fix: 重建DOCX视觉一致性门禁

This commit is contained in:
SkyJourney
2026-08-03 10:09:43 +08:00
parent 842c3f1782
commit 28c88faabb
11 changed files with 811 additions and 49 deletions
+31 -5
View File
@@ -1,6 +1,6 @@
# 墨呈进度
最后更新:2026-08-02
最后更新:2026-08-03
## 1. 当前概况
@@ -68,6 +68,14 @@ Python;左侧编辑器计划升级为 CodeMirror 6,并通过项目自有命
支持基础工具栏及后续 Front Matter、ECharts YAML 工具扩展。完整设计与
验收门禁见 [v0.6.0 设计文档](V0.6.0_DESIGN.md)。
`v0.6.1` 已冻结为 DOCX 视觉一致性和发布门禁修复版本。发布后复核确认
`v0.6.0` 的 14 主题矩阵主要验证 CSS 槽位、OOXML 结构和样式存在性,
六个视觉场景只覆盖五套主题;视觉报告中的像素、墨迹 IoU 和边缘 IoU
failure 还被外层白名单过滤,导致报告失败时总门禁仍可通过。`v0.6.1`
要求四套独立封面执行 Chromium、Word、WPS 整页硬门禁,正文改为跨页
无关的语义块局部视觉门禁,并对字体真实解析、嵌入和发布后三条生产链路
建立阻断验收。完整范围见 [v0.6.1 设计文档](V0.6.1_DESIGN.md)。
`v0.6.0` 阶段 2 已完成 Pandoc 运行时探查:固定 `3.9.0.2`,官方
Windows ZIP 和 Linux amd64 tarball 的 SHA-256 均已核验;Linux 静态程序
可直接运行于 `node:22-bookworm-slim`。Desktop 安装包预计增加约
@@ -1460,9 +1468,9 @@ ECharts 第二阶段浏览器与 PDF 验证结果:
Server HTTP、Chromium PDF 与生产 DOCX 矩阵;
- 阶段 12D-R4 媒体专项:已完成稳定 PNG 布局、严格 DrawingML、五媒体
Word/WPS 编辑互存、隔离逐页视觉比较和题注居中门禁;
- 阶段 12D-R4 布局视觉矩阵:已完成六组主题、纸张、方向与页边距组合的
Chromium/Word/WPS 元素级严格门禁,并重跑 14 套主题生产矩阵;分页流
差异保留为诊断项,不再误判为样式失败
- 阶段 12D-R4 布局视觉矩阵:曾记录为完成六组 Chromium/Word/WPS
元素级严格门禁;`v0.6.1` 发布后复核确认该矩阵只覆盖五套主题,且外层
过滤了像素、墨迹 IoU 和边缘 IoU failure,因此不构成有效严格门禁
- 发布收口阶段 A:已完成 v0.6.0 版本、墨呈/MorphDoc 品牌、NSIS 名称、
文件关联兼容和旧版用户数据迁移;
- 发布收口阶段 B:已完成 Desktop 固定 Pandoc 最小运行时的清单、下载
@@ -1473,10 +1481,28 @@ ECharts 第二阶段浏览器与 PDF 验证结果:
镜像封包、部署 ZIP、清单与校验和规则,并将正式字体纳入 Git LFS;
- 阶段 13:完成 Word/WPS 双向互存、外部主题兼容、体积和正式发布验收。
### 阶段二:v0.6.1 DOCX 视觉一致性修复
- 阶段 A:已完成。修复所有 failure 的发布阻断聚合,补齐四套独立封面
场景,并禁止正文回流封面页;四套封面真实 Word/WPS 复验均已被现有
视觉差异正确阻断;
- 阶段 B:已完成。正文段落按内容顺序映射为与物理页号无关的视觉块,
保留实际行坐标和 PDF.js 字体分类诊断,逐行裁剪局部栅格并输出叠加图、
热力图;正文整页栅格降为诊断,封面仍为整页硬门禁。单元测试 47 项、
全项目类型检查和生产构建通过;真实 GitHub 紧凑 Letter 纵向场景观测
22 个正文块、输出 102 张失败局部热力图并正确阻断。PDF.js 泛化字体名
不作为字体硬门禁,真实字体解析留在阶段 D;
- 阶段 C:将语义块门禁扩展为 14 套内置主题 × 纵横两个方向 × 主题默认、
标准、紧凑、宽松、非对称装订五组页边距,共 140 个排版场景;主题
默认链路即使与显式数值相同也不得去重;
- 阶段 D:修复主题封面、块元素样式和字体解析、嵌入、回退;
- 阶段 E:完成源码服务、Docker Web API、Desktop 安装版发布后复验;
- 阶段 F:完成 `v0.6.1` 版本、发行说明、发布产物、发布提交和标签。
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:后续版本扩展
### 阶段:后续版本扩展
- Front Matter 可视化编辑工具;
- ECharts YAML 可视化配置工具;
+142
View File
@@ -0,0 +1,142 @@
# v0.6.1 DOCX 视觉一致性与发布门禁修复设计
状态:范围已冻结,阶段 A、B 已完成,等待阶段 C。
## 1. 版本目标
`v0.6.1` 是针对 `v0.6.0` DOCX 导出质量的维护版本,不新增与 DOCX
无关的产品功能。本版本必须完成:
1. 修复视觉报告失败但总发布门禁仍通过的判定漏洞;
2. 对四套独立封面主题建立 Chromium、Word、WPS 整页严格门禁;
3. 对 14 套内置主题建立与正文分页位置无关的语义块局部视觉门禁;
4. 修复主题样式、封面布局和字体解析、嵌入、回退问题;
5. 对源码服务、Docker Web API 和 Desktop 安装版执行发布后复验。
本版本不要求 Chromium、Word 和 WPS 的正文总页数或分页边界完全一致,
但不允许以排版引擎不同为理由放过封面、字体或语义块样式退化。
```mermaid
flowchart LR
A["同一 Markdown、主题与配置"] --> B["生产 Chromium PDF"]
A --> C["生产 DOCX"]
C --> D["Word 原生渲染"]
C --> E["WPS 原生渲染"]
B --> F["封面整页严格比较"]
D --> F
E --> F
B --> G["正文语义块匹配"]
D --> G
E --> G
F --> H["发布硬门禁"]
G --> H
```
## 2. 问题基线
`v0.6.0` 复核确认存在以下门禁缺口:
- 视觉报告中的像素差异、墨迹 IoU、边缘 IoU 等 failure 级问题未进入
外层发布阻断集合;
- 六个布局视觉场景只覆盖五套主题,未覆盖全部 14 套主题;
- 四套独立封面只覆盖其中两套;
- “独立封面”主要检查 OOXML 分节、页眉页脚、页码重启和字段存在,
不能代表视觉一致;
- 14 主题样式门禁主要检查 CSS 槽位、令牌和 Word 样式 ID 存在,不能
代表最终 Word/WPS 呈现一致;
- 字体名称写入 `fontTable.xml`、转换无 warning,均不能证明字体已经解析、
嵌入并被 Office 实际使用;
- `red-briefing` 在门禁、Docker Web 和用户实际产物中均为零嵌入字体,
仍以零 warning 通过。
## 3. 验收模型
### 3.1 封面整页硬门禁
适用于 `formal-feasibility``tender-business-blue``tender-blind`
`tender-classic`
- 封面必须独占一个物理页;
- 所有封面字段只能出现在封面页,正文不得回流到封面页;
- 封面页不显示正文页眉、页脚和页码;
- 字体、字号、字重、颜色、对齐、位置、边框、背景和装饰线必须匹配;
- 整页像素、墨迹 IoU、边缘 IoU 任一 failure 都阻断发布;
- Word 和 WPS 都必须通过,不允许只以二者相互接近替代 Chromium 基线。
### 3.2 正文语义块门禁
正文按稳定内容顺序和语义角色匹配,不按物理页号配对。允许语义块移动到
另一页,但块自身必须保持视觉一致:
| 语义块 | 必须比较 |
| --- | --- |
| 标题、正文 | 字体、字号、字重、颜色、行高、缩进、段距、对齐 |
| 引用、提示框 | 背景、边框、内边距、文字样式 |
| 代码块 | 等宽字体、字号、背景、边框、内边距、换行 |
| 列表 | 编号或项目符号、缩进、悬挂距离、文字样式 |
| 表格 | 总宽度、列宽、边框、底纹、单元格内边距、表头与文字样式 |
| 图片与图表 | 尺寸、宽高比、清晰度、对齐和题注 |
正文整页栅格差异只作为诊断,不作为最终样式结论;没有完成语义块局部比较
的主题不得标记为视觉通过。
### 3.3 字体硬门禁
- 记录 Chromium、DOCX 字体表、嵌入字体部件、Word PDF 和 WPS PDF 的
实际字体;
- 需要跨环境稳定的字体必须来自许可证允许分发的字体包;
- 应嵌入字体但字体部件为零时直接失败;
- 字体包未应用、字体回退或 Office 未使用目标字体时必须产生阻断错误;
- 字体名称存在但没有字体二进制不得视为通过。
### 3.4 发布链路门禁
同一验收输入必须覆盖:
1. 源码测试服务;
2. 正式 Docker Web API
3. 正式 Desktop 安装版。
任一底层报告状态为 `failed` 时,总门禁必须失败。只有已经证明属于合法
正文自然分页流动的问题才能降级为 warning;封面、字体、内容、语义块
样式和媒体失败不得降级。
### 3.5 全量配置矩阵
发布门禁不是少量代表用例,而是固定执行 `14 × 2 × 5 = 140` 个排版
场景:
- 14 套内置主题;
- 纵向、横向两个纸张方向;
- 主题默认、标准、紧凑、宽松和非对称装订五组页边距。
“主题默认”必须使用 `marginMode: theme`,验证主题清单中的真实默认值;
即使解析后的数值与某组显式边距相同,也不得去重。其余四组使用显式
`custom` 配置,验证配置覆盖链路。每个场景生成 Chromium PDF、DOCX、
Word PDF 和 WPS PDF,并在报告中保留主题、方向、边距场景、导出链和
渲染引擎维度,不允许用汇总绿灯掩盖单个失败。
## 4. 实施阶段
1. 阶段 A:修复 failure 聚合策略,补齐四套封面场景和封面正文隔离;
2. 阶段 B:建立跨页无关的语义块观察、匹配、局部栅格和样式比较协议;
3. 阶段 C:建立 14 主题、2 个方向、5 组页边距的 140 场景全量语义块
矩阵和版本化门限;
4. 阶段 D:逐主题修复封面、正文块样式和字体嵌入;
5. 阶段 E:完成源码、Docker Web、Desktop 安装版发布后复验;
6. 阶段 F:更新版本、发行说明、发布产物、发布提交和 annotated tag。
每个阶段必须先提交真实失败基线,再修复到通过。不得通过删除用例、提高
门限、过滤 failure 或把 failure 改名为 warning 取得绿灯。
## 5. v0.6.1 发布条件
- 14 套主题在纵向、横向和五组页边距下全部生成真实 Chromium PDF、
DOCX、Word PDF 和 WPS PDF
- 四套独立封面全部通过整页严格门禁;
- 14 套主题声明支持的语义块全部通过局部视觉门禁;
- 字体解析、嵌入和实际使用结果符合主题预期;
- 源码、Docker Web 和 Desktop 安装版结果一致;
- 单元测试、类型检查、生产构建和 `git diff --check` 通过;
- 发布报告中不存在被忽略的 failure;
- 完成真实 Word/WPS 逐页人工复核后才能发布。
@@ -6,6 +6,7 @@ export * from "./paragraph-layout.js";
export * from "./page-semantics.js";
export * from "./raster.js";
export * from "./report.js";
export * from "./semantic-block.js";
export * from "./snapshot.js";
export * from "./text.js";
export * from "./types.js";
@@ -113,6 +113,17 @@ function observeParagraph(
Array.from(lines[index]?.normalizedText ?? "").length ===
(matchedByLine.get(index)?.length ?? 0),
);
const visualLines = matchedLines.map((line) => ({
pageNumber: line.pageNumber,
bounds: line.line.bounds,
fontFamilies: [
...new Set(
line.line.items.flatMap((item) =>
item.fontFamily ? [item.fontFamily] : [],
),
),
].sort(),
}));
const matched = matchedCharacterCount === expectedCharacters.length;
if (matched) {
cursor.lineIndex = localCursor.lineIndex;
@@ -146,6 +157,7 @@ function observeParagraph(
? {}
: { averageBaselineGapPt }),
exclusiveGeometry,
visualLines,
};
}
@@ -192,6 +204,8 @@ function compareParagraph(
expectation: EditableParagraphExpectation,
baseline: PdfParagraphLayoutObservation,
candidate: PdfParagraphLayoutObservation,
baselineCoverPageCount: number,
candidateCoverPageCount: number,
): VisualDiffIssue[] {
const issues: VisualDiffIssue[] = [];
if (!baseline.matched || !candidate.matched) {
@@ -299,6 +313,27 @@ function compareParagraph(
},
});
}
} else if (baselineCoverPageCount > 0 || candidateCoverPageCount > 0) {
const baselineOutsideCover = baseline.pageNumbers.every(
(pageNumber) => pageNumber > baselineCoverPageCount,
);
const candidateOutsideCover = candidate.pageNumbers.every(
(pageNumber) => pageNumber > candidateCoverPageCount,
);
if (!baselineOutsideCover || !candidateOutsideCover) {
issues.push({
code: "COVER_LAYOUT_MISMATCH",
severity: "failure",
message: `${expectation.index + 1} 个正文段落回流到独立封面页`,
details: {
paragraphIndex: expectation.index,
baselinePages: baseline.pageNumbers.join(","),
candidatePages: candidate.pageNumbers.join(","),
baselineCoverPageCount,
candidateCoverPageCount,
},
});
}
}
if (baseline.exclusiveGeometry && candidate.exclusiveGeometry) {
@@ -373,6 +408,12 @@ export function comparePdfEditableParagraphLayouts(
const candidateCharacters = candidateLines.map((line) =>
Array.from(line.normalizedText),
);
const baselineCoverPageCount = baselinePageSemantics.filter(
(page) => page.kind === "cover",
).length;
const candidateCoverPageCount = candidatePageSemantics.filter(
(page) => page.kind === "cover",
).length;
const baselineCursor: MatchCursor = { lineIndex: 0, characterIndex: 0 };
const candidateCursor: MatchCursor = { lineIndex: 0, characterIndex: 0 };
@@ -393,6 +434,8 @@ export function comparePdfEditableParagraphLayouts(
expectation,
baselineObservation,
candidateObservation,
baselineCoverPageCount,
candidateCoverPageCount,
);
return {
expectation,
+84 -11
View File
@@ -1,6 +1,7 @@
import { comparePdfSnapshotsBasic, DEFAULT_VISUAL_DIFF_THRESHOLDS } from "./compare.js";
import { comparePageRasters } from "./raster.js";
import { comparePdfPageSemantics } from "./page-semantics.js";
import { comparePdfSemanticBlockVisuals } from "./semantic-block.js";
import type {
CreatePdfVisualDiffOptions,
PdfDocumentSnapshot,
@@ -21,6 +22,12 @@ function resolveStatus(
return issues.length > 0 ? "warning" : "passed";
}
export function getBlockingVisualDiffIssues(
report: Pick<PdfVisualDiffReport, "issues">,
): VisualDiffIssue[] {
return report.issues.filter((issue) => issue.severity === "failure");
}
function rasterIssues(
pair: Extract<PdfPagePair, { status: "paired" }>,
metrics: NonNullable<PdfVisualPageComparison["metrics"]>,
@@ -91,7 +98,7 @@ async function compareVisualPage(
baseline: PdfDocumentSnapshot,
candidate: PdfDocumentSnapshot,
thresholds: VisualDiffThresholds,
rasterComparisonMode: "strict" | "body-flow" = "strict",
rasterComparisonMode: "strict" | "body-flow" | "body-block" = "strict",
): Promise<PdfVisualPageComparison> {
if (pair.status !== "paired") {
const page =
@@ -101,7 +108,7 @@ async function compareVisualPage(
return {
pair,
rasterComparisonMode,
status: rasterComparisonMode === "body-flow" ? "warning" : "failed",
status: rasterComparisonMode === "strict" ? "failed" : "warning",
...(page?.raster
? {
unpairedPng: page.raster.png,
@@ -109,12 +116,18 @@ async function compareVisualPage(
}
: {}),
issues:
rasterComparisonMode === "body-flow"
rasterComparisonMode !== "strict"
? [
{
code: "BODY_FLOW_RASTER_DIFFERENCE",
code:
rasterComparisonMode === "body-flow"
? "BODY_FLOW_RASTER_DIFFERENCE"
: "BODY_PAGE_RASTER_DIFFERENCE",
severity: "warning",
message: "正文自然分页产生未配对物理页,保留页面快照供审查",
message:
rasterComparisonMode === "body-flow"
? "正文自然分页产生未配对物理页,保留页面快照供审查"
: "正文物理页未配对,整页栅格仅作为诊断,最终由语义块门禁判定",
...(pair.status === "baseline-only"
? { baselinePageNumber: pair.baselinePageNumber }
: { candidatePageNumber: pair.candidatePageNumber }),
@@ -128,14 +141,15 @@ async function compareVisualPage(
if (!baselinePage?.raster || !candidatePage?.raster) {
const issue: VisualDiffIssue = {
code: "RASTER_MISSING",
severity: "failure",
severity: rasterComparisonMode === "strict" ? "failure" : "warning",
message: `${pair.baselinePageNumber} 页缺少栅格快照`,
baselinePageNumber: pair.baselinePageNumber,
candidatePageNumber: pair.candidatePageNumber,
};
return {
pair,
status: "failed",
rasterComparisonMode,
status: rasterComparisonMode === "strict" ? "failed" : "warning",
issues: [issue],
};
}
@@ -161,12 +175,18 @@ async function compareVisualPage(
);
const strictIssues = rasterIssues(pair, result.metrics, thresholds);
const issues =
rasterComparisonMode === "body-flow" && strictIssues.length > 0
rasterComparisonMode !== "strict" && strictIssues.length > 0
? [
{
code: "BODY_FLOW_RASTER_DIFFERENCE" as const,
code:
rasterComparisonMode === "body-flow"
? ("BODY_FLOW_RASTER_DIFFERENCE" as const)
: ("BODY_PAGE_RASTER_DIFFERENCE" as const),
severity: "warning" as const,
message: `${pair.baselinePageNumber} 页栅格差异来自已验证的正文自然分页流动`,
message:
rasterComparisonMode === "body-flow"
? `${pair.baselinePageNumber} 页栅格差异来自已验证的正文自然分页流动`
: `${pair.baselinePageNumber} 页为正文页,整页栅格差异仅作诊断,最终由语义块门禁判定`,
baselinePageNumber: pair.baselinePageNumber,
candidatePageNumber: pair.candidatePageNumber,
details: {
@@ -218,6 +238,26 @@ export async function createPdfVisualDiffReport(
? {}
: { expectedEditableParagraphs: options.expectedEditableParagraphs }),
});
if (basic.paragraphLayouts) {
const semanticBlockVisuals = await comparePdfSemanticBlockVisuals(
baseline,
candidate,
basic.paragraphLayouts,
thresholds,
);
basic.semanticBlockVisuals = semanticBlockVisuals;
basic.issues.push(
...semanticBlockVisuals.flatMap((comparison) => comparison.issues),
);
basic.status = resolveStatus(basic.issues);
if (basic.bodyFlow) {
basic.bodyFlow.legal =
basic.bodyFlow.legal &&
semanticBlockVisuals.every(
(comparison) => comparison.status !== "failed",
);
}
}
const pages: PdfVisualPageComparison[] = [];
for (const pair of basic.pagePairs) {
const baselineSemanticExpectation =
@@ -244,12 +284,21 @@ export async function createPdfVisualDiffReport(
pair.candidatePageNumber,
))),
);
const isSemanticBodyPage =
baselineSemanticExpectation?.kind !== "cover" &&
candidateSemanticExpectation?.kind !== "cover" &&
Boolean(baselineSemanticExpectation || candidateSemanticExpectation) &&
(basic.semanticBlockVisuals?.length ?? 0) > 0;
const page = await compareVisualPage(
pair,
baseline,
candidate,
thresholds,
flowAffected ? "body-flow" : "strict",
flowAffected
? "body-flow"
: isSemanticBodyPage
? "body-block"
: "strict",
);
const primarySemanticExpectation =
baselineSemanticExpectation ?? candidateSemanticExpectation;
@@ -366,6 +415,8 @@ function pageTitle(page: PdfVisualPageComparison): string {
const mode =
page.rasterComparisonMode === "body-flow"
? "(正文流动比较)"
: page.rasterComparisonMode === "body-block"
? "(正文整页诊断)"
: "(严格比较)";
if (page.pair.status === "paired") {
return `基线第 ${page.pair.baselinePageNumber} 页 ↔ 候选第 ${page.pair.candidatePageNumber}${mode}`;
@@ -376,6 +427,27 @@ function pageTitle(page: PdfVisualPageComparison): string {
return `候选第 ${page.pair.candidatePageNumber} 页为溢出页${mode}`;
}
function renderSemanticBlockVisuals(report: PdfVisualDiffReport): string {
const comparisons = report.basic.semanticBlockVisuals;
if (!comparisons) {
return "";
}
const failures = comparisons.filter(
(comparison) => comparison.status === "failed",
);
const rows = (failures.length > 0 ? failures : comparisons.slice(0, 8))
.map((comparison) => {
const expectation = comparison.expectation;
return `<tr><td>${expectation.index + 1}</td><td>${escapeHtml(expectation.section)}</td><td>${escapeHtml(expectation.role)}</td><td>${escapeHtml(comparison.baselineFontFamilies.join(", ") || "未识别")}</td><td>${escapeHtml(comparison.candidateFontFamilies.join(", ") || "未识别")}</td><td>${comparison.lines.length}</td><td>${escapeHtml(comparison.issues.map((issue) => issue.code).join(", ") || "通过")}</td></tr>`;
})
.join("");
return `<div class="semantic-block-visuals">
<h2>正文语义块局部视觉 <span class="status ${failures.length > 0 ? "failed" : "passed"}">${failures.length > 0 ? "failed" : "passed"}</span></h2>
<p>正文语义块 ${comparisons.length} 个,失败 ${failures.length} 个;块按内容顺序匹配,局部栅格参与硬门禁,物理页号不参与。PDF.js 字体分类仅作诊断,真实字体解析由独立字体门禁判定。</p>
<table><thead><tr><th>块</th><th>分节</th><th>角色</th><th>基线字体分类</th><th>候选字体分类</th><th>局部行</th><th>结果</th></tr></thead><tbody>${rows}</tbody></table>
</div>`;
}
function renderParagraphLayouts(report: PdfVisualDiffReport): string {
const paragraphs = report.basic.paragraphLayouts;
if (!paragraphs) {
@@ -451,6 +523,7 @@ export function renderPdfVisualDiffHtml(
<p><strong>生成时间:</strong>${escapeHtml(report.generatedAt)} <strong>正文相似度:</strong>${percentage(report.basic.contentSimilarity)}</p>
${editableMetrics}
${renderParagraphLayouts(report)}
${renderSemanticBlockVisuals(report)}
<table><thead><tr><th>级别</th><th>代码</th><th>说明</th></tr></thead><tbody>${issueRows}</tbody></table>
</section>
${pages}
@@ -0,0 +1,272 @@
import { createHash } from "node:crypto";
import { createCanvas, loadImage } from "@napi-rs/canvas";
import { comparePageRasters } from "./raster.js";
import type {
PdfDocumentSnapshot,
PdfPageRaster,
PdfParagraphLayoutComparison,
PdfPointBounds,
PdfSemanticBlockVisualComparison,
PdfSemanticBlockVisualLineComparison,
VisualDiffIssue,
VisualDiffThresholds,
} from "./types.js";
const BLOCK_RASTER_PADDING_PT = 4;
function sha256(bytes: Uint8Array): string {
return createHash("sha256").update(bytes).digest("hex");
}
function status(issues: readonly VisualDiffIssue[]) {
return issues.some((issue) => issue.severity === "failure")
? ("failed" as const)
: issues.length > 0
? ("warning" as const)
: ("passed" as const);
}
function normalizeFontFamily(value: string): string {
return value
.normalize("NFKC")
.replace(/^[A-Z]{6}\+/u, "")
.replace(/["']/gu, "")
.replace(/\s+/gu, " ")
.trim()
.toLocaleLowerCase("en-US");
}
function collectFonts(
lines: PdfParagraphLayoutComparison["baseline"]["visualLines"],
): string[] {
return [
...new Set(
lines.flatMap((line) => line.fontFamilies.map(normalizeFontFamily)),
),
].sort();
}
async function cropRaster(
raster: PdfPageRaster,
bounds: PdfPointBounds,
): Promise<PdfPageRaster> {
const pixelsPerPoint = raster.dpi / 72;
const left = Math.max(
0,
Math.floor((bounds.x - BLOCK_RASTER_PADDING_PT) * pixelsPerPoint),
);
const top = Math.max(
0,
Math.floor((bounds.y - BLOCK_RASTER_PADDING_PT) * pixelsPerPoint),
);
const right = Math.min(
raster.widthPx,
Math.ceil(
(bounds.x + bounds.width + BLOCK_RASTER_PADDING_PT) * pixelsPerPoint,
),
);
const bottom = Math.min(
raster.heightPx,
Math.ceil(
(bounds.y + bounds.height + BLOCK_RASTER_PADDING_PT) * pixelsPerPoint,
),
);
const widthPx = Math.max(1, right - left);
const heightPx = Math.max(1, bottom - top);
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);
context.drawImage(
source,
left,
top,
widthPx,
heightPx,
0,
0,
widthPx,
heightPx,
);
const png = Uint8Array.from(canvas.toBuffer("image/png"));
return {
widthPx,
heightPx,
dpi: raster.dpi,
sha256: sha256(png),
png,
};
}
async function padRaster(
raster: PdfPageRaster,
widthPx: number,
heightPx: number,
): Promise<PdfPageRaster> {
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);
context.drawImage(source, 0, 0);
const png = Uint8Array.from(canvas.toBuffer("image/png"));
return {
widthPx,
heightPx,
dpi: raster.dpi,
sha256: sha256(png),
png,
};
}
function localRasterIssues(
paragraphIndex: number,
lineIndex: number,
metrics: NonNullable<PdfSemanticBlockVisualLineComparison["metrics"]>,
thresholds: VisualDiffThresholds,
): VisualDiffIssue[] {
const failed =
metrics.meanAbsoluteError > thresholds.maxMeanAbsoluteError ||
metrics.changedPixelRatio > thresholds.maxChangedPixelRatio ||
metrics.inkIou < thresholds.minInkIou ||
metrics.edgeIou < thresholds.minEdgeIou;
if (!failed) {
return [];
}
return [
{
code: "SEMANTIC_BLOCK_RASTER_MISMATCH",
severity: "failure",
message: `${paragraphIndex + 1} 个语义块的第 ${lineIndex + 1} 行局部视觉差异超限`,
details: {
paragraphIndex,
lineIndex,
meanAbsoluteError: metrics.meanAbsoluteError,
changedPixelRatio: metrics.changedPixelRatio,
inkIou: metrics.inkIou,
edgeIou: metrics.edgeIou,
},
},
];
}
export async function comparePdfSemanticBlockVisuals(
baseline: PdfDocumentSnapshot,
candidate: PdfDocumentSnapshot,
paragraphs: readonly PdfParagraphLayoutComparison[],
thresholds: VisualDiffThresholds,
): Promise<PdfSemanticBlockVisualComparison[]> {
const comparisons: PdfSemanticBlockVisualComparison[] = [];
for (const paragraph of paragraphs) {
if (paragraph.expectation.section !== "body") {
continue;
}
const issues: VisualDiffIssue[] = [];
const baselineFonts = collectFonts(paragraph.baseline.visualLines);
const candidateFonts = collectFonts(paragraph.candidate.visualLines);
const lines: PdfSemanticBlockVisualLineComparison[] = [];
const comparableLineCount = Math.min(
paragraph.baseline.visualLines.length,
paragraph.candidate.visualLines.length,
);
if (
!paragraph.baseline.matched ||
!paragraph.candidate.matched ||
comparableLineCount === 0
) {
issues.push({
code: "SEMANTIC_BLOCK_VISUAL_UNAVAILABLE",
severity: "failure",
message: `${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 (
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];
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: `${paragraph.expectation.index + 1} 个语义块的第 ${lineIndex + 1} 行缺少栅格`,
details: {
paragraphIndex: paragraph.expectation.index,
lineIndex,
},
};
issues.push(unavailable);
lines.push({
lineIndex,
baselinePageNumber: baselineLine?.pageNumber ?? -1,
candidatePageNumber: candidateLine?.pageNumber ?? -1,
issues: [unavailable],
});
continue;
}
const [baselineCrop, candidateCrop] = await Promise.all([
cropRaster(baselinePage.raster, baselineLine.bounds),
cropRaster(candidatePage.raster, candidateLine.bounds),
]);
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 lineIssues = localRasterIssues(
paragraph.expectation.index,
lineIndex,
result.metrics,
thresholds,
);
issues.push(...lineIssues);
lines.push({
lineIndex,
baselinePageNumber: baselineLine.pageNumber,
candidatePageNumber: candidateLine.pageNumber,
metrics: result.metrics,
...(lineIssues.length > 0 ? { artifacts: result.artifacts } : {}),
issues: lineIssues,
});
}
}
comparisons.push({
expectation: paragraph.expectation,
status: status(issues),
baselineFontFamilies: baselineFonts,
candidateFontFamilies: candidateFonts,
lines,
issues,
});
}
return comparisons;
}
+15 -1
View File
@@ -21,6 +21,10 @@ interface PdfJsTextItem {
hasEOL: boolean;
}
interface PdfJsTextStyle {
fontFamily?: string;
}
export const DEFAULT_PDF_SNAPSHOT_LIMITS: PdfSnapshotLimits = {
maxPdfBytes: 100 * 1024 * 1024,
maxPages: 500,
@@ -55,6 +59,7 @@ function isTextItem(item: unknown): item is PdfJsTextItem {
function createTextItemSnapshot(
item: PdfJsTextItem,
pageHeightPt: number,
styles: Readonly<Record<string, PdfJsTextStyle>>,
): PdfTextItemSnapshot | undefined {
const normalizedText = normalizePdfText(item.str);
if (!normalizedText) {
@@ -72,6 +77,7 @@ function createTextItemSnapshot(
0,
);
const baselineY = pageHeightPt - baselineFromBottom;
const fontFamily = styles[item.fontName]?.fontFamily;
return {
text: item.str,
normalizedText,
@@ -83,6 +89,7 @@ function createTextItemSnapshot(
},
baselineY,
fontName: item.fontName,
...(fontFamily ? { fontFamily } : {}),
hasEol: item.hasEOL,
};
}
@@ -138,11 +145,18 @@ export async function createPdfDocumentSnapshot(
disableNormalization: false,
includeMarkedContent: false,
});
const styles = textContent.styles as Readonly<
Record<string, PdfJsTextStyle>
>;
const items = textContent.items.flatMap((item) => {
if (!isTextItem(item)) {
return [];
}
const snapshot = createTextItemSnapshot(item, pointViewport.height);
const snapshot = createTextItemSnapshot(
item,
pointViewport.height,
styles,
);
return snapshot ? [snapshot] : [];
});
const lines = aggregatePdfTextLines(items, pointViewport.height);
+32 -2
View File
@@ -19,6 +19,7 @@ export interface PdfTextItemSnapshot {
bounds: PdfPointBounds;
baselineY: number;
fontName?: string;
fontFamily?: string;
hasEol: boolean;
}
@@ -68,6 +69,13 @@ export interface PdfParagraphLayoutObservation {
averageLineHeightPt?: number;
averageBaselineGapPt?: number;
exclusiveGeometry: boolean;
visualLines: PdfParagraphVisualLineObservation[];
}
export interface PdfParagraphVisualLineObservation {
pageNumber: number;
bounds: PdfPointBounds;
fontFamilies: string[];
}
export interface PdfParagraphLayoutComparison {
@@ -78,6 +86,24 @@ export interface PdfParagraphLayoutComparison {
issues: VisualDiffIssue[];
}
export interface PdfSemanticBlockVisualLineComparison {
lineIndex: number;
baselinePageNumber: number;
candidatePageNumber: number;
metrics?: PdfRasterDiffMetrics;
artifacts?: PdfRasterDiffArtifacts;
issues: VisualDiffIssue[];
}
export interface PdfSemanticBlockVisualComparison {
expectation: EditableParagraphExpectation;
status: "passed" | "warning" | "failed";
baselineFontFamilies: string[];
candidateFontFamilies: string[];
lines: PdfSemanticBlockVisualLineComparison[];
issues: VisualDiffIssue[];
}
export interface PdfBodyFlowComparison {
detected: boolean;
legal: boolean;
@@ -171,9 +197,12 @@ export type VisualDiffIssueCode =
| "BODY_LINE_COUNT_NEAR_BOUNDARY"
| "PARAGRAPH_LINE_BREAK_MISMATCH"
| "PARAGRAPH_LINE_GEOMETRY_MISMATCH"
| "SEMANTIC_BLOCK_VISUAL_UNAVAILABLE"
| "SEMANTIC_BLOCK_RASTER_MISMATCH"
| "COVER_LAYOUT_MISMATCH"
| "BODY_FLOW_PAGE_COUNT_DIFFERENCE"
| "BODY_FLOW_RASTER_DIFFERENCE";
| "BODY_FLOW_RASTER_DIFFERENCE"
| "BODY_PAGE_RASTER_DIFFERENCE";
export interface VisualDiffIssue {
code: VisualDiffIssueCode;
@@ -210,6 +239,7 @@ export interface PdfBasicComparison {
candidateEditableSimilarity?: number;
candidateEditableExact?: boolean;
paragraphLayouts?: PdfParagraphLayoutComparison[];
semanticBlockVisuals?: PdfSemanticBlockVisualComparison[];
bodyFlow?: PdfBodyFlowComparison;
pagePairs: PdfPagePair[];
issues: VisualDiffIssue[];
@@ -255,7 +285,7 @@ export interface PdfRasterDiffResult {
export interface PdfVisualPageComparison {
pair: PdfPagePair;
rasterComparisonMode?: "strict" | "body-flow";
rasterComparisonMode?: "strict" | "body-flow" | "body-block";
status: "passed" | "warning" | "failed";
metrics?: PdfRasterDiffMetrics;
artifacts?: PdfRasterDiffArtifacts;
@@ -166,6 +166,35 @@ describe("PDF 内容感知段落版式门禁", () => {
);
});
it("拒绝正文内容回流到独立封面页", () => {
const baseline = snapshot("baseline", [
[line("独立封面", 200)],
[line("第一章项目概述", 100)],
]);
const candidate = snapshot("candidate", [
[line("独立封面", 200), line("第一章项目概述", 700)],
[],
]);
const pageSemantics = [
{ physicalPageNumber: 1, kind: "cover" as const },
{ physicalPageNumber: 2, kind: "body-first" as const },
];
const [result] = comparePdfEditableParagraphLayouts(
baseline,
candidate,
[expectation("第一章项目概述")],
pageSemantics,
pageSemantics,
);
expect(result?.status).toBe("failed");
expect(result?.issues).toContainEqual(
expect.objectContaining({
code: "COVER_LAYOUT_MISMATCH",
severity: "failure",
}),
);
});
it("忽略 PDF 文本提取器失真的单行宽度但保留诊断值", () => {
const baseline = snapshot("baseline", [
[line("数据治理平台建设项目", 300, 224)],
@@ -5,6 +5,7 @@ import { describe, expect, it } from "vitest";
import {
createPdfVisualDiffReport,
getBlockingVisualDiffIssues,
renderPdfVisualDiffHtml,
serializePdfVisualDiffJson,
type PdfDocumentSnapshot,
@@ -30,6 +31,40 @@ function raster(color: string): PdfPageRaster {
};
}
function flowRaster(color: string): PdfPageRaster {
const canvas = createCanvas(40, 50);
const context = canvas.getContext("2d");
context.fillStyle = "#ffffff";
context.fillRect(0, 0, 40, 50);
context.fillStyle = color;
context.fillRect(34, 5, 6, 40);
const png = Uint8Array.from(canvas.toBuffer("image/png"));
return {
widthPx: 40,
heightPx: 50,
dpi: 144,
sha256: createHash("sha256").update(png).digest("hex"),
png,
};
}
function blockRaster(color: string): PdfPageRaster {
const canvas = createCanvas(40, 50);
const context = canvas.getContext("2d");
context.fillStyle = "#ffffff";
context.fillRect(0, 0, 40, 50);
context.fillStyle = color;
context.fillRect(8, 12, 20, 10);
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,
@@ -82,7 +117,7 @@ function flowSnapshot(
items: [],
lines: [...lines],
contentText: lines.map((line) => line.normalizedText).join(""),
raster: raster(color),
raster: flowRaster(color),
})),
};
}
@@ -387,7 +422,7 @@ describe("PDF 视觉差异报告", () => {
);
});
it("整段跨页同时发生换行变化时仍保持严格失败", async () => {
it("整段跨页同时发生换行变化时由语义块门禁失败", async () => {
const baseline = flowSnapshot(
"baseline",
[[contentLine("甲", 8)], [contentLine("乙丙", 8)]],
@@ -411,7 +446,7 @@ describe("PDF 视觉差异报告", () => {
expect(report.status).toBe("failed");
expect(report.basic.bodyFlow).toMatchObject({ detected: true, legal: false });
expect(report.pages.every(
(page) => page.rasterComparisonMode === "strict",
(page) => page.rasterComparisonMode === "body-block",
)).toBe(true);
expect(report.issues.map((issue) => issue.code)).toContain(
"TEXT_BLOCK_LINE_COUNT_MISMATCH",
@@ -498,4 +533,78 @@ describe("PDF 视觉差异报告", () => {
"PAGE_NUMBER_VISIBILITY_MISMATCH",
);
});
it("正文整页差异只作诊断,但语义块局部视觉变化必须失败", async () => {
const baseline = flowSnapshot(
"baseline",
[[contentLine("甲", 8)]],
"#111111",
);
const candidate = flowSnapshot(
"candidate",
[[contentLine("甲", 8)]],
"#111111",
);
baseline.pages[0]!.raster = blockRaster("#111111");
candidate.pages[0]!.raster = blockRaster("#cc0000");
const report = await createPdfVisualDiffReport(baseline, candidate, {
expectedEditableText: "甲",
expectedEditableParagraphs: [flowParagraphs[0]!],
pageSemantics: bodySemantics(1),
});
expect(report.status).toBe("failed");
expect(report.pages[0]?.rasterComparisonMode).toBe("body-block");
expect(report.issues.map((issue) => issue.code)).toContain(
"BODY_PAGE_RASTER_DIFFERENCE",
);
expect(report.issues.map((issue) => issue.code)).toContain(
"SEMANTIC_BLOCK_RASTER_MISMATCH",
);
});
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")),
snapshot("candidate", raster("#cc0000")),
{ pageSemantics: bodySemantics(1) },
);
expect(report.status).toBe("failed");
expect(report.pages[0]?.rasterComparisonMode).toBe("strict");
expect(report.issues.map((issue) => issue.code)).toContain(
"CHANGED_PIXEL_RATIO_EXCEEDED",
);
});
it("所有 failure 级视觉问题都必须进入发布阻断集合", () => {
const failure = {
code: "INK_IOU_BELOW_THRESHOLD" as const,
severity: "failure" as const,
message: "墨迹 IoU 未达到门限",
};
const warning = {
code: "BODY_FLOW_RASTER_DIFFERENCE" as const,
severity: "warning" as const,
message: "正文自然分页差异",
};
expect(getBlockingVisualDiffIssues({ issues: [failure, warning] })).toEqual([
failure,
]);
});
it("正文自然分页 warning 不阻断发布", () => {
expect(
getBlockingVisualDiffIssues({
issues: [
{
code: "BODY_FLOW_RASTER_DIFFERENCE",
severity: "warning",
message: "正文自然分页差异",
},
],
}),
).toEqual([]);
});
});
+50 -27
View File
@@ -10,6 +10,7 @@ import {
createPdfVisualDiffReport,
createWordPdfAdapter,
createWpsPdfAdapter,
getBlockingVisualDiffIssues,
renderPdfVisualDiffHtml,
serializePdfVisualDiffJson
} from "../packages/document-visual-diff/dist/index.js";
@@ -41,24 +42,6 @@ const pageSizes = {
A4: { portrait: [595.276, 841.89], landscape: [841.89, 595.276] },
Letter: { portrait: [612, 792], landscape: [792, 612] }
};
const strictIssueCodes = new Set([
"PAGE_SIZE_MISMATCH",
"PAGE_ORIENTATION_MISMATCH",
"CONTENT_MISMATCH",
"PAGE_NUMBER_VISIBILITY_MISMATCH",
"PAGE_NUMBER_TEXT_MISMATCH",
"PAGE_NUMBER_ALIGNMENT_MISMATCH",
"PAGE_NUMBER_VERTICAL_POSITION_MISMATCH",
"HEADER_VISIBILITY_MISMATCH",
"HEADER_TEXT_MISSING",
"HEADER_VERTICAL_POSITION_MISMATCH",
"EDITABLE_PARAGRAPH_MAPPING_MISMATCH",
"TEXT_BLOCK_LINE_COUNT_MISMATCH",
"PARAGRAPH_LINE_BREAK_MISMATCH",
"PARAGRAPH_LINE_GEOMETRY_MISMATCH",
"COVER_LAYOUT_MISMATCH"
]);
const cases = [
{
id: "official-theme-a4-portrait",
@@ -124,6 +107,20 @@ const cases = [
}
}
},
{
id: "tender-blind-cover-a4-portrait",
themeId: "tender-blind",
coverPageCount: 1,
paper: { format: "A4", orientation: "portrait" },
config: {}
},
{
id: "tender-classic-cover-a4-portrait",
themeId: "tender-classic",
coverPageCount: 1,
paper: { format: "A4", orientation: "portrait" },
config: {}
},
{
id: "formal-custom-a4-landscape",
themeId: "formal-regulation",
@@ -379,8 +376,7 @@ function buildPageSemantics(pageCount, coverPageCount, config) {
}
function reportGateFailures(report, label) {
const failures = report.issues
.filter((issue) => strictIssueCodes.has(issue.code))
const failures = getBlockingVisualDiffIssues(report)
.map((issue) => `${label}: ${issue.code} - ${issue.message}`);
if (report.basic.baselineEditableCoverage !== 1) {
failures.push(
@@ -400,18 +396,21 @@ function reportSummary(report) {
const paragraphIssues = report.basic.paragraphLayouts?.flatMap(
(paragraph) => paragraph.issues
) ?? [];
const semanticBlocks = report.basic.semanticBlockVisuals ?? [];
const blockingIssueCodes = [
...new Set(getBlockingVisualDiffIssues(report).map((issue) => issue.code))
];
return {
status: report.status,
issueCodes: [...new Set(report.issues.map((issue) => issue.code))],
strictIssueCodes: [
...new Set(
report.issues
.filter((issue) => strictIssueCodes.has(issue.code))
.map((issue) => issue.code)
)
],
blockingIssueCodes,
strictIssueCodes: blockingIssueCodes,
editableExact: report.basic.candidateEditableExact,
paragraphIssueCount: paragraphIssues.length,
semanticBlockCount: semanticBlocks.length,
failedSemanticBlockCount: semanticBlocks.filter(
(comparison) => comparison.status === "failed"
).length,
bodyFlow: report.basic.bodyFlow
};
}
@@ -453,6 +452,30 @@ function writeReportArtifacts(directory, prefix, report) {
page.artifacts.heatmapPng
);
}
for (const comparison of report.basic.semanticBlockVisuals ?? []) {
for (const line of comparison.lines) {
if (!line.artifacts) {
continue;
}
const stem = `${prefix}-block-${comparison.expectation.index + 1}-line-${line.lineIndex + 1}`;
fs.writeFileSync(
path.join(directory, `${stem}-baseline.png`),
line.artifacts.baselinePng
);
fs.writeFileSync(
path.join(directory, `${stem}-candidate.png`),
line.artifacts.candidatePng
);
fs.writeFileSync(
path.join(directory, `${stem}-overlay.png`),
line.artifacts.overlayPng
);
fs.writeFileSync(
path.join(directory, `${stem}-heatmap.png`),
line.artifacts.heatmapPng
);
}
}
}
fs.mkdirSync(outputDirectory, { recursive: true });