kordoc 4.18.3 → 4.18.4

package/dist/{parser-NKP2NCSV.js → parser-SOWTUOMK.js}

@@ -5,6 +5,8 @@ import {
   CONT_PARTS,
   EMPTY_PARTS,
   FILLER_CELLS,
+  FRAME_READING_UNITS,
+  FRAME_TITLE_BLOCKS,
   IMAGE_CELLS,
   NO_EDGES,
   PART_COLXS,
@@ -15,6 +17,7 @@ import {
   TOC_BLOCKS,
   WrapLexicon,
   collapseEvenSpacing,
+  computeBBox,
   computeMedianFontSizeFromFreq,
   detectDocumentStyleHeadings,
   detectHeadings,
@@ -22,33 +25,39 @@ import {
   detectMarkerHeadings,
   detectPageLeadHeadings,
   detectRepeatedPageLabels,
+  detectRightArrowRegions,
   detectSiblingStyleHeadings,
   detectTableCaptions,
   detectTypographyHeadings,
+  dominantStyle,
   extractImageRegions,
   extractLines,
   extractPageBlocksWithLines,
   filterHiddenText,
+  groupByY,
   joinCellEdges,
   joinPageBreakWraps,
   latinSoftWrap,
+  mergeLineSimple,
   mergeStackedHeadingLines,
   normalizeItems,
+  recordFrameReadingUnit,
   refineDocumentStyleHeadings,
   removeHeaderFooterBlocks,
   removeSideTabs,
+  restoreImageBullets,
   startsNewItem,
   trimTrailingEmptyTableCols,
   wrapJoiner
-} from "./chunk-J2A46VCA.js";
+} from "./chunk-KUUSP6OS.js";
 import {
   ocrModelsCached
-} from "./chunk-5X7XWYR2.js";
-import "./chunk-RDBTOE5X.js";
-import "./chunk-AZQWESRH.js";
+} from "./chunk-Z2PPDQK2.js";
+import "./chunk-HRJJDNRD.js";
+import "./chunk-QVCHFCT2.js";
 import {
   blocksToPages
-} from "./chunk-6GLTGOP6.js";
+} from "./chunk-4IBY5LTJ.js";
 import {
   encodePng
 } from "./chunk-7FCNOKOR.js";
@@ -58,7 +67,7 @@ import {
   escapeLiteralTags,
   stripScriptTags,
   unframeLayoutTables
-} from "./chunk-UVUKT3IQ.js";
+} from "./chunk-TMJVD6FW.js";
 import {
   hasRequestedPagesAfter,
   parsePageRange
@@ -67,11 +76,11 @@ import {
   WINGDINGS,
   blocksToMarkdown,
   escapeLiteralDollar
-} from "./chunk-W3RADOLO.js";
+} from "./chunk-VTVK2JMI.js";
 import {
   KordocError,
   sanitizeHref
-} from "./chunk-VGDTONTB.js";
+} from "./chunk-F5YVCJVF.js";
 
 // src/pdf/ocr-region-merge.ts
 function mergeOcrImageRegions(blocks, page, regions, ocrBlocks) {
@@ -84,8 +93,10 @@ function mergeOcrImageRegions(blocks, page, regions, ocrBlocks) {
       const overlapH = Math.max(0, Math.min(b.y + b.height, region.y2) - Math.max(b.y, region.y1));
       return overlapW * overlapH >= b.width * b.height * 0.8;
     });
+    const labels = candidates.filter((b) => b.type === "paragraph" && b.bbox.height <= 24 && (b.text?.match(/[\p{L}\p{N}]/gu)?.length ?? 0) >= 2);
+    const axisFragments = rotatedAxisFragments(candidates);
     const accepted = candidates.filter((b) => {
-      if (b.type === "paragraph") return (b.text?.match(/[\p{L}\p{N}]/gu)?.length ?? 0) >= 2;
+      if (b.type === "paragraph") return (b.text?.match(/[\p{L}\p{N}]/gu)?.length ?? 0) >= 2 || !axisFragments.has(b) && supportedDiagramLabel(b, labels, region);
       const t = b.table;
       if (b.type !== "table" || !t) return false;
       if (t.rows === 1 && t.cols === 1) {
@@ -96,29 +107,89 @@ function mergeOcrImageRegions(blocks, page, regions, ocrBlocks) {
       return t.rows >= 2 && t.cols >= 2 && headerLabels >= t.cols * 0.75 && t.cells.slice(1).some((row) => row.filter((c) => c.text.trim()).length >= 2);
     });
     const selected = candidates.flatMap((b) => accepted.includes(b) ? [b] : b.type === "table" && b.table ? rowParagraphs(b).filter((p) => /[\p{L}\p{N}]{2}/u.test(p.text ?? "")) : []);
-    for (const block of selected) {
+    const hasOriginal = (block) => blocks.some((existing) => {
       const b = block.bbox;
-      const hasOriginal = blocks.some((existing) => {
-        if (existing.pageNumber !== page || !existing.bbox || existing.type === "image") return false;
-        const e = existing.bbox;
-        const x = Math.max(0, Math.min(e.x + e.width, b.x + b.width) - Math.max(e.x, b.x));
-        const y = Math.max(0, Math.min(e.y + e.height, b.y + b.height) - Math.max(e.y, b.y));
-        return x * y > b.width * b.height * 0.2;
-      });
-      if (hasOriginal) continue;
-      const index = blocks.findIndex((existing) => {
-        const e = existing.bbox;
-        if (existing.pageNumber !== page || !e) return false;
-        const beside = e.y < region.y2 && e.y + e.height > region.y1 && e.x >= region.x2 - 1;
-        return e.y < b.y || beside;
-      });
-      const placed = block.type === "paragraph" ? { ...block, style: void 0 } : block;
-      blocks.splice(index < 0 ? blocks.length : index, 0, placed);
-      added++;
+      if (existing.pageNumber !== page || !existing.bbox || existing.type === "image") return false;
+      const e = existing.bbox;
+      const x = Math.max(0, Math.min(e.x + e.width, b.x + b.width) - Math.max(e.x, b.x));
+      const y = Math.max(0, Math.min(e.y + e.height, b.y + b.height) - Math.max(e.y, b.y));
+      return x * y > b.width * b.height * 0.2;
+    });
+    const selectedSet = new Set(selected), seen = /* @__PURE__ */ new Set();
+    for (const block of selected) {
+      if (seen.has(block)) continue;
+      const frame = FRAME_READING_UNITS.get(block);
+      const source = frame?.blocks.every((member) => selectedSet.has(member)) ? frame.blocks : [block];
+      source.forEach((member) => seen.add(member));
+      const available = source.filter((member) => !hasOriginal(member));
+      const units = frame && available.length === source.length && source === frame.blocks ? [{ blocks: available, bbox: frame.bbox, atomic: true }] : available.map((member) => ({ blocks: [member], bbox: member.bbox, atomic: false }));
+      for (const unit of units) {
+        const b = unit.bbox;
+        const index = blocks.findIndex((existing) => {
+          const e = FRAME_READING_UNITS.get(existing)?.bbox ?? existing.bbox;
+          if (existing.pageNumber !== page || !e) return false;
+          const beside = e.y < region.y2 && e.y + e.height > region.y1 && e.x >= region.x2 - 1;
+          return e.y < b.y || beside;
+        });
+        const placed = unit.blocks.map((member) => member.type === "paragraph" ? { ...member, style: void 0 } : member);
+        if (unit.atomic) recordFrameReadingUnit(placed, unit.bbox);
+        blocks.splice(index < 0 ? blocks.length : index, 0, ...placed);
+        added += placed.length;
+      }
     }
   }
   return added;
 }
+function supportedDiagramLabel(block, labels, region) {
+  const b = block.bbox, text = block.text?.trim() ?? "";
+  if (!b || !/^[\p{L}\p{N}]$/u.test(text) || b.height < 4 || b.width < 1 || b.width > b.height * 2 || b.x < region.x1 || b.y < region.y1 || b.x + b.width > region.x2 || b.y + b.height > region.y2 || b.width * b.height > (region.x2 - region.x1) * (region.y2 - region.y1) * 5e-3 || labels.length < 3 || !labels.some((l) => /\p{L}{3}/u.test(l.text ?? "")) || labels.filter((l) => !/\p{L}/u.test(l.text ?? "") && (l.text?.match(/\p{N}/gu)?.length ?? 0) >= 2).length < 2) return false;
+  let above = false, below = false;
+  for (const label of labels) {
+    const a = label.bbox, em = Math.max(b.height, a.height);
+    if (b.x + b.width < a.x - em * 2 || b.x > a.x + a.width + em * 2) continue;
+    const upperGap = a.y - (b.y + b.height), lowerGap = b.y - (a.y + a.height);
+    if (upperGap >= 0 && upperGap <= em * 4) above = true;
+    if (lowerGap >= 0 && lowerGap <= em * 4) below = true;
+  }
+  return above && below;
+}
+function rotatedAxisFragments(candidates) {
+  const paragraphs = candidates.filter((b) => b.type === "paragraph" && b.bbox.height <= 24);
+  const axes = [];
+  for (const block of paragraphs) {
+    const b = block.bbox, text = block.text?.trim() ?? "";
+    if (!/^[\d\s.,%+−-]+$/u.test(text) || !/\d/u.test(text) || b.width > b.height * 6) continue;
+    const axis = axes.find((a) => Math.abs(a[0].bbox.x + a[0].bbox.width - b.x - b.width) <= Math.max(a[0].bbox.height, b.height) * 0.5);
+    if (axis) axis.push(block);
+    else axes.push([block]);
+  }
+  const fragments = /* @__PURE__ */ new Set();
+  const aligned = (a, b) => {
+    const x = a.bbox, y = b.bbox, tolerance = Math.max(x.height, y.height) * 0.5;
+    return Math.abs(x.x - y.x) <= tolerance || Math.abs(x.x + x.width - y.x - y.width) <= tolerance;
+  };
+  for (const axis of axes) {
+    if (axis.length < 3) continue;
+    const bottom = Math.min(...axis.map((a) => a.bbox.y)), top = Math.max(...axis.map((a) => a.bbox.y + a.bbox.height));
+    const em = Math.max(...axis.map((a) => a.bbox.height)), right = axis[0].bbox.x + axis[0].bbox.width;
+    if (top - bottom < em * 6) continue;
+    const left = paragraphs.filter((block) => {
+      const b = block.bbox, text = block.text?.trim() ?? "";
+      return /\p{Script=Latin}/u.test(text) && (text.match(/[\p{L}\p{N}]/gu)?.length ?? 0) <= 3 && b.width <= b.height * 3 && b.x + b.width >= right - em * 6 && b.x + b.width <= right - Math.max(em, b.height) && b.y >= bottom && b.y + b.height <= top;
+    });
+    const stack = left.filter((a) => left.some((b) => {
+      if (a === b || !aligned(a, b)) return false;
+      const x = a.bbox, y = b.bbox, gap = Math.max(x.y, y.y) - Math.min(x.y + x.height, y.y + y.height);
+      return gap >= -Math.min(x.height, y.height) * 0.5 && gap <= Math.max(x.height, y.height);
+    }));
+    if (stack.length < 2) continue;
+    for (const block of paragraphs) {
+      const b = block.bbox;
+      if (b.y >= bottom && b.y + b.height <= top && b.x + b.width >= right - em * 6 && b.x + b.width <= right - Math.max(em, b.height) && /^[\p{L}\p{N}]$/u.test(block.text?.trim() ?? "") && stack.some((s) => aligned(s, block))) fragments.add(block);
+    }
+  }
+  return fragments;
+}
 function rowParagraphs(block) {
   const t = block.table, b = block.bbox;
   const h = b.height / t.rows;
@@ -1128,7 +1199,7 @@ function sameHeadShape(a, b) {
   let tall = false;
   for (let c = 0; c < a.length; c++) {
     if (a[c].colSpan !== b[c].colSpan || a[c].rowSpan !== b[c].rowSpan) return false;
-    if (a[c].rowSpan >= 2 && a[c].text.trim()) tall = true;
+    if (a[c].rowSpan >= 2 && a[c].text.trim() && b[c].text.trim()) tall = true;
   }
   return tall;
 }
@@ -1165,9 +1236,10 @@ function mergeCrossPageTables(blocks, pageHeights, lex) {
     if (px && cx && !shiftedSame(px, cx, !CLIP_TABLES.has(prev.table) && !CLIP_TABLES.has(curr.table))) continue;
     if (!rowTextsEqual(prev.table.cells[0], curr.table.cells[0]) && sameHeadShape(prev.table.cells[0], curr.table.cells[0])) continue;
     let currCells = curr.table.cells;
-    if (currCells.length > 1 && prev.table.cells.length > 0 && rowTextsEqual(prev.table.cells[0], currCells[0])) {
-      currCells = currCells.slice(1);
-    }
+    const headRows = repeatedHeaderRows(prev.table, curr.table);
+    if (headRows < 0) continue;
+    if (headRows) currCells = currCells.slice(headRows);
+    if (currCells.length) continueRowSpans(prev.table.cells, currCells);
     if (currCells.length === 0) {
       blocks.splice(j, 1);
       continue;
@@ -1369,6 +1441,45 @@ function chainHead(blocks, i, pageHeights) {
   }
   return k;
 }
+var ROW_CONTINUATIONS = /* @__PURE__ */ new WeakMap();
+function continueRowSpans(prevCells, currCells) {
+  const first = currCells[0];
+  const last = prevCells.length;
+  for (let c = 0; c < first.length; c++) {
+    const cell = first[c];
+    if (hasContent(cell) || (cell.rowSpan || 1) < 2) continue;
+    let anchor;
+    for (let r = last - 1; r >= 0 && !anchor; r--) {
+      const a = prevCells[r]?.[c];
+      if (a && r + (a.rowSpan || 1) === last && (a.colSpan || 1) === (cell.colSpan || 1) && (hasContent(a) || a.rowSpan >= 2)) anchor = a;
+    }
+    if (!anchor) continue;
+    if (!hasContent(anchor)) {
+      ROW_CONTINUATIONS.set(anchor, cell);
+      continue;
+    }
+    for (let part = cell; part; ) {
+      const next = ROW_CONTINUATIONS.get(part);
+      anchor.rowSpan = (anchor.rowSpan || 1) + part.rowSpan;
+      part.rowSpan = 1;
+      part.colSpan = 1;
+      ROW_CONTINUATIONS.delete(part);
+      part = next;
+    }
+  }
+}
+function repeatedHeaderRows(prev, curr) {
+  if (!rowTextsEqual(prev.cells[0], curr.cells[0])) return 0;
+  const pa = anchorsOf(prev), ca = anchorsOf(curr);
+  let depth = 1;
+  for (let r = 0; r < depth && r < prev.rows; r++) {
+    for (const a of pa) if (a.r === r) depth = Math.max(depth, r + a.rs);
+  }
+  if (depth >= curr.rows || depth > prev.rows) return -1;
+  const p = pa.filter((a) => a.r < depth), d = ca.filter((a) => a.r < depth);
+  const norm = (text) => flowText(text).replace(/\s+/g, "");
+  return p.length === d.length && p.every((a, i) => a.r === d[i].r && a.c === d[i].c && a.rs === d[i].rs && a.cs === d[i].cs && norm(a.cell.text) === norm(d[i].cell.text)) ? depth : -1;
+}
 function rowTextsEqual(a, b) {
   if (a.length !== b.length) return false;
   const norm = (t) => flowText(t).replace(/\s+/g, "");
@@ -1463,6 +1574,34 @@ function fillNestedContinuation(blocks, at) {
   blocks.splice(at, 1);
 }
 
+// src/pdf/selection-context.ts
+function selectionContextPages(selected, pageCount) {
+  const out = /* @__PURE__ */ new Set();
+  for (const page of selected) {
+    for (let distance = 1; distance <= 4; distance++) {
+      for (const neighbor of [page - distance, page + distance]) {
+        if (neighbor < 1 || neighbor > pageCount || selected.has(neighbor)) continue;
+        out.add(neighbor);
+      }
+    }
+  }
+  return [...out].sort((a, b) => a - b);
+}
+function marginContextBlocks(raw, page, view) {
+  const [x1, y1, x2, y2] = view;
+  const width = x2 - x1, height = y2 - y1;
+  const items = filterHiddenText(normalizeItems(raw), width, height, x1, y1).visible;
+  for (const item of items) {
+    item.x -= x1;
+    item.y -= y1;
+  }
+  return groupByY(items).flatMap((line) => {
+    const bbox = computeBBox(line, page);
+    if (bbox.y + bbox.height < height * 0.88 && bbox.y > height * 0.12) return [];
+    return [{ type: "paragraph", text: mergeLineSimple(line), pageNumber: page, bbox, style: dominantStyle(line) }];
+  });
+}
+
 // src/pdf/symbol-fonts.ts
 var WINANSI_REVERSE = {
   8364: 128,
@@ -1770,6 +1909,9 @@ var NAMED = {
   quotedblright: "\u201D",
   quotesingle: "'",
   quotedbl: '"',
+  bullet: "\u2022",
+  openbullet: "\u25E6",
+  whitebullet: "\u25E6",
   question: "?",
   exclam: "!",
   percent: "%",
@@ -1786,7 +1928,9 @@ function glyphNameText(name) {
   let out = "";
   for (const part of base.split("_")) {
     const uni = /^uni([0-9A-F]{4})$/.exec(part) ?? /^u([0-9A-F]{4,6})$/.exec(part);
-    const ch = /^[A-Za-z]$/.test(part) ? part : NAMED[part] ?? (uni ? String.fromCodePoint(parseInt(uni[1], 16)) : void 0);
+    const code = uni ? parseInt(uni[1], 16) : void 0;
+    if (code !== void 0 && (code > 1114111 || code >= 55296 && code <= 57343)) return void 0;
+    const ch = /^[A-Za-z]$/.test(part) ? part : NAMED[part] ?? (code !== void 0 ? String.fromCodePoint(code) : void 0);
     if (ch === void 0) return void 0;
     out += ch;
   }
@@ -2472,6 +2616,35 @@ function demoteNonHeadingRoles(blocks, pageHeights, faceNames) {
       }
     }
   }
+  const citedProse = /* @__PURE__ */ new Set();
+  const numberedProse = /* @__PURE__ */ new Set();
+  for (const page of byPage.values()) {
+    const numbered = page.filter((b) => (b.type === "heading" || b.type === "paragraph") && /^[①-⑮]\s*/.test(b.text?.trim() ?? "") && b.bbox && b.style?.fontName && b.style.fontSize && !/Bold|Black|Heavy|Semibold/i.test(faceNames?.get(b.style.fontName) ?? ""));
+    let chain = [];
+    const finish = () => {
+      if (chain.length < 3 || !chain.some((b) => (b.text?.length ?? 0) >= 40 && b.bbox.height >= b.style.fontSize * 2.4)) return;
+      for (const b of chain) numberedProse.add(b);
+      const first = chain[0], size = first.style.fontSize;
+      for (const b of page) if (/^[○●•◦]\s/.test(b.text?.trim() ?? "") && b.style?.fontName === first.style?.fontName && b.style?.fontSize === size && b.bbox && Math.abs(b.bbox.x - first.bbox.x) <= size * 0.5 && (b.text?.length ?? 0) >= 15) numberedProse.add(b);
+    };
+    for (const b of numbered) {
+      const prev = chain.at(-1), size = b.style.fontSize;
+      if (!prev || b.text.trim().charCodeAt(0) !== prev.text.trim().charCodeAt(0) + 1 || b.style.fontName !== prev.style.fontName || size !== prev.style.fontSize || Math.abs(b.bbox.x - prev.bbox.x) > size * 0.5 || prev.bbox.y - (b.bbox.y + b.bbox.height) > size * 8) {
+        finish();
+        chain = [];
+      }
+      chain.push(b);
+    }
+    finish();
+  }
+  for (let i = 0; i + 1 < blocks.length; i++) {
+    const prose = blocks[i], reference = blocks[i + 1];
+    const a = prose.bbox, b = reference.bbox, size = prose.style?.fontSize ?? 0;
+    const source = reference.text?.replace(/<[^>]+>/g, "").trim() ?? "";
+    if (prose.type !== "heading" && prose.type !== "paragraph" && prose.type !== "list" || reference.type !== "heading" && reference.type !== "paragraph" || (prose.text?.length ?? 0) < 40 || !a || !b || size <= 0 || a.height < size * 2.4 || prose.pageNumber !== reference.pageNumber || prose.style?.fontName !== reference.style?.fontName || prose.style?.fontSize !== reference.style?.fontSize || Math.abs(a.x - b.x) > size && !(/^[○●•◦]\s/.test(prose.text ?? "") && b.x >= a.x && b.x - a.x <= size * 2) || a.y - (b.y + b.height) < 0 || a.y - (b.y + b.height) > size || !/^[（(].*(?:제\s*\d+\s*조|(?:Article|Section)\s+\d+).*[）)]$/i.test(source)) continue;
+    citedProse.add(prose);
+    citedProse.add(reference);
+  }
   for (let i = 0; i < blocks.length; i++) {
     const block = blocks[i];
     if (block.type !== "heading" || !block.text) continue;
@@ -2519,7 +2692,7 @@ function demoteNonHeadingRoles(blocks, pageHeights, faceNames) {
     const byline = /Italic|Oblique/i.test(face) && !/Bold|Black|Heavy|Semibold/i.test(face) && !!bodyFace && size <= Number(bodySize) + 0.5 && blocks[i - 1]?.type === "heading" && blocks[i - 1].pageNumber === block.pageNumber;
     const ocrFragment = block.style?.fontName === "ocr" && (/^[.,;:…·•-]/.test(text) || text.includes("\u2026") || (text.match(/\p{L}/gu)?.length ?? 0) < 2 || text.split(/\s+/).length > 7);
     const crowdedOcr = block.style?.fontName === "ocr" && !!box && [prev, next].every((o) => o?.type === "paragraph" && o.pageNumber === block.pageNumber && o.style?.fontName === "ocr" && !!o.bbox && (o.text?.length ?? 0) >= 15 && Math.abs(o.bbox.x - box.x) <= size * 0.25 && Math.min(o.bbox.y + o.bbox.height, box.y + box.height) - Math.max(o.bbox.y, box.y) >= o.bbox.height * 0.2) && prev.bbox.y > box.y && next.bbox.y < box.y;
-    if (ocrFragment || crowdedOcr || tocEntry || proseStyle || kicker || byline || tiny || !/\p{L}/u.test(text) && !chapterNumber || /^[a-z]/.test(text) || CAPTION.test(text) || EQUATION_NUMBER.test(block.text) || DISPLAY_MATH.test(text) || // 닫는 괄호가 여는 괄호보다 많으면 앞 줄에서 이어진 문장 조각이다 ("Fact-checking) and is used …") — "1)"·"가)" 앞머리 번호는 빼고 센다
+    if (numberedProse.has(block) || citedProse.has(block) || ocrFragment || crowdedOcr || tocEntry || proseStyle || kicker || byline || tiny || !/\p{L}/u.test(text) && !chapterNumber || /^[a-z]/.test(text) || CAPTION.test(text) || EQUATION_NUMBER.test(block.text) || DISPLAY_MATH.test(text) || // 닫는 괄호가 여는 괄호보다 많으면 앞 줄에서 이어진 문장 조각이다 ("Fact-checking) and is used …") — "1)"·"가)" 앞머리 번호는 빼고 센다
     unbalancedClose(text.replace(/^\s*[\dA-Za-z가-힣ⅰ-ⅹ]{1,3}\)\s*/, "")) || isRunningHead(block, page, pageHeights.get(block.pageNumber ?? 0))) {
       block.type = "paragraph";
       block.level = void 0;
@@ -2560,6 +2733,11 @@ function splitSingleCellTables(blocks) {
   const out = [];
   for (const b of blocks) {
     const t = b.type === "table" ? b.table : void 0;
+    const title = t && FRAME_TITLE_BLOCKS.get(t);
+    if (title) {
+      out.push({ ...title, text: `${t.cells[0][0].text.trim()} ${t.cells[0][2].text.trim()}` });
+      continue;
+    }
     const only = t && t.rows === 1 && t.cols === 1 && t.cells[0]?.[0]?.blocks?.length === 1 ? t.cells[0][0].blocks[0] : void 0;
     const squash = (s) => s.replace(/\s/g, "");
     const inner = only && (!squash(t.cells[0][0].text) || only.type === "table" && squash(t.cells[0][0].text) === squash(only.table.cells.flat().map((c) => c.text).join(""))) ? only : void 0;
@@ -2569,12 +2747,16 @@ function splitSingleCellTables(blocks) {
       continue;
     }
     const cell = t && t.rows === 1 && t.cols === 1 ? t.cells[0]?.[0] : void 0;
-    if (!cell || cell.blocks?.some((x) => x.type === "table")) {
+    if (!cell) {
       out.push(b);
       continue;
     }
     const lines = (cell.text ?? "").split(/\n/).map((l) => l.trim()).filter(Boolean);
     if (t?.caption?.trim()) out.push({ type: "paragraph", text: t.caption.trim(), pageNumber: b.pageNumber, bbox: b.bbox });
+    if (cell.blocks?.length) {
+      out.push(...splitSingleCellTables(cell.blocks));
+      continue;
+    }
     for (const text of lines) out.push({ type: "paragraph", text, pageNumber: b.pageNumber, bbox: b.bbox });
   }
   return out;
@@ -2699,7 +2881,7 @@ function mergeLinkRuns(markdown) {
 
 // src/pdf/formula-ocr.ts
 async function applyFormulaOcr(buffer, blocks, pageFilter, effectivePageCount, warnings, _onProgress) {
-  const formulaMod = await import("./formula-QZAOKPBX.js");
+  const formulaMod = await import("./formula-7ZOZ5QJ6.js");
   const { FormulaPipeline, ensureFormulaModels } = formulaMod;
   await ensureFormulaModels((p) => {
     if (p.phase === "download" && p.total) {
@@ -2922,12 +3104,29 @@ async function parsePdfDocument(buffer, options) {
     const uncoveredImageRegions = /* @__PURE__ */ new Map();
     const largeImageRegions = /* @__PURE__ */ new Set();
     const imageState = createPdfImageState();
+    const imageBulletShapes = /* @__PURE__ */ new Map();
+    const imageArrowDirections = /* @__PURE__ */ new Map();
     const extractedImages = [];
     const pageImageBlocks = /* @__PURE__ */ new Map();
     const carry = {};
     const vectorPageOps = /* @__PURE__ */ new Map();
     const wrapLexicon = new WrapLexicon();
     const noteMarks = /* @__PURE__ */ new Map();
+    const headerContext = [];
+    if (pageFilter && options?.removeHeaderFooter !== false && pageFilter.size < effectivePageCount) {
+      for (const number of selectionContextPages(pageFilter, effectivePageCount)) {
+        let contextPage;
+        try {
+          contextPage = await doc.getPage(number);
+          const content = await contextPage.getTextContent();
+          pageHeights.set(number, contextPage.view[3] - contextPage.view[1]);
+          headerContext.push(...marginContextBlocks(content.items, number, contextPage.view));
+        } catch {
+        } finally {
+          contextPage?.cleanup();
+        }
+      }
+    }
     let parsedPages = 0;
     for (let i = 1; i <= effectivePageCount; i++) {
       if (pageFilter && !pageFilter.has(i)) continue;
@@ -2981,6 +3180,7 @@ async function parsePdfDocument(buffer, options) {
         } : rawOps;
         remapSymbolFontItems(visible, (loadedName) => fontObj(loadedName)?.name);
         remapControlGlyphs(visible, differencesOf);
+        await restoreImageBullets(visible, page, opList.fnArray, opList.argsArray, imageBulletShapes);
         const faces = /* @__PURE__ */ new Map();
         for (const it of visible) {
           if (!it.fontName) continue;
@@ -3027,7 +3227,8 @@ async function parsePdfDocument(buffer, options) {
         if (marks.length) noteMarks.set(i, { marks, seps: footnoteSeparators(extractLines(opList.fnArray, opList.argsArray).horizontals, pageW, pageH) });
         const stamp = marginStamp(visible);
         const flow = stamp.length ? visible.filter((it) => !stamp.includes(it)) : visible;
-        const pageBlocks = orderTwoUpPage(extractPageBlocksWithLines(flow, i, opList, pageW, pageH, void 0, options?.tables !== false, carry, wrapLexicon), pageW, pageH);
+        const rightArrows = options?.tables !== false ? await detectRightArrowRegions(page, opList.fnArray, opList.argsArray, imageArrowDirections) : [];
+        const pageBlocks = orderTwoUpPage(extractPageBlocksWithLines(flow, i, opList, pageW, pageH, void 0, options?.tables !== false, carry, wrapLexicon, rightArrows), pageW, pageH);
         if (stamp.length) pageBlocks.unshift({ type: "paragraph", text: [...stamp].sort((a, b) => a.y - b.y).map((it) => it.text).join(" "), pageNumber: i });
         for (const b of pageBlocks) blocks.push(b);
         try {
@@ -3083,7 +3284,7 @@ async function parsePdfDocument(buffer, options) {
       }
       if (targets.size > 0) {
         try {
-          const { runPdfOcr } = await import("./pdf-ocr-VJL7ON2Q.js");
+          const { runPdfOcr } = await import("./pdf-ocr-7IOOVR2I.js");
           const mode = typeof options?.ocr === "function" ? options.ocr : "builtin";
           const regionPages = new Map([...ocrRegions].filter(([p]) => options?.ocr !== "force" && !isImageBased && !pageQuality.find((q) => q.page === p)?.needsOcr));
           const ocrPageBlocks = await runPdfOcr(ocrBuffer, targets, mode, warnings, options?.onProgress, options?.tables !== false, vectorPageOps, regionPages);
@@ -3150,8 +3351,8 @@ async function parsePdfDocument(buffer, options) {
         warnings.push({ page, message: `${count}\uAC1C \uC774\uBBF8\uC9C0 \uC601\uC5ED\uC5D0 \uCD94\uCD9C \uAC00\uB2A5\uD55C \uD14D\uC2A4\uD2B8 \uC5C6\uC74C (\uADF8\uB9BC/\uCC28\uD2B8/\uB3C4\uC7A5 \uB0B4\uC6A9 \uB204\uB77D \uAC00\uB2A5 \u2014 \uADF8\uB9BC \uC18D \uAE00\uC740 ocr: true / CLI --ocr \uB85C \uC77D\uC2B5\uB2C8\uB2E4)`, code: "SKIPPED_IMAGE" });
       }
     }
-    if (options?.removeHeaderFooter !== false && parsedPageCount >= 3) {
-      const removed = removeHeaderFooterBlocks(blocks, pageHeights, warnings, noteMarks);
+    if (options?.removeHeaderFooter !== false && (parsedPageCount >= 3 || headerContext.length)) {
+      const removed = removeHeaderFooterBlocks(blocks, pageHeights, warnings, noteMarks, false, headerContext);
       for (let ri = removed.length - 1; ri >= 0; ri--) {
         blocks.splice(removed[ri], 1);
       }
@@ -3199,6 +3400,10 @@ async function parsePdfDocument(buffer, options) {
     const outline = blocks.filter((b) => b.type === "heading" && b.level && b.text).map((b) => ({ level: b.level, text: b.text, pageNumber: b.pageNumber }));
     sanitizeBlockControlChars(blocks);
     let outBlocks = splitSingleCellTables(blocks);
+    if (options?.layoutTables !== "keep") {
+      blocks.length = 0;
+      for (const block of outBlocks) blocks.push(block);
+    }
     outBlocks = inlineFootnotes(outBlocks, noteMarks);
     outBlocks = relocateEndnotes(outBlocks);
     if (options?.removeHeaderFooter !== false) outBlocks = outBlocks.filter((b) => {
@@ -3315,4 +3520,4 @@ export {
   parsePdfDocument,
   removeHeaderFooterBlocks
 };
-//# sourceMappingURL=parser-NKP2NCSV.js.map
+//# sourceMappingURL=parser-SOWTUOMK.js.map