@dsh-libing/knowledge-base 0.1.0

package/lib/document/pdf-worker.js

@@ -0,0 +1,210 @@
+/** Isolated PDF.js engine: original strings and bounded raster previews, without shell dependencies. */
+import { parentPort, workerData } from 'node:worker_threads';
+import { readFile } from 'node:fs/promises';
+import { dirname, join } from 'node:path';
+import { createRequire } from 'node:module';
+import { getDocument, OPS, Util } from 'pdfjs-dist/legacy/build/pdf.mjs';
+import { createCanvas, Path2D } from '@napi-rs/canvas';
+import { z } from 'zod';
+import { bboxSchema } from "../data.js";
+import { message, required } from "../util.js";
+const port = required(parentPort, 'PDF worker port');
+const input = z.object({ file: z.string(), config: z.object({ maxInputBytes: z.number().positive(), pdfMaxPages: z.number().positive(), pdfMaxPagePixels: z.number().positive() }) }).parse(workerData);
+const matrixSchema = z.tuple([z.number(), z.number(), z.number(), z.number(), z.number(), z.number()]);
+const markTag = (value) => z.union([z.string(), z.object({ name: z.string() })]).transform(value => typeof value === 'string' ? value : value.name).parse(value);
+const numeric = (value) => z.array(z.number().finite()).parse(value instanceof Float32Array || value instanceof Float64Array ? Array.from(value) : value);
+// PDF.js uses napi-rs on Node. Correct only subpixel clip skew in this isolated worker's supported canvas prototype.
+const proto = Object.getPrototypeOf(createCanvas(1, 1).getContext('2d')), nativeClip = proto.clip;
+proto.clip = function (path, rule) {
+    const original = this.getTransform(), tiny = (value) => value !== 0 && Math.abs(value) < 1e-7;
+    if (path instanceof Path2D && (tiny(original.b) || tiny(original.c))) {
+        const matrix = this.getTransform(), transformed = new Path2D();
+        if (tiny(matrix.b))
+            matrix.b = 0;
+        if (tiny(matrix.c))
+            matrix.c = 0;
+        transformed.addPath(path, matrix);
+        this.resetTransform();
+        try {
+            Reflect.apply(nativeClip, this, rule === undefined ? [transformed] : [transformed, rule]);
+        }
+        finally {
+            this.setTransform(original);
+        }
+    }
+    else
+        Reflect.apply(nativeClip, this, path === undefined ? [] : rule === undefined ? [path] : [path, rule]);
+};
+const require = createRequire(import.meta.url), base = dirname(require.resolve('pdfjs-dist/package.json')), config = input.config;
+let doc, activePage = 0;
+const engineWarnings = new Map();
+console.warn = (...parts) => {
+    const value = parts.map(String).join(' ').replace(/^Warning:\s*/u, '').slice(0, 1000), list = engineWarnings.get(activePage) || [];
+    if (list.length < 100 && !list.includes(value))
+        list.push(value);
+    engineWarnings.set(activePage, list);
+};
+const point = (m, x, y) => [m[0] * x + m[2] * y + m[4], m[1] * x + m[3] * y + m[5]];
+function box(m, b) { const p = [point(m, b[0], b[1]), point(m, b[2], b[1]), point(m, b[0], b[3]), point(m, b[2], b[3])]; return [Math.min(...p.map(v => v[0])), Math.min(...p.map(v => v[1])), Math.max(...p.map(v => v[0])), Math.max(...p.map(v => v[1]))]; }
+function prune(node) {
+    if (!node)
+        return null;
+    if ('type' in node)
+        return node.type === 'content' ? { type: 'content', id: node.id } : null;
+    const children = node.children.flatMap(child => { const value = prune(child); return value ? [value] : []; });
+    return children.length ? { role: node.role, children } : null;
+}
+async function info() {
+    const data = await readFile(input.file);
+    if (data.length > config.maxInputBytes)
+        throw new Error('PDF 文件超过大小上限');
+    if (!data.subarray(0, 1024).includes(Buffer.from('%PDF-')))
+        throw new Error('PDF 文件头无效');
+    const loading = getDocument({ data: new Uint8Array(data), verbosity: 1, stopAtErrors: true, maxImageSize: config.pdfMaxPagePixels, cMapUrl: join(base, 'cmaps') + '/', cMapPacked: true, standardFontDataUrl: join(base, 'standard_fonts') + '/', wasmUrl: join(base, 'wasm') + '/', useSystemFonts: true });
+    loading.onPassword = () => { void loading.destroy(); port.postMessage({ fatal: 'PDF 已加密，请提供解除密码保护的副本' }); };
+    doc = await loading.promise;
+    if (doc.numPages > config.pdfMaxPages)
+        throw new Error(`PDF 共 ${doc.numPages} 页，超过 pdfMaxPages=${config.pdfMaxPages}`);
+    const outline = [];
+    async function visit(items, level = 1) {
+        for (const entry of items || []) {
+            if (outline.length >= 10000 || level > 32)
+                throw new Error('PDF 书签层级或数量超过上限');
+            let page = null, y = null;
+            try {
+                const raw = typeof entry.dest === 'string' ? await doc.getDestination(entry.dest) : entry.dest;
+                if (Array.isArray(raw) && raw.length) {
+                    const ref = z.union([z.number().int().nonnegative(), z.object({ num: z.number().int(), gen: z.number().int() })]).parse(raw[0]);
+                    page = (typeof ref === 'number' ? ref : await doc.getPageIndex(ref)) + 1;
+                    const kind = z.object({ name: z.string() }).safeParse(raw[1]);
+                    if (kind.success && kind.data.name === 'XYZ' && typeof raw[3] === 'number')
+                        y = raw[3];
+                }
+            }
+            catch { /* An unresolved bookmark remains visible with no invented page location. */ }
+            outline.push({ title: String(entry.title || '').slice(0, 1000), level, page, y });
+            await visit(entry.items, level + 1);
+        }
+    }
+    await visit(await doc.getOutline());
+    return { warnings: engineWarnings.get(0) || [], pageCount: doc.numPages, pageLabels: await doc.getPageLabels(), outline };
+}
+async function nativePage(n) {
+    activePage = n;
+    const page = await doc.getPage(n);
+    try {
+        const viewport = page.getViewport({ scale: 1 }), content = await page.getTextContent({ includeMarkedContent: true }), tree = prune(await page.getStructTree());
+        const spans = [], marks = [];
+        let seq = 0;
+        for (const item of content.items) {
+            if ('type' in item) {
+                if (item.type.startsWith('beginMarkedContent'))
+                    marks.push({ id: item.id, artifact: 'tag' in item && item.tag === 'Artifact' });
+                else if (item.type === 'endMarkedContent')
+                    marks.pop();
+                continue;
+            }
+            if (!item.str.trim())
+                continue;
+            const t = matrixSchema.parse(Util.transform(viewport.transform, item.transform)), h = Math.hypot(t[2], t[3]) || item.height || 10, style = content.styles[item.fontName], a = style?.ascent ?? .8;
+            const a0 = [t[4] + t[2] * a, t[5] + t[3] * a], b0 = [t[4] + t[2] * (a - 1), t[5] + t[3] * (a - 1)];
+            const angle = Math.atan2(t[1], t[0]), w = item.width || h / 2, dx = Math.cos(angle) * w, dy = Math.sin(angle) * w;
+            const points = [a0, b0, [a0[0] + dx, a0[1] + dy], [b0[0] + dx, b0[1] + dy]], bbox = [Math.min(...points.map(p => p[0])), Math.min(...points.map(p => p[1])), Math.max(...points.map(p => p[0])), Math.max(...points.map(p => p[1]))];
+            const rotation = ((page.rotate % 360) + 360) % 360, W = viewport.width, H = viewport.height;
+            const readingBox = rotation === 90 ? [bbox[1], W - bbox[2], bbox[3], W - bbox[0]] : rotation === 180 ? [W - bbox[2], H - bbox[3], W - bbox[0], H - bbox[1]] : rotation === 270 ? [H - bbox[3], bbox[0], H - bbox[1], bbox[2]] : bbox;
+            spans.push({ id: `s${++seq}`, text: item.str, mcid: [...marks].reverse().find(m => m.id)?.id || null, artifact: marks.some(m => m.artifact), bbox, readingBox, fontSize: h, fontName: style?.fontFamily || item.fontName });
+            if (spans.length > 100000)
+                throw new Error(`PDF 第 ${n} 页文字对象过多`);
+        }
+        const operators = await page.getOperatorList(), drawings = [], stack = [], mc = [], prefix = `p${page.ref?.num || n}R_mc`;
+        let matrix = [1, 0, 0, 1, 0, 0];
+        const add = (b, kind, shape = null) => {
+            const bounds = box(matrixSchema.parse(Util.transform(viewport.transform, matrix)), b);
+            if (mc.some(x => x.tag === 'Artifact'))
+                return;
+            const id = [...mc].reverse().find(x => Number.isInteger(x.id))?.id;
+            drawings.push({ bbox: bounds, kind, shape, mcid: id === undefined ? null : prefix + id });
+        };
+        for (let i = 0; i < operators.fnArray.length; i++) {
+            const op = required(operators.fnArray[i]), args = operators.argsArray[i] || [];
+            if (op === OPS.save)
+                stack.push([...matrix]);
+            else if (op === OPS.restore)
+                matrix = stack.pop() || matrix;
+            else if (op === OPS.transform)
+                matrix = matrixSchema.parse(Util.transform(matrix, matrixSchema.parse(args)));
+            else if (op === OPS.paintFormXObjectBegin) {
+                stack.push([...matrix]);
+                if (args[0])
+                    matrix = matrixSchema.parse(Util.transform(matrix, matrixSchema.parse(args[0])));
+            }
+            else if (op === OPS.paintFormXObjectEnd)
+                matrix = stack.pop() || matrix;
+            else if (op === OPS.beginMarkedContentProps)
+                mc.push({ tag: markTag(args[0]), ...Number.isInteger(args[1]) ? { id: z.number().parse(args[1]) } : {} });
+            else if (op === OPS.beginMarkedContent)
+                mc.push({ tag: markTag(args[0]) });
+            else if (op === OPS.endMarkedContent)
+                mc.pop();
+            else if (op === OPS.constructPath && [OPS.stroke, OPS.fill, OPS.eoFill, OPS.fillStroke, OPS.eoFillStroke].includes(z.number().parse(args[0])) && args[2]) {
+                const paths = Array.isArray(args[1]) ? args[1] : [], path = paths[0] === undefined ? [] : numeric(paths[0]);
+                add(bboxSchema.parse(numeric(args[2])), 'path', path.length === 13 && path[0] === 0 && path[3] === 1 && path[6] === 1 && path[9] === 1 && path[12] === 4 ? 'rect' : null);
+            }
+            else if ([OPS.paintImageXObject, OPS.paintInlineImageXObject, OPS.paintImageMaskXObject].includes(op))
+                add([0, 0, 1, 1], 'image');
+            if (drawings.length > 100000)
+                throw new Error(`PDF 第 ${n} 页图形对象过多`);
+        }
+        return { warnings: engineWarnings.get(n) || [], page: n, width: viewport.width, height: viewport.height, rotation: page.rotate, viewportTransform: viewport.transform, spans, tree, drawings };
+    }
+    finally {
+        page.cleanup();
+    }
+}
+async function render(n, bbox) {
+    activePage = n;
+    const page = await doc.getPage(n), v = page.getViewport({ scale: 1 }), b = bbox || [0, 0, v.width, v.height];
+    if (b[0] < 0 || b[1] < 0 || b[2] > v.width + .1 || b[3] > v.height + .1 || b[2] <= b[0] || b[3] <= b[1])
+        throw new Error('PDF 裁切区域无效');
+    const scale = Math.min(2, Math.sqrt(config.pdfMaxPagePixels / (v.width * v.height)));
+    if (scale < .5)
+        throw new Error(`PDF 第 ${n} 页页面尺寸过大`);
+    const viewport = page.getViewport({ scale }), canvas = createCanvas(Math.ceil(viewport.width), Math.ceil(viewport.height)), crop = createCanvas(Math.max(1, Math.ceil((b[2] - b[0]) * scale)), Math.max(1, Math.ceil((b[3] - b[1]) * scale)));
+    try {
+        // PDF.js publishes DOM-only render types for its supported Node napi-rs canvas. This bridge uses that same tested Canvas2D implementation.
+        await page.render({ canvas: null, canvasContext: canvas.getContext('2d'), viewport, background: 'rgb(255,255,255)' }).promise;
+        crop.getContext('2d').drawImage(canvas, b[0] * scale, b[1] * scale, crop.width, crop.height, 0, 0, crop.width, crop.height);
+        return { data: new Uint8Array(crop.toBuffer('image/png')), scale, bbox: b, warnings: engineWarnings.get(n) || [] };
+    }
+    finally {
+        canvas.width = canvas.height = 1;
+        crop.width = crop.height = 1;
+        page.cleanup();
+    }
+}
+try {
+    port.postMessage({ ready: await info() });
+}
+catch (error) {
+    port.postMessage({ fatal: message(error) });
+}
+const commandSchema = z.object({ id: z.number().int().positive(), command: z.enum(['page', 'render']), page: z.number().int().positive(), bbox: bboxSchema.optional() });
+let tail = Promise.resolve();
+port.on('message', (raw) => {
+    tail = tail.then(async () => {
+        let id = 0;
+        try {
+            const request = commandSchema.parse(raw);
+            id = request.id;
+            if (request.command === 'page')
+                port.postMessage({ id, value: await nativePage(request.page) });
+            else {
+                const value = await render(request.page, request.bbox);
+                port.postMessage({ id, value }, [value.data.buffer]);
+            }
+        }
+        catch (error) {
+            port.postMessage({ id, error: message(error) });
+        }
+    });
+});

package/lib/document/pdf.d.ts

@@ -0,0 +1,3 @@
+import type { Document, Job } from '../data.ts';
+import type { Models } from '../models.ts';
+export declare function convertPdf(file: string, destination: string, job: Job, models: Models, signal: AbortSignal): Promise<Document>;

package/lib/document/pdf.js

@@ -0,0 +1,218 @@
+import { block as makeBlock, documentSchema, pageReceiptSchema } from "../data.js";
+import { message, required } from "../util.js";
+import { nativePageSchema, pdfBlock } from "./pdf-data.js";
+import { join } from 'node:path';
+import { PdfSession, pageRange } from "./pdf-session.js";
+import { nativeLayout, PDF_PARSER_VERSION, tableMarkdown, normalize } from "./pdf-layout.js";
+import { restoreLayout, PDF_VISION_VERSION } from "./pdf-vision.js";
+import { atomicWrite, checkAbort, hash, readJson } from "../util.js";
+import { blockMarkdown } from "./convert.js";
+import { BudgetError } from "../models.js";
+export async function convertPdf(file, destination, job, models, signal) {
+    let session = new PdfSession(file, job.config, signal);
+    const assets = new Map(), pages = [], blocks = [], warnings = [];
+    job.parserVersion = PDF_PARSER_VERSION;
+    const addAsset = async (data, origin, source) => {
+        const assetId = `a-${hash(data).slice(0, 20)}`, asset = { assetId, mime: 'image/png', path: `assets/${assetId}.png`, bytes: data.length, origin, source };
+        if (!assets.has(assetId)) {
+            await atomicWrite(join(destination, asset.path), Buffer.from(data));
+            assets.set(assetId, asset);
+        }
+        return asset;
+    };
+    try {
+        const info = await session.ready, range = pageRange(job.pageStart, job.pageEnd, info.pageCount);
+        warnings.push(...(info.warnings || []).map(message => ({ code: 'pdf-engine-warning', message: `PDF 打开警告：${message}`, blocking: true })));
+        job.pdf = { pageCount: info.pageCount, selectedPages: range, outlineCount: info.outline.length };
+        job.progress = { ...job.progress, totalPages: range.count, parsedPages: 0, ocrPages: 0, tables: 0 };
+        await job.checkpoint();
+        for (let n = range.start; n <= range.end; n++) {
+            checkAbort(signal);
+            job.progress.currentPage = n;
+            await job.checkpoint();
+            const key = hash({ source: job.sourceHash, page: n, parser: PDF_PARSER_VERSION, visionVersion: PDF_VISION_VERSION, route: job.routes.vision, skip: job.config.skipImages, schema: job.schemaHash, pixels: job.config.pdfMaxPagePixels });
+            const receipt = join(destination, 'pdf-pages', `${n}-${key}.json`);
+            let result = await readJson(receipt, pageReceiptSchema, null);
+            if (!result) {
+                let page, layout, previewAsset;
+                try {
+                    if (session.closed) {
+                        session = new PdfSession(file, job.config, signal);
+                        await session.ready;
+                    }
+                    const nativeFile = join(destination, 'pdf-pages', `${n}-native-${hash({ source: job.sourceHash, parser: PDF_PARSER_VERSION })}.json`);
+                    page = (await readJson(nativeFile, nativePageSchema, null)) ?? undefined;
+                    if (!page) {
+                        page = await session.call('page', n);
+                        await atomicWrite(nativeFile, page);
+                    }
+                    layout = nativeLayout(page, info);
+                    const render = async (bbox) => {
+                        const r = await session.call('render', n, bbox);
+                        for (const message of r.warnings || [])
+                            if (!required(layout).warnings.some(w => w.message.endsWith(message)))
+                                required(layout).warnings.push({ code: 'pdf-engine-warning', page: n, message: `PDF 第 ${n} 页渲染警告：${message}`, blocking: true });
+                        return r;
+                    };
+                    if (layout.needsLayout) {
+                        previewAsset = await addAsset((await render()).data, `PDF 第 ${n} 页预览`, layout.source([0, 0, page.width, page.height]));
+                        try {
+                            const restored = await restoreLayout(page, layout, join(destination, previewAsset.path), job, models, signal);
+                            layout = { ...layout, ...restored, warnings: [...layout.warnings, ...restored.warnings] };
+                        }
+                        catch (error) {
+                            if (error instanceof BudgetError || signal?.aborted)
+                                throw error;
+                            required(layout).warnings.push({ code: 'pdf-layout-gap', page: n, message: `PDF 第 ${n} 页版面/OCR 未完成：${message(error)}`, blocking: true });
+                            layout.blocks.push({ ...pdfBlock({ type: 'figure', source: layout.source([0, 0, page.width, page.height]), caption: `PDF 第 ${n} 页未解析区域（整页参考）`, text: '', nativeLabels: [], needsReview: true }), previewAsset: required(previewAsset) });
+                        }
+                    }
+                    const localAssets = [];
+                    for (const block of layout.blocks) {
+                        if (block.type === 'figure' || block.type === 'table') {
+                            const asset = block.previewAsset || await addAsset((await render(block.source.bbox)).data, `PDF 第 ${n} 页 ${block.caption || block.type}`, block.source);
+                            if (block.type === 'figure')
+                                block.assetId = asset.assetId;
+                            else
+                                block.sourcePreviewAssetId = asset.assetId;
+                            localAssets.push(asset);
+                            delete block.previewAsset;
+                        }
+                    }
+                    if (previewAsset)
+                        localAssets.push(previewAsset);
+                    // Caption association uses temporary object links, not persistent cyclic references.
+                    const pageBlocks = layout.blocks.map((b, i) => {
+                        const { captionFor, captionBlock: _captionBlock, ...value } = b;
+                        if (captionFor?.type === 'figure') {
+                            value.type = 'caption';
+                            value.captionFigureIndex = required(layout).blocks.indexOf(captionFor);
+                        }
+                        if (captionFor?.type === 'table') {
+                            value.type = 'caption';
+                            value.tableCaption = true;
+                        }
+                        return { ...value, localIndex: i };
+                    });
+                    result = { blocks: pageBlocks, assets: localAssets, warnings: layout.warnings, audit: { page: n, width: page.width, height: page.height, pageLabel: info.pageLabels?.[n - 1] || null, status: layout.warnings.some(w => w.blocking) ? 'needs-review' : layout.navigation ? 'navigation' : 'parsed', ocr: layout.ocr, spanCount: layout.spanCount, artifactCount: layout.artifactCount, excluded: layout.excluded, tables: pageBlocks.filter(b => b.type === 'table').length, figures: pageBlocks.filter(b => b.type === 'figure').length, headings: pageBlocks.filter(b => b.type === 'heading').map(b => ({ text: b.text, basis: b.headingBasis, level: b.level })), previewAssetId: previewAsset?.assetId || null } };
+                }
+                catch (error) {
+                    if (error instanceof BudgetError || signal?.aborted)
+                        throw error;
+                    const source = { page: n, width: page?.width ?? null, height: page?.height ?? null, pageLabel: info.pageLabels?.[n - 1] || null, bbox: page ? [0, 0, page.width, page.height] : null, coordinateSystem: 'viewport-points-top-left', method: 'unprocessed' };
+                    // A page render failure does not erase native text or earlier page checkpoints.
+                    const fallback = (layout?.blocks || []).map(({ captionFor: _captionFor, captionBlock: _captionBlock, previewAsset: _previewAsset, ...b }) => b).filter(b => b.type !== 'figure' || b.assetId);
+                    fallback.push(makeBlock({ type: 'note', text: `[PDF 第 ${n} 页处理未完成：${message(error)}]`, source, needsReview: true }));
+                    result = { blocks: fallback, assets: [], warnings: [{ code: 'pdf-page-failed', page: n, message: `PDF 第 ${n} 页处理失败：${message(error)}`, blocking: true }], audit: { page: n, pageLabel: source.pageLabel, status: 'failed', error: message(error), ocr: layout?.ocr || false, tables: fallback.filter(b => b.type === 'table').length, figures: 0, headings: [] } };
+                }
+                // Failed/gap pages are deliberately retried; successful pages and model receipts are durable.
+                if (result.audit.status !== 'failed' && !result.warnings.some(w => w.code === 'pdf-layout-gap'))
+                    await atomicWrite(receipt, result);
+            }
+            for (const a of result.assets)
+                assets.set(a.assetId, a);
+            if (result.blocks.some(b => b.openEnd) && n === range.end && range.partial)
+                result.audit.status = 'needs-review';
+            pages.push(result.audit);
+            warnings.push(...result.warnings);
+            if (result.audit.status !== 'navigation') {
+                const figures = new Map();
+                for (const b of result.blocks)
+                    if (b.type === 'figure') {
+                        const occurrenceId = `fig-p${String(n).padStart(5, '0')}-${String(b.localIndex ?? result.blocks.indexOf(b)).padStart(4, '0')}`;
+                        figures.set(b.localIndex, occurrenceId);
+                        b.occurrenceId = occurrenceId;
+                    }
+                for (const b of result.blocks) {
+                    if (b.openEnd) {
+                        b.text += '\n[表格可能延续到下一页，当前页记录不是完整表；请按来源页及关联片段核对。]';
+                        b.markdown = (b.markdown || '') + '\n\n> 表格可能延续到下一页，当前页记录不是完整表；请按来源页及关联片段核对。';
+                        warnings.push({ code: 'pdf-table-continuation', page: n, message: `PDF 第 ${n} 页表格下边界未闭合${n === range.end && range.partial ? '；所选范围在此结束，请补充下一页或接受覆盖缺口' : ''}`, blocking: n === range.end && range.partial });
+                    }
+                    if (b.captionFigureIndex !== undefined)
+                        b.figureRef = required(figures.get(b.captionFigureIndex));
+                    delete b.captionFigureIndex;
+                    delete b.localIndex;
+                    blocks.push(b);
+                }
+            }
+            job.progress.parsedPages = (job.progress.parsedPages || 0) + 1;
+            if (result.audit.ocr)
+                job.progress.ocrPages = (job.progress.ocrPages || 0) + 1;
+            job.progress.tables = (job.progress.tables || 0) + result.audit.tables;
+            await job.checkpoint();
+        }
+        if (range.partial)
+            warnings.push({ code: 'pdf-selected-pages', message: `仅导入 PDF 实际第 ${range.start}–${range.end} 页 / ${range.total}；首尾章节、段落和表格可能延续到范围外，请按原件核对`, blocking: false });
+        // Only repeated marginal paragraphs are excluded; headings and body text remain even when repeated.
+        if (pages.length >= 3) {
+            const repeated = new Map();
+            for (const b of blocks) {
+                const box = b.source?.bbox, p = pages.find(p => p.page === b.source?.page);
+                if (b.type !== 'paragraph' || !box || b.text.length > 160)
+                    continue;
+                const key = normalize(b.text).replace(/\d+/g, '#'), edge = box[1] < 40 ? 'top' : box[1] > (p?.height || 842) - 45 ? 'bottom' : null;
+                if (edge) {
+                    const k = edge + ':' + key;
+                    repeated.set(k, [...(repeated.get(k) || []), b]);
+                }
+            }
+            for (const candidates of repeated.values())
+                if (new Set(candidates.map(b => required(b.source).page)).size >= Math.max(3, Math.ceil(pages.length * .6)))
+                    for (const b of candidates) {
+                        (required(pages.find(p => p.page === required(b.source).page)).excluded ||= []).push({ kind: 'repeated-header-footer', text: b.text, bbox: required(b.source).bbox });
+                        blocks.splice(blocks.indexOf(b), 1);
+                    }
+        }
+        // Across-page tables remain separate unless table number, header and column structure all agree.
+        for (let i = 1; i < blocks.length; i++) {
+            const b = required(blocks[i]);
+            if (b.type !== 'table' || !b.source?.bbox)
+                continue;
+            let previous = i - 1;
+            const repeated = [];
+            while (previous >= 0) {
+                const item = required(blocks[previous]);
+                if (!item.mergeRemove && !(item.tableCaption && normalize(item.text) === normalize(b.caption)))
+                    break;
+                if (!item.mergeRemove)
+                    repeated.push(item);
+                previous--;
+            }
+            const a = blocks[previous], last = a?.source?.locations?.at(-1) || a?.source;
+            const tableId = (t) => /^(?:表|Table)\s*([A-Z\d][\w.\-–]*)/iu.exec(t.caption)?.[1];
+            if (!a || a.type !== 'table' || !a.source || !last?.bbox || !a.rows[0] || !b.rows[0])
+                continue;
+            const height = pages.find(p => p.page === last.page)?.height || 842, nextHeight = pages.find(p => p.page === b.source?.page)?.height || 842;
+            if (b.source.page === last.page + 1 && tableId(a) && tableId(a) === tableId(b) && last.bbox[3] > height * .8 && b.source.bbox[1] < nextHeight * .25 && !a.needsReview && !b.needsReview && JSON.stringify(a.rows[0].map(c => normalize(c.text))) === JSON.stringify(b.rows[0].map(c => normalize(c.text))) && a.rows[0].length === b.rows[0].length && [...a.rows, ...b.rows].every(r => r.length === required(a.rows[0]).length && r.every(c => c.rowspan === 1 && c.colspan === 1))) {
+                a.rows.push(...b.rows.slice(1));
+                const { locations, ...first } = a.source;
+                a.source.locations = [...(locations || [first]), b.source];
+                a.sourcePreviewAssetIds = [...(a.sourcePreviewAssetIds || [a.sourcePreviewAssetId]), b.sourcePreviewAssetId].flatMap(id => id ? [id] : []);
+                a.markdown = tableMarkdown(a.rows);
+                a.text = a.rows.map(r => r.map(c => c.text).join(' | ')).join('\n');
+                b.mergeRemove = true;
+                for (const caption of repeated)
+                    caption.mergeRemove = true;
+            }
+        }
+        for (let i = blocks.length - 1; i >= 0; i--)
+            if (required(blocks[i]).mergeRemove)
+                blocks.splice(i, 1);
+        for (let i = 0; i < blocks.length; i++)
+            required(blocks[i]).blockId = `b-${String(i + 1).padStart(6, '0')}`;
+        if (!blocks.some(b => b.type !== 'heading' && (b.text?.trim() || b.type === 'figure')))
+            throw new Error('所选 PDF 页没有可导入内容，请核对页范围及质量报告');
+        if (!blocks.some(b => b.type === 'heading'))
+            warnings.push({ code: 'pdf-no-headings', message: '所选 PDF 页未匹配可靠正文标题，按正文分组，请核对目录层级', blocking: false });
+        const document = documentSchema.parse({ formatVersion: 1, sourceFormat: 'pdf', pageCount: info.pageCount, selectedPages: range, parserVersion: PDF_PARSER_VERSION, sourceHash: job.sourceHash, blocks, assets: [...assets.values()], internalRefs: [], warnings, audit: { pages, outlineCount: info.outline.length, matchedBookmarks: blocks.filter(b => b.headingBasis === 'bookmark').length, convertedFigures: blocks.filter(b => b.type === 'figure').length, convertedTables: blocks.filter(b => b.type === 'table').length } });
+        await atomicWrite(join(destination, 'document.json'), document);
+        await atomicWrite(join(destination, 'converted.md'), `> PDF 实际页序 ${range.start}–${range.end} / ${range.total}。${range.partial ? '仅覆盖所选范围。' : ''}\n\n` + blocks.map(b => blockMarkdown(b, document.assets)).join('\n\n') + '\n');
+        delete job.progress.currentPage;
+        await job.checkpoint();
+        return document;
+    }
+    finally {
+        await session.close();
+    }
+}

package/lib/document/zip.d.ts

@@ -0,0 +1,3 @@
+import type { Settings } from '../config.ts';
+/** Inspect central-directory limits before any decompression, including non-XML media. */
+export declare function unpackDocx(buffer: Buffer, config: Settings): Record<string, Uint8Array>;

package/lib/document/zip.js

@@ -0,0 +1,59 @@
+import { unzipSync } from 'fflate';
+/** Inspect central-directory limits before any decompression, including non-XML media. */
+export function unpackDocx(buffer, config) {
+    if (buffer.length > config.maxInputBytes)
+        throw new Error('Word 文件超过大小限制');
+    let end = -1;
+    for (let i = buffer.length - 22; i >= Math.max(0, buffer.length - 65557); i--) {
+        if (buffer.readUInt32LE(i) === 0x06054b50 && i + 22 + buffer.readUInt16LE(i + 20) === buffer.length) {
+            end = i;
+            break;
+        }
+    }
+    if (end < 0)
+        throw new Error('无效 DOCX ZIP 文件');
+    const count = buffer.readUInt16LE(end + 10), size = buffer.readUInt32LE(end + 12), start = buffer.readUInt32LE(end + 16);
+    if (count === 65535 || count > config.maxArchiveEntries || start + size > end || buffer.readUInt16LE(end + 4) !== 0)
+        throw new Error('不支持的分卷/ZIP64 或过多 ZIP 条目');
+    let pos = start, total = 0;
+    const entries = new Map();
+    for (let n = 0; n < count; n++) {
+        if (pos + 46 > end || buffer.readUInt32LE(pos) !== 0x02014b50)
+            throw new Error('损坏的 ZIP 目录');
+        const flags = buffer.readUInt16LE(pos + 8), method = buffer.readUInt16LE(pos + 10);
+        const compressed = buffer.readUInt32LE(pos + 20), uncompressed = buffer.readUInt32LE(pos + 24);
+        const nameLength = buffer.readUInt16LE(pos + 28), extraLength = buffer.readUInt16LE(pos + 30), commentLength = buffer.readUInt16LE(pos + 32);
+        const next = pos + 46 + nameLength + extraLength + commentLength;
+        if (next > end || uncompressed === 0xffffffff || compressed === 0xffffffff)
+            throw new Error('损坏或不支持的 ZIP 条目');
+        const name = buffer.subarray(pos + 46, pos + 46 + nameLength).toString('utf8');
+        if (!name || (name.includes('\\') || name.includes('\0')) || name.startsWith('/') || /^[A-Za-z]:/.test(name) || name.split('/').some(x => x === '..') || entries.has(name))
+            throw new Error('ZIP 条目路径非法或重复');
+        if (flags & 1 || ![0, 8].includes(method) || ((buffer.readUInt32LE(pos + 38) >>> 16) & 0xf000) === 0xa000)
+            throw new Error('不支持加密、特殊压缩或符号链接 ZIP');
+        const local = buffer.readUInt32LE(pos + 42);
+        if (local + 30 > start || buffer.readUInt32LE(local) !== 0x04034b50)
+            throw new Error('损坏的 ZIP 本地条目');
+        const localNameLength = buffer.readUInt16LE(local + 26), localExtra = buffer.readUInt16LE(local + 28);
+        if (buffer.subarray(local + 30, local + 30 + localNameLength).toString('utf8') !== name || local + 30 + localNameLength + localExtra + compressed > start)
+            throw new Error('ZIP 条目不一致');
+        total += uncompressed;
+        if (total > config.maxUncompressedBytes || (name.endsWith('.xml') && uncompressed > 16 * 1024 * 1024))
+            throw new Error('DOCX 解压大小超过限制');
+        entries.set(name, uncompressed);
+        pos = next;
+    }
+    if (pos !== start + size)
+        throw new Error('ZIP 目录长度不一致');
+    const files = unzipSync(buffer, { filter: file => {
+            if (!entries.has(file.name) || entries.get(file.name) !== file.originalSize)
+                throw new Error('ZIP 解压元数据不一致');
+            return true;
+        } });
+    for (const [name, data] of Object.entries(files))
+        if (data.length !== entries.get(name))
+            throw new Error('ZIP 解压长度不一致');
+    if (!files['word/document.xml'] || !files['[Content_Types].xml'])
+        throw new Error('文件不是有效的 Word DOCX');
+    return files;
+}