result = finalizeGeneratedDocxStructure(
       generatedFixture(),
@@ -882,6 +1142,36 @@ describe("生成 DOCX 结构收口", () => {
     );
   });
 
+  it("将 Pandoc 拆分的标准 Alert 正文恢复为引用块样式", () => {
+    const source = readGeneratedDocxPackage(generatedFixture());
+    const entries = new Map(source.entries);
+    const documentXml = decoder.decode(entries.get("word/document.xml")!);
+    entries.set(
+      "word/document.xml",
+      encoder.encode(documentXml.replace(
+        "<w:sectPr>",
+        `<w:p><w:pPr><w:pStyle w:val="FirstParagraph"/></w:pPr><w:r><w:t>Caution</w:t></w:r></w:p>` +
+        `<w:p><w:pPr><w:pStyle w:val="BodyText"/></w:pPr><w:r><w:t>警告正文</w:t></w:r></w:p>` +
+        `<w:sectPr>`
+      ))
+    );
+
+    const result = finalizeGeneratedDocxStructure(
+      writeGeneratedDocxPackage(entries),
+      plan,
+      tokens
+    );
+    const finalizedXml = decoder.decode(
+      readGeneratedDocxPackage(result.content).entries.get(
+        "word/document.xml"
+      )!
+    );
+
+    expect(finalizedXml).toMatch(
+      /<w:pPr><w:pStyle w:val="BlockText"\/>[\s\S]*?<w:t>警告正文<\/w:t>/u
+    );
+  });
+
   it("按最终页面内容区重算右置百分比容器缩进", () => {
     const relativeTokens: DocxThemeTokenSet = {
       ...tokens,

packages/docx-engine/tests/pandoc-converter.test.ts

@@ -127,7 +127,7 @@ describe("Pandoc DOCX 转换器", () => {
         expect(arguments_).toContain("--data-dir");
         expect(arguments_).toContain("--resource-path");
         expect(argumentAfter(arguments_, "--from")).toBe(
-          "markdown+yaml_metadata_block+pipe_tables+footnotes+task_lists+tex_math_dollars-raw_html"
+          "commonmark_x-yaml_metadata_block+pipe_tables+footnotes+task_lists+tex_math_dollars+raw_html"
         );
         expect(
           options.env?.MD_TO_PDF_DOCX_MEDIA_MAP
@@ -219,4 +219,29 @@ describe("Pandoc DOCX 转换器", () => {
       code: "DOCX_OUTPUT_INVALID"
     } satisfies Partial<PandocDocxConversionError>);
   });
+
+  it("进程失败时保留内部诊断但保持稳定的对外错误", async () => {
+    const failedRunner = vi.fn<PandocProcessRunner>(async () => ({
+      outcome: "completed",
+      exitCode: 64,
+      stdout: new Uint8Array(),
+      stderr: "pandoc: 媒体映射失败"
+    }));
+
+    await expect(
+      new PandocDocxConverter({
+        runtime,
+        runner: failedRunner,
+        temporaryRoot
+      }).convert(input())
+    ).rejects.toMatchObject({
+      code: "DOCX_GENERATION_FAILED",
+      message: "Pandoc 未能生成 DOCX",
+      diagnostics: {
+        outcome: "completed",
+        exitCode: 64,
+        stderr: "pandoc: 媒体映射失败"
+      }
+    } satisfies Partial<PandocDocxConversionError>);
+  });
 });

packages/docx-engine/tests/reference-builder.test.ts

@@ -404,11 +404,12 @@ describe("动态 reference.docx", () => {
     const sourceCodeStyle = stylesXml.match(
       /<w:style\b[^>]*w:styleId="SourceCode"[\s\S]*?<\/w:style>/u
     )?.[0];
+    expect(sourceCodeStyle).not.toContain("<w:shd");
     expect(sourceCodeStyle).toContain(
       '<w:left w:val="single" w:sz="6" w:space="7" w:color="E7EAED"/>'
     );
     expect(sourceCodeStyle).toMatch(
-      /<w:ind\b[^>]*w:left="195"[^>]*w:right="75"[^>]*\/>/u
+      /<w:ind\b[^>]*w:left="195"[^>]*w:right="0"[^>]*\/>/u
     );
     expect(sourceCodeStyle).toContain('w:ascii="Code Text Face"');
     expect(stylesXml).toMatch(
@@ -451,6 +452,49 @@ describe("动态 reference.docx", () => {
     );
   });
 
+  it("按代码容器令牌清除或写入 SourceCode 段落底纹", () => {
+    const createWithBackground = (backgroundColor?: string) => {
+      const result = createDynamicReferenceDocx(
+        createBaselineReference(),
+        {
+          ...createOptions(defaultExportConfig),
+          themeTokens: themeTokens("f".repeat(64), [
+            {
+              slot: "code-block",
+              source: "computed-css",
+              confidence: "exact",
+              style: {
+                fontCandidates: ["Consolas"],
+                fontSizePt: 10,
+                ...(backgroundColor ? { backgroundColor } : {})
+              }
+            },
+            {
+              slot: "code-block-text",
+              source: "computed-css",
+              confidence: "exact",
+              style: {
+                fontCandidates: ["Consolas"],
+                fontSizePt: 10
+              }
+            }
+          ])
+        }
+      );
+      const stylesXml = decoder.decode(
+        unzipSync(result.content)["word/styles.xml"]
+      );
+      return stylesXml.match(
+        /<w:style\b[^>]*w:styleId="SourceCode"[\s\S]*?<\/w:style>/u
+      )?.[0];
+    };
+
+    expect(createWithBackground()).not.toContain("<w:shd");
+    expect(createWithBackground("#eef2f6")).toContain(
+      '<w:shd w:val="clear" w:color="auto" w:fill="EEF2F6"/>'
+    );
+  });
+
   it("生成横向、自定义页边距和奇偶首页页眉页脚", () => {
     const exportConfig: ExportConfig = {
       ...defaultExportConfig,
@@ -506,6 +550,7 @@ describe("动态 reference.docx", () => {
     expect(documentXml.match(/w:headerReference/gu)).toHaveLength(3);
     expect(documentXml.match(/w:footerReference/gu)).toHaveLength(3);
     expect(settingsXml).toContain("<w:evenAndOddHeaders");
+    expect(settingsXml).toContain("<w:doNotExpandShiftReturn");
     expect(stylesXml).toContain('<w:kern w:val="2"');
     expect(headerXml).toContain("年度 &lt;报告&gt;");
     expect(headerXml).toContain("报告 &amp; 计划.md");

packages/docx-engine/tests/token-style-map.test.ts

@@ -89,6 +89,15 @@ describe("DOCX 令牌样式映射", () => {
   it("将 Pandoc 语法样式映射到引擎级代码语义槽位", () => {
     expect(resolvePandocSyntaxStyleSlot("DataTypeTok", '"key"'))
       .toBe("code-token-attribute");
+    expect(resolvePandocSyntaxStyleSlot("DataTypeTok", "<"))
+      .toBe("code-token-punctuation");
+    expect(
+      resolvePandocSyntaxStyleSlot(
+        "KeywordTok",
+        "section",
+        { htmlTagName: true }
+      )
+    ).toBe("code-token-name");
     expect(resolvePandocSyntaxStyleSlot("StringTok", '"value"'))
       .toBe("code-token-string");
     expect(resolvePandocSyntaxStyleSlot("FunctionTok", ":"))
@@ -97,6 +106,13 @@ describe("DOCX 令牌样式映射", () => {
       .toBe("code-token-title");
     expect(resolvePandocSyntaxStyleSlot("NormalTok", " "))
       .toBe("code-block-text");
+    expect(
+      resolvePandocSyntaxStyleSlot(
+        "NormalTok",
+        "  method",
+        { objectPropertyKey: true }
+      )
+    ).toBe("code-token-attribute");
     expect(resolvePandocSyntaxStyleSlot("UnknownTok", "x"))
       .toBeUndefined();
   });

packages/docx-theme-engine/src/collector.ts

@@ -41,6 +41,9 @@ export function readDocxComputedStyle(
     fontStyle: style.fontStyle,
     color: style.color,
     backgroundColor: style.backgroundColor,
+    ...(style.backgroundImage
+      ? { backgroundImage: style.backgroundImage }
+      : {}),
     lineHeight: style.lineHeight,
     letterSpacing: style.letterSpacing,
     textAlign: style.textAlign,

packages/docx-theme-engine/src/css-values.ts

@@ -79,6 +79,25 @@ export function parseCssColor(
   );
 }
 
+export function parseCssLinearGradientFallbackColor(
+  input: string
+): string | undefined {
+  const value = input.trim();
+  if (!/^(?:repeating-)?linear-gradient\(/iu.test(value)) {
+    return undefined;
+  }
+  const colorValues = value.match(
+    /rgba?\([^)]*\)|#[\da-f]{3,8}/giu
+  ) ?? [];
+  for (const colorValue of colorValues) {
+    const parsed = parseCssColor(colorValue);
+    if (parsed) {
+      return parsed;
+    }
+  }
+  return undefined;
+}
+
 function parseAlpha(value: string | undefined): number | undefined {
   if (value === undefined) {
     return undefined;

packages/docx-theme-engine/src/normalizer.ts

@@ -7,6 +7,7 @@ import {
   millimetersToPoints,
   parseCssBorder,
   parseCssColor,
+  parseCssLinearGradientFallbackColor,
   parseCssFontFamilies,
   parseCssLengthToPt,
   roundDocxValue
@@ -292,6 +293,20 @@ function normalizeComputedSlot(
   );
   if (background) {
     style.backgroundColor = background;
+  } else if (computed.backgroundImage) {
+    const gradientBackground = parseCssLinearGradientFallbackColor(
+      computed.backgroundImage
+    );
+    if (gradientBackground) {
+      style.backgroundColor = gradientBackground;
+      diagnostic(context, {
+        severity: "info",
+        code: "layout-approximated",
+        property: "background-image",
+        message: "CSS 线性渐变已使用首个非透明色近似为 Word 段落底纹"
+      });
+      context.approximate = true;
+    }
   }
   const letterSpacing = computed.letterSpacing.toLowerCase() === "normal"
     ? 0

packages/docx-theme-engine/src/probe.ts

@@ -35,7 +35,7 @@ export function createDocxStyleProbeMarkup(): string {
     <ul ${attribute("unordered-list")}><li ${attribute("list-item")}>无序列表</li></ul>
     <ol ${attribute("ordered-list")}><li>有序列表</li></ol>
     <blockquote ${attribute("block-quote")}><p ${attribute("block-quote-text")}>引用内容</p></blockquote>
-    <pre class="md-fences" ${attribute("code-block")}><code class="language-javascript" ${attribute("code-block-text")}><span class="hljs-keyword" ${attribute("code-token-keyword")}>const</span> <span class="hljs-attr" ${attribute("code-token-attribute")}>key</span> <span class="hljs-operator" ${attribute("code-token-operator")}>=</span> <span class="hljs-string" ${attribute("code-token-string")}>&quot;value&quot;</span>; <span class="hljs-literal" ${attribute("code-token-literal")}>true</span> <span class="hljs-title" ${attribute("code-token-title")}>title</span> <span class="hljs-number" ${attribute("code-token-number")}>1</span> <span class="hljs-comment" ${attribute("code-token-comment")}>// comment</span> <span class="hljs-built_in" ${attribute("code-token-built-in")}>Array</span> <span class="hljs-meta" ${attribute("code-token-meta")}>@meta</span> <span class="hljs-variable" ${attribute("code-token-variable")}>value</span> <span class="hljs-regexp" ${attribute("code-token-regexp")}>/x/</span> <span class="hljs-punctuation" ${attribute("code-token-punctuation")}>{}</span></code></pre>
+    <pre class="md-fences" ${attribute("code-block")}><code class="language-javascript" ${attribute("code-block-text")}><span class="hljs-keyword" ${attribute("code-token-keyword")}>const</span> <span class="hljs-attr" ${attribute("code-token-attribute")}>key</span> <span class="hljs-operator" ${attribute("code-token-operator")}>=</span> <span class="hljs-string" ${attribute("code-token-string")}>&quot;value&quot;</span>; <span class="hljs-literal" ${attribute("code-token-literal")}>true</span> <span class="hljs-title" ${attribute("code-token-title")}>title</span> <span class="hljs-number" ${attribute("code-token-number")}>1</span> <span class="hljs-comment" ${attribute("code-token-comment")}>// comment</span> <span class="hljs-built_in" ${attribute("code-token-built-in")}>Array</span> <span class="hljs-meta" ${attribute("code-token-meta")}>@meta</span> <span class="hljs-variable" ${attribute("code-token-variable")}>value</span> <span class="hljs-regexp" ${attribute("code-token-regexp")}>/x/</span> <span class="hljs-punctuation" ${attribute("code-token-punctuation")}>{}</span> <span class="hljs-name" ${attribute("code-token-name")}>section</span></code></pre>
     <table ${attribute("table")}>
       <thead><tr><th ${attribute("table-header")}>表头</th></tr></thead>
       <tbody><tr><td ${attribute("table-cell")}>单元格</td></tr></tbody>

packages/docx-theme-engine/src/runtime.ts

@@ -55,6 +55,7 @@ const computedProperties = [
   "fontStyle",
   "color",
   "backgroundColor",
+  "backgroundImage",
   "lineHeight",
   "letterSpacing",
   "textAlign",

packages/docx-theme-engine/src/slots.ts

@@ -35,6 +35,7 @@ export const DOCX_STYLE_SLOT_NAMES = [
   "code-token-operator",
   "code-token-regexp",
   "code-token-punctuation",
+  "code-token-name",
   "table",
   "table-header",
   "table-cell",
@@ -153,6 +154,7 @@ const kinds: Record<DocxStyleSlotName, DocxStyleSlotKind> = {
   "code-token-operator": "inline",
   "code-token-regexp": "inline",
   "code-token-punctuation": "inline",
+  "code-token-name": "inline",
   table: "table",
   "table-header": "table",
   "table-cell": "table",

packages/docx-theme-engine/src/snapshot.ts

@@ -10,6 +10,7 @@ export const docxComputedStyleSchema = z.object({
   fontStyle: cssValueSchema,
   color: cssValueSchema,
   backgroundColor: cssValueSchema,
+  backgroundImage: cssValueSchema.optional(),
   lineHeight: cssValueSchema,
   letterSpacing: cssValueSchema,
   textAlign: cssValueSchema,

packages/docx-theme-engine/tests/normalizer.test.ts

@@ -4,6 +4,7 @@ import {
   millimetersToPoints,
   parseCssBorder,
   parseCssColor,
+  parseCssLinearGradientFallbackColor,
   parseCssFontFamilies,
   parseCssLengthToPt,
   resolveDocxThemeTokens,
@@ -106,6 +107,12 @@ describe("CSS 到 Word 基础值归一化", () => {
     expect(parseCssColor("rgba(0, 0, 0, 0)")).toBeUndefined();
     expect(parseCssColor("rgba(255, 0, 0, 0.5)")).toBe("#ff8080");
     expect(parseCssColor("#00000080")).toBe("#7f7f7f");
+    expect(parseCssLinearGradientFallbackColor(
+      "linear-gradient(90deg, rgb(239, 246, 255), rgba(0, 0, 0, 0))"
+    )).toBe("#eff6ff");
+    expect(parseCssLinearGradientFallbackColor(
+      "radial-gradient(rgb(239, 246, 255), transparent)"
+    )).toBeUndefined();
     expect(
       parseCssFontFamilies(
         '"Source Han Serif SC", "Microsoft YaHei", serif'
@@ -185,6 +192,32 @@ describe("DOCX 主题令牌归一化", () => {
     expect(tokens.slots).toHaveLength(DOCX_STYLE_SLOT_NAMES.length);
   });
 
+  it("将线性渐变首个实色近似为 Word 段落底纹", () => {
+    const tokens = resolveDocxThemeTokens({
+      snapshot: createSnapshot({
+        "heading-2": {
+          backgroundColor: "rgba(0, 0, 0, 0)",
+          backgroundImage:
+            "linear-gradient(90deg, rgb(239, 246, 255), rgba(0, 0, 0, 0))"
+        }
+      }),
+      config: {
+        mode: "auto",
+        basePreset: "general",
+        overrides: {},
+        legacyPreset: false
+      }
+    });
+
+    expect(findSlot(tokens, "heading-2").style.backgroundColor)
+      .toBe("#eff6ff");
+    expect(tokens.diagnostics).toContainEqual(expect.objectContaining({
+      slot: "heading-2",
+      code: "layout-approximated",
+      property: "background-image"
+    }));
+  });
+
   it("按 border-box 文档的内容区宽度归一化全宽表格", () => {
     const tokens = resolveDocxThemeTokens({
       snapshot: createSnapshot({

packages/preview-engine/src/code-block-pagination.ts

@@ -3,6 +3,7 @@ const CHUNK_CLASS = "md-code-pagination-chunk";
 const CHUNK_POSITION_ATTRIBUTE = "data-code-pagination-position";
 const LINE_GROUP_CLASS = "md-code-line-group";
 const LINE_CLASS = "md-code-line";
+const LINE_INDENT_CLASS = "md-code-line-indent";
 
 function splitNodeIntoLines(node: Node): Node[][] {
   const documentRef = node.ownerDocument;
@@ -47,6 +48,56 @@ function lineHasContent(nodes: Node[]) {
   return nodes.some((node) => (node.textContent ?? "").length > 0);
 }
 
+function stabilizeLeadingWhitespace(
+  documentRef: Document,
+  nodes: Node[]
+) {
+  let indent = "";
+  let index = 0;
+  for (; index < nodes.length; index += 1) {
+    const node = nodes[index];
+    if (!node || node.nodeType !== node.TEXT_NODE) {
+      break;
+    }
+    const value = node.textContent ?? "";
+    const match = value.match(/^[\t ]+/u);
+    if (!match) {
+      if (value.length === 0) {
+        continue;
+      }
+      break;
+    }
+    indent += match[0];
+    const remainder = value.slice(match[0].length);
+    if (remainder) {
+      node.textContent = remainder;
+      break;
+    }
+  }
+  if (!indent) {
+    return nodes;
+  }
+  const indentElement = documentRef.createElement("span");
+  indentElement.className = LINE_INDENT_CLASS;
+  let columns = 0;
+  for (const character of indent) {
+    columns = character === "\t"
+      ? columns + (4 - columns % 4)
+      : columns + 1;
+  }
+  indentElement.dataset.codeIndentColumns = String(columns);
+  indentElement.style.width = `${columns}ch`;
+  indentElement.textContent = indent;
+  return [
+    indentElement,
+    ...nodes.slice(index).filter(
+      (node) =>
+        node.nodeType !== node.TEXT_NODE ||
+        (node.textContent ?? "").length > 0
+    )
+  ];
+}
+
 function createLineGroups(
   documentRef: Document,
   lines: Node[][]
@@ -68,7 +119,12 @@ function createLineGroups(
     ) {
       const line = documentRef.createElement("span");
       line.className = LINE_CLASS;
-      line.append(...(lines[lineIndex] ?? []));
+      line.append(
+        ...stabilizeLeadingWhitespace(
+          documentRef,
+          lines[lineIndex] ?? []
+        )
+      );
       group.append(line);
     }
 

packages/preview-engine/src/docx-media-runtime.ts

@@ -1,13 +1,14 @@
 import {
   DOCX_MEDIA_RASTER_SCALE,
   MAXIMUM_DOCX_MEDIA_EDGE_PIXELS,
+  MAXIMUM_DOCX_MEDIA_COUNT,
   MAXIMUM_DOCX_MEDIA_PIXELS,
-  MAXIMUM_DOCX_RESOURCE_COUNT,
   type DocxMediaCapturePlan,
   type DocxMediaCaptureTarget,
   type DocxMediaAlignment,
   type DocxMediaKind,
   type DocxDocumentLayoutPlan,
+  type DocxEmojiRunLayout,
   type DocxInlineCodeLayout,
   type DocxListItemLayout,
   type DocxTableLayout,
@@ -15,12 +16,19 @@ import {
   type PagedDocumentPayload
 } from "@md-to-pdf/core";
 import { PagedDocumentRuntime } from "./paged-document-runtime.js";
+import { stabilizePagedTableColumns } from "./paged-table-handler.js";
 
 export interface DocxMediaRenderDimensions {
   contentWidthPx: number;
   contentHeightPx: number;
 }
 
+export type DocxTableGeometry = Pick<
+  DocxTableLayout,
+  "ordinal" | "widthPercent" | "leftOffsetPercent" |
+    "columnWidthPercents"
+>;
+
 declare global {
   interface Window {
     __mdToPdfRenderDocxMedia?: (
@@ -157,6 +165,23 @@ function average(values: readonly number[], fallback: number) {
 }
 
 function collectTableColumnWidths(table: HTMLTableElement, tableRect: DOMRect) {
+  const stabilizedColumns = Array.from(
+    table.querySelectorAll<HTMLTableColElement>(
+      ":scope > colgroup[data-stabilized-columns=\"true\"] > col"
+    )
+  );
+  const stabilizedWidths = stabilizedColumns.map((column) =>
+    Number(column.dataset.stabilizedWidthPx)
+  );
+  if (
+    stabilizedWidths.length > 0 &&
+    stabilizedWidths.every((width) => Number.isFinite(width) && width > 0)
+  ) {
+    // 这组轨道来自分页前的完整自动布局。后续为稳定 Paged.js 克隆
+    // 写入的 col/cell 辅助宽度可能反过来扰动源表二次测量，DOCX 必须
+    // 使用最初与 Chromium 分页基线一致的列轨，而不是辅助样式后的值。
+    return stabilizedWidths;
+  }
   const rows = Array.from(table.rows);
   const columnCount = rows.reduce(
     (maximum, row) =>
@@ -255,7 +280,7 @@ function cellAlignment(value: string): "left" | "center" | "right" | "justify" {
 
 const DOCX_TEXT_BLOCK_SELECTOR =
   "h1, h2, h3, h4, h5, h6, p, li, th, td, figcaption, " +
-  ".doc-classification, .doc-issue-row, .doc-printing-row, " +
+  ".md-alert-text, .doc-classification, .doc-issue-row, .doc-printing-row, " +
   ".doc-briefing-meta";
 
 function textNodeCharacters(root: HTMLElement) {
@@ -293,9 +318,19 @@ function characterRect(
   return rect.width > 0 && rect.height > 0 ? rect : undefined;
 }
 
+function hasDistributedLastLine(
+  _element: HTMLElement,
+  computed: CSSStyleDeclaration,
+  _lastLine: readonly DOMRect[]
+) {
+  const textAlignLast = computed.textAlignLast.trim().toLowerCase();
+  return textAlignLast === "justify" || textAlignLast === "distribute";
+}
+
 export function collectDocxTextBlockLayouts(
   article: HTMLElement
 ): DocxTextBlockLayout[] {
+  const articleRect = article.getBoundingClientRect();
   const candidates = Array.from(
     article.querySelectorAll<HTMLElement>(DOCX_TEXT_BLOCK_SELECTOR)
   ).filter((element) => {
@@ -350,39 +385,33 @@ export function collectDocxTextBlockLayouts(
     }
     const text = output.join("").trim();
     const computed = getComputedStyle(element);
+    const alert = element.closest<HTMLElement>(".md-alert");
+    const alertRole = element.matches(".md-alert-text")
+      ? "title" as const
+      : alert
+        ? "body" as const
+        : undefined;
+    const alertRect = alert?.getBoundingClientRect();
+    const alertComputed = alert ? getComputedStyle(alert) : undefined;
+    const alertLeft = alertRect && alertComputed
+      ? alertRect.left +
+        Number.parseFloat(alertComputed.borderLeftWidth || "0") +
+        Number.parseFloat(alertComputed.paddingLeft || "0")
+      : undefined;
+    const alertRight = alertRect && alertComputed
+      ? alertRect.right -
+        Number.parseFloat(alertComputed.borderRightWidth || "0") -
+        Number.parseFloat(alertComputed.paddingRight || "0")
+      : undefined;
     const letterSpacingPx = computed.letterSpacing === "normal"
       ? 0
       : Number.parseFloat(computed.letterSpacing);
-    const elementRect = element.getBoundingClientRect();
-    const contentWidth = Math.max(
-      0,
-      elementRect.width -
-        Number.parseFloat(computed.paddingLeft || "0") -
-        Number.parseFloat(computed.paddingRight || "0") -
-        Number.parseFloat(computed.borderLeftWidth || "0") -
-        Number.parseFloat(computed.borderRightWidth || "0")
-    );
     const lastLine = lineCharacterRects.at(-1) ?? [];
-    const sortedLastLine = [...lastLine].sort(
+    const distributed = hasDistributedLastLine(
-      (left, right) => left.left - right.left
+      element,
+      computed,
+      lastLine
     );
-    const lastLineWidth = sortedLastLine.length > 0
-      ? sortedLastLine.at(-1)!.right - sortedLastLine[0]!.left
-      : 0;
-    const maximumGap = sortedLastLine.slice(1).reduce(
-      (largest, rect, rectIndex) => Math.max(
-        largest,
-        rect.left - sortedLastLine[rectIndex]!.right
-      ),
-      0
-    );
-    const fontSizePx = Number.parseFloat(computed.fontSize);
-    const distributed =
-      sortedLastLine.length > 1 &&
-      contentWidth > 0 &&
-      lastLineWidth / contentWidth >= 0.85 &&
-      Number.isFinite(fontSizePx) &&
-      maximumGap >= fontSizePx * 0.5;
     const lineTops = lineCharacterRects
       .map((rects) => rects.length > 0
         ? Math.min(...rects.map((rect) => rect.top))
@@ -401,6 +430,19 @@ export function collectDocxTextBlockLayouts(
           letterSpacingPt: Number.isFinite(letterSpacingPx)
             ? letterSpacingPx * 0.75
             : 0,
+          ...(alertRole && alertLeft !== undefined && alertRight !== undefined
+            ? {
+                alertRole,
+                fontSizePt: cssPixelsToPoints(computed.fontSize),
+                bold: computed.fontWeight === "bold" ||
+                  Number.parseInt(computed.fontWeight, 10) >= 600,
+                italic: computed.fontStyle === "italic" ||
+                  computed.fontStyle === "oblique",
+                color: cssColor(computed.color) ?? "#000000",
+                leftIndentPt: Math.max(0, (alertLeft - articleRect.left) * 0.75),
+                rightIndentPt: Math.max(0, (articleRect.right - alertRight) * 0.75)
+              }
+            : {}),
           alignment: distributed
             ? "distribute" as const
             : cellAlignment(computed.textAlign),
@@ -447,6 +489,18 @@ export function collectDocxListItemLayouts(
       if (!rect) {
         return [];
       }
+      const visibleRects = characters.flatMap((character) => {
+        const characterBounds = characterRect(range, character);
+        return characterBounds ? [characterBounds] : [];
+      });
+      const lastLineTop = visibleRects.length > 0
+        ? Math.max(...visibleRects.map((bounds) => bounds.top))
+        : 0;
+      const lastLine = visibleRects.filter(
+        (bounds) => Math.abs(bounds.top - lastLineTop) <= 1
+      );
+      const computed = getComputedStyle(item);
+      const distributed = hasDistributedLastLine(item, computed, lastLine);
       let listDepth = -1;
       for (
         let ancestor: Element | null = item.parentElement;
@@ -464,7 +518,8 @@ export function collectDocxListItemLayouts(
         textStartPt: Math.max(
           0,
           (rect.left - articleLeft) * 0.75
-        )
+        ),
+        ...(distributed ? { alignment: "distribute" as const } : {})
       }];
     }
   );
@@ -475,6 +530,86 @@ function cssPixelsToPoints(value: string) {
   return Number.isFinite(pixels) ? pixels * 0.75 : 0;
 }
 
+const EMOJI_ONLY_TEXT = /^(?:\p{Extended_Pictographic}|\p{Emoji_Modifier}|\u200d|\ufe0f|\s)+$/u;
+
+function sampledEmojiColor(text: string, computed: CSSStyleDeclaration) {
+  const canvas = document.createElement("canvas");
+  const context = canvas.getContext("2d", { willReadFrequently: true });
+  if (!context) {
+    return undefined;
+  }
+  context.font = computed.font;
+  const metrics = context.measureText(text);
+  const fontSize = Math.max(12, Number.parseFloat(computed.fontSize) || 16);
+  canvas.width = Math.max(1, Math.ceil(metrics.width + fontSize));
+  canvas.height = Math.max(1, Math.ceil(fontSize * 2));
+  const renderContext = canvas.getContext("2d", { willReadFrequently: true });
+  if (!renderContext) {
+    return undefined;
+  }
+  renderContext.clearRect(0, 0, canvas.width, canvas.height);
+  renderContext.font = computed.font;
+  renderContext.textBaseline = "middle";
+  renderContext.fillStyle = computed.color;
+  renderContext.fillText(text, fontSize / 2, canvas.height / 2);
+  const pixels = renderContext.getImageData(
+    0,
+    0,
+    canvas.width,
+    canvas.height
+  ).data;
+  const histogram = new Map<string, number>();
+  for (let index = 0; index < pixels.length; index += 4) {
+    const alpha = pixels[index + 3] ?? 0;
+    if (alpha < 96) {
+      continue;
+    }
+    const red = pixels[index] ?? 0;
+    const green = pixels[index + 1] ?? 0;
+    const blue = pixels[index + 2] ?? 0;
+    if (Math.max(red, green, blue) - Math.min(red, green, blue) < 16) {
+      continue;
+    }
+    const key = [red, green, blue]
+      .map((value) => Math.round(value / 8) * 8)
+      .join(",");
+    histogram.set(key, (histogram.get(key) ?? 0) + alpha);
+  }
+  const dominant = Array.from(histogram.entries()).sort(
+    ([, left], [, right]) => right - left
+  )[0]?.[0];
+  if (!dominant) {
+    return undefined;
+  }
+  const channels = dominant.split(",").map(Number);
+  if (channels.length !== 3 || channels.some((value) => !Number.isFinite(value))) {
+    return undefined;
+  }
+  return `#${channels
+    .map((value) => Math.min(255, value).toString(16).padStart(2, "0"))
+    .join("")}`;
+}
+
+export function collectDocxEmojiRunLayouts(
+  article: HTMLElement
+): DocxEmojiRunLayout[] {
+  const walker = document.createTreeWalker(article, NodeFilter.SHOW_TEXT);
+  const layouts: DocxEmojiRunLayout[] = [];
+  for (let node = walker.nextNode(); node; node = walker.nextNode()) {
+    const text = (node.textContent ?? "").normalize("NFC").trim();
+    const parent = node.parentElement;
+    if (!text || !parent || !EMOJI_ONLY_TEXT.test(text)) {
+      continue;
+    }
+    const color = sampledEmojiColor(text, getComputedStyle(parent));
+    if (!color) {
+      continue;
+    }
+    layouts.push({ ordinal: layouts.length + 1, text, color });
+  }
+  return layouts;
+}
+
 export function collectDocxInlineCodeLayouts(
   article: HTMLElement
 ): DocxInlineCodeLayout[] {
@@ -545,7 +680,7 @@ export function collectDocxDocumentLayoutPlan(
     );
     return {
       ordinal: index + 1,
-      widthPercent: Math.min(100, (width / contentWidth) * 100),
+      widthPercent: Math.min(300, (width / contentWidth) * 100),
       leftOffsetPercent: Math.max(
         0,
         Math.min(100, ((rect.left - articleRect.left) / contentWidth) * 100)
@@ -570,11 +705,112 @@ export function collectDocxDocumentLayoutPlan(
       }))
     };
   });
+  const emojiRuns = collectDocxEmojiRunLayouts(article);
   return {
     tables,
     textBlocks: collectDocxTextBlockLayouts(article),
     listItems: collectDocxListItemLayouts(article),
-    inlineCodes: collectDocxInlineCodeLayouts(article)
+    inlineCodes: collectDocxInlineCodeLayouts(article),
+    ...(emojiRuns.length > 0 ? { emojiRuns } : {})
+  };
+}
+
+export function collectPagedDocxTableGeometries(
+  root: ParentNode,
+  dimensions: DocxMediaRenderDimensions
+): DocxTableGeometry[] {
+  const tablesByReference = new Map<string, HTMLTableElement[]>();
+  let anonymousTableIndex = 0;
+  for (const table of root.querySelectorAll<HTMLTableElement>(
+    ".pagedjs_pages table"
+  )) {
+    const reference = table.dataset.ref ??
+      `__anonymous_table_${anonymousTableIndex += 1}`;
+    const tables = tablesByReference.get(reference) ?? [];
+    tables.push(table);
+    tablesByReference.set(reference, tables);
+  }
+  const geometries: DocxTableGeometry[] = [];
+  for (const tables of tablesByReference.values()) {
+    const table = tables.reduce((best, candidate) => {
+      const visibleArea = (element: HTMLTableElement) => {
+        const content = element.closest<HTMLElement>(
+          ".pagedjs_area, #write"
+        );
+        const contentRect = content?.getBoundingClientRect();
+        const rect = element.getBoundingClientRect();
+        if (!contentRect) {
+          return Math.max(0, rect.width) * Math.max(0, rect.height);
+        }
+        const width = Math.max(
+          0,
+          Math.min(rect.right, contentRect.right) -
+            Math.max(rect.left, contentRect.left)
+        );
+        const height = Math.max(
+          0,
+          Math.min(rect.bottom, contentRect.bottom) -
+            Math.max(rect.top, contentRect.top)
+        );
+        return width * height;
+      };
+      return visibleArea(candidate) > visibleArea(best)
+        ? candidate
+        : best;
+    });
+    const content = table.closest<HTMLElement>(
+      ".pagedjs_area, #write"
+    );
+    const contentRect = content?.getBoundingClientRect();
+    const contentWidth = finitePositive(
+      contentRect?.width ?? 0,
+      dimensions.contentWidthPx
+    );
+    const contentLeft = contentRect?.left ?? 0;
+    const rect = table.getBoundingClientRect();
+    const width = finitePositive(rect.width, contentWidth);
+    const columnWidths = collectTableColumnWidths(table, rect);
+    const columnTotal = columnWidths.reduce(
+      (total, value) => total + value,
+      0
+    );
+    geometries.push({
+      ordinal: geometries.length + 1,
+      widthPercent: Math.min(300, (width / contentWidth) * 100),
+      leftOffsetPercent: Math.max(
+        0,
+        Math.min(100, ((rect.left - contentLeft) / contentWidth) * 100)
+      ),
+      columnWidthPercents: columnWidths.map(
+        (value) => (value / columnTotal) * 100
+      )
+    });
+  }
+  return geometries;
+}
+
+export function mergePagedTableGeometries(
+  layout: DocxDocumentLayoutPlan,
+  pagedGeometries: readonly DocxTableGeometry[]
+): DocxDocumentLayoutPlan {
+  return {
+    ...layout,
+    tables: layout.tables.map((table, index) => {
+      const geometry = pagedGeometries[index];
+      if (
+        !geometry ||
+        geometry.columnWidthPercents.length !==
+          table.columnWidthPercents.length
+      ) {
+        return table;
+      }
+      return {
+        ...table,
+        widthPercent: geometry.widthPercent,
+        leftOffsetPercent: geometry.leftOffsetPercent,
+        columnWidthPercents: [...geometry.columnWidthPercents]
+      };
+    })
   };
 }
 
@@ -595,9 +831,9 @@ export function collectDocxMediaCaptureTargets(
       ].join(",")
     )
   );
-  if (candidates.length > MAXIMUM_DOCX_RESOURCE_COUNT) {
+  if (candidates.length > MAXIMUM_DOCX_MEDIA_COUNT) {
     throw new Error(
-      `DOCX 媒体数量不能超过 ${MAXIMUM_DOCX_RESOURCE_COUNT}`
+      `DOCX 媒体数量不能超过 ${MAXIMUM_DOCX_MEDIA_COUNT}`
     );
   }
 
@@ -658,6 +894,20 @@ export async function renderDocxMediaCapturePlan(
   payload: PagedDocumentPayload,
   dimensions: DocxMediaRenderDimensions
 ): Promise<DocxMediaCapturePlan> {
+  const pagedResult = await runtime.render(payload, {
+    target: "pdf",
+    // DOCX 只消费分页表格的宽度、偏移和列宽比例；完整行结构与样式随后
+    // 从连续 DOM 重新采集。Paged.js 偶发丢失边界行时允许几何降级，
+    // 正常 PDF 渲染仍保持严格的正文行完整性门禁。
+    allowIncompleteTableGeometry: true
+  });
+  if (!pagedResult) {
+    throw new Error("DOCX 分页布局渲染已取消");
+  }
+  const pagedTableGeometries = collectPagedDocxTableGeometries(
+    root,
+    dimensions
+  );
   const renderResult = await runtime.renderContinuous(payload, {
     geometryCss: createGeometryCss(dimensions),
     themeMedia: "print"
@@ -665,10 +915,24 @@ export async function renderDocxMediaCapturePlan(
   if (!renderResult) {
     throw new Error("DOCX 媒体渲染已取消");
   }
+  stabilizePagedTableColumns(root, dimensions.contentHeightPx);
+  const continuousLayout = collectDocxDocumentLayoutPlan(
+    root,
+    dimensions
+  );
   return {
     targets: collectDocxMediaCaptureTargets(root, dimensions),
-    echartsErrors: renderResult.echartsErrors,
+    echartsErrors: Array.from(new Set([
-    mermaidErrors: renderResult.mermaidErrors,
+      ...pagedResult.echartsErrors,
-    documentLayout: collectDocxDocumentLayoutPlan(root, dimensions)
+      ...renderResult.echartsErrors
+    ])),
+    mermaidErrors: Array.from(new Set([
+      ...pagedResult.mermaidErrors,
+      ...renderResult.mermaidErrors
+    ])),
+    documentLayout: mergePagedTableGeometries(
+      continuousLayout,
+      pagedTableGeometries
+    )
   };
 }

packages/preview-engine/src/index.ts

@@ -17,5 +17,6 @@ export * from "./paged-page-sequence.js";
 export * from "./semantic-cover-fit.js";
 export * from "./paged-page-decorations.js";
 export * from "./paged-render-target.js";
+export * from "./paged-table-handler.js";
 export * from "./pdf-document-links.js";
 export * from "./preview-styles.js";

packages/preview-engine/src/paged-document-runtime.ts

@@ -37,7 +37,11 @@ import { renderMermaidDefinitions } from "./mermaid-renderer.js";
 import { replaceMermaidSvgWithImages } from "./mermaid-static-image.js";
 import type { MermaidOutputMode } from "./mermaid-static-image.js";
 import { enablePrintMediaForPreview } from "./preview-styles.js";
-import "./paged-table-handler.js";
+import {
+  forcePagedTableRowsToNextPage,
+  missingPagedTableRowIds,
+  stabilizePagedTableColumns
+} from "./paged-table-handler.js";
 import {
   buildPagedMediaCss,
   continuousDocumentGeometryCss,
@@ -61,6 +65,7 @@ export interface PagedDocumentRenderOptions {
   target: PagedRenderTarget;
   shouldContinue?: () => boolean;
   mermaidOutput?: MermaidOutputMode;
+  allowIncompleteTableGeometry?: boolean;
 }
 
 export interface ContinuousDocumentRenderOptions {
@@ -75,6 +80,33 @@ export interface PreviewEngineStyles {
   echartsCss: string;
 }
 
+export function protectFittableInlineCodeParagraphs(
+  root: ParentNode,
+  pageContentHeightPx: number
+) {
+  if (!Number.isFinite(pageContentHeightPx) || pageContentHeightPx <= 0) {
+    return 0;
+  }
+  let protectedCount = 0;
+  for (const paragraph of root.querySelectorAll<HTMLElement>(
+    "#write > p.md-inline-code-paragraph"
+  )) {
+    const height = paragraph.getBoundingClientRect().height;
+    if (!(height > 0 && height <= pageContentHeightPx)) {
+      continue;
+    }
+    paragraph.style.setProperty("break-inside", "avoid", "important");
+    paragraph.style.setProperty(
+      "page-break-inside",
+      "avoid",
+      "important"
+    );
+    paragraph.dataset.mdtpInlineCodeKeepWhole = "true";
+    protectedCount += 1;
+  }
+  return protectedCount;
+}
+
 export const DEFAULT_IMAGE_READY_TIMEOUT_MS = 10_000;
 
 function waitForImage(
@@ -699,6 +731,7 @@ export class PagedDocumentRuntime {
       payload.features.includes("echarts") ||
       content.querySelector(".mermaid svg") ||
       content.querySelector("img.md-document-image") ||
+      content.querySelector("table") ||
       content.querySelector('[data-semantic-region="cover"]')
     ) {
       const measurement = mountMeasurementContainer(
@@ -710,6 +743,14 @@ export class PagedDocumentRuntime {
       );
       try {
         await documentRef.fonts.ready;
+        protectFittableInlineCodeParagraphs(
+          measurement.host,
+          getPageContentDimensions(payload.exportConfig).height
+        );
+        stabilizePagedTableColumns(
+          measurement.host,
+          getPageContentDimensions(payload.exportConfig).height
+        );
         constrainSemanticCoversToPage(
           measurement.host,
           getPageContentDimensions(payload.exportConfig).height
@@ -816,6 +857,51 @@ export class PagedDocumentRuntime {
         stylesheets,
         this.root
       );
+      const maximumTableRowRecoveryAttempts = 3;
+      for (
+        let attempt = 0;
+        attempt < maximumTableRowRecoveryAttempts;
+        attempt += 1
+      ) {
+        const missingRowIds = missingPagedTableRowIds(
+          repaginationSource,
+          this.root
+        );
+        if (missingRowIds.length === 0) {
+          break;
+        }
+        const recoveredCount = forcePagedTableRowsToNextPage(
+          repaginationSource,
+          missingRowIds
+        );
+        if (recoveredCount !== missingRowIds.length) {
+          throw new Error(
+            `分页表格行恢复标记不完整：缺失 ${missingRowIds.length} 行，` +
+              `仅定位 ${recoveredCount} 行`
+          );
+        }
+        previewer.chunker.destroy();
+        previewer.polisher.destroy();
+        previewer = new Previewer();
+        this.activePreviewer = previewer;
+        flow = await previewer.preview(
+          repaginationSource.cloneNode(true) as DocumentFragment,
+          stylesheets,
+          this.root
+        );
+      }
+      const unresolvedTableRowIds = missingPagedTableRowIds(
+        repaginationSource,
+        this.root
+      );
+      if (
+        unresolvedTableRowIds.length > 0 &&
+        !options.allowIncompleteTableGeometry
+      ) {
+        throw new Error(
+          `分页后表格正文行缺失：${unresolvedTableRowIds.join(", ")}`
+        );
+      }
       const mediaIds = getMediaBackfillIdsInDocumentOrder(
         repaginationSource
       );

packages/preview-engine/src/paged-preview.ts

@@ -127,9 +127,21 @@ svg {
   table-layout: auto;
 }
 
+#write table[data-stabilized-columns="true"] {
+  table-layout: fixed !important;
+}
+
 #write th,
 #write td {
   overflow-wrap: anywhere;
+  word-break: break-word;
+}
+
+#write th code,
+#write td code {
+  white-space: pre-wrap;
+  overflow-wrap: anywhere;
+  word-break: break-word;
 }
 
 .mermaid {
@@ -421,6 +433,15 @@ ${buildFooterCss(config)}
   widows: 3;
 }
 
+#write > p.md-inline-code-paragraph {
+  orphans: 1;
+  widows: 1;
+}
+
+#write li {
+  text-align-last: left;
+}
+
 #write h1,
 #write h2,
 #write h3,
@@ -489,6 +510,11 @@ ${buildFooterCss(config)}
   word-break: inherit;
 }
 
+#write .md-code-line-indent {
+  display: inline-block;
+  white-space: pre;
+}
+
 #write table[data-empty-split-table="true"] {
   display: none;
 }

packages/preview-engine/src/paged-table-handler.ts

@@ -9,8 +9,352 @@ interface PagedChunker {
   source: ParentNode;
 }
 
-interface PagedHandlerContext {
+function measuredColumnWidths(table: HTMLTableElement) {
-  chunker: PagedChunker;
+  const rows = Array.from(table.rows);
+  const columnCount = rows.reduce(
+    (maximum, row) => Math.max(
+      maximum,
+      Array.from(row.cells).reduce(
+        (total, cell) => total + Math.max(1, cell.colSpan),
+        0
+      )
+    ),
+    0
+  );
+  const tableRect = table.getBoundingClientRect();
+  if (columnCount < 1 || tableRect.width <= 0) {
+    return [];
+  }
+  const boundaries = Array.from(
+    { length: columnCount + 1 },
+    () => [] as number[]
+  );
+  boundaries[0]!.push(0);
+  boundaries[columnCount]!.push(tableRect.width);
+  for (const row of rows) {
+    let column = 0;
+    for (const cell of Array.from(row.cells)) {
+      const span = Math.max(1, cell.colSpan);
+      const rect = cell.getBoundingClientRect();
+      boundaries[column]?.push(
+        Math.max(0, Math.min(tableRect.width, rect.left - tableRect.left))
+      );
+      column = Math.min(columnCount, column + span);
+      boundaries[column]?.push(
+        Math.max(0, Math.min(tableRect.width, rect.right - tableRect.left))
+      );
+    }
+  }
+  const resolved = boundaries.map((samples, index) =>
+    samples.length > 0
+      ? samples.reduce((sum, value) => sum + value, 0) / samples.length
+      : tableRect.width * index / columnCount
+  );
+  resolved[0] = 0;
+  resolved[columnCount] = tableRect.width;
+  for (let index = 1; index < resolved.length; index += 1) {
+    resolved[index] = Math.max(resolved[index - 1]! + 0.01, resolved[index]!);
+  }
+  return resolved.slice(1).map(
+    (boundary, index) => boundary - resolved[index]!
+  );
+}
+
+function applyColumnWidthWeights(
+  columns: readonly HTMLTableColElement[],
+  weights: readonly number[]
+) {
+  for (const [index, column] of columns.entries()) {
+    column.style.width = `${Math.max(0, weights[index] ?? 0)}px`;
+  }
+}
+
+function applyCellWidthTracks(
+  table: HTMLTableElement,
+  widths: readonly number[]
+) {
+  for (const row of Array.from(table.rows)) {
+    let columnIndex = 0;
+    for (const cell of Array.from(row.cells)) {
+      const span = Math.max(1, cell.colSpan);
+      const width = widths
+        .slice(columnIndex, columnIndex + span)
+        .reduce((sum, value) => sum + value, 0);
+      if (width > 0) {
+        cell.style.width = `${width}px`;
+      }
+      columnIndex += span;
+    }
+  }
+}
+
+export function stabilizePagedTableColumns(
+  root: ParentNode,
+  pageContentHeightPx?: number
+) {
+  let stabilizedCount = 0;
+  for (const [tableIndex, table] of Array.from(
+    root.querySelectorAll<HTMLTableElement>("table")
+  ).entries()) {
+    let bodyRowIndex = 0;
+    if (pageContentHeightPx && pageContentHeightPx > 0) {
+      for (const row of Array.from(table.rows)) {
+        const rowHeight = row.getBoundingClientRect().height;
+        const keepWhole = rowHeight > 0 && rowHeight <= pageContentHeightPx;
+        if (row.parentElement?.tagName === "TBODY") {
+          row.dataset.mdtpTableRowId ??=
+            `table-${tableIndex + 1}-row-${bodyRowIndex + 1}`;
+          row.dataset.mdtpTableRowKeepWhole = String(keepWhole);
+          bodyRowIndex += 1;
+        }
+        for (const cell of Array.from(row.cells)) {
+          cell.style.setProperty(
+            "break-inside",
+            keepWhole ? "avoid" : "auto",
+            "important"
+          );
+          cell.style.setProperty(
+            "page-break-inside",
+            keepWhole ? "avoid" : "auto",
+            "important"
+          );
+        }
+      }
+    }
+    if (
+      table.dataset.stabilizedColumns === "true" ||
+      table.querySelector(":scope > colgroup")
+    ) {
+      continue;
+    }
+    const widths = measuredColumnWidths(table);
+    const total = widths.reduce((sum, value) => sum + value, 0);
+    if (widths.length === 0 || total <= 0) {
+      continue;
+    }
+    const colgroup = table.ownerDocument.createElement("colgroup");
+    colgroup.dataset.stabilizedColumns = "true";
+    for (const width of widths) {
+      const column = table.ownerDocument.createElement("col");
+      column.dataset.stabilizedWidthPx = String(width);
+      colgroup.append(column);
+    }
+    table.insertBefore(colgroup, table.firstChild);
+    table.dataset.stabilizedColumns = "true";
+    table.style.tableLayout = "fixed";
+    const columns = Array.from(colgroup.children) as HTMLTableColElement[];
+    let weights = widths;
+    applyColumnWidthWeights(columns, weights);
+    for (let iteration = 0; iteration < 3; iteration += 1) {
+      const measured = measuredColumnWidths(table);
+      if (
+        measured.length !== widths.length ||
+        measured.some((width) => width <= 0)
+      ) {
+        break;
+      }
+      weights = weights.map(
+        (weight, index) => weight * widths[index]! / measured[index]!
+      );
+      applyColumnWidthWeights(columns, weights);
+    }
+    applyCellWidthTracks(table, widths);
+    stabilizedCount += 1;
+  }
+  return stabilizedCount;
+}
+
+export function missingPagedTableRowIds(
+  source: ParentNode,
+  rendered: ParentNode
+) {
+  const isVisiblePagedRow = (row: HTMLElement) => {
+    const pageArea = row.closest<HTMLElement>(".pagedjs_area");
+    if (!pageArea) {
+      return true;
+    }
+    const rowRect = row.getBoundingClientRect();
+    const areaRect = pageArea.getBoundingClientRect();
+    const intersectionWidth = Math.min(rowRect.right, areaRect.right) -
+      Math.max(rowRect.left, areaRect.left);
+    const intersectionHeight = Math.min(rowRect.bottom, areaRect.bottom) -
+      Math.max(rowRect.top, areaRect.top);
+    return intersectionWidth > 0.5 && intersectionHeight > 0.5;
+  };
+  const renderedIds = new Set(
+    Array.from(rendered.querySelectorAll<HTMLElement>(
+      "[data-mdtp-table-row-id]"
+    )).filter(isVisiblePagedRow)
+      .map((row) => row.dataset.mdtpTableRowId)
+      .filter((id): id is string => Boolean(id))
+  );
+  return Array.from(source.querySelectorAll<HTMLTableRowElement>(
+    'tbody > tr[data-mdtp-table-row-id][data-mdtp-table-row-keep-whole="true"]'
+  )).map((row) => row.dataset.mdtpTableRowId)
+    .filter((id): id is string => Boolean(id))
+    .filter((id) => !renderedIds.has(id));
+}
+
+function splitPagedTableBeforeRow(row: HTMLTableRowElement) {
+  const body = row.parentElement as HTMLTableSectionElement | null;
+  const table = row.closest<HTMLTableElement>("table");
+  if (
+    !body ||
+    body.tagName !== "TBODY" ||
+    !table ||
+    body.querySelector(":scope > tr") === row
+  ) {
+    return table;
+  }
+
+  const continuation = table.cloneNode(false) as HTMLTableElement;
+  continuation.dataset.mdtpTableRecoveryContinuation = "true";
+  for (const child of Array.from(table.children)) {
+    if (child.tagName === "COLGROUP" || child.tagName === "THEAD") {
+      continuation.append(child.cloneNode(true));
+    }
+  }
+
+  const continuationBody = body.cloneNode(false) as HTMLTableSectionElement;
+  let movingRow: HTMLTableRowElement | null = row;
+  while (movingRow) {
+    const nextRow = movingRow.nextElementSibling as HTMLTableRowElement | null;
+    continuationBody.append(movingRow);
+    movingRow = nextRow;
+  }
+  continuation.append(continuationBody);
+
+  let followingSection = body.nextElementSibling;
+  while (followingSection) {
+    const nextSection = followingSection.nextElementSibling;
+    continuation.append(followingSection);
+    followingSection = nextSection;
+  }
+  table.after(continuation);
+  return continuation;
+}
+
+export function forcePagedTableRowsToNextPage(
+  source: ParentNode,
+  rowIds: readonly string[]
+) {
+  const requested = new Set(rowIds);
+  let count = 0;
+  for (const row of source.querySelectorAll<HTMLTableRowElement>(
+    "tbody > tr[data-mdtp-table-row-id]"
+  )) {
+    const id = row.dataset.mdtpTableRowId;
+    if (!id || !requested.has(id)) {
+      continue;
+    }
+    const table = splitPagedTableBeforeRow(row);
+    let recoveryTarget: HTMLElement = row;
+    if (table?.parentNode) {
+      const previous = table.previousElementSibling as HTMLElement | null;
+      const marker = previous?.dataset.mdtpTableRecoveryMarker === "true"
+        ? previous
+        : table.ownerDocument.createElement("div");
+      if (marker !== previous) {
+        marker.dataset.mdtpTableRecoveryMarker = "true";
+        marker.setAttribute("aria-hidden", "true");
+        marker.textContent = "\u00a0";
+        marker.style.setProperty("height", "1px", "important");
+        marker.style.setProperty("line-height", "1px", "important");
+        marker.style.setProperty("margin", "0 0 -1px", "important");
+        marker.style.setProperty("padding", "0", "important");
+        marker.style.setProperty("border", "0", "important");
+        marker.style.setProperty("overflow", "hidden", "important");
+        marker.style.setProperty("opacity", "0", "important");
+        table.parentNode.insertBefore(marker, table);
+      }
+      recoveryTarget = marker;
+      table.dataset.mdtpTableRecovery = "true";
+    }
+    recoveryTarget.style.setProperty("break-before", "page", "important");
+    recoveryTarget.style.setProperty(
+      "page-break-before",
+      "always",
+      "important"
+    );
+    row.dataset.mdtpTableRowRecovery = "true";
+    count += 1;
+  }
+  return count;
+}
+
+export function restorePagedTableStructure(
+  sourceTable: HTMLTableElement,
+  renderedTable: HTMLTableElement
+) {
+  restorePagedTableColumns(sourceTable, renderedTable);
+
+  if (!renderedTable.querySelector("thead")) {
+    const sourceHeader = sourceTable.querySelector("thead");
+    if (sourceHeader) {
+      const firstNonColumnGroup = Array.from(renderedTable.children).find(
+        (child) => child.tagName !== "COLGROUP"
+      ) ?? null;
+      renderedTable.insertBefore(
+        sourceHeader.cloneNode(true),
+        firstNonColumnGroup
+      );
+    }
+  }
+}
+
+export function restorePagedTableColumns(
+  sourceTable: HTMLTableElement,
+  renderedTable: HTMLTableElement
+) {
+  const sourceGroups = Array.from(
+    sourceTable.querySelectorAll<HTMLTableColElement>(
+      ":scope > colgroup"
+    )
+  );
+  if (sourceGroups.length === 0) {
+    return;
+  }
+  const renderedGroups = Array.from(
+    renderedTable.querySelectorAll<HTMLTableColElement>(
+      ":scope > colgroup"
+    )
+  );
+  const matchingStructure =
+    renderedGroups.length === sourceGroups.length &&
+    renderedGroups.every(
+      (group, index) =>
+        group.children.length === sourceGroups[index]?.children.length
+    );
+  if (!matchingStructure) {
+    for (const group of renderedGroups) {
+      group.remove();
+    }
+    const firstChild = renderedTable.firstChild;
+    for (const group of sourceGroups) {
+      renderedTable.insertBefore(group.cloneNode(true), firstChild);
+    }
+  } else {
+    for (const [groupIndex, sourceGroup] of sourceGroups.entries()) {
+      const renderedGroup = renderedGroups[groupIndex]!;
+      for (const attribute of Array.from(sourceGroup.attributes)) {
+        renderedGroup.setAttribute(attribute.name, attribute.value);
+      }
+      for (const [columnIndex, sourceColumn] of Array.from(
+        sourceGroup.children
+      ).entries()) {
+        const renderedColumn = renderedGroup.children[
+          columnIndex
+        ] as HTMLTableColElement;
+        for (const attribute of Array.from(sourceColumn.attributes)) {
+          renderedColumn.setAttribute(attribute.name, attribute.value);
+        }
+      }
+    }
+  }
+  if (sourceTable.dataset.stabilizedColumns === "true") {
+    renderedTable.dataset.stabilizedColumns = "true";
+    renderedTable.style.tableLayout = "fixed";
+  }
 }
 
 function elementAncestors(
@@ -48,6 +392,21 @@ class RepeatTableHeadersHandler extends Handler {
     breakToken?: PagedBreakToken
   ) {
     this.splitTableRefs = [];
+    for (const renderedTable of pageElement.querySelectorAll<HTMLTableElement>(
+      "table[data-ref]"
+    )) {
+      const ref = renderedTable.getAttribute("data-ref");
+      if (!ref) {
+        continue;
+      }
+      const sourceTable =
+        this.chunker.source.querySelector<HTMLTableElement>(
+          `table[data-ref="${CSS.escape(ref)}"]`
+        );
+      if (sourceTable) {
+        restorePagedTableColumns(sourceTable, renderedTable);
+      }
+    }
     const element = resolveBreakTokenElement(breakToken?.node);
     if (!element) {
       return;
@@ -81,6 +440,21 @@ class RepeatTableHeadersHandler extends Handler {
   }
 
   layout(rendered: HTMLElement) {
+    for (const renderedTable of rendered.querySelectorAll<HTMLTableElement>(
+      "table[data-ref]"
+    )) {
+      const ref = renderedTable.getAttribute("data-ref");
+      if (!ref) {
+        continue;
+      }
+      const sourceTable =
+        this.chunker.source.querySelector<HTMLTableElement>(
+          `table[data-ref="${CSS.escape(ref)}"]`
+        );
+      if (sourceTable) {
+        restorePagedTableColumns(sourceTable, renderedTable);
+      }
+    }
     for (const ref of this.splitTableRefs) {
       const renderedTable =
         rendered.querySelector<HTMLTableElement>(
@@ -101,23 +475,7 @@ class RepeatTableHeadersHandler extends Handler {
         continue;
       }
 
-      const firstChild = renderedTable.firstChild;
+      restorePagedTableStructure(sourceTable, renderedTable);
-      for (const colgroup of sourceTable.querySelectorAll("colgroup")) {
-        renderedTable.insertBefore(
-          colgroup.cloneNode(true),
-          firstChild
-        );
-      }
-
-      if (!renderedTable.querySelector("thead")) {
-        const sourceHeader = sourceTable.querySelector("thead");
-        if (sourceHeader) {
-          renderedTable.insertBefore(
-            sourceHeader.cloneNode(true),
-            renderedTable.firstChild
-          );
-        }
-      }
 
       renderedTable.setAttribute("data-repeated-header", "true");
     }

packages/preview-engine/tests/code-block-pagination.test.ts

@@ -98,6 +98,32 @@ describe("代码块分页预处理", () => {
     ).toBe("return");
   });
 
+  it("将高亮代码行首缩进封装为分页稳定节点", () => {
+    const root = createRoot(
+      '<span class="hljs-punctuation">{</span>\n  ' +
+        '<span class="hljs-attr">"nested"</span>: true\n' +
+        '<span class="hljs-punctuation">}</span>\n'
+    );
+
+    prepareCodeBlockPagination(root);
+
+    const lines = getLines(root);
+    const indent = lines[1]?.querySelector(
+      ":scope > .md-code-line-indent"
+    );
+    expect(lines.map((line) => line.textContent)).toEqual([
+      "{",
+      '  "nested": true',
+      "}"
+    ]);
+    expect(indent?.textContent).toBe("  ");
+    expect((indent as HTMLElement | null)?.style.width).toBe("2ch");
+    expect((indent as HTMLElement | null)?.dataset.codeIndentColumns)
+      .toBe("2");
+    expect(indent?.nextElementSibling?.classList.contains("hljs-attr"))
+      .toBe(true);
+  });
+
   it("重复调用不会再次包装已经准备的代码块", () => {
     const root = createRoot("line-1\nline-2\n");
 

packages/preview-engine/tests/docx-media-runtime.test.ts

@@ -2,13 +2,21 @@
 
 import { beforeEach, describe, expect, it, vi } from "vitest";
 import {
+  collectPagedDocxTableGeometries,
   collectDocxDocumentLayoutPlan,
   collectDocxInlineCodeLayouts,
   collectDocxListItemLayouts,
   collectDocxTextBlockLayouts,
   collectDocxMediaCaptureTargets,
+  forcePagedTableRowsToNextPage,
+  mergePagedTableGeometries,
+  missingPagedTableRowIds,
+  protectFittableInlineCodeParagraphs,
   removeInheritedPagedSplitJustification,
-  renderDocxMediaCapturePlan
+  renderDocxMediaCapturePlan,
+  restorePagedTableColumns,
+  restorePagedTableStructure,
+  stabilizePagedTableColumns
 } from "../src/index.js";
 
 describe("DOCX 媒体捕获计划", () => {
@@ -18,6 +26,10 @@ describe("DOCX 媒体捕获计划", () => {
   });
 
   it("连续布局使用原始打印媒体主题 CSS", async () => {
+    const render = vi.fn(async () => ({
+      echartsErrors: [],
+      mermaidErrors: []
+    }));
     const renderContinuous = vi.fn(async () => ({
       echartsErrors: [],
       mermaidErrors: []
@@ -26,7 +38,7 @@ describe("DOCX 媒体捕获计划", () => {
     root.innerHTML = '<article id="write"></article>';
 
     await renderDocxMediaCapturePlan(
-      { renderContinuous } as never,
+      { render, renderContinuous } as never,
       root,
       {
         articleHtml: '<article id="write"></article>',
@@ -57,6 +69,155 @@ describe("DOCX 媒体捕获计划", () => {
       expect.anything(),
       expect.objectContaining({ themeMedia: "print" })
     );
+    expect(render).toHaveBeenCalledWith(
+      expect.anything(),
+      {
+        target: "pdf",
+        allowIncompleteTableGeometry: true
+      }
+    );
+  });
+
+  it("仅将能完整放入一页的顶层行内代码段落保持为整体", () => {
+    document.body.innerHTML = `
+      <article id="write">
+        <p id="short" class="md-inline-code-paragraph">短段落</p>
+        <p id="long" class="md-inline-code-paragraph">超长段落</p>
+        <table><tbody><tr><td>
+          <p id="cell" class="md-inline-code-paragraph">单元格段落</p>
+        </td></tr></tbody></table>
+      </article>
+    `;
+    const short = document.querySelector<HTMLElement>("#short")!;
+    const long = document.querySelector<HTMLElement>("#long")!;
+    const cell = document.querySelector<HTMLElement>("#cell")!;
+    short.getBoundingClientRect = () => ({ height: 240 }) as DOMRect;
+    long.getBoundingClientRect = () => ({ height: 1200 }) as DOMRect;
+    cell.getBoundingClientRect = () => ({ height: 120 }) as DOMRect;
+
+    expect(protectFittableInlineCodeParagraphs(document, 900)).toBe(1);
+    expect(short.style.getPropertyValue("break-inside")).toBe("avoid");
+    expect(short.style.getPropertyPriority("break-inside")).toBe("important");
+    expect(short.dataset.mdtpInlineCodeKeepWhole).toBe("true");
+    expect(long.style.getPropertyValue("break-inside")).toBe("");
+    expect(cell.style.getPropertyValue("break-inside")).toBe("");
+  });
+
+  it("按逻辑表格去重采集分页后的真实列轨", () => {
+    document.body.innerHTML = `
+      <main class="pagedjs_pages">
+        <section class="pagedjs_page">
+          <div class="pagedjs_area">
+            <table data-ref="table-a">
+              <colgroup data-stabilized-columns="true">
+                <col data-stabilized-width-px="120">
+                <col data-stabilized-width-px="280">
+              </colgroup>
+              <tbody><tr><td>A</td><td>B</td></tr></tbody>
+            </table>
+          </div>
+        </section>
+        <section class="pagedjs_page">
+          <div class="pagedjs_area">
+            <table data-ref="table-a">
+              <colgroup data-stabilized-columns="true">
+                <col data-stabilized-width-px="120">
+                <col data-stabilized-width-px="280">
+              </colgroup>
+              <tbody><tr><td>C</td><td>D</td></tr></tbody>
+            </table>
+          </div>
+        </section>
+      </main>
+    `;
+    const areas = Array.from(
+      document.querySelectorAll<HTMLElement>(".pagedjs_area")
+    );
+    const tables = Array.from(
+      document.querySelectorAll<HTMLTableElement>("table")
+    );
+    areas.forEach((area) => {
+      area.getBoundingClientRect = () => ({
+        left: 100, right: 900, top: 0, bottom: 900,
+        width: 800, height: 900
+      }) as DOMRect;
+    });
+    tables.forEach((table, index) => {
+      table.getBoundingClientRect = () => ({
+        left: index === 0 ? 900 : 140,
+        right: index === 0 ? 1300 : 540,
+        top: 20, bottom: 220,
+        width: 400, height: 200
+      }) as DOMRect;
+    });
+
+    expect(collectPagedDocxTableGeometries(document, {
+      contentWidthPx: 800,
+      contentHeightPx: 900
+    })).toEqual([{
+      ordinal: 1,
+      widthPercent: 50,
+      leftOffsetPercent: 5,
+      columnWidthPercents: [30, 70]
+    }]);
+  });
+
+  it("只覆盖列数一致的分页表格几何并保留连续布局行样式", () => {
+    const row = {
+      cells: [{
+        columnSpan: 1,
+        backgroundColor: "#ffffff",
+        color: "#000000",
+        bold: true,
+        italic: false,
+        alignment: "left" as const
+      }]
+    };
+    const layout = {
+      tables: [
+        {
+          ordinal: 1,
+          widthPercent: 100,
+          leftOffsetPercent: 0,
+          columnWidthPercents: [40, 60],
+          rows: [row]
+        },
+        {
+          ordinal: 2,
+          widthPercent: 100,
+          leftOffsetPercent: 0,
+          columnWidthPercents: [100],
+          rows: [row]
+        }
+      ],
+      textBlocks: [],
+      listItems: [],
+      inlineCodes: []
+    };
+
+    const merged = mergePagedTableGeometries(layout, [
+      {
+        ordinal: 1,
+        widthPercent: 90,
+        leftOffsetPercent: 5,
+        columnWidthPercents: [30, 70]
+      },
+      {
+        ordinal: 2,
+        widthPercent: 80,
+        leftOffsetPercent: 10,
+        columnWidthPercents: [20, 80]
+      }
+    ]);
+
+    expect(merged.tables[0]).toEqual({
+      ...layout.tables[0],
+      widthPercent: 90,
+      leftOffsetPercent: 5,
+      columnWidthPercents: [30, 70]
+    });
+    expect(merged.tables[0]?.rows).toBe(layout.tables[0]?.rows);
+    expect(merged.tables[1]).toBe(layout.tables[1]);
   });
 
   it("只移除会把 Paged.js 末行两端对齐继承给子块的容器标记", () => {
@@ -334,6 +495,382 @@ describe("DOCX 媒体捕获计划", () => {
     });
   });
 
+  it("保留主题表格超出文档内容盒的实测宽度", () => {
+    document.body.innerHTML = `
+      <article id="write"><table><tbody><tr><td>内容</td></tr></tbody></table></article>
+    `;
+    const article = document.querySelector<HTMLElement>("#write")!;
+    const table = document.querySelector<HTMLTableElement>("table")!;
+    const cell = document.querySelector<HTMLTableCellElement>("td")!;
+    article.getBoundingClientRect = () => ({
+      left: 100,
+      right: 900,
+      top: 0,
+      bottom: 900,
+      width: 800,
+      height: 900,
+    }) as DOMRect;
+    table.getBoundingClientRect = () => ({
+      left: 100,
+      right: 980,
+      top: 20,
+      bottom: 120,
+      width: 880,
+      height: 100,
+    }) as DOMRect;
+    cell.getBoundingClientRect = table.getBoundingClientRect;
+
+    const layout = collectDocxDocumentLayoutPlan(document, {
+      contentWidthPx: 800,
+      contentHeightPx: 900,
+    });
+
+    expect(layout.tables[0]?.widthPercent).toBeCloseTo(110, 8);
+    expect(layout.tables[0]?.columnWidthPercents).toEqual([100]);
+  });
+
+  it("分页前冻结整表实测列宽供拆分页片段复用", () => {
+    document.body.innerHTML = `
+      <article id="write"><table><tbody><tr><td>A</td><td>B</td></tr></tbody></table></article>
+    `;
+    const table = document.querySelector<HTMLTableElement>("table")!;
+    const cells = Array.from(table.rows[0]!.cells);
+    table.getBoundingClientRect = () => ({
+      left: 100,
+      right: 500,
+      top: 0,
+      bottom: 100,
+      width: 400,
+      height: 100
+    }) as DOMRect;
+    cells[0]!.getBoundingClientRect = () => ({
+      left: 100,
+      right: 220,
+      top: 0,
+      bottom: 100,
+      width: 120,
+      height: 100
+    }) as DOMRect;
+    cells[1]!.getBoundingClientRect = () => ({
+      left: 220,
+      right: 500,
+      top: 0,
+      bottom: 100,
+      width: 280,
+      height: 100
+    }) as DOMRect;
+
+    expect(stabilizePagedTableColumns(document)).toBe(1);
+    expect(table.dataset.stabilizedColumns).toBe("true");
+    expect(table.style.tableLayout).toBe("fixed");
+    expect(Array.from(table.querySelectorAll("col")).map(
+      (column) => column.style.width
+    )).toEqual(["120px", "280px"]);
+    expect(Array.from(table.querySelectorAll("col")).map(
+      (column) => (column as HTMLTableColElement).dataset.stabilizedWidthPx
+    )).toEqual(["120", "280"]);
+    expect(Array.from(table.querySelectorAll("th, td")).map(
+      (cell) => (cell as HTMLElement).style.width
+    )).toEqual(["120px", "280px"]);
+    expect(stabilizePagedTableColumns(document)).toBe(0);
+  });
+
+  it("DOCX 布局采集保持分页前原始列轨而不受辅助样式二次测量影响", () => {
+    document.body.innerHTML = `
+      <article id="write">
+        <table data-stabilized-columns="true">
+          <colgroup data-stabilized-columns="true">
+            <col data-stabilized-width-px="120" style="width: 140px">
+            <col data-stabilized-width-px="280" style="width: 260px">
+          </colgroup>
+          <tbody><tr><td>A</td><td>B</td></tr></tbody>
+        </table>
+      </article>
+    `;
+    const article = document.querySelector<HTMLElement>("#write")!;
+    const table = document.querySelector<HTMLTableElement>("table")!;
+    const cells = Array.from(table.rows[0]!.cells);
+    article.getBoundingClientRect = () => ({
+      left: 100, right: 500, top: 0, bottom: 100,
+      width: 400, height: 100
+    }) as DOMRect;
+    table.getBoundingClientRect = article.getBoundingClientRect;
+    cells[0]!.getBoundingClientRect = () => ({
+      left: 100, right: 240, top: 0, bottom: 100,
+      width: 140, height: 100
+    }) as DOMRect;
+    cells[1]!.getBoundingClientRect = () => ({
+      left: 240, right: 500, top: 0, bottom: 100,
+      width: 260, height: 100
+    }) as DOMRect;
+
+    const layout = collectDocxDocumentLayoutPlan(document, {
+      contentWidthPx: 400,
+      contentHeightPx: 900
+    });
+
+    expect(layout.tables[0]?.columnWidthPercents).toEqual([30, 70]);
+  });
+
+  it("冻结跨列单元格为对应列轨宽度之和", () => {
+    document.body.innerHTML = `
+      <table><tbody>
+        <tr><td colspan="2">跨列</td></tr>
+        <tr><td>甲</td><td>乙</td></tr>
+      </tbody></table>
+    `;
+    const table = document.querySelector("table")!;
+    const cells = Array.from(table.querySelectorAll("td"));
+    table.getBoundingClientRect = () => ({
+      left: 100, right: 500, top: 0, bottom: 100,
+      width: 400, height: 100
+    }) as DOMRect;
+    cells[0]!.getBoundingClientRect = () => ({
+      left: 100, right: 500, top: 0, bottom: 50,
+      width: 400, height: 50
+    }) as DOMRect;
+    cells[1]!.getBoundingClientRect = () => ({
+      left: 100, right: 220, top: 50, bottom: 100,
+      width: 120, height: 50
+    }) as DOMRect;
+    cells[2]!.getBoundingClientRect = () => ({
+      left: 220, right: 500, top: 50, bottom: 100,
+      width: 280, height: 50
+    }) as DOMRect;
+
+    expect(stabilizePagedTableColumns(document)).toBe(1);
+    expect(cells.map((cell) => cell.style.width)).toEqual([
+      "400px", "120px", "280px"
+    ]);
+  });
+
+  it("分页片段已有冻结列定义时不重复插入 colgroup", () => {
+    document.body.innerHTML = `
+      <table id="source">
+        <colgroup data-stabilized-columns="true"><col style="width: 30%"><col style="width: 70%"></colgroup>
+        <thead><tr><th>A</th><th>B</th></tr></thead>
+        <tbody><tr><td>甲</td><td>乙</td></tr></tbody>
+      </table>
+      <table id="rendered">
+        <colgroup data-stabilized-columns="true"><col style="width: 30%"><col style="width: 70%"></colgroup>
+        <tbody><tr><td>丙</td><td>丁</td></tr></tbody>
+      </table>
+    `;
+    const source = document.querySelector<HTMLTableElement>("#source")!;
+    const rendered = document.querySelector<HTMLTableElement>("#rendered")!;
+
+    restorePagedTableStructure(source, rendered);
+
+    expect(rendered.querySelectorAll(":scope > colgroup")).toHaveLength(1);
+    expect(rendered.querySelectorAll(":scope > colgroup > col")).toHaveLength(2);
+    expect(rendered.querySelectorAll(":scope > thead")).toHaveLength(1);
+    expect(Array.from(rendered.children).map((child) => child.tagName)).toEqual([
+      "COLGROUP",
+      "THEAD",
+      "TBODY"
+    ]);
+  });
+
+  it("分页前仅对可容纳于单页的表格行单元格启用整行保护", () => {
+    document.body.innerHTML = `
+      <table><tbody>
+        <tr id="normal"><td>普通行</td><td>内容</td></tr>
+        <tr id="tall"><td>超高行</td><td>内容</td></tr>
+      </tbody></table>
+    `;
+    const table = document.querySelector<HTMLTableElement>("table")!;
+    const normal = document.querySelector<HTMLTableRowElement>("#normal")!;
+    const tall = document.querySelector<HTMLTableRowElement>("#tall")!;
+    table.getBoundingClientRect = () => ({
+      width: 400, height: 1020, top: 0, right: 400,
+      bottom: 1020, left: 0
+    }) as DOMRect;
+    normal.getBoundingClientRect = () => ({
+      width: 400, height: 120, top: 0, right: 400,
+      bottom: 120, left: 0
+    }) as DOMRect;
+    tall.getBoundingClientRect = () => ({
+      width: 400, height: 900, top: 120, right: 400,
+      bottom: 1020, left: 0
+    }) as DOMRect;
+    Array.from(table.rows).flatMap((row) => Array.from(row.cells))
+      .forEach((cell, index) => {
+        cell.getBoundingClientRect = () => ({
+          width: 200,
+          height: index < 2 ? 120 : 900,
+          top: index < 2 ? 0 : 120,
+          right: index % 2 === 0 ? 200 : 400,
+          bottom: index < 2 ? 120 : 1020,
+          left: index % 2 === 0 ? 0 : 200
+        }) as DOMRect;
+      });
+
+    stabilizePagedTableColumns(document, 800);
+
+    expect(Array.from(normal.cells).map(
+      (cell) => cell.style.getPropertyValue("break-inside")
+    )).toEqual(["avoid", "avoid"]);
+    expect(Array.from(tall.cells).map(
+      (cell) => cell.style.getPropertyValue("break-inside")
+    )).toEqual(["auto", "auto"]);
+    expect(Array.from(normal.cells).every(
+      (cell) => cell.style.getPropertyPriority("break-inside") === "important"
+    )).toBe(true);
+    expect(normal.dataset.mdtpTableRowId).toBe("table-1-row-1");
+    expect(normal.dataset.mdtpTableRowKeepWhole).toBe("true");
+    expect(tall.dataset.mdtpTableRowId).toBe("table-1-row-2");
+    expect(tall.dataset.mdtpTableRowKeepWhole).toBe("false");
+  });
+
+  it("检测分页后缺失的可容纳正文行并施加恢复断点", () => {
+    document.body.innerHTML = `
+      <main id="source"><table><tbody>
+        <tr data-mdtp-table-row-id="row-1" data-mdtp-table-row-keep-whole="true"><td>甲</td></tr>
+        <tr data-mdtp-table-row-id="row-2" data-mdtp-table-row-keep-whole="true"><td>乙</td></tr>
+        <tr data-mdtp-table-row-id="row-tall" data-mdtp-table-row-keep-whole="false"><td>超高</td></tr>
+      </tbody></table></main>
+      <main id="rendered"><table><tbody>
+        <tr data-mdtp-table-row-id="row-2"><td>乙</td></tr>
+      </tbody></table></main>
+    `;
+    const source = document.querySelector("#source")!;
+    const rendered = document.querySelector("#rendered")!;
+
+    expect(missingPagedTableRowIds(source, rendered)).toEqual(["row-1"]);
+    expect(forcePagedTableRowsToNextPage(source, ["row-1"])).toBe(1);
+    const row = source.querySelector<HTMLTableRowElement>(
+      '[data-mdtp-table-row-id="row-1"]'
+    )!;
+    const table = row.closest("table")!;
+    const marker = table.previousElementSibling as HTMLElement;
+    expect(marker.dataset.mdtpTableRecoveryMarker).toBe("true");
+    expect(marker.textContent).toBe("\u00a0");
+    expect(marker.style.getPropertyValue("height")).toBe("1px");
+    expect(marker.style.getPropertyValue("margin-bottom")).toBe("-1px");
+    expect(marker.style.getPropertyValue("opacity")).toBe("0");
+    expect(marker.style.getPropertyValue("break-before")).toBe("page");
+    expect(marker.style.getPropertyPriority("break-before")).toBe("important");
+    expect(table.dataset.mdtpTableRecovery).toBe("true");
+    expect(row.dataset.mdtpTableRowRecovery).toBe("true");
+    expect(forcePagedTableRowsToNextPage(source, ["row-1"])).toBe(1);
+    expect(source.querySelectorAll("[data-mdtp-table-recovery-marker]"))
+      .toHaveLength(1);
+
+  });
+
+  it("在非首行缺失时拆分续表并将恢复断点放到表格外", () => {
+    document.body.innerHTML = `
+      <main id="source"><table id="source-table">
+        <colgroup><col><col></colgroup>
+        <thead><tr><th>甲</th><th>乙</th></tr></thead>
+        <tbody>
+          <tr data-mdtp-table-row-id="row-1"><td>一</td><td>壹</td></tr>
+          <tr data-mdtp-table-row-id="row-2"><td>二</td><td>贰</td></tr>
+          <tr data-mdtp-table-row-id="row-3"><td>三</td><td>叁</td></tr>
+        </tbody>
+        <tfoot><tr><td>尾</td><td>末</td></tr></tfoot>
+      </table></main>
+    `;
+    const source = document.querySelector("#source")!;
+
+    expect(forcePagedTableRowsToNextPage(source, ["row-2"])).toBe(1);
+
+    const tables = source.querySelectorAll<HTMLTableElement>("table");
+    expect(tables).toHaveLength(2);
+    expect(Array.from(tables[0]!.querySelectorAll("tbody > tr")).map(
+      (row) => (row as HTMLElement).dataset.mdtpTableRowId
+    )).toEqual(["row-1"]);
+    expect(Array.from(tables[1]!.querySelectorAll("tbody > tr")).map(
+      (row) => (row as HTMLElement).dataset.mdtpTableRowId
+    )).toEqual(["row-2", "row-3"]);
+    expect(tables[1]!.dataset.mdtpTableRecoveryContinuation).toBe("true");
+    expect(tables[1]!.querySelector("colgroup")).not.toBeNull();
+    expect(tables[1]!.querySelector("thead")).not.toBeNull();
+    expect(tables[0]!.querySelector("tfoot")).toBeNull();
+    expect(tables[1]!.querySelector("tfoot")).not.toBeNull();
+    const marker = tables[1]!.previousElementSibling as HTMLElement;
+    expect(marker.dataset.mdtpTableRecoveryMarker).toBe("true");
+    expect(marker.style.getPropertyValue("break-before")).toBe("page");
+    expect(marker.style.getPropertyPriority("break-before")).toBe("important");
+  });
+
+  it("仅将分页可视区域内真实可见的表格行视为已渲染", () => {
+    document.body.innerHTML = `
+      <main id="source"><table><tbody>
+        <tr data-mdtp-table-row-id="row-clipped" data-mdtp-table-row-keep-whole="true"><td>裁切</td></tr>
+        <tr data-mdtp-table-row-id="row-visible" data-mdtp-table-row-keep-whole="true"><td>可见</td></tr>
+      </tbody></table></main>
+      <main id="rendered">
+        <section class="pagedjs_area" id="page-1">
+          <table><tbody>
+            <tr id="clipped" data-mdtp-table-row-id="row-clipped"><td>裁切</td></tr>
+            <tr id="visible" data-mdtp-table-row-id="row-visible"><td>可见</td></tr>
+          </tbody></table>
+        </section>
+        <section class="pagedjs_area" id="page-2">
+          <table><tbody>
+            <tr id="visible-duplicate" data-mdtp-table-row-id="row-visible"><td>可见副本</td></tr>
+          </tbody></table>
+        </section>
+      </main>
+    `;
+    const rect = (
+      left: number,
+      top: number,
+      right: number,
+      bottom: number
+    ) => ({
+      left, top, right, bottom,
+      width: right - left,
+      height: bottom - top,
+      x: left,
+      y: top,
+      toJSON: () => ({})
+    }) as DOMRect;
+    document.querySelector<HTMLElement>("#page-1")!.getBoundingClientRect =
+      () => rect(0, 0, 600, 800);
+    document.querySelector<HTMLElement>("#page-2")!.getBoundingClientRect =
+      () => rect(0, 900, 600, 1700);
+    document.querySelector<HTMLElement>("#clipped")!.getBoundingClientRect =
+      () => rect(0, 810, 600, 850);
+    document.querySelector<HTMLElement>("#visible")!.getBoundingClientRect =
+      () => rect(0, 780, 600, 820);
+    document.querySelector<HTMLElement>("#visible-duplicate")!
+      .getBoundingClientRect = () => rect(0, 920, 600, 960);
+
+    expect(missingPagedTableRowIds(
+      document.querySelector("#source")!,
+      document.querySelector("#rendered")!
+    )).toEqual(["row-clipped"]);
+  });
+
+  it("分页片段同步源表冻结列宽而不重复 colgroup", () => {
+    document.body.innerHTML = `
+      <table id="source" data-stabilized-columns="true">
+        <colgroup data-stabilized-columns="true"><col data-stabilized-width-px="118" style="width: 120px"><col data-stabilized-width-px="282" style="width: 280px"></colgroup>
+        <tbody><tr><td>甲</td><td>乙</td></tr></tbody>
+      </table>
+      <table id="rendered">
+        <colgroup><col style="width: 100px"><col style="width: 300px"></colgroup>
+        <tbody><tr><td>甲</td><td>乙</td></tr></tbody>
+      </table>
+    `;
+    const source = document.querySelector<HTMLTableElement>("#source")!;
+    const rendered = document.querySelector<HTMLTableElement>("#rendered")!;
+
+    restorePagedTableColumns(source, rendered);
+
+    expect(rendered.querySelectorAll(":scope > colgroup")).toHaveLength(1);
+    expect(Array.from(rendered.querySelectorAll("col")).map(
+      (column) => column.style.width
+    )).toEqual(["120px", "280px"]);
+    expect(Array.from(rendered.querySelectorAll("col")).map(
+      (column) => (column as HTMLTableColElement).dataset.stabilizedWidthPx
+    )).toEqual(["118", "282"]);
+    expect(rendered.dataset.stabilizedColumns).toBe("true");
+    expect(rendered.style.tableLayout).toBe("fixed");
+  });
+
   it("按真实上下文采集行内代码字号、颜色和盒模型", () => {
     document.body.innerHTML = `
       <article id="write" style="font-size: 20px">
@@ -419,6 +956,42 @@ describe("DOCX 媒体捕获计划", () => {
     }
   });
 
+  it("不把普通两端对齐列表的自然末行误判为分散对齐", () => {
+    document.body.innerHTML = `
+      <article id="write">
+        <ul><li style="text-align: justify; text-align-last: left">末行保持左对齐</li></ul>
+      </article>
+    `;
+    const article = document.querySelector<HTMLElement>("#write")!;
+    article.getBoundingClientRect = () => ({
+      left: 0,
+      right: 800,
+      top: 0,
+      bottom: 900,
+      width: 800,
+      height: 900
+    }) as DOMRect;
+    const rangePrototype = Object.getPrototypeOf(document.createRange()) as {
+      getBoundingClientRect?: () => DOMRect;
+    };
+    const original = rangePrototype.getBoundingClientRect;
+    rangePrototype.getBoundingClientRect = function (this: Range) {
+      return {
+        left: this.startOffset * 80,
+        right: this.startOffset * 80 + 10,
+        top: 20,
+        bottom: 32,
+        width: 10,
+        height: 12
+      } as DOMRect;
+    };
+    try {
+      expect(collectDocxListItemLayouts(article)[0]?.alignment).toBeUndefined();
+    } finally {
+      rangePrototype.getBoundingClientRect = original;
+    }
+  });
+
   it("采集正文文本块在 Chromium 中的实际行断点", () => {
     document.body.innerHTML = `
       <article id="write">

packages/preview-engine/tests/paged-preview.test.ts

@@ -40,6 +40,14 @@ const payload: PagedPreviewPayload = {
 };
 
 describe("分页预览协议", () => {
+  it("含行内代码的段落避免被 Paged.js 跨页拆分丢失片段", () => {
+    const css = buildPagedMediaCss(defaultExportConfig, payload);
+
+    expect(css).toContain("#write > p.md-inline-code-paragraph");
+    expect(css).toContain("orphans: 1");
+    expect(css).toContain("widows: 1");
+  });
+
   it("生成真实纸张尺寸和页边距 CSS", () => {
     const css = buildPagedMediaCss(defaultExportConfig, payload);
 
@@ -278,6 +286,12 @@ describe("分页预览协议", () => {
     expect(documentBaseCss).toContain("white-space: pre-wrap;");
     expect(documentBaseCss).toContain("overflow-wrap: anywhere;");
     expect(documentBaseCss).toContain("word-break: break-word;");
+    expect(documentBaseCss).toMatch(
+      /#write th,[\s\S]*#write td \{[\s\S]*word-break: break-word;/u,
+    );
+    expect(documentBaseCss).toMatch(
+      /#write th code,[\s\S]*#write td code \{[\s\S]*white-space: pre-wrap;[\s\S]*overflow-wrap: anywhere;[\s\S]*word-break: break-word;/u,
+    );
 
     const css = buildPagedMediaCss(defaultExportConfig, payload);
     expect(css).toContain("#write .md-code-line");

packages/renderer/src/render-markdown.ts

@@ -23,6 +23,7 @@ import hljs from "highlight.js";
 import MarkdownIt from "markdown-it";
 import type { Options as MarkdownItOptions } from "markdown-it";
 import type Renderer from "markdown-it/lib/renderer.mjs";
+import type StateInline from "markdown-it/lib/rules_inline/state_inline.mjs";
 import type Token from "markdown-it/lib/token.mjs";
 import sanitizeHtml from "sanitize-html";
 import { renderDocumentStructure } from "./render-document-structure.js";
@@ -49,6 +50,7 @@ export interface RenderMarkdownOptions {
 export interface RenderedMarkdown
   extends Omit<RenderedMarkdownDocument, "rendererVersion"> {
   rendererVersion: typeof RENDERER_VERSION;
+  markdownBody: string;
 }
 
 const markdownOptions: MarkdownItOptions = {
@@ -93,6 +95,104 @@ const markdown = new MarkdownIt(markdownOptions)
     trust: false
   });
 
+const tableBreakMarkupPattern = /^(?:<br>|<br\/>|<br \/>)/iu;
+
+function tableBreakCandidateRule(
+  state: StateInline,
+  silent: boolean
+): boolean {
+  const match = state.src.slice(state.pos).match(tableBreakMarkupPattern);
+  if (!match) {
+    return false;
+  }
+  if (!silent) {
+    const token = state.push("table_break_candidate", "", 0);
+    token.content = match[0];
+  }
+  state.pos += match[0].length;
+  return true;
+}
+
+markdown.inline.ruler.before(
+  "html_inline",
+  "table_break_candidate",
+  tableBreakCandidateRule
+);
+markdown.core.ruler.after("inline", "scope_table_breaks", (state) => {
+  let tableCellDepth = 0;
+  for (const token of state.tokens) {
+    if (token.type === "td_open" || token.type === "th_open") {
+      tableCellDepth += 1;
+      continue;
+    }
+    if (token.type === "td_close" || token.type === "th_close") {
+      tableCellDepth = Math.max(0, tableCellDepth - 1);
+      continue;
+    }
+    if (token.type !== "inline") {
+      continue;
+    }
+    for (const child of token.children ?? []) {
+      if (child.type === "table_break_candidate" && tableCellDepth > 0) {
+        child.type = "table_break";
+      }
+    }
+  }
+});
+markdown.renderer.rules.table_break_candidate = (tokens, index) =>
+  markdown.utils.escapeHtml(tokens[index]?.content ?? "");
+markdown.renderer.rules.table_break = () => "<br>";
+
+const markdownAlertTitles = new Map([
+  ["note", "Note"],
+  ["tip", "Tip"],
+  ["important", "Important"],
+  ["warning", "Warning"],
+  ["caution", "Caution"]
+]);
+
+markdown.core.ruler.after("scope_table_breaks", "github_alerts", (state) => {
+  for (let index = 0; index < state.tokens.length - 2; index += 1) {
+    const blockquote = state.tokens[index];
+    const paragraph = state.tokens[index + 1];
+    const inline = state.tokens[index + 2];
+    if (
+      blockquote?.type !== "blockquote_open" ||
+      paragraph?.type !== "paragraph_open" ||
+      inline?.type !== "inline"
+    ) {
+      continue;
+    }
+    const [marker, bodyBreak] = inline.children ?? [];
+    const match = marker?.type === "text"
+      ? marker.content.match(/^\[!(NOTE|TIP|IMPORTANT|WARNING|CAUTION)\]$/iu)
+      : undefined;
+    if (!marker || !match) {
+      continue;
+    }
+    const alertType = match[1]?.toLowerCase();
+    const title = alertType ? markdownAlertTitles.get(alertType) : undefined;
+    if (!alertType || !title) {
+      continue;
+    }
+    blockquote.attrJoin("class", `md-alert md-alert-${alertType}`);
+    marker.type = "markdown_alert_title";
+    marker.content = title;
+    marker.meta = { alertType };
+    if (bodyBreak?.type === "softbreak") {
+      bodyBreak.type = "markdown_alert_body_break";
+    }
+  }
+});
+markdown.renderer.rules.markdown_alert_title = (tokens, index) => {
+  const token = tokens[index];
+  const alertType = typeof token?.meta?.alertType === "string"
+    ? token.meta.alertType
+    : "note";
+  return `<span class="md-alert-text md-alert-text-${escapeAttribute(alertType)}">${markdown.utils.escapeHtml(token?.content ?? "")}</span>`;
+};
+markdown.renderer.rules.markdown_alert_body_break = () => "</p>\n<p>";
+
 const defaultValidateLink = markdown.validateLink.bind(markdown);
 markdown.validateLink = (href) =>
   defaultValidateLink(href) ||
@@ -141,6 +241,13 @@ markdown.renderer.rules.paragraph_open = (
   if (standaloneImage) {
     return '<figure class="md-document-image-block">\n';
   }
+  const inlineToken = tokens[index + 1];
+  if (
+    inlineToken?.type === "inline" &&
+    (inlineToken.children ?? []).some((child) => child.type === "code_inline")
+  ) {
+    tokens[index]?.attrJoin("class", "md-inline-code-paragraph");
+  }
   return defaultParagraphOpenRenderer
     ? defaultParagraphOpenRenderer(
         tokens,
@@ -367,6 +474,7 @@ export function renderMarkdown(
 
   return {
     rendererVersion: RENDERER_VERSION,
+    markdownBody: parsed.content,
     articleHtml,
     bodyHtml,
     metadata,

packages/renderer/tests/render-markdown.test.ts

@@ -6,6 +6,38 @@ import {
 } from "../src/render-markdown.js";
 
 describe("renderMarkdown", () => {
+  it("将 GFM Alert 渲染为与 Pandoc 一致的独立标题引用块", () => {
+    const result = renderMarkdown(
+      "> [!CAUTION]\n> **关键约束**：必须执行。",
+    );
+
+    expect(result.bodyHtml).toContain(
+      '<blockquote class="md-alert md-alert-caution">',
+    );
+    expect(result.bodyHtml).toContain(
+      '<p><span class="md-alert-text md-alert-text-caution">Caution</span></p>\n<p><strong>关键约束</strong>：必须执行。</p>',
+    );
+    expect(result.bodyHtml).not.toContain("[!CAUTION]");
+  });
+
+  it("为含行内代码的普通段落标记分页保护类", () => {
+    const result = renderMarkdown("段落前 `inline_code` 段落后");
+
+    expect(result.articleHtml).toContain(
+      '<p class="md-inline-code-paragraph">段落前 <code>inline_code</code> 段落后</p>',
+    );
+  });
+
+  it("向下游暴露已剥离文首 Front Matter 的正文 Markdown", () => {
+    const rendered = renderMarkdown(
+      "---\ntitle: 测试标题\n---\n# 正文\n\n---\n\n后续内容"
+    );
+
+    expect(rendered.markdownBody).toBe(
+      "# 正文\n\n---\n\n后续内容"
+    );
+    expect(rendered.metadata.title).toBe("测试标题");
+  });
   it("渲染常用 Markdown 扩展并识别能力", () => {
     const result = renderMarkdown(`
 # 示例文档
@@ -80,6 +112,39 @@ const answer = 42;
     expect(result.bodyHtml).toContain('<td style="text-align:right">C</td>');
   });
 
+  it("只在 GFM 表格单元格内解释无属性 br 换行", () => {
+    const result = renderMarkdown(`
+| 场景 | 内容 |
+| --- | --- |
+| 标准 | 第一行<br>第二行<BR/>第三行<br />第四行 |
+
+正文中的 <br>、<br/> 和 <br /> 保持文本。
+
+| 转义与代码 | 内容 |
+| --- | --- |
+| 转义 | \\<br> 与 &lt;br&gt; |
+| 代码 | \`<br>\` |
+| 属性 | <br class="unsafe"> |
+
+\`\`\`html
+<br>
+\`\`\`
+`);
+
+    expect(result.bodyHtml).toContain(
+      "第一行<br />第二行<br />第三行<br />第四行"
+    );
+    expect(result.bodyHtml).toContain(
+      "正文中的 &lt;br&gt;、&lt;br/&gt; 和 &lt;br /&gt; 保持文本"
+    );
+    expect(result.bodyHtml).toContain("&lt;br&gt; 与 &lt;br&gt;");
+    expect(result.bodyHtml).toContain("<code>&lt;br&gt;</code>");
+    expect(result.bodyHtml).toContain("&lt;br class=\"unsafe\"&gt;");
+    expect(result.bodyHtml).toContain(
+      '<pre class="md-fences" lang="html"><code class="language-html">'
+    );
+  });
+
   it("读取并规范化 Front Matter 元数据", () => {
     const result = renderMarkdown(`---
 title: 项目报告

scripts/build-font-packs.mjs

@@ -11,7 +11,7 @@ const repositoryRoot = resolve(dirname(fileURLToPath(import.meta.url)), "..");
 const bundlePath = join(repositoryRoot, "font-packs", "bundle.json");
 const bundle = JSON.parse(await readFile(bundlePath, "utf8"));
 const outputRoot = join(repositoryRoot, "output", "font-packs", "root");
-const appVersion = process.env.npm_package_version || "0.6.1";
+const appVersion = process.env.npm_package_version || "0.6.2";
 
 if (
   bundle.schemaVersion !== 1 ||

scripts/docx-layout-visual-matrix-cases.mjs

@@ -71,6 +71,29 @@ function hasIndependentCover(manifest) {
   return profiles.includes("project-report") || profiles.includes("tender");
 }
 
+export function resolveEffectiveCoverPageCount({
+  caseDefinition,
+  scope,
+  sourceMetadata
+}) {
+  assert(caseDefinition && Number.isInteger(caseDefinition.coverPageCount),
+    "矩阵场景缺少有效封面页数");
+  if (scope !== "corpus") {
+    return caseDefinition.coverPageCount;
+  }
+  const profile = sourceMetadata?.document?.profile;
+  return profile === "project-report" || profile === "tender" ? 1 : 0;
+}
+
+export function isCorpusTextColorEquivalent(metrics) {
+  return Boolean(
+    metrics &&
+    metrics.inkIou >= 0.98 &&
+    metrics.edgeIou >= 0.98 &&
+    metrics.meanAbsoluteError <= 8
+  );
+}
+
 export function readBundledThemeMatrixDefinitions(themesDirectory) {
   assert(
     typeof themesDirectory === "string" && themesDirectory.length > 0,

scripts/docx-layout-visual-matrix-cases.test.mjs

@@ -8,7 +8,9 @@ import {
   DOCX_LAYOUT_MATRIX_MARGIN_SCENARIOS,
   DOCX_LAYOUT_MATRIX_ORIENTATIONS,
   getDocxFontGateFailures,
+  isCorpusTextColorEquivalent,
   readBundledThemeMatrixDefinitions,
+  resolveEffectiveCoverPageCount,
   summarizeDocxLayoutVisualMatrix
 } from "./docx-layout-visual-matrix-cases.mjs";
 
@@ -156,3 +158,40 @@ test("四套独立封面主题扩展为 40 个严格封面场景", () => {
     ]
   );
 });
+
+test("真实语料按文档语义而不是主题能力判定独立封面", () => {
+  const caseDefinition = { coverPageCount: 1 };
+  assert.equal(resolveEffectiveCoverPageCount({
+    caseDefinition,
+    scope: "full",
+    sourceMetadata: {}
+  }), 1);
+  assert.equal(resolveEffectiveCoverPageCount({
+    caseDefinition,
+    scope: "corpus",
+    sourceMetadata: {}
+  }), 0);
+  assert.equal(resolveEffectiveCoverPageCount({
+    caseDefinition,
+    scope: "corpus",
+    sourceMetadata: { document: { profile: "tender" } }
+  }), 1);
+});
+
+test("短文本仅在墨迹边缘等价且像素误差低时忽略抗锯齿色差", () => {
+  assert.equal(isCorpusTextColorEquivalent({
+    inkIou: 1,
+    edgeIou: 1,
+    meanAbsoluteError: 6.9
+  }), true);
+  assert.equal(isCorpusTextColorEquivalent({
+    inkIou: 1,
+    edgeIou: 0.999,
+    meanAbsoluteError: 23
+  }), false);
+  assert.equal(isCorpusTextColorEquivalent({
+    inkIou: 0.8,
+    edgeIou: 1,
+    meanAbsoluteError: 4
+  }), false);
+});

scripts/docx-real-world-corpus.mjs

@@ -0,0 +1,100 @@
+import fs from "node:fs";
+import path from "node:path";
+
+export const DOCX_REAL_WORLD_CORPUS = Object.freeze([
+  {
+    id: "changqing-tender-requirements",
+    label: "长庆油田智能健康技术服务招标技术要求",
+    markdownPath: "tmp/长庆油田智能健康技术服务招标技术要求.md",
+    requiredFeatures: ["inline-code", "table-math-alignment"]
+  },
+  {
+    id: "health-data-schema",
+    label: "运动健康类和营养饮食数据结构升级",
+    markdownPath: "tmp/运动健康类和营养饮食数据数据结构升级.md",
+    requiredFeatures: ["json-code-indentation"],
+    codeIndentationProbe: {
+      rootLine: "{",
+      nestedLineIncludes: '"heart_rate_daily"',
+      minimumIndentPt: 4
+    }
+  },
+  {
+    id: "m4n-table-design",
+    label: "M4N 全量表设计模块化讲解",
+    markdownPath: "tmp/数据中台项目周报_2026_W30_附件_M4N全量表设计模块化讲解.md",
+    requiredFeatures: ["mermaid", "long-document"]
+  }
+]);
+
+export function resolveDocxRealWorldCorpus(repositoryDirectory) {
+  return DOCX_REAL_WORLD_CORPUS.map((entry) => ({
+    ...entry,
+    absoluteMarkdownPath: path.resolve(
+      repositoryDirectory,
+      entry.markdownPath
+    ),
+    byteLength: fs.existsSync(path.resolve(repositoryDirectory, entry.markdownPath))
+      ? fs.statSync(path.resolve(repositoryDirectory, entry.markdownPath)).size
+      : 0
+  }));
+}
+
+export function sortDocxRealWorldCorpusBySize(repositoryDirectory) {
+  return resolveDocxRealWorldCorpus(repositoryDirectory).sort(
+    (first, second) =>
+      first.byteLength - second.byteLength ||
+      first.id.localeCompare(second.id, "en")
+  );
+}
+
+export function validateDocxRealWorldCorpus(repositoryDirectory) {
+  const corpus = sortDocxRealWorldCorpusBySize(repositoryDirectory);
+  const failures = [];
+  for (const entry of corpus) {
+    if (!fs.existsSync(entry.absoluteMarkdownPath)) {
+      failures.push(`${entry.id} 缺少 Markdown：${entry.markdownPath}`);
+      continue;
+    }
+    const source = fs.readFileSync(entry.absoluteMarkdownPath, "utf8");
+    if (!source.trim()) {
+      failures.push(`${entry.id} Markdown 为空`);
+    }
+    if (
+      entry.requiredFeatures.includes("json-code-indentation") &&
+      !/```json\s*[\s\S]*?\n[\t ]+\S/gu.test(source)
+    ) {
+      failures.push(`${entry.id} 缺少带缩进的 JSON 围栏探针`);
+    }
+    if (
+      entry.requiredFeatures.includes("inline-code") &&
+      !/(^|[^`])`[^`\n]+`([^`]|$)/mu.test(source)
+    ) {
+      failures.push(`${entry.id} 缺少行内代码探针`);
+    }
+    if (
+      entry.requiredFeatures.includes("table-math-alignment") &&
+      !/^\|.*\$[^$]+\$.*\|/mu.test(source)
+    ) {
+      failures.push(`${entry.id} 缺少表格公式对齐探针`);
+    }
+  }
+  return { corpus, failures };
+}
+
+export function summarizeDocxReleaseGateMatrix({
+  baselineCaseCount,
+  corpusCaseCounts
+}) {
+  const corpusCaseCount = Object.values(corpusCaseCounts).reduce(
+    (sum, value) => sum + value,
+    0
+  );
+  return {
+    baselineCaseCount,
+    corpusDocumentCount: Object.keys(corpusCaseCounts).length,
+    corpusCaseCounts,
+    corpusCaseCount,
+    totalCaseCount: baselineCaseCount + corpusCaseCount
+  };
+}

scripts/docx-real-world-corpus.test.mjs

@@ -0,0 +1,60 @@
+import assert from "node:assert/strict";
+import path from "node:path";
+import test from "node:test";
+
+import {
+  DOCX_REAL_WORLD_CORPUS,
+  sortDocxRealWorldCorpusBySize,
+  summarizeDocxReleaseGateMatrix,
+  validateDocxRealWorldCorpus
+} from "./docx-real-world-corpus.mjs";
+
+const repositoryDirectory = path.resolve(import.meta.dirname, "..");
+
+test("冻结三份真实 Markdown 语料及必需特征", () => {
+  const validation = validateDocxRealWorldCorpus(repositoryDirectory);
+  assert.deepEqual(validation.failures, []);
+  assert.equal(validation.corpus.length, 3);
+  assert.equal(new Set(DOCX_REAL_WORLD_CORPUS.map((entry) => entry.id)).size, 3);
+  assert.ok(
+    DOCX_REAL_WORLD_CORPUS.some((entry) =>
+      entry.requiredFeatures.includes("json-code-indentation")
+    )
+  );
+  assert.ok(
+    DOCX_REAL_WORLD_CORPUS.some((entry) =>
+      entry.requiredFeatures.includes("table-math-alignment")
+    )
+  );
+});
+
+test("真实 Markdown 按文件字节数稳定升序执行", () => {
+  const corpus = sortDocxRealWorldCorpusBySize(repositoryDirectory);
+  assert.deepEqual(
+    corpus.map((entry) => entry.byteLength),
+    corpus.map((entry) => entry.byteLength).toSorted((a, b) => a - b)
+  );
+  assert.ok(corpus.every((entry) => entry.byteLength > 0));
+});
+
+test("发布门禁矩阵聚合为 140+420=560", () => {
+  assert.deepEqual(
+    summarizeDocxReleaseGateMatrix({
+      baselineCaseCount: 140,
+      corpusCaseCounts: Object.fromEntries(
+        DOCX_REAL_WORLD_CORPUS.map((entry) => [entry.id, 140])
+      )
+    }),
+    {
+      baselineCaseCount: 140,
+      corpusDocumentCount: 3,
+      corpusCaseCounts: {
+        "changqing-tender-requirements": 140,
+        "health-data-schema": 140,
+        "m4n-table-design": 140
+      },
+      corpusCaseCount: 420,
+      totalCaseCount: 560
+    }
+  );
+});

scripts/docx-release-gate-suites.mjs

@@ -0,0 +1,232 @@
+import crypto from "node:crypto";
+import fs from "node:fs";
+import path from "node:path";
+
+import {
+  createDocxLayoutVisualMatrixCases,
+  readBundledThemeMatrixDefinitions
+} from "./docx-layout-visual-matrix-cases.mjs";
+import { validateDocxRealWorldCorpus } from "./docx-real-world-corpus.mjs";
+
+export const DOCX_RELEASE_GATE_CASE_COUNT = 140;
+export const DOCX_RELEASE_GATE_SCHEMA_VERSION = 2;
+
+function assert(condition, message) {
+  if (!condition) {
+    throw new Error(message);
+  }
+}
+
+export function createDocxReleaseGateSuites(repositoryDirectory) {
+  const validation = validateDocxRealWorldCorpus(repositoryDirectory);
+  assert(
+    validation.failures.length === 0,
+    `真实语料无效：${validation.failures.join("；")}`
+  );
+  const themes = readBundledThemeMatrixDefinitions(
+    path.join(repositoryDirectory, "themes")
+  );
+  const cases = createDocxLayoutVisualMatrixCases(themes);
+  assert(
+    cases.length === DOCX_RELEASE_GATE_CASE_COUNT,
+    `单套矩阵应为 ${DOCX_RELEASE_GATE_CASE_COUNT}，实际 ${cases.length}`
+  );
+  const caseIds = cases.map((entry) => entry.id);
+  return [
+    {
+      id: "synthetic-baseline-140",
+      label: "标准合成基线 140",
+      kind: "synthetic",
+      order: 0,
+      outputDirectoryName: "synthetic-baseline-140",
+      expectedCaseCount: caseIds.length,
+      caseIds
+    },
+    ...validation.corpus.map((entry, index) => ({
+      id: entry.id,
+      label: `${entry.label}（${entry.byteLength} bytes）`,
+      kind: "corpus",
+      order: index + 1,
+      outputDirectoryName:
+        `corpus-${String(index + 1).padStart(2, "0")}-${entry.byteLength}-${entry.id}`,
+      expectedCaseCount: caseIds.length,
+      caseIds,
+      markdownPath: entry.markdownPath,
+      absoluteMarkdownPath: entry.absoluteMarkdownPath,
+      markdownByteLength: entry.byteLength,
+      requiredFeatures: entry.requiredFeatures
+    }))
+  ];
+}
+
+export function isSuiteAggregatePassed(aggregate, suite) {
+  return Boolean(
+    aggregate &&
+    aggregate.suite?.id === suite.id &&
+    aggregate.execution?.expectedCaseCount === suite.expectedCaseCount &&
+    aggregate.execution?.completedCaseCount === suite.expectedCaseCount &&
+    aggregate.execution?.passedCaseCount === suite.expectedCaseCount &&
+    aggregate.gatePassed === true
+  );
+}
+
+export function resolveDocxReleaseGateSuite({
+  suites,
+  requestedId,
+  aggregates,
+  allowOutOfOrder = false
+}) {
+  assert(Array.isArray(suites) && suites.length === 4, "发布门禁必须包含四套独立矩阵");
+  const target = !requestedId || requestedId === "next"
+    ? suites.find((suite) => !isSuiteAggregatePassed(aggregates.get(suite.id), suite))
+    : suites.find((suite) => suite.id === requestedId);
+  if (!target) {
+    if (!requestedId || requestedId === "next") {
+      return undefined;
+    }
+    throw new Error(`未知发布门禁套件：${requestedId}`);
+  }
+  if (!allowOutOfOrder || !requestedId || requestedId === "next") {
+    const unmetPrerequisite = suites
+      .slice(0, target.order)
+      .find((suite) => !isSuiteAggregatePassed(aggregates.get(suite.id), suite));
+    assert(
+      !unmetPrerequisite,
+      `必须先通过前置套件 ${unmetPrerequisite?.id}`
+    );
+  }
+  return target;
+}
+
+function collectFingerprintFiles(repositoryDirectory, suite) {
+  const roots = [
+    "themes",
+    "apps/server/scripts",
+    "apps/server/src",
+    "packages/application/src",
+    "packages/core/src",
+    "packages/document-visual-diff/src",
+    "packages/docx-engine/assets",
+    "packages/docx-engine/src",
+    "packages/preview-engine/src",
+    "packages/renderer/src"
+  ];
+  const files = [
+    "package-lock.json",
+    "scripts/docx-layout-visual-matrix-cases.mjs",
+    "scripts/docx-real-world-corpus.mjs",
+    "scripts/docx-release-gate-suites.mjs",
+    "scripts/verify-docx-layout-visual-matrix.mjs",
+    "scripts/verify-docx-release-gate-suite.mjs"
+  ];
+  const allowedExtension = /\.(?:css|js|json|lua|mjs|ts)$/iu;
+  const visit = (relativeDirectory) => {
+    const absoluteDirectory = path.join(repositoryDirectory, relativeDirectory);
+    if (!fs.existsSync(absoluteDirectory)) {
+      return;
+    }
+    for (const entry of fs.readdirSync(absoluteDirectory, { withFileTypes: true })) {
+      const relativePath = path.join(relativeDirectory, entry.name);
+      if (entry.isDirectory()) {
+        visit(relativePath);
+      } else if (allowedExtension.test(entry.name)) {
+        files.push(relativePath);
+      }
+    }
+  };
+  for (const root of roots) {
+    visit(root);
+  }
+  if (suite.markdownPath) {
+    files.push(suite.markdownPath);
+  }
+  return [...new Set(files)].sort((left, right) => left.localeCompare(right, "en"));
+}
+
+export function createDocxReleaseGateFingerprint(repositoryDirectory, suite) {
+  const hash = crypto.createHash("sha256");
+  hash.update(`schema:${DOCX_RELEASE_GATE_SCHEMA_VERSION}\n`);
+  hash.update(`suite:${suite.id}\n`);
+  for (const relativePath of collectFingerprintFiles(repositoryDirectory, suite)) {
+    const absolutePath = path.join(repositoryDirectory, relativePath);
+    if (!fs.existsSync(absolutePath)) {
+      hash.update(`${relativePath}:missing\n`);
+      continue;
+    }
+    hash.update(`${relativePath}\0`);
+    hash.update(fs.readFileSync(absolutePath));
+    hash.update("\0");
+  }
+  return hash.digest("hex");
+}
+
+export function readJsonIfExists(filePath) {
+  if (!fs.existsSync(filePath)) {
+    return undefined;
+  }
+  try {
+    return JSON.parse(fs.readFileSync(filePath, "utf8"));
+  } catch {
+    return undefined;
+  }
+}
+
+const WINDOWS_RENAME_RETRY_CODES = new Set(["EACCES", "EBUSY", "EPERM"]);
+
+function sleepSync(milliseconds) {
+  Atomics.wait(
+    new Int32Array(new SharedArrayBuffer(Int32Array.BYTES_PER_ELEMENT)),
+    0,
+    0,
+    milliseconds
+  );
+}
+
+export function writeJsonAtomic(filePath, value, options = {}) {
+  fs.mkdirSync(path.dirname(filePath), { recursive: true });
+  const temporaryPath = `${filePath}.${process.pid}.${crypto.randomUUID()}.tmp`;
+  fs.writeFileSync(temporaryPath, `${JSON.stringify(value, null, 2)}\n`, "utf8");
+  const rename = options.rename ?? fs.renameSync;
+  const sleep = options.sleep ?? sleepSync;
+  const retryDeadline = Date.now() + (options.retryWindowMs ?? 3_000);
+  let attempt = 0;
+  while (true) {
+    try {
+      rename(temporaryPath, filePath);
+      return;
+    } catch (error) {
+      if (
+        !WINDOWS_RENAME_RETRY_CODES.has(error?.code) ||
+        Date.now() >= retryDeadline
+      ) {
+        fs.rmSync(temporaryPath, { force: true });
+        throw error;
+      }
+      attempt += 1;
+      sleep(Math.min(100, 10 * attempt));
+    }
+  }
+}
+
+export function summarizeSuiteProgress({ suite, fingerprint, records }) {
+  const currentRecords = suite.caseIds
+    .map((caseId) => records[caseId])
+    .filter((record) => record?.fingerprint === fingerprint);
+  const passedCaseCount = currentRecords.filter((record) => record.status === "passed").length;
+  const gateFailedCaseCount = currentRecords.filter((record) => record.status === "gate-failed").length;
+  const errorCaseCount = currentRecords.filter((record) => record.status === "error").length;
+  const completedCaseCount = currentRecords.length;
+  return {
+    expectedCaseCount: suite.expectedCaseCount,
+    completedCaseCount,
+    pendingCaseCount: suite.expectedCaseCount - completedCaseCount,
+    passedCaseCount,
+    gateFailedCaseCount,
+    errorCaseCount,
+    complete: completedCaseCount === suite.expectedCaseCount,
+    gatePassed:
+      passedCaseCount === suite.expectedCaseCount &&
+      gateFailedCaseCount === 0 &&
+      errorCaseCount === 0
+  };
+}

scripts/docx-release-gate-suites.test.mjs

@@ -0,0 +1,135 @@
+import assert from "node:assert/strict";
+import fs from "node:fs";
+import os from "node:os";
+import path from "node:path";
+import test from "node:test";
+import { fileURLToPath } from "node:url";
+
+import {
+  createDocxReleaseGateSuites,
+  isSuiteAggregatePassed,
+  resolveDocxReleaseGateSuite,
+  summarizeSuiteProgress,
+  writeJsonAtomic
+} from "./docx-release-gate-suites.mjs";
+
+const repositoryDirectory = path.resolve(
+  path.dirname(fileURLToPath(import.meta.url)),
+  ".."
+);
+
+function passedAggregate(suite) {
+  return {
+    suite: { id: suite.id },
+    execution: {
+      expectedCaseCount: 140,
+      completedCaseCount: 140,
+      passedCaseCount: 140
+    },
+    gatePassed: true
+  };
+}
+
+test("发布门禁拆分为四套独立 140 且真实文档按字节升序", () => {
+  const suites = createDocxReleaseGateSuites(repositoryDirectory);
+  assert.equal(suites.length, 4);
+  assert.equal(suites[0].id, "synthetic-baseline-140");
+  assert.ok(suites.every((suite) => suite.expectedCaseCount === 140));
+  assert.deepEqual(
+    suites.slice(1).map((suite) => suite.markdownByteLength),
+    suites.slice(1).map((suite) => suite.markdownByteLength).toSorted((a, b) => a - b)
+  );
+  assert.ok(suites.slice(1).every((suite) => suite.outputDirectoryName.includes(String(suite.markdownByteLength))));
+});
+
+test("next 严格遵守合成基线和真实文档大小顺序", () => {
+  const suites = createDocxReleaseGateSuites(repositoryDirectory);
+  const aggregates = new Map();
+  assert.equal(resolveDocxReleaseGateSuite({ suites, requestedId: "next", aggregates }).id, suites[0].id);
+  aggregates.set(suites[0].id, passedAggregate(suites[0]));
+  assert.equal(resolveDocxReleaseGateSuite({ suites, requestedId: "next", aggregates }).id, suites[1].id);
+  assert.throws(
+    () => resolveDocxReleaseGateSuite({ suites, requestedId: suites[2].id, aggregates }),
+    /必须先通过前置套件/u
+  );
+  for (const suite of suites) {
+    aggregates.set(suite.id, passedAggregate(suite));
+  }
+  assert.equal(resolveDocxReleaseGateSuite({ suites, requestedId: "next", aggregates }), undefined);
+});
+
+test("只有显式指定套件时才允许受控越过前置门禁", () => {
+  const suites = createDocxReleaseGateSuites(repositoryDirectory);
+  const aggregates = new Map();
+  assert.equal(
+    resolveDocxReleaseGateSuite({
+      suites,
+      requestedId: suites[3].id,
+      aggregates,
+      allowOutOfOrder: true
+    }).id,
+    suites[3].id
+  );
+  assert.equal(
+    resolveDocxReleaseGateSuite({
+      suites,
+      requestedId: "next",
+      aggregates,
+      allowOutOfOrder: true
+    }).id,
+    suites[0].id
+  );
+});
+
+test("套件进度独立统计通过、门禁失败和基础设施错误", () => {
+  const suite = createDocxReleaseGateSuites(repositoryDirectory)[0];
+  const records = {
+    [suite.caseIds[0]]: { fingerprint: "same", status: "passed" },
+    [suite.caseIds[1]]: { fingerprint: "same", status: "gate-failed" },
+    [suite.caseIds[2]]: { fingerprint: "same", status: "error" },
+    [suite.caseIds[3]]: { fingerprint: "old", status: "passed" }
+  };
+  const progress = summarizeSuiteProgress({ suite, fingerprint: "same", records });
+  assert.deepEqual(progress, {
+    expectedCaseCount: 140,
+    completedCaseCount: 3,
+    pendingCaseCount: 137,
+    passedCaseCount: 1,
+    gateFailedCaseCount: 1,
+    errorCaseCount: 1,
+    complete: false,
+    gatePassed: false
+  });
+  assert.equal(isSuiteAggregatePassed({ suite: { id: suite.id }, execution: progress, gatePassed: false }, suite), false);
+});
+
+test("Windows 瞬时文件锁不会中断原子进度写入", () => {
+  const directory = fs.mkdtempSync(path.join(os.tmpdir(), "mdtp-gate-atomic-"));
+  const filePath = path.join(directory, "progress.json");
+  let attempts = 0;
+  try {
+    writeJsonAtomic(filePath, { completed: 9 }, {
+      rename(source, destination) {
+        attempts += 1;
+        if (attempts < 3) {
+          const error = new Error("temporary Windows lock");
+          error.code = "EPERM";
+          throw error;
+        }
+        fs.renameSync(source, destination);
+      },
+      sleep() {}
+    });
+    assert.equal(attempts, 3);
+    assert.deepEqual(
+      JSON.parse(fs.readFileSync(filePath, "utf8")),
+      { completed: 9 }
+    );
+    assert.deepEqual(
+      fs.readdirSync(directory).filter((name) => name.endsWith(".tmp")),
+      []
+    );
+  } finally {
+    fs.rmSync(directory, { recursive: true, force: true });
+  }
+});

scripts/verify-docx-layout-visual-matrix.mjs

@@ -4,6 +4,7 @@ import path from "node:path";
 import { promisify } from "node:util";
 import { DOMParser } from "@xmldom/xmldom";
 import { unzipSync } from "fflate";
+import matter from "gray-matter";
 
 import {
   createPdfDocumentSnapshot,
@@ -11,6 +12,8 @@ import {
   createWordPdfAdapter,
   createWpsPdfAdapter,
   getBlockingVisualDiffIssues,
+  isCrossEngineRasterEquivalent,
+  normalizePdfEditableText,
   renderPdfVisualDiffHtml,
   serializePdfVisualDiffJson
 } from "../packages/document-visual-diff/dist/index.js";
@@ -21,9 +24,14 @@ import {
 import {
   createDocxLayoutVisualMatrixCases,
   getDocxFontGateFailures,
+  isCorpusTextColorEquivalent,
   readBundledThemeMatrixDefinitions,
+  resolveEffectiveCoverPageCount,
   summarizeDocxLayoutVisualMatrix
 } from "./docx-layout-visual-matrix-cases.mjs";
+import {
+  DOCX_REAL_WORLD_CORPUS
+} from "./docx-real-world-corpus.mjs";
 
 const execFileAsync = promisify(execFile);
 const WORD_NAMESPACE =
@@ -32,10 +40,10 @@ const MATH_NAMESPACE =
   "http://schemas.openxmlformats.org/officeDocument/2006/math";
 const decoder = new TextDecoder();
 const repositoryDirectory = path.resolve(import.meta.dirname, "..");
-const outputDirectory = path.join(
+const outputDirectory = path.resolve(
   repositoryDirectory,
-  "output",
+  process.env.MD_TO_PDF_LAYOUT_VISUAL_OUTPUT_DIR?.trim() ||
-  "docx-layout-visual-matrix"
+    path.join("output", "docx-layout-visual-matrix")
 );
 const matrixScript = path.join(
   repositoryDirectory,
@@ -64,6 +72,23 @@ const selectedOrientation =
   process.env.MD_TO_PDF_LAYOUT_VISUAL_ORIENTATION?.trim() || undefined;
 const selectedMarginScenarioId =
   process.env.MD_TO_PDF_LAYOUT_VISUAL_MARGIN?.trim() || undefined;
+const expectInlineCodeProbes =
+  process.env.MD_TO_PDF_R4_APPEND_INLINE_CODE_PROBES?.trim() !== "0";
+const selectedCorpusId =
+  process.env.MD_TO_PDF_CORPUS_ID?.trim() || undefined;
+const corpusDefinition = selectedCorpusId
+  ? DOCX_REAL_WORLD_CORPUS.find((entry) => entry.id === selectedCorpusId)
+  : undefined;
+const corpusSourceMetadata = selectedScope === "corpus" &&
+  process.env.MD_TO_PDF_R4_MARKDOWN_PATH?.trim()
+  ? matter(fs.readFileSync(
+      path.resolve(
+        repositoryDirectory,
+        process.env.MD_TO_PDF_R4_MARKDOWN_PATH.trim()
+      ),
+      "utf8"
+    )).data
+  : undefined;
 const scopedCases = selectedCaseId
   ? cases.filter((entry) => entry.id === selectedCaseId)
   : selectedScope === "cover"
@@ -185,6 +210,43 @@ function paragraphBlockKind(paragraph, styleId) {
   return "paragraph";
 }
 
+function directWordChildren(node, localName) {
+  return Array.from(node?.childNodes ?? []).filter(
+    (child) =>
+      child.nodeType === 1 &&
+      child.localName === localName &&
+      child.namespaceURI === WORD_NAMESPACE
+  );
+}
+
+function nearestWordAncestor(node, localName) {
+  let current = node?.parentNode;
+  while (current) {
+    if (
+      current.localName === localName &&
+      current.namespaceURI === WORD_NAMESPACE
+    ) {
+      return current;
+    }
+    current = current.parentNode;
+  }
+  return undefined;
+}
+
+function paragraphTableCoordinates(paragraph, tableIndexes) {
+  const cell = nearestWordAncestor(paragraph, "tc");
+  const row = nearestWordAncestor(paragraph, "tr");
+  const table = nearestWordAncestor(paragraph, "tbl");
+  if (!cell || !row || !table) {
+    return undefined;
+  }
+  return {
+    tableGroupId: `table-${tableIndexes.get(table)}`,
+    tableRowIndex: directWordChildren(table, "tr").indexOf(row),
+    tableColumnIndex: directWordChildren(row, "tc").indexOf(cell)
+  };
+}
+
 function extractEditableContract(docxPath) {
   const entries = unzipSync(Uint8Array.from(fs.readFileSync(docxPath)));
   const documentXml = entries["word/document.xml"];
@@ -193,13 +255,83 @@ function extractEditableContract(docxPath) {
     decoder.decode(documentXml),
     "application/xml"
   );
+  const tableIndexes = new Map(
+    Array.from(document.getElementsByTagName("*")).filter(
+      (node) => node.localName === "tbl" && node.namespaceURI === WORD_NAMESPACE
+    ).map((table, index) => [table, index])
+  );
   let section = "cover";
   let hasCoverSection = false;
   const paragraphs = [];
+  let inlineCodeParagraphCount = 0;
+  let inlineCodeExactLineRuleCount = 0;
+  let pureMathTableParagraphCount = 0;
+  let stabilizedPureMathTableParagraphCount = 0;
   for (const paragraph of Array.from(document.getElementsByTagName("*")).filter(
     (node) => node.localName === "p" && node.namespaceURI === WORD_NAMESPACE
   )) {
     const descendants = Array.from(paragraph.getElementsByTagName("*"));
+    const inlineCode = descendants.some(
+      (node) =>
+        node.localName === "rStyle" &&
+        node.namespaceURI === WORD_NAMESPACE &&
+        (node.getAttribute("w:val") || node.getAttribute("val")) ===
+          "VerbatimChar"
+    );
+    if (inlineCode) {
+      inlineCodeParagraphCount += 1;
+      if (
+        descendants.some(
+          (node) =>
+            node.localName === "spacing" &&
+            node.namespaceURI === WORD_NAMESPACE &&
+            (node.getAttribute("w:lineRule") ||
+              node.getAttribute("lineRule")) === "exact"
+        )
+      ) {
+        inlineCodeExactLineRuleCount += 1;
+      }
+    }
+    if (hasAncestor(paragraph, "tc")) {
+      const directContent = Array.from(paragraph.childNodes).filter(
+        (node) =>
+          node.nodeType === 1 &&
+          !(
+            node.localName === "pPr" &&
+            node.namespaceURI === WORD_NAMESPACE
+          )
+      );
+      const directMath = directContent.filter(
+        (node) =>
+          node.localName === "oMath" &&
+          node.namespaceURI === MATH_NAMESPACE
+      );
+      const stabilizers = directContent.filter(
+        (node) =>
+          node.localName === "r" &&
+          node.namespaceURI === WORD_NAMESPACE &&
+          Array.from(node.getElementsByTagName("*")).some(
+            (descendant) =>
+              descendant.localName === "noProof" &&
+              descendant.namespaceURI === WORD_NAMESPACE
+          ) &&
+          Array.from(node.getElementsByTagName("*")).some(
+            (descendant) =>
+              descendant.localName === "t" &&
+              descendant.namespaceURI === WORD_NAMESPACE &&
+              descendant.textContent === "\u200B"
+          )
+      );
+      if (
+        directMath.length > 0 &&
+        directContent.length === directMath.length + stabilizers.length
+      ) {
+        pureMathTableParagraphCount += 1;
+        if (stabilizers.length > 0) {
+          stabilizedPureMathTableParagraphCount += 1;
+        }
+      }
+    }
     const hasSectionBreak = descendants.some(
       (node) =>
         node.localName === "sectPr" && node.namespaceURI === WORD_NAMESPACE
@@ -209,20 +341,53 @@ function extractEditableContract(docxPath) {
       section = "body";
       continue;
     }
-    const text = descendants.flatMap((node) =>
+    const hardBreakSegments = [""];
+    const mathCharacterIndexes = [];
+    let normalizedCharacterOffset = 0;
+    for (const node of descendants) {
+      if (
         node.localName === "t" &&
         (node.namespaceURI === WORD_NAMESPACE ||
           node.namespaceURI === MATH_NAMESPACE)
-        ? [node.textContent ?? ""]
+      ) {
-        : []
+        const nodeText = (node.textContent ?? "").replace(
-    ).join("");
+          /[\u200B\u2060\uFEFF]/gu,
+          ""
+        );
+        hardBreakSegments[hardBreakSegments.length - 1] += nodeText;
+        const normalizedNodeLength = Array.from(
+          normalizePdfEditableText(nodeText)
+        ).length;
+        if (node.namespaceURI === MATH_NAMESPACE) {
+          for (let index = 0; index < normalizedNodeLength; index += 1) {
+            mathCharacterIndexes.push(normalizedCharacterOffset + index);
+          }
+        }
+        normalizedCharacterOffset += normalizedNodeLength;
+      } else if (
+        node.localName === "br" &&
+        node.namespaceURI === WORD_NAMESPACE
+      ) {
+        hardBreakSegments.push("");
+      }
+    }
+    const text = hardBreakSegments.join("");
     if (!text || isInternalLayoutSpacerParagraph(descendants, text)) {
       continue;
     }
     const styleId = paragraphStyleId(paragraph);
+    const tableCoordinates = paragraphTableCoordinates(paragraph, tableIndexes);
     paragraphs.push({
       index: paragraphs.length,
       text,
+      ...(mathCharacterIndexes.length > 0
+        ? { mathCharacterIndexes }
+        : {}),
+      ...(hardBreakSegments.length > 1
+        ? { hardBreakSegments }
+        : {}),
+      ...(inlineCode ? { hasInlineCode: true } : {}),
+      ...(tableCoordinates ?? {}),
       ...(styleId ? { styleId } : {}),
       role: paragraphRole(styleId),
       blockKind: paragraphBlockKind(paragraph, styleId),
@@ -236,7 +401,13 @@ function extractEditableContract(docxPath) {
   }
   return {
     text: paragraphs.map((paragraph) => paragraph.text).join(""),
-    paragraphs
+    paragraphs,
+    translationInvariants: {
+      inlineCodeParagraphCount,
+      inlineCodeExactLineRuleCount,
+      pureMathTableParagraphCount,
+      stabilizedPureMathTableParagraphCount
+    }
   };
 }
 
@@ -338,6 +509,82 @@ function inspectInlineCodeContinuity(snapshot, label) {
   };
 }
 
+function inspectCodeIndentation(snapshot, label, probe) {
+  if (!probe) {
+    return {
+      label,
+      applicable: false,
+      passed: true,
+      failures: []
+    };
+  }
+  for (const page of snapshot.pages) {
+    const nestedIndex = page.lines.findIndex((line) =>
+      line.text.includes(probe.nestedLineIncludes)
+    );
+    if (nestedIndex < 0) {
+      continue;
+    }
+    const nested = page.lines[nestedIndex];
+    const root = page.lines
+      .slice(0, nestedIndex)
+      .reverse()
+      .find(
+        (line) =>
+          line.text.trim() === probe.rootLine &&
+          nested.baselineY - line.baselineY <= 30
+      );
+    if (!root) {
+      continue;
+    }
+    const indentPt = nested.bounds.x - root.bounds.x;
+    const passed = indentPt >= probe.minimumIndentPt;
+    return {
+      label,
+      applicable: true,
+      passed,
+      pageNumber: page.pageNumber,
+      rootText: root.text,
+      nestedText: nested.text,
+      rootXPt: root.bounds.x,
+      nestedXPt: nested.bounds.x,
+      indentPt,
+      failures: passed
+        ? []
+        : [
+            `${label}: CODE_INDENTATION_MISMATCH - JSON 嵌套行缩进 ${indentPt.toFixed(2)}pt，小于 ${probe.minimumIndentPt}pt`
+          ]
+    };
+  }
+  return {
+    label,
+    applicable: true,
+    passed: false,
+    failures: [
+      `${label}: CODE_INDENTATION_PROBE_UNAVAILABLE - 未定位 JSON 根行与嵌套行`
+    ]
+  };
+}
+
+function docxTranslationInvariantFailures(invariants) {
+  const failures = [];
+  if (invariants.inlineCodeExactLineRuleCount > 0) {
+    failures.push(
+      `DOCX: INLINE_CODE_EXACT_LINE_BOX - ${invariants.inlineCodeExactLineRuleCount} 个行内代码段落仍使用 exact 行盒`
+    );
+  }
+  if (
+    invariants.pureMathTableParagraphCount !==
+    invariants.stabilizedPureMathTableParagraphCount
+  ) {
+    failures.push(
+      "DOCX: TABLE_MATH_ALIGNMENT_UNSTABLE - " +
+        `${invariants.pureMathTableParagraphCount - invariants.stabilizedPureMathTableParagraphCount} 个纯公式表格段落缺少对齐稳定结构`
+    );
+  }
+  return failures;
+}
+
 function buildPageSemantics(pageCount, coverPageCount, config) {
   assert(
     pageCount > coverPageCount,
@@ -423,6 +670,108 @@ function reportCoverGateFailures(report, label) {
     .map((issue) => `${label}: ${issue.code} - ${issue.message}`);
 }
 
+function metricRatio(first, second) {
+  const maximum = Math.max(first, second);
+  return maximum > 0 ? Math.min(first, second) / maximum : 1;
+}
+
+const MAXIMUM_CORPUS_ELEMENT_COLOR_DELTA = 55;
+
+function reportCorpusGateFailures(
+  report,
+  label,
+  coverPageCount
+) {
+  const failures = [];
+  if ((report.basic.candidateEditableSimilarity ?? 0) < 0.995) {
+    failures.push(
+      `${label}: EDITABLE_CONTENT_SIMILARITY - Office 可编辑正文相似度 ` +
+        `${((report.basic.candidateEditableSimilarity ?? 0) * 100).toFixed(3)}% 低于 99.5%`
+    );
+  }
+  const strictKinds = new Set([
+    "heading",
+    "title",
+    "caption",
+    "table-header",
+    "code-block"
+  ]);
+  for (const comparison of report.basic.semanticBlockVisuals ?? []) {
+    const index = comparison.expectation.index + 1;
+    const visualUnavailable =
+      comparison.status === "unavailable" ||
+      comparison.lines.length === 0 ||
+      comparison.issues?.some(
+        (issue) => issue.code === "SEMANTIC_BLOCK_VISUAL_UNAVAILABLE"
+      );
+    if (visualUnavailable) {
+      failures.push(
+        `${label}: SEMANTIC_BLOCK_VISUAL_UNAVAILABLE - 第 ${index} 个元素块无法建立视觉观测`
+      );
+      continue;
+    }
+    for (const line of comparison.lines) {
+      const metrics = line.metrics;
+      if (!metrics) {
+        continue;
+      }
+      const heightRatio = metricRatio(
+        metrics.baselineHeightPx,
+        metrics.candidateHeightPx
+      );
+      if (heightRatio < 0.78) {
+        failures.push(
+          `${label}: ELEMENT_LINE_BOX_MISMATCH - 第 ${index} 个元素块第 ${line.lineIndex + 1} 行高度比例 ${heightRatio.toFixed(3)} 低于 0.78`
+        );
+      }
+      const widthRatio = metricRatio(
+        metrics.baselineWidthPx,
+        metrics.candidateWidthPx
+      );
+      if (comparison.lines.length === 1 && widthRatio < 0.85) {
+        failures.push(
+          `${label}: ELEMENT_WIDTH_MISMATCH - 第 ${index} 个单行元素块宽度比例 ${widthRatio.toFixed(3)} 低于 0.85`
+        );
+      }
+      const minimumIou = strictKinds.has(
+        comparison.expectation.blockKind
+      ) ? 0.45 : 0.3;
+      if (
+        comparison.lines.length === 1 &&
+        metrics.inkIou < minimumIou &&
+        metrics.edgeIou < minimumIou
+      ) {
+        failures.push(
+          `${label}: ELEMENT_RASTER_MISMATCH - 第 ${index} 个元素块第 ${line.lineIndex + 1} 行墨迹与边缘 IoU 均低于 ${minimumIou}`
+        );
+      }
+      const textReflow = line.issues?.some(
+        (issue) => issue.details?.textReflow === true
+      );
+      const crossEngineRasterEquivalent = isCrossEngineRasterEquivalent(
+        metrics,
+        textReflow,
+        comparison.expectation.blockKind,
+        comparison.expectation.hasInlineCode === true
+      );
+      if (!textReflow &&
+        !crossEngineRasterEquivalent &&
+        !isCorpusTextColorEquivalent(metrics) && (
+        metrics.backgroundColorDelta > MAXIMUM_CORPUS_ELEMENT_COLOR_DELTA ||
+        metrics.foregroundColorDelta > MAXIMUM_CORPUS_ELEMENT_COLOR_DELTA
+      )) {
+        failures.push(
+          `${label}: ELEMENT_COLOR_MISMATCH - 第 ${index} 个元素块第 ${line.lineIndex + 1} 行颜色差异超限`
+        );
+      }
+    }
+  }
+  if (coverPageCount > 0) {
+    failures.push(...reportCoverGateFailures(report, label));
+  }
+  return [...new Set(failures)];
+}
+
 function reportSummary(report) {
   const paragraphIssues = report.basic.paragraphLayouts?.flatMap(
     (paragraph) => paragraph.issues
@@ -599,9 +948,18 @@ assert(
   `视觉矩阵用例不存在：${selectedCaseId}`
 );
 assert(
-  selectedScope === "full" || selectedScope === "cover",
+  ["full", "cover", "corpus"].includes(selectedScope),
   `不支持的视觉矩阵作用域：${selectedScope}`
 );
+assert(
+  selectedScope !== "corpus" || corpusDefinition,
+  `真实语料门禁缺少有效 MD_TO_PDF_CORPUS_ID：${selectedCorpusId}`
+);
+assert(
+  selectedScope !== "corpus" ||
+    Boolean(process.env.MD_TO_PDF_R4_MARKDOWN_PATH?.trim()),
+  "真实语料门禁必须显式提供 MD_TO_PDF_R4_MARKDOWN_PATH"
+);
 assert(
   !selectedOrientation || ["portrait", "landscape"].includes(selectedOrientation),
   `不支持的视觉矩阵方向：${selectedOrientation}`
@@ -632,6 +990,11 @@ for (const caseDefinition of selectedCases) {
   );
   const docxPath = path.resolve(repositoryDirectory, result.docx.outputFile);
   const editable = extractEditableContract(docxPath);
+  const effectiveCoverPageCount = resolveEffectiveCoverPageCount({
+    caseDefinition,
+    scope: selectedScope,
+    sourceMetadata: corpusSourceMetadata
+  });
   const wordGeneration = await wordAdapter.generate({ docxPath });
   const wpsGeneration = await wpsAdapter.generate({ docxPath });
   const wordPdfPath = path.join(caseDirectory, "word.pdf");
@@ -668,17 +1031,17 @@ for (const caseDefinition of selectedCases) {
   });
   const chromiumPageSemantics = buildPageSemantics(
     chromium.pageCount,
-    caseDefinition.coverPageCount,
+    effectiveCoverPageCount,
     result.exportConfig
   );
   const wordPageSemantics = buildPageSemantics(
     word.pageCount,
-    caseDefinition.coverPageCount,
+    effectiveCoverPageCount,
     result.exportConfig
   );
   const wpsPageSemantics = buildPageSemantics(
     wps.pageCount,
-    caseDefinition.coverPageCount,
+    effectiveCoverPageCount,
     result.exportConfig
   );
   const reportOptions = {
@@ -715,6 +1078,23 @@ for (const caseDefinition of selectedCases) {
     word: inspectInlineCodeContinuity(word, "Microsoft Word"),
     wps: inspectInlineCodeContinuity(wps, "WPS Writer")
   };
+  const codeIndentation = {
+    chromium: inspectCodeIndentation(
+      chromium,
+      "Chromium",
+      corpusDefinition?.codeIndentationProbe
+    ),
+    word: inspectCodeIndentation(
+      word,
+      "Microsoft Word",
+      corpusDefinition?.codeIndentationProbe
+    ),
+    wps: inspectCodeIndentation(
+      wps,
+      "WPS Writer",
+      corpusDefinition?.codeIndentationProbe
+    )
+  };
 
   const rasterDirectory = path.join(caseDirectory, "raster-pages");
   fs.rmSync(rasterDirectory, { recursive: true, force: true });
@@ -742,9 +1122,13 @@ for (const caseDefinition of selectedCases) {
       renderedFonts: wps.fonts,
       engine: "wps"
     }),
+    ...(expectInlineCodeProbes
+      ? [
           ...inlineCodeContinuity.chromium.failures,
           ...inlineCodeContinuity.word.failures,
-    ...inlineCodeContinuity.wps.failures,
+          ...inlineCodeContinuity.wps.failures
+        ]
+      : []),
     ...reportGateFailures(wordReport, "Chromium/Word"),
     ...reportGateFailures(wpsReport, "Chromium/WPS"),
     ...reportGateFailures(officeReport, "Word/WPS")
@@ -755,6 +1139,45 @@ for (const caseDefinition of selectedCases) {
         ...reportCoverGateFailures(wpsReport, "Chromium/WPS"),
         ...reportCoverGateFailures(officeReport, "Word/WPS")
       ])]
+    : selectedScope === "corpus"
+      ? [...new Set([
+          ...pageGeometryFailures(chromium, caseDefinition, "Chromium"),
+          ...pageGeometryFailures(word, caseDefinition, "Microsoft Word"),
+          ...pageGeometryFailures(wps, caseDefinition, "WPS Writer"),
+          ...getDocxFontGateFailures({
+            theme: themeDefinitions.find((theme) => theme.id === caseDefinition.themeId),
+            inspection: result.docx.inspection,
+            renderedFonts: word.fonts,
+            engine: "word"
+          }),
+          ...getDocxFontGateFailures({
+            theme: themeDefinitions.find((theme) => theme.id === caseDefinition.themeId),
+            inspection: result.docx.inspection,
+            renderedFonts: wps.fonts,
+            engine: "wps"
+          }),
+          ...docxTranslationInvariantFailures(
+            editable.translationInvariants
+          ),
+          ...codeIndentation.chromium.failures,
+          ...codeIndentation.word.failures,
+          ...codeIndentation.wps.failures,
+          ...reportCorpusGateFailures(
+            wordReport,
+            "Chromium/Word",
+            effectiveCoverPageCount
+          ),
+          ...reportCorpusGateFailures(
+            wpsReport,
+            "Chromium/WPS",
+            effectiveCoverPageCount
+          ),
+          ...reportCorpusGateFailures(
+            officeReport,
+            "Word/WPS",
+            effectiveCoverPageCount
+          )
+        ])]
       : allCurrentFailures;
   gateFailures.push(
     ...currentFailures.map((failure) => `${caseDefinition.id}: ${failure}`)
@@ -772,7 +1195,7 @@ for (const caseDefinition of selectedCases) {
     themeDefaultMargins: caseDefinition.themeDefaultMargins,
     paper: caseDefinition.paper,
     exportConfig: result.exportConfig,
-    coverPageCount: caseDefinition.coverPageCount,
+    coverPageCount: effectiveCoverPageCount,
     pages: {
       chromium: chromium.pageCount,
       word: word.pageCount,
@@ -787,7 +1210,9 @@ for (const caseDefinition of selectedCases) {
       wps: wps.fonts
     },
     editableParagraphCount: editable.paragraphs.length,
+    translationInvariants: editable.translationInvariants,
     inlineCodeContinuity,
+    codeIndentation,
     reports: {
       word: reportSummary(wordReport),
       wps: reportSummary(wpsReport),
@@ -811,6 +1236,14 @@ const summary = {
   matrixDefinition,
   selectedCaseId,
   selectedScope,
+  corpus: corpusDefinition
+    ? {
+        id: corpusDefinition.id,
+        label: corpusDefinition.label,
+        markdownPath: corpusDefinition.markdownPath,
+        requiredFeatures: corpusDefinition.requiredFeatures
+      }
+    : undefined,
   selectedOrientation,
   selectedMarginScenarioId,
   execution: {

scripts/verify-docx-release-gate-560.mjs

@@ -0,0 +1,4 @@
+process.stderr.write(
+  "verify-docx-release-gate-560 已改为兼容入口：每次只执行下一套独立 140，不再启动整批 560。\n"
+);
+await import("./verify-docx-release-gate-suite.mjs");

scripts/verify-docx-release-gate-suite.mjs

@@ -0,0 +1,277 @@
+import { execFile } from "node:child_process";
+import fs from "node:fs";
+import path from "node:path";
+import { promisify } from "node:util";
+
+import {
+  createDocxReleaseGateFingerprint,
+  createDocxReleaseGateSuites,
+  DOCX_RELEASE_GATE_SCHEMA_VERSION,
+  readJsonIfExists,
+  resolveDocxReleaseGateSuite,
+  summarizeSuiteProgress,
+  writeJsonAtomic
+} from "./docx-release-gate-suites.mjs";
+
+const execFileAsync = promisify(execFile);
+const repositoryDirectory = path.resolve(import.meta.dirname, "..");
+const matrixScript = path.join(repositoryDirectory, "scripts", "verify-docx-layout-visual-matrix.mjs");
+const outputRoot = path.resolve(
+  repositoryDirectory,
+  process.env.MD_TO_PDF_RELEASE_GATE_OUTPUT_DIR?.trim() ||
+    "output/docx-release-gate-v0.6.2"
+);
+const requestedSuiteId = process.env.MD_TO_PDF_RELEASE_GATE_SUITE?.trim() || "next";
+const selectedCaseId = process.env.MD_TO_PDF_RELEASE_GATE_CASE?.trim() || undefined;
+const allowOutOfOrder =
+  process.env.MD_TO_PDF_RELEASE_GATE_ALLOW_OUT_OF_ORDER?.trim() === "1";
+const requestedCaseLimitText =
+  process.env.MD_TO_PDF_RELEASE_GATE_CASE_LIMIT?.trim() || undefined;
+const requestedCaseLimit = requestedCaseLimitText
+  ? Number.parseInt(requestedCaseLimitText, 10)
+  : undefined;
+if (
+  requestedCaseLimitText &&
+  (!Number.isInteger(requestedCaseLimit) || requestedCaseLimit <= 0)
+) {
+  throw new Error("MD_TO_PDF_RELEASE_GATE_CASE_LIMIT 必须是正整数");
+}
+
+function escapeHtml(value) {
+  return String(value)
+    .replaceAll("&", "&amp;")
+    .replaceAll("<", "&lt;")
+    .replaceAll(">", "&gt;")
+    .replaceAll('"', "&quot;");
+}
+
+function readAggregates(suites) {
+  return new Map(suites.map((suite) => {
+    const aggregate = readJsonIfExists(
+      path.join(outputRoot, suite.outputDirectoryName, "aggregate.json")
+    );
+    const currentFingerprint = createDocxReleaseGateFingerprint(
+      repositoryDirectory,
+      suite
+    );
+    return [
+      suite.id,
+      aggregate?.fingerprint === currentFingerprint ? aggregate : undefined
+    ];
+  }));
+}
+
+function renderSuiteHtml(aggregate) {
+  const rows = aggregate.cases.map((entry) =>
+    `<tr><td>${escapeHtml(entry.caseId)}</td><td class="${escapeHtml(entry.status)}">${escapeHtml(entry.status)}</td><td>${escapeHtml(entry.finishedAt || "")}</td><td>${entry.summaryPath ? `<a href="${escapeHtml(entry.summaryPath)}">JSON</a>` : "-"}</td><td>${escapeHtml(entry.error || entry.gateFailures?.join("；") || "")}</td></tr>`
+  ).join("");
+  return `<!doctype html><html lang="zh-CN"><head><meta charset="utf-8"><title>${escapeHtml(aggregate.suite.label)}</title><style>body{font-family:Inter,"Microsoft YaHei",sans-serif;margin:32px;color:#172033}table{border-collapse:collapse;width:100%}th,td{border:1px solid #d0d5dd;padding:8px;text-align:left;vertical-align:top}.passed{background:#e6f7ec}.gate-failed,.error{background:#ffe4e4}</style></head><body><h1>${escapeHtml(aggregate.suite.label)}</h1><p>${aggregate.execution.completedCaseCount}/${aggregate.execution.expectedCaseCount}；通过 ${aggregate.execution.passedCaseCount}；门禁失败 ${aggregate.execution.gateFailedCaseCount}；基础设施错误 ${aggregate.execution.errorCaseCount}。</p><table><thead><tr><th>场景</th><th>状态</th><th>完成时间</th><th>报告</th><th>问题</th></tr></thead><tbody>${rows}</tbody></table></body></html>`;
+}
+
+function buildAggregate({ suite, fingerprint, records }) {
+  const execution = summarizeSuiteProgress({ suite, fingerprint, records });
+  return {
+    schemaVersion: DOCX_RELEASE_GATE_SCHEMA_VERSION,
+    generatedAt: new Date().toISOString(),
+    fingerprint,
+    suite: {
+      id: suite.id,
+      label: suite.label,
+      kind: suite.kind,
+      order: suite.order,
+      markdownPath: suite.markdownPath,
+      markdownByteLength: suite.markdownByteLength
+    },
+    execution,
+    gatePassed: execution.gatePassed,
+    cases: suite.caseIds.map((caseId) => records[caseId])
+      .filter((record) => record?.fingerprint === fingerprint)
+  };
+}
+
+function writeSuiteReports({ suiteDirectory, suite, fingerprint, records }) {
+  const aggregate = buildAggregate({ suite, fingerprint, records });
+  writeJsonAtomic(path.join(suiteDirectory, "progress.json"), {
+    schemaVersion: DOCX_RELEASE_GATE_SCHEMA_VERSION,
+    updatedAt: aggregate.generatedAt,
+    fingerprint,
+    records
+  });
+  writeJsonAtomic(path.join(suiteDirectory, "aggregate.json"), aggregate);
+  fs.writeFileSync(path.join(suiteDirectory, "aggregate.html"), renderSuiteHtml(aggregate), "utf8");
+  return aggregate;
+}
+
+function writeTopLevelIndex(suites) {
+  const entries = suites.map((suite) => ({
+    suite: {
+      id: suite.id,
+      label: suite.label,
+      kind: suite.kind,
+      order: suite.order,
+      outputDirectoryName: suite.outputDirectoryName,
+      markdownByteLength: suite.markdownByteLength
+    },
+    aggregate: readJsonIfExists(path.join(outputRoot, suite.outputDirectoryName, "aggregate.json"))
+  }));
+  const index = {
+    schemaVersion: DOCX_RELEASE_GATE_SCHEMA_VERSION,
+    generatedAt: new Date().toISOString(),
+    strategy: "four-independent-140-suites",
+    totalExpectedCaseCount: 560,
+    entries
+  };
+  writeJsonAtomic(path.join(outputRoot, "index.json"), index);
+  const rows = entries.map(({ suite, aggregate }) =>
+    `<tr><td>${suite.order + 1}</td><td>${escapeHtml(suite.label)}</td><td>${aggregate?.execution?.completedCaseCount ?? 0}/140</td><td class="${aggregate?.gatePassed ? "passed" : "pending"}">${aggregate?.gatePassed ? "通过" : aggregate ? "未通过/未完成" : "未开始"}</td><td><a href="${escapeHtml(suite.outputDirectoryName)}/aggregate.html">报告</a></td></tr>`
+  ).join("");
+  fs.writeFileSync(path.join(outputRoot, "index.html"), `<!doctype html><html lang="zh-CN"><head><meta charset="utf-8"><title>DOCX 4×140 发布门禁</title><style>body{font-family:Inter,"Microsoft YaHei",sans-serif;margin:32px;color:#172033}table{border-collapse:collapse;min-width:900px}th,td{border:1px solid #d0d5dd;padding:10px;text-align:left}.passed{background:#e6f7ec}.pending{background:#fff4cc}</style></head><body><h1>DOCX 4×140 独立发布门禁</h1><p>先通过标准合成基线，再按 Markdown 字节数从小到大逐套执行。</p><table><thead><tr><th>顺序</th><th>套件</th><th>进度</th><th>状态</th><th>独立报告</th></tr></thead><tbody>${rows}</tbody></table></body></html>`, "utf8");
+}
+
+function parseChildSummary(stdout) {
+  const text = Buffer.isBuffer(stdout) ? stdout.toString("utf8") : String(stdout || "");
+  return JSON.parse(text.trim());
+}
+
+fs.mkdirSync(outputRoot, { recursive: true });
+const suites = createDocxReleaseGateSuites(repositoryDirectory);
+const suite = resolveDocxReleaseGateSuite({
+  suites,
+  requestedId: requestedSuiteId,
+  aggregates: readAggregates(suites),
+  allowOutOfOrder
+});
+if (!suite) {
+  writeTopLevelIndex(suites);
+  process.stdout.write("四套独立 140 门禁均已通过，无需重复执行。\n");
+  process.exit(0);
+}
+if (selectedCaseId && !suite.caseIds.includes(selectedCaseId)) {
+  throw new Error(`套件 ${suite.id} 不包含场景 ${selectedCaseId}`);
+}
+
+const suiteDirectory = path.join(outputRoot, suite.outputDirectoryName);
+const logsDirectory = path.join(suiteDirectory, "logs");
+fs.mkdirSync(logsDirectory, { recursive: true });
+fs.writeFileSync(
+  path.join(suiteDirectory, "night-run.pid"),
+  `${process.pid}\n`,
+  "utf8"
+);
+const fingerprint = createDocxReleaseGateFingerprint(repositoryDirectory, suite);
+const previousProgress = readJsonIfExists(path.join(suiteDirectory, "progress.json"));
+const records = previousProgress?.records && typeof previousProgress.records === "object"
+  ? previousProgress.records
+  : {};
+const runStartedAt = new Date().toISOString();
+const runPath = path.join(suiteDirectory, "run.json");
+writeJsonAtomic(runPath, {
+  schemaVersion: DOCX_RELEASE_GATE_SCHEMA_VERSION,
+  pid: process.pid,
+  processStartedAt: runStartedAt,
+  command: process.argv,
+  suiteId: suite.id,
+  suiteFingerprint: fingerprint,
+  allowOutOfOrder,
+  ...(requestedCaseLimit ? { requestedCaseLimit } : {}),
+  status: "running"
+});
+
+const requestedCaseIds = selectedCaseId
+  ? [selectedCaseId]
+  : requestedCaseLimit
+    ? suite.caseIds.slice(0, requestedCaseLimit)
+    : suite.caseIds;
+for (const caseId of requestedCaseIds) {
+  if (records[caseId]?.fingerprint === fingerprint && records[caseId]?.status === "passed") {
+    process.stderr.write(`[DOCX gate ${suite.id}] ${caseId} 已通过，跳过\n`);
+    continue;
+  }
+  process.stderr.write(`[DOCX gate ${suite.id}] 执行 ${caseId}\n`);
+  const environment = {
+    ...process.env,
+    MD_TO_PDF_LAYOUT_VISUAL_CASE: caseId,
+    MD_TO_PDF_LAYOUT_VISUAL_OUTPUT_DIR: suiteDirectory,
+    MD_TO_PDF_LAYOUT_VISUAL_SCOPE: suite.kind === "synthetic" ? "full" : "corpus",
+    MD_TO_PDF_R4_APPEND_INLINE_CODE_PROBES: suite.kind === "synthetic" ? "1" : "0"
+  };
+  if (suite.kind === "corpus") {
+    environment.MD_TO_PDF_CORPUS_ID = suite.id;
+    environment.MD_TO_PDF_R4_MARKDOWN_PATH = suite.markdownPath;
+  } else {
+    delete environment.MD_TO_PDF_CORPUS_ID;
+    delete environment.MD_TO_PDF_R4_MARKDOWN_PATH;
+  }
+  const startedAt = new Date().toISOString();
+  let record;
+  try {
+    const { stdout, stderr } = await execFileAsync(process.execPath, [matrixScript], {
+      cwd: repositoryDirectory,
+      windowsHide: true,
+      timeout: 24 * 60 * 60 * 1000,
+      maxBuffer: 256 * 1024 * 1024,
+      env: environment
+    });
+    fs.writeFileSync(path.join(logsDirectory, `${caseId}.stdout.log`), stdout, "utf8");
+    fs.writeFileSync(path.join(logsDirectory, `${caseId}.stderr.log`), stderr, "utf8");
+    const summary = parseChildSummary(stdout);
+    record = {
+      caseId,
+      fingerprint,
+      status: summary.gatePassed ? "passed" : "gate-failed",
+      startedAt,
+      finishedAt: new Date().toISOString(),
+      gateFailures: summary.gateFailures ?? [],
+      summaryPath: `summary-${caseId}.json`,
+      matrixPath: `matrix-${caseId}.html`
+    };
+  } catch (error) {
+    const stdout = Buffer.isBuffer(error?.stdout) ? error.stdout.toString("utf8") : String(error?.stdout || "");
+    const stderr = Buffer.isBuffer(error?.stderr) ? error.stderr.toString("utf8") : String(error?.stderr || "");
+    fs.writeFileSync(path.join(logsDirectory, `${caseId}.stdout.log`), stdout, "utf8");
+    fs.writeFileSync(path.join(logsDirectory, `${caseId}.stderr.log`), stderr || String(error?.stack || error), "utf8");
+    try {
+      const summary = parseChildSummary(stdout);
+      record = {
+        caseId,
+        fingerprint,
+        status: "gate-failed",
+        startedAt,
+        finishedAt: new Date().toISOString(),
+        gateFailures: summary.gateFailures ?? [String(error?.message || error)],
+        summaryPath: `summary-${caseId}.json`,
+        matrixPath: `matrix-${caseId}.html`
+      };
+    } catch {
+      record = {
+        caseId,
+        fingerprint,
+        status: "error",
+        startedAt,
+        finishedAt: new Date().toISOString(),
+        error: String(error?.message || error)
+      };
+    }
+  }
+  records[caseId] = record;
+  writeSuiteReports({ suiteDirectory, suite, fingerprint, records });
+  writeTopLevelIndex(suites);
+}
+
+const aggregate = writeSuiteReports({ suiteDirectory, suite, fingerprint, records });
+writeTopLevelIndex(suites);
+writeJsonAtomic(runPath, {
+  schemaVersion: DOCX_RELEASE_GATE_SCHEMA_VERSION,
+  pid: process.pid,
+  processStartedAt: runStartedAt,
+  completedAt: new Date().toISOString(),
+  command: process.argv,
+  suiteId: suite.id,
+  suiteFingerprint: fingerprint,
+  status: aggregate.gatePassed ? "passed" : "completed-with-failures",
+  execution: aggregate.execution
+});
+process.stdout.write(`${JSON.stringify(aggregate, null, 2)}\n`);
+if (!selectedCaseId && !aggregate.gatePassed) {
+  process.exitCode = 1;
+}