kordoc 2.0.3 → 2.2.0

package/dist/index.cjs

@@ -85,6 +85,7 @@ async function ocrPages(doc, provider, pageFilter, effectivePageCount) {
         blocks.push({ type: "paragraph", text: text.trim(), pageNumber: i });
       }
     } catch {
+      blocks.push({ type: "paragraph", text: `[OCR \uC2E4\uD328: \uD398\uC774\uC9C0 ${i}]` });
     }
   }
   return blocks;
@@ -182,7 +183,7 @@ var import_zlib = require("zlib");
 var import_xmldom = require("@xmldom/xmldom");
 
 // src/utils.ts
-var VERSION = true ? "2.0.3" : "0.0.0-dev";
+var VERSION = true ? "2.2.0" : "0.0.0-dev";
 function toArrayBuffer(buf) {
   if (buf.byteOffset === 0 && buf.byteLength === buf.buffer.byteLength) {
     return buf.buffer;
@@ -198,7 +199,8 @@ var KordocError = class extends Error {
 function isPathTraversal(name) {
   if (name.includes("\0")) return true;
   const normalized = name.replace(/\\/g, "/");
-  return normalized.includes("..") || normalized.startsWith("/") || /^[A-Za-z]:/.test(normalized);
+  const segments = normalized.split("/");
+  return segments.some((s) => s === "..") || normalized.startsWith("/") || /^[A-Za-z]:/.test(normalized);
 }
 function precheckZipSize(buffer, maxUncompressedSize = 100 * 1024 * 1024, maxEntries = 500) {
   try {
@@ -238,12 +240,25 @@ function precheckZipSize(buffer, maxUncompressedSize = 100 * 1024 * 1024, maxEnt
     return { totalUncompressed: 0, entryCount: 0 };
   }
 }
+function stripDtd(xml) {
+  return xml.replace(/<!DOCTYPE\s[^[>]*(\[[\s\S]*?\])?\s*>/gi, "");
+}
 var SAFE_HREF_RE = /^(?:https?:|mailto:|tel:|#)/i;
 function sanitizeHref(href) {
   const trimmed = href.trim();
   if (!trimmed || !SAFE_HREF_RE.test(trimmed)) return null;
   return trimmed;
 }
+function safeMin(arr) {
+  let min = Infinity;
+  for (let i = 0; i < arr.length; i++) if (arr[i] < min) min = arr[i];
+  return min;
+}
+function safeMax(arr) {
+  let max = -Infinity;
+  for (let i = 0; i < arr.length; i++) if (arr[i] > max) max = arr[i];
+  return max;
+}
 function classifyError(err) {
   if (!(err instanceof Error)) return "PARSE_ERROR";
   const msg = err.message;
@@ -318,6 +333,7 @@ function buildTableDirect(rows, numRows) {
       if (end > maxCols) maxCols = end;
     }
   }
+  if (maxCols > MAX_COLS) maxCols = MAX_COLS;
   if (maxCols === 0) return { rows: 0, cols: 0, cells: [], hasHeader: false };
   const grid = Array.from(
     { length: numRows },
@@ -327,7 +343,7 @@ function buildTableDirect(rows, numRows) {
     for (const cell of row) {
       const r = cell.rowAddr ?? 0;
       const c = cell.colAddr ?? 0;
-      if (r >= numRows || c >= maxCols) continue;
+      if (r >= numRows || c >= maxCols || r < 0 || c < 0) continue;
       grid[r][c] = { text: cell.text.trim(), colSpan: cell.colSpan, rowSpan: cell.rowSpan };
       for (let dr = 0; dr < cell.rowSpan; dr++) {
         for (let dc = 0; dc < cell.colSpan; dc++) {
@@ -607,7 +623,12 @@ function parseCharProperties(doc, map) {
       if (!id) continue;
       const prop = {};
       const height = el.getAttribute("height");
-      if (height) prop.fontSize = parseInt(height, 10) / 100;
+      if (height) {
+        const parsedHeight = parseInt(height, 10);
+        if (!isNaN(parsedHeight) && parsedHeight > 0) {
+          prop.fontSize = parsedHeight / 100;
+        }
+      }
       const bold = el.getAttribute("bold");
       if (bold === "true" || bold === "1") prop.bold = true;
       const italic = el.getAttribute("italic");
@@ -642,9 +663,6 @@ function parseStyleElements(doc, map) {
     }
   }
 }
-function stripDtd(xml) {
-  return xml.replace(/<!DOCTYPE\s[^[>]*(\[[\s\S]*?\])?\s*>/gi, "");
-}
 async function parseHwpxDocument(buffer, options) {
   precheckZipSize(buffer, MAX_DECOMPRESS_SIZE, MAX_ZIP_ENTRIES);
   let zip;
@@ -747,7 +765,7 @@ async function extractImagesFromZip(zip, blocks, decompressed, warnings) {
         const data = await file.async("uint8array");
         decompressed.total += data.length;
         if (decompressed.total > MAX_DECOMPRESS_SIZE) throw new KordocError("ZIP \uC555\uCD95 \uD574\uC81C \uD06C\uAE30 \uCD08\uACFC (ZIP bomb \uC758\uC2EC)");
-        const ext = ref.includes(".") ? ref.split(".").pop() : "png";
+        const ext = ref.includes(".") ? ref.split(".").pop() || "png" : "png";
         const mimeType = imageExtToMime(ext);
         imageIndex++;
         const filename = `image_${String(imageIndex).padStart(3, "0")}.${mimeToExt(mimeType)}`;
@@ -994,7 +1012,8 @@ function walkSection(node, blocks, tableCtx, tableStack, styleMap, warnings, sec
         if (newTable.rows.length > 0) {
           if (tableStack.length > 0) {
             const parentTable = tableStack.pop();
-            const nestedCols = Math.max(...newTable.rows.map((r) => r.length));
+            let nestedCols = 0;
+            for (const r of newTable.rows) if (r.length > nestedCols) nestedCols = r.length;
             if (newTable.rows.length >= 3 && nestedCols >= 2) {
               blocks.push({ type: "table", table: buildTable(newTable.rows), pageNumber: sectionNum });
             } else {
@@ -1041,8 +1060,10 @@ function walkSection(node, blocks, tableCtx, tableStack, styleMap, warnings, sec
         break;
       case "cellSpan":
         if (tableCtx?.cell) {
-          const cs = parseInt(el.getAttribute("colSpan") || "1", 10);
-          const rs = parseInt(el.getAttribute("rowSpan") || "1", 10);
+          const rawCs = parseInt(el.getAttribute("colSpan") || "1", 10);
+          const cs = isNaN(rawCs) ? 1 : rawCs;
+          const rawRs = parseInt(el.getAttribute("rowSpan") || "1", 10);
+          const rs = isNaN(rawRs) ? 1 : rawRs;
           tableCtx.cell.colSpan = clampSpan(cs, MAX_COLS);
           tableCtx.cell.rowSpan = clampSpan(rs, MAX_ROWS);
         }
@@ -1101,7 +1122,8 @@ function walkParagraphChildren(node, blocks, tableCtx, tableStack, styleMap, war
         if (newTable.rows.length > 0) {
           if (tableStack.length > 0) {
             const parentTable = tableStack.pop();
-            const nestedCols = Math.max(...newTable.rows.map((r) => r.length));
+            let nestedCols = 0;
+            for (const r of newTable.rows) if (r.length > nestedCols) nestedCols = r.length;
             if (newTable.rows.length >= 3 && nestedCols >= 2) {
               blocks.push({ type: "table", table: buildTable(newTable.rows), pageNumber: sectionNum });
             } else {
@@ -1134,6 +1156,8 @@ function walkParagraphChildren(node, blocks, tableCtx, tableStack, styleMap, war
         extractDrawTextBlocks(el, blocks, styleMap, sectionNum);
       } else if (localTag === "r" || localTag === "run" || localTag === "ctrl" || localTag === "rect" || localTag === "ellipse" || localTag === "polygon" || localTag === "line" || localTag === "arc" || localTag === "curve" || localTag === "connectLine" || localTag === "container") {
         walkChildren(el, d + 1);
+      } else if (localTag === "run") {
+        tableCtx = walkParagraphChildren(el, blocks, tableCtx, tableStack, styleMap, warnings, sectionNum, depth + 1);
       }
     }
   };
@@ -2197,6 +2221,7 @@ function parseLenientCfb(data) {
   if (miniSectorSizeShift > 16) throw new Error("\uC720\uD6A8\uD558\uC9C0 \uC54A\uC740 \uBBF8\uB2C8 \uC139\uD130 \uD06C\uAE30 \uC2DC\uD504\uD2B8: " + miniSectorSizeShift);
   const miniSectorSize = 1 << miniSectorSizeShift;
   const fatSectorCount = data.readUInt32LE(44);
+  if (fatSectorCount > 1e4) throw new Error("FAT \uC139\uD130 \uC218\uAC00 \uB108\uBB34 \uB9CE\uC2B5\uB2C8\uB2E4: " + fatSectorCount);
   const firstDirSector = data.readUInt32LE(48);
   const miniStreamCutoff = data.readUInt32LE(56);
   const firstMiniFatSector = data.readUInt32LE(60);
@@ -2573,10 +2598,14 @@ function findSections(cfb) {
 }
 function findSectionsLenient(lcfb, compressed) {
   const sections = [];
+  let totalDecompressed = 0;
   for (let i = 0; i < MAX_SECTIONS; i++) {
     const raw = lcfb.findStream(`/BodyText/Section${i}`) ?? lcfb.findStream(`Section${i}`);
     if (!raw) break;
-    sections.push({ idx: i, content: compressed ? decompressStream(raw) : raw });
+    const content = compressed ? decompressStream(raw) : raw;
+    totalDecompressed += content.length;
+    if (totalDecompressed > MAX_TOTAL_DECOMPRESS) throw new KordocError("\uCD1D \uC555\uCD95 \uD574\uC81C \uD06C\uAE30 \uCD08\uACFC (decompression bomb \uC758\uC2EC)");
+    sections.push({ idx: i, content });
   }
   if (sections.length === 0) {
     for (const e of lcfb.entries()) {
@@ -2584,7 +2613,12 @@ function findSectionsLenient(lcfb, compressed) {
       if (e.name.startsWith("Section")) {
         const idx = parseInt(e.name.replace("Section", ""), 10) || 0;
         const raw = lcfb.findStream(e.name);
-        if (raw) sections.push({ idx, content: compressed ? decompressStream(raw) : raw });
+        if (raw) {
+          const content = compressed ? decompressStream(raw) : raw;
+          totalDecompressed += content.length;
+          if (totalDecompressed > MAX_TOTAL_DECOMPRESS) throw new KordocError("\uCD1D \uC555\uCD95 \uD574\uC81C \uD06C\uAE30 \uCD08\uACFC (decompression bomb \uC758\uC2EC)");
+          sections.push({ idx, content });
+        }
       }
     }
   }
@@ -2592,11 +2626,15 @@ function findSectionsLenient(lcfb, compressed) {
 }
 function findViewTextSectionsLenient(lcfb, compressed) {
   const sections = [];
+  let totalDecompressed = 0;
   for (let i = 0; i < MAX_SECTIONS; i++) {
     const raw = lcfb.findStream(`/ViewText/Section${i}`) ?? lcfb.findStream(`Section${i}`);
     if (!raw) break;
     try {
-      sections.push({ idx: i, content: decryptViewText(raw, compressed) });
+      const content = decryptViewText(raw, compressed);
+      totalDecompressed += content.length;
+      if (totalDecompressed > MAX_TOTAL_DECOMPRESS) throw new KordocError("\uCD1D \uC555\uCD95 \uD574\uC81C \uD06C\uAE30 \uCD08\uACFC (decompression bomb \uC758\uC2EC)");
+      sections.push({ idx: i, content });
     } catch {
       break;
     }
@@ -2998,10 +3036,14 @@ init_page_range();
 // src/pdf/line-detector.ts
 var import_pdf = require("pdfjs-dist/legacy/build/pdf.mjs");
 var ORIENTATION_TOL = 2;
-var MIN_LINE_LENGTH = 10;
-var COORD_MERGE_TOL = 3;
+var MIN_LINE_LENGTH = 15;
+var MAX_LINE_WIDTH = 5;
 var CONNECT_TOL = 5;
 var CELL_PADDING = 2;
+var MIN_COL_WIDTH = 15;
+var MIN_ROW_HEIGHT = 6;
+var VERTEX_MERGE_FACTOR = 4;
+var MIN_COORD_MERGE_TOL = 8;
 function extractLines(fnArray, argsArray) {
   const horizontals = [];
   const verticals = [];
@@ -3153,6 +3195,55 @@ function classifyAndAdd(seg, lineWidth, horizontals, verticals) {
     verticals.push({ x1: x, y1, x2: x, y2, lineWidth });
   }
 }
+function preprocessLines(horizontals, verticals) {
+  let h = horizontals.filter((l) => l.lineWidth <= MAX_LINE_WIDTH);
+  let v = verticals.filter((l) => l.lineWidth <= MAX_LINE_WIDTH);
+  h = mergeParallelLines(h, "h");
+  v = mergeParallelLines(v, "v");
+  return { horizontals: h, verticals: v };
+}
+function mergeParallelLines(lines, dir) {
+  if (lines.length <= 1) return lines;
+  const sorted = [...lines].sort((a, b) => {
+    const posA = dir === "h" ? a.y1 : a.x1;
+    const posB = dir === "h" ? b.y1 : b.x1;
+    if (Math.abs(posA - posB) > 0.1) return posA - posB;
+    return dir === "h" ? a.x1 - b.x1 : a.y1 - b.y1;
+  });
+  const MERGE_TOL = 3;
+  const result = [sorted[0]];
+  for (let i = 1; i < sorted.length; i++) {
+    const prev = result[result.length - 1];
+    const curr = sorted[i];
+    const prevPos = dir === "h" ? prev.y1 : prev.x1;
+    const currPos = dir === "h" ? curr.y1 : curr.x1;
+    if (Math.abs(prevPos - currPos) <= MERGE_TOL) {
+      const prevStart = dir === "h" ? prev.x1 : prev.y1;
+      const prevEnd = dir === "h" ? prev.x2 : prev.y2;
+      const currStart = dir === "h" ? curr.x1 : curr.y1;
+      const currEnd = dir === "h" ? curr.x2 : curr.y2;
+      const overlap = Math.min(prevEnd, currEnd) - Math.max(prevStart, currStart);
+      const minLen = Math.min(prevEnd - prevStart, currEnd - currStart);
+      if (overlap > minLen * 0.3) {
+        if (dir === "h") {
+          prev.x1 = Math.min(prev.x1, curr.x1);
+          prev.x2 = Math.max(prev.x2, curr.x2);
+          prev.y1 = (prev.y1 + curr.y1) / 2;
+          prev.y2 = prev.y1;
+        } else {
+          prev.y1 = Math.min(prev.y1, curr.y1);
+          prev.y2 = Math.max(prev.y2, curr.y2);
+          prev.x1 = (prev.x1 + curr.x1) / 2;
+          prev.x2 = prev.x1;
+        }
+        prev.lineWidth = Math.max(prev.lineWidth, curr.lineWidth);
+        continue;
+      }
+    }
+    result.push(curr);
+  }
+  return result;
+}
 function filterPageBorderLines(horizontals, verticals, pageWidth, pageHeight) {
   const margin = 5;
   return {
@@ -3164,8 +3255,49 @@ function filterPageBorderLines(horizontals, verticals, pageWidth, pageHeight) {
     )
   };
 }
+function buildVertices(horizontals, verticals) {
+  const vertices = [];
+  const tol = CONNECT_TOL;
+  for (const h of horizontals) {
+    for (const v of verticals) {
+      if (v.x1 >= h.x1 - tol && v.x1 <= h.x2 + tol && h.y1 >= v.y1 - tol && h.y1 <= v.y2 + tol) {
+        const radius = Math.max(h.lineWidth, v.lineWidth, 1);
+        vertices.push({ x: v.x1, y: h.y1, radius });
+      }
+    }
+  }
+  return vertices;
+}
+function mergeVertices(vertices) {
+  if (vertices.length <= 1) return vertices;
+  const merged = [];
+  const used = new Array(vertices.length).fill(false);
+  for (let i = 0; i < vertices.length; i++) {
+    if (used[i]) continue;
+    let sumX = vertices[i].x, sumY = vertices[i].y;
+    let maxRadius = vertices[i].radius;
+    let count = 1;
+    for (let j = i + 1; j < vertices.length; j++) {
+      if (used[j]) continue;
+      const mergeTol = VERTEX_MERGE_FACTOR * Math.max(maxRadius, vertices[j].radius);
+      if (Math.abs(vertices[i].x - vertices[j].x) <= mergeTol && Math.abs(vertices[i].y - vertices[j].y) <= mergeTol) {
+        sumX += vertices[j].x;
+        sumY += vertices[j].y;
+        maxRadius = Math.max(maxRadius, vertices[j].radius);
+        count++;
+        used[j] = true;
+      }
+    }
+    merged.push({ x: sumX / count, y: sumY / count, radius: maxRadius });
+  }
+  return merged;
+}
 function buildTableGrids(horizontals, verticals) {
   if (horizontals.length < 2 || verticals.length < 2) return [];
+  const allVertices = buildVertices(horizontals, verticals);
+  const vertices = mergeVertices(allVertices);
+  if (vertices.length < 4) return [];
+  const globalRadius = vertices.reduce((max, v) => Math.max(max, v.radius), 1);
   const allLines = [
     ...horizontals.map((l, i) => ({ ...l, type: "h", id: i })),
     ...verticals.map((l, i) => ({ ...l, type: "v", id: i + horizontals.length }))
@@ -3176,21 +3308,74 @@ function buildTableGrids(horizontals, verticals) {
     const hLines = group.filter((l) => l.type === "h");
     const vLines = group.filter((l) => l.type === "v");
     if (hLines.length < 2 || vLines.length < 2) continue;
-    const rawYs = hLines.map((l) => l.y1);
-    const rowYs = clusterCoordinates(rawYs).sort((a, b) => b - a);
-    const rawXs = vLines.map((l) => l.x1);
-    const colXs = clusterCoordinates(rawXs).sort((a, b) => a - b);
+    let gx1 = Infinity, gy1 = Infinity, gx2 = -Infinity, gy2 = -Infinity;
+    for (const l of vLines) {
+      if (l.x1 < gx1) gx1 = l.x1;
+      if (l.x1 > gx2) gx2 = l.x1;
+    }
+    for (const l of hLines) {
+      if (l.y1 < gy1) gy1 = l.y1;
+      if (l.y1 > gy2) gy2 = l.y1;
+    }
+    const groupBbox = {
+      x1: gx1 - CONNECT_TOL,
+      y1: gy1 - CONNECT_TOL,
+      x2: gx2 + CONNECT_TOL,
+      y2: gy2 + CONNECT_TOL
+    };
+    const groupVertices = vertices.filter(
+      (v) => v.x >= groupBbox.x1 && v.x <= groupBbox.x2 && v.y >= groupBbox.y1 && v.y <= groupBbox.y2
+    );
+    const groupRadius = groupVertices.length > 0 ? groupVertices.reduce((max, v) => Math.max(max, v.radius), 1) : globalRadius;
+    const coordMergeTol = Math.max(VERTEX_MERGE_FACTOR * groupRadius, MIN_COORD_MERGE_TOL);
+    const rawYs = [
+      ...hLines.map((l) => l.y1),
+      ...groupVertices.map((v) => v.y)
+    ];
+    const rowYs = clusterCoordinates(rawYs, coordMergeTol).sort((a, b) => b - a);
+    const rawXs = [
+      ...vLines.map((l) => l.x1),
+      ...groupVertices.map((v) => v.x)
+    ];
+    const colXs = clusterCoordinates(rawXs, coordMergeTol).sort((a, b) => a - b);
     if (rowYs.length < 2 || colXs.length < 2) continue;
+    const validColXs = enforceMinWidth(colXs, MIN_COL_WIDTH);
+    const validRowYs = enforceMinHeight(rowYs, MIN_ROW_HEIGHT);
+    if (validRowYs.length < 2 || validColXs.length < 2) continue;
     const bbox = {
-      x1: colXs[0],
-      y1: rowYs[rowYs.length - 1],
-      x2: colXs[colXs.length - 1],
-      y2: rowYs[0]
+      x1: validColXs[0],
+      y1: validRowYs[validRowYs.length - 1],
+      x2: validColXs[validColXs.length - 1],
+      y2: validRowYs[0]
     };
-    grids.push({ rowYs, colXs, bbox });
+    grids.push({ rowYs: validRowYs, colXs: validColXs, bbox, vertexRadius: groupRadius });
   }
   return mergeAdjacentGrids(grids);
 }
+function enforceMinWidth(colXs, minWidth) {
+  if (colXs.length <= 2) return colXs;
+  const result = [colXs[0]];
+  for (let i = 1; i < colXs.length; i++) {
+    const prevX = result[result.length - 1];
+    if (colXs[i] - prevX < minWidth && i < colXs.length - 1) {
+      continue;
+    }
+    result.push(colXs[i]);
+  }
+  return result;
+}
+function enforceMinHeight(rowYs, minHeight) {
+  if (rowYs.length <= 2) return rowYs;
+  const result = [rowYs[0]];
+  for (let i = 1; i < rowYs.length; i++) {
+    const prevY = result[result.length - 1];
+    if (prevY - rowYs[i] < minHeight && i < rowYs.length - 1) {
+      continue;
+    }
+    result.push(rowYs[i]);
+  }
+  return result;
+}
 function mergeAdjacentGrids(grids) {
   if (grids.length <= 1) return grids;
   const sorted = [...grids].sort((a, b) => b.bbox.y2 - a.bbox.y2);
@@ -3199,9 +3384,10 @@ function mergeAdjacentGrids(grids) {
     const prev = merged[merged.length - 1];
     const curr = sorted[i];
     if (prev.colXs.length === curr.colXs.length) {
-      const colMatch = prev.colXs.every((x, ci) => Math.abs(x - curr.colXs[ci]) <= COORD_MERGE_TOL * 3);
+      const mergeTol = Math.max(VERTEX_MERGE_FACTOR * Math.max(prev.vertexRadius, curr.vertexRadius), 6) * 3;
+      const colMatch = prev.colXs.every((x, ci) => Math.abs(x - curr.colXs[ci]) <= mergeTol);
       const verticalGap = prev.bbox.y1 - curr.bbox.y2;
-      if (colMatch && verticalGap >= -COORD_MERGE_TOL && verticalGap <= 20) {
+      if (colMatch && verticalGap >= -CONNECT_TOL && verticalGap <= 20) {
         const allRowYs = [.../* @__PURE__ */ new Set([...prev.rowYs, ...curr.rowYs])].sort((a, b) => b - a);
         merged[merged.length - 1] = {
           rowYs: allRowYs,
@@ -3211,7 +3397,8 @@ function mergeAdjacentGrids(grids) {
             y1: Math.min(prev.bbox.y1, curr.bbox.y1),
             x2: Math.max(prev.bbox.x2, curr.bbox.x2),
             y2: Math.max(prev.bbox.y2, curr.bbox.y2)
-          }
+          },
+          vertexRadius: Math.max(prev.vertexRadius, curr.vertexRadius)
         };
         continue;
       }
@@ -3220,14 +3407,14 @@ function mergeAdjacentGrids(grids) {
   }
   return merged;
 }
-function clusterCoordinates(values) {
+function clusterCoordinates(values, tolerance) {
   if (values.length === 0) return [];
   const sorted = [...values].sort((a, b) => a - b);
   const clusters = [{ sum: sorted[0], count: 1 }];
   for (let i = 1; i < sorted.length; i++) {
     const last = clusters[clusters.length - 1];
     const avg = last.sum / last.count;
-    if (Math.abs(sorted[i] - avg) <= COORD_MERGE_TOL) {
+    if (Math.abs(sorted[i] - avg) <= tolerance) {
       last.sum += sorted[i];
       last.count++;
     } else {
@@ -3284,6 +3471,20 @@ function extractCells(grid, horizontals, verticals) {
   const numRows = rowYs.length - 1;
   const numCols = colXs.length - 1;
   if (numRows <= 0 || numCols <= 0) return [];
+  const vBorders = Array.from(
+    { length: numRows },
+    (_, r) => Array.from(
+      { length: numCols + 1 },
+      (_2, c) => hasVerticalLine(verticals, colXs[c], rowYs[r], rowYs[r + 1], grid.vertexRadius)
+    )
+  );
+  const hBorders = Array.from(
+    { length: numRows + 1 },
+    (_, r) => Array.from(
+      { length: numCols },
+      (_2, c) => hasHorizontalLine(horizontals, rowYs[r], colXs[c], colXs[c + 1], grid.vertexRadius)
+    )
+  );
   const occupied = Array.from({ length: numRows }, () => Array(numCols).fill(false));
   const cells = [];
   for (let r = 0; r < numRows; r++) {
@@ -3291,18 +3492,26 @@ function extractCells(grid, horizontals, verticals) {
       if (occupied[r][c]) continue;
       let colSpan = 1;
       let rowSpan = 1;
-      while (c + colSpan < numCols) {
-        const borderX = colXs[c + colSpan];
-        const topY = rowYs[r];
-        const botY = rowYs[r + 1];
-        if (hasVerticalLine(verticals, borderX, topY, botY)) break;
+      while (c + colSpan < numCols && !vBorders[r][c + colSpan]) {
+        let canExpand = true;
+        for (let dr = 0; dr < rowSpan; dr++) {
+          if (vBorders[r + dr][c + colSpan]) {
+            canExpand = false;
+            break;
+          }
+        }
+        if (!canExpand) break;
         colSpan++;
       }
       while (r + rowSpan < numRows) {
-        const borderY = rowYs[r + rowSpan];
-        const leftX = colXs[c];
-        const rightX = colXs[c + colSpan];
-        if (hasHorizontalLine(horizontals, borderY, leftX, rightX)) break;
+        let hasLine = false;
+        for (let dc = 0; dc < colSpan; dc++) {
+          if (hBorders[r + rowSpan][c + dc]) {
+            hasLine = true;
+            break;
+          }
+        }
+        if (hasLine) break;
         rowSpan++;
       }
       for (let dr = 0; dr < rowSpan; dr++) {
@@ -3326,28 +3535,30 @@ function extractCells(grid, horizontals, verticals) {
   }
   return cells;
 }
-function hasVerticalLine(verticals, x, topY, botY) {
-  const tol = COORD_MERGE_TOL + 1;
+function hasVerticalLine(verticals, x, topY, botY, vertexRadius) {
+  const tol = Math.max(VERTEX_MERGE_FACTOR * vertexRadius, 4);
   for (const v of verticals) {
     if (Math.abs(v.x1 - x) <= tol) {
       const cellH = Math.abs(topY - botY);
+      if (cellH < 0.1) continue;
       const overlapTop = Math.min(v.y2, topY);
       const overlapBot = Math.max(v.y1, botY);
       const overlap = overlapTop - overlapBot;
-      if (overlap >= cellH * 0.5) return true;
+      if (overlap >= cellH * 0.75) return true;
     }
   }
   return false;
 }
-function hasHorizontalLine(horizontals, y, leftX, rightX) {
-  const tol = COORD_MERGE_TOL + 1;
+function hasHorizontalLine(horizontals, y, leftX, rightX, vertexRadius) {
+  const tol = Math.max(VERTEX_MERGE_FACTOR * vertexRadius, 4);
   for (const h of horizontals) {
     if (Math.abs(h.y1 - y) <= tol) {
       const cellW = Math.abs(rightX - leftX);
+      if (cellW < 0.1) continue;
       const overlapLeft = Math.max(h.x1, leftX);
       const overlapRight = Math.min(h.x2, rightX);
       const overlap = overlapRight - overlapLeft;
-      if (overlap >= cellW * 0.5) return true;
+      if (overlap >= cellW * 0.75) return true;
     }
   }
   return false;
@@ -3358,23 +3569,24 @@ function mapTextToCells(items, cells) {
     result.set(cell, []);
   }
   for (const item of items) {
-    const cx = item.x + item.w / 2;
-    const cy = item.y;
     const pad = CELL_PADDING;
     let bestCell = null;
-    let bestDist = Infinity;
+    let bestScore = 0;
     for (const cell of cells) {
-      if (cx >= cell.bbox.x1 - pad && cx <= cell.bbox.x2 + pad && cy >= cell.bbox.y1 - pad && cy <= cell.bbox.y2 + pad) {
-        const cellCx = (cell.bbox.x1 + cell.bbox.x2) / 2;
-        const cellCy = (cell.bbox.y1 + cell.bbox.y2) / 2;
-        const dist = Math.abs(cx - cellCx) + Math.abs(cy - cellCy);
-        if (dist < bestDist) {
-          bestDist = dist;
-          bestCell = cell;
-        }
+      const ix1 = Math.max(item.x, cell.bbox.x1 - pad);
+      const ix2 = Math.min(item.x + item.w, cell.bbox.x2 + pad);
+      const iy1 = Math.max(item.y, cell.bbox.y1 - pad);
+      const iy2 = Math.min(item.y + (item.h || item.fontSize), cell.bbox.y2 + pad);
+      if (ix1 >= ix2 || iy1 >= iy2) continue;
+      const intersectArea = (ix2 - ix1) * (iy2 - iy1);
+      const itemArea = Math.max(item.w, 1) * Math.max(item.h || item.fontSize, 1);
+      const score = intersectArea / itemArea;
+      if (score > bestScore) {
+        bestScore = score;
+        bestCell = cell;
       }
     }
-    if (bestCell) {
+    if (bestCell && bestScore > 0.3) {
       result.get(bestCell).push(item);
     }
   }
@@ -3401,8 +3613,13 @@ function cellTextToString(items) {
   const textLines = lines.map((line) => {
     const s = line.sort((a, b) => a.x - b.x);
     if (s.length === 1) return s[0].text;
+    const evenSpaced = detectEvenSpacedItems(s);
     let result = s[0].text;
     for (let j = 1; j < s.length; j++) {
+      if (evenSpaced[j]) {
+        result += s[j].text;
+        continue;
+      }
       const gap = s[j].x - (s[j - 1].x + s[j - 1].w);
       const avgFs = (s[j].fontSize + s[j - 1].fontSize) / 2;
       const prevIsKorean = /[가-힣]$/.test(result);
@@ -3417,6 +3634,57 @@ function cellTextToString(items) {
     }
     return result;
   });
+  return mergeCellTextLines(textLines);
+}
+function detectEvenSpacedItems(items) {
+  const result = new Array(items.length).fill(false);
+  if (items.length < 3) return result;
+  let runStart = -1;
+  for (let i = 0; i < items.length; i++) {
+    const isShortKorean = /^[가-힣]{1}$/.test(items[i].text) || /^[\d]{1}$/.test(items[i].text);
+    if (isShortKorean && runStart >= 0 && i > 0) {
+      const gap = items[i].x - (items[i - 1].x + items[i - 1].w);
+      const maxRunGap = Math.max(items[i].fontSize * 3, 30);
+      if (gap > maxRunGap) {
+        if (i - runStart >= 3) markEvenRun(items, result, runStart, i);
+        runStart = i;
+        continue;
+      }
+    }
+    if (isShortKorean) {
+      if (runStart < 0) runStart = i;
+    } else {
+      if (runStart >= 0 && i - runStart >= 3) {
+        markEvenRun(items, result, runStart, i);
+      }
+      runStart = -1;
+    }
+  }
+  if (runStart >= 0 && items.length - runStart >= 3) {
+    markEvenRun(items, result, runStart, items.length);
+  }
+  return result;
+}
+function markEvenRun(items, result, start, end) {
+  const gaps = [];
+  for (let i = start + 1; i < end; i++) {
+    gaps.push(items[i].x - (items[i - 1].x + items[i - 1].w));
+  }
+  const posGaps = gaps.filter((g2) => g2 > 0);
+  if (posGaps.length < 2) return;
+  let minGap = Infinity, maxGap = -Infinity;
+  for (const g2 of posGaps) {
+    if (g2 < minGap) minGap = g2;
+    if (g2 > maxGap) maxGap = g2;
+  }
+  const avgFs = items[start].fontSize;
+  if (minGap >= avgFs * 0.1 && maxGap <= avgFs * 3 && maxGap / Math.max(minGap, 0.1) <= 3) {
+    for (let i = start + 1; i < end; i++) {
+      result[i] = true;
+    }
+  }
+}
+function mergeCellTextLines(textLines) {
   if (textLines.length <= 1) return textLines[0] || "";
   const merged = [textLines[0]];
   for (let i = 1; i < textLines.length; i++) {
@@ -3442,24 +3710,172 @@ var Y_TOL = 3;
 var COL_CLUSTER_TOL = 15;
 var MIN_ROWS = 3;
 var MIN_COLS = 2;
-var MIN_GAP_FACTOR = 1.5;
-var MIN_COL_FILL_RATIO = 0.3;
+var MIN_GAP_FACTOR = 2;
+var MIN_GAP_ABSOLUTE = 20;
+var MIN_COL_FILL_RATIO = 0.4;
 function detectClusterTables(items, pageNum) {
   if (items.length < MIN_ROWS * MIN_COLS) return [];
-  const rows = groupByBaseline(items);
+  const { merged, originMap } = mergeEvenSpacedClusters(items);
+  const rows = groupByBaseline(merged);
   if (rows.length < MIN_ROWS) return [];
-  const suspiciousRows = rows.filter((row) => hasSuspiciousGaps(row));
-  if (suspiciousRows.length < MIN_ROWS) return [];
-  const columns = extractColumnClusters(suspiciousRows);
-  if (columns.length < MIN_COLS) return [];
-  const tableRegions = findTableRegions(rows, columns);
   const results = [];
-  for (const region of tableRegions) {
-    const table = buildClusterTable(region.rows, columns, pageNum);
-    if (table) results.push(table);
+  const headerResult = detectHeaderRow(rows);
+  if (headerResult) {
+    const { columns, headerIdx } = headerResult;
+    const headerRow = rows[headerIdx];
+    const headerItems = [...headerRow.items].sort((a, b) => a.x - b.x);
+    const headerAndBelow = rows.slice(headerIdx);
+    const mergedRows = mergeMultiLineRows(headerAndBelow, columns);
+    const tableRegions = findTableRegionsByHeader(mergedRows, columns, headerItems);
+    for (const region of tableRegions) {
+      const table = buildClusterTable(region.rows, columns, pageNum);
+      if (table) {
+        expandUsedItems(table.usedItems, originMap);
+        results.push(table);
+      }
+    }
+  }
+  if (results.length === 0) {
+    const suspiciousRows = rows.filter((row) => hasSuspiciousGaps(row));
+    if (suspiciousRows.length >= MIN_ROWS) {
+      const columns = extractColumnClusters(suspiciousRows);
+      if (columns.length >= MIN_COLS) {
+        const tableRegions = findTableRegions(rows, columns);
+        for (const region of tableRegions) {
+          const mergedRows = mergeMultiLineRows(region.rows, columns);
+          const table = buildClusterTable(mergedRows, columns, pageNum);
+          if (table) {
+            expandUsedItems(table.usedItems, originMap);
+            results.push(table);
+          }
+        }
+      }
+    }
   }
   return results;
 }
+function mergeEvenSpacedClusters(items) {
+  const originMap = /* @__PURE__ */ new Map();
+  const rows = groupByBaseline(items);
+  const merged = [];
+  for (const row of rows) {
+    const sorted = [...row.items].sort((a, b) => a.x - b.x);
+    let i = 0;
+    while (i < sorted.length) {
+      if (/^[가-힣\d]$/.test(sorted[i].text)) {
+        let runEnd = i + 1;
+        while (runEnd < sorted.length && /^[가-힣\d]$/.test(sorted[runEnd].text)) {
+          const gap = sorted[runEnd].x - (sorted[runEnd - 1].x + sorted[runEnd - 1].w);
+          const fs = sorted[runEnd].fontSize;
+          if (gap < fs * 0.1 || gap > fs * 3) break;
+          runEnd++;
+        }
+        if (runEnd - i >= 3) {
+          const gaps = [];
+          for (let g2 = i + 1; g2 < runEnd; g2++) {
+            gaps.push(sorted[g2].x - (sorted[g2 - 1].x + sorted[g2 - 1].w));
+          }
+          let minG = Infinity, maxG = -Infinity;
+          for (const g2 of gaps) {
+            if (g2 < minG) minG = g2;
+            if (g2 > maxG) maxG = g2;
+          }
+          if (minG > 0 && maxG / minG <= 3) {
+            const run = sorted.slice(i, runEnd);
+            const text = run.map((r) => r.text).join("");
+            const first = run[0], last = run[runEnd - i - 1];
+            const item = {
+              text,
+              x: first.x,
+              y: first.y,
+              w: last.x + last.w - first.x,
+              h: first.h,
+              fontSize: first.fontSize,
+              fontName: first.fontName
+            };
+            originMap.set(item, run);
+            merged.push(item);
+            i = runEnd;
+            continue;
+          }
+        }
+      }
+      merged.push(sorted[i]);
+      i++;
+    }
+  }
+  return { merged, originMap };
+}
+function expandUsedItems(usedItems, originMap) {
+  const toAdd = [];
+  for (const item of usedItems) {
+    const origins = originMap.get(item);
+    if (origins) for (const o of origins) toAdd.push(o);
+  }
+  for (const a of toAdd) usedItems.add(a);
+}
+function detectHeaderRow(rows) {
+  const allItems = rows.flatMap((r) => r.items);
+  if (allItems.length === 0) return null;
+  let allMinX = Infinity, allMaxX = -Infinity;
+  for (const i of allItems) {
+    if (i.x < allMinX) allMinX = i.x;
+    const r = i.x + i.w;
+    if (r > allMaxX) allMaxX = r;
+  }
+  const pageSpan = allMaxX - allMinX;
+  if (pageSpan <= 0) return null;
+  for (let ri = 0; ri < rows.length; ri++) {
+    const row = rows[ri];
+    if (row.items.length < MIN_COLS || row.items.length > 6) continue;
+    if (row.items.some((i) => i.text.length > 8)) continue;
+    if (!row.items.some((i) => /[가-힣]/.test(i.text))) continue;
+    if (row.items.some((i) => /^[□■○●·※▶▷◆◇\-]/.test(i.text))) continue;
+    const sorted = [...row.items].sort((a, b) => a.x - b.x);
+    const xSpan = sorted[sorted.length - 1].x + sorted[sorted.length - 1].w - sorted[0].x;
+    if (xSpan / pageSpan < 0.4) continue;
+    const avgFs = sorted.reduce((s, i) => s + i.fontSize, 0) / sorted.length;
+    let hasLargeGap = false;
+    for (let i = 1; i < sorted.length; i++) {
+      const gap = sorted[i].x - (sorted[i - 1].x + sorted[i - 1].w);
+      if (gap >= avgFs * 2.5) {
+        hasLargeGap = true;
+        break;
+      }
+    }
+    if (!hasLargeGap) continue;
+    const columns = sorted.map((item) => ({ x: item.x, count: 0 }));
+    let matchCount = 0;
+    for (let j = ri + 1; j < rows.length && matchCount < MIN_ROWS + 2; j++) {
+      const matched = countMatchedColumnsRange(rows[j], columns, sorted);
+      if (matched >= MIN_COLS) matchCount++;
+    }
+    if (matchCount < MIN_ROWS) continue;
+    return { columns, headerIdx: ri };
+  }
+  return null;
+}
+function mergeMultiLineRows(rows, columns) {
+  if (rows.length <= 1) return rows;
+  const result = [rows[0]];
+  const allFontSizes = rows.flatMap((r) => r.items).map((i) => i.fontSize);
+  const avgFontSize = allFontSizes.length > 0 ? allFontSizes.reduce((s, v) => s + v, 0) / allFontSizes.length : 12;
+  for (let i = 1; i < rows.length; i++) {
+    const prev = result[result.length - 1];
+    const curr = rows[i];
+    const yGap = Math.abs(prev.y - curr.y);
+    const matchedCols = countMatchedColumns(curr, columns);
+    if (yGap < avgFontSize * 1.8 && curr.items.length <= 2 && (matchedCols < MIN_COLS || curr.items.length === 1)) {
+      result[result.length - 1] = {
+        y: prev.y,
+        items: [...prev.items, ...curr.items]
+      };
+    } else {
+      result.push(curr);
+    }
+  }
+  return result;
+}
 function groupByBaseline(items) {
   if (items.length === 0) return [];
   const sorted = [...items].sort((a, b) => b.y - a.y || a.x - b.x);
@@ -3481,8 +3897,9 @@ function groupByBaseline(items) {
 function hasSuspiciousGaps(row) {
   if (row.items.length < 2) return false;
   const sorted = [...row.items].sort((a, b) => a.x - b.x);
+  if (sorted.length === 2 && sorted[1].text.length > 20) return false;
   const avgFontSize = sorted.reduce((s, i) => s + i.fontSize, 0) / sorted.length;
-  const minGap = avgFontSize * MIN_GAP_FACTOR;
+  const minGap = Math.max(avgFontSize * MIN_GAP_FACTOR, MIN_GAP_ABSOLUTE);
   for (let i = 1; i < sorted.length; i++) {
     const gap = sorted[i].x - (sorted[i - 1].x + sorted[i - 1].w);
     if (gap >= minGap) return true;
@@ -3509,6 +3926,41 @@ function extractColumnClusters(rows) {
   const minCount = Math.max(2, Math.floor(rows.length * MIN_COL_FILL_RATIO));
   return clusters.filter((c) => c.count >= minCount).sort((a, b) => a.x - b.x);
 }
+function findTableRegionsByHeader(allRows, columns, headerItems) {
+  const regions = [];
+  let currentRegion = [];
+  let missStreak = 0;
+  for (const row of allRows) {
+    const matchedCols = countMatchedColumnsRange(row, columns, headerItems);
+    if (matchedCols >= MIN_COLS) {
+      currentRegion.push(row);
+      missStreak = 0;
+    } else if (currentRegion.length > 0 && (row.items.length <= 2 || missStreak === 0)) {
+      currentRegion.push(row);
+      missStreak++;
+    } else {
+      while (currentRegion.length > 0) {
+        const last = currentRegion[currentRegion.length - 1];
+        if (countMatchedColumnsRange(last, columns, headerItems) >= MIN_COLS) break;
+        currentRegion.pop();
+      }
+      if (currentRegion.length >= MIN_ROWS) {
+        regions.push({ rows: [...currentRegion] });
+      }
+      currentRegion = [];
+      missStreak = 0;
+    }
+  }
+  while (currentRegion.length > 0) {
+    const last = currentRegion[currentRegion.length - 1];
+    if (countMatchedColumnsRange(last, columns, headerItems) >= MIN_COLS) break;
+    currentRegion.pop();
+  }
+  if (currentRegion.length >= MIN_ROWS) {
+    regions.push({ rows: currentRegion });
+  }
+  return regions;
+}
 function findTableRegions(allRows, columns) {
   const regions = [];
   let currentRegion = [];
@@ -3544,18 +3996,81 @@ function countMatchedColumns(row, columns) {
   }
   return matched.size;
 }
-function assignToColumn(item, columns) {
-  const MAX_DIST = COL_CLUSTER_TOL * 3;
-  let bestCol = -1;
-  let bestDist = Infinity;
-  for (let ci = 0; ci < columns.length; ci++) {
-    const dist = Math.abs(item.x - columns[ci].x);
-    if (dist < bestDist) {
-      bestDist = dist;
-      bestCol = ci;
+function countMatchedColumnsRange(row, columns, headerItems) {
+  const boundaries = [];
+  for (let ci = 0; ci < headerItems.length; ci++) {
+    const left = ci === 0 ? 0 : (headerItems[ci - 1].x + headerItems[ci - 1].w + headerItems[ci].x) / 2;
+    const right = ci === headerItems.length - 1 ? Infinity : (headerItems[ci].x + headerItems[ci].w + headerItems[ci + 1].x) / 2;
+    boundaries.push({ left, right });
+  }
+  const matched = /* @__PURE__ */ new Set();
+  for (const item of row.items) {
+    for (let ci = 0; ci < boundaries.length; ci++) {
+      if (item.x >= boundaries[ci].left && item.x < boundaries[ci].right) {
+        matched.add(ci);
+        break;
+      }
     }
   }
-  return bestDist <= MAX_DIST ? bestCol : -1;
+  return matched.size;
+}
+function assignRowItems(items, columns, numCols) {
+  if (items.length === 0) return [];
+  const sorted = [...items].sort((a, b) => a.x - b.x);
+  const colCenters = columns.map((c) => c.x);
+  const gaps = [];
+  for (let i = 1; i < sorted.length; i++) {
+    gaps.push({ idx: i, size: sorted[i].x - (sorted[i - 1].x + sorted[i - 1].w) });
+  }
+  const gapSizes = gaps.map((g2) => g2.size).sort((a, b) => a - b);
+  const medianGap = gapSizes.length > 0 ? gapSizes[Math.floor(gapSizes.length / 2)] : 0;
+  const gapThreshold = sorted.length <= numCols + 1 ? 12 : Math.max(medianGap * 2.5, 12);
+  const significantGaps = gaps.filter((g2) => g2.size >= gapThreshold).sort((a, b) => b.size - a.size).slice(0, numCols - 1).sort((a, b) => a.idx - b.idx);
+  const groups = [];
+  let start = 0;
+  for (const gap of significantGaps) {
+    groups.push(sorted.slice(start, gap.idx));
+    start = gap.idx;
+  }
+  groups.push(sorted.slice(start));
+  const result = [];
+  const usedCols = /* @__PURE__ */ new Set();
+  const groupCenters = groups.map((g2) => {
+    let minX = Infinity, maxX = -Infinity;
+    for (const i of g2) {
+      if (i.x < minX) minX = i.x;
+      const r = i.x + i.w;
+      if (r > maxX) maxX = r;
+    }
+    return (minX + maxX) / 2;
+  });
+  const assignments = [];
+  for (let gi = 0; gi < groups.length; gi++) {
+    for (let ci = 0; ci < numCols; ci++) {
+      assignments.push({ gi, ci, dist: Math.abs(groupCenters[gi] - colCenters[ci]) });
+    }
+  }
+  assignments.sort((a, b) => a.dist - b.dist);
+  const assignedGroups = /* @__PURE__ */ new Set();
+  for (const { gi, ci } of assignments) {
+    if (assignedGroups.has(gi) || usedCols.has(ci)) continue;
+    result.push({ col: ci, items: groups[gi] });
+    assignedGroups.add(gi);
+    usedCols.add(ci);
+  }
+  for (let gi = 0; gi < groups.length; gi++) {
+    if (assignedGroups.has(gi)) continue;
+    let bestCol = 0, bestDist = Infinity;
+    for (let ci = 0; ci < numCols; ci++) {
+      const d = Math.abs(groupCenters[gi] - colCenters[ci]);
+      if (d < bestDist) {
+        bestDist = d;
+        bestCol = ci;
+      }
+    }
+    result.push({ col: bestCol, items: groups[gi] });
+  }
+  return result;
 }
 function buildClusterTable(rows, columns, pageNum) {
   const numCols = columns.length;
@@ -3573,12 +4088,12 @@ function buildClusterTable(rows, columns, pageNum) {
       usedItems.add(row.items[0]);
       continue;
     }
-    for (const item of row.items) {
-      const col = assignToColumn(item, columns);
-      if (col < 0) continue;
+    const assignments = assignRowItems(row.items, columns, numCols);
+    for (const { col, items } of assignments) {
+      const text = items.map((i) => i.text).join(" ");
       const existing = cells[r][col].text;
-      cells[r][col].text = existing ? existing + " " + item.text : item.text;
-      usedItems.add(item);
+      cells[r][col].text = existing ? existing + " " + text : text;
+      for (const item of items) usedItems.add(item);
     }
   }
   let emptyRows = 0;
@@ -3590,11 +4105,48 @@ function buildClusterTable(rows, columns, pageNum) {
     const hasValue = cells.some((row) => row[c].text !== "");
     if (!hasValue) return null;
   }
+  for (let r = numRows - 1; r >= 1; r--) {
+    const nonEmptyCols = cells[r].filter((c) => c.text.trim()).length;
+    if (nonEmptyCols !== 1) continue;
+    if (cells[r][0].text.trim() !== "") continue;
+    const contentText = cells[r].find((c) => c.text.trim())?.text.trim() || "";
+    if (/^[○●▶\-·]/.test(contentText)) continue;
+    for (let pr = r - 1; pr >= 0; pr--) {
+      if (cells[pr].some((c) => c.text.trim())) {
+        for (let c = 0; c < numCols; c++) {
+          const prev = cells[pr][c].text.trim();
+          const curr = cells[r][c].text.trim();
+          if (curr) cells[pr][c].text = prev ? prev + " " + curr : curr;
+        }
+        for (let c = 0; c < numCols; c++) cells[r][c].text = "";
+        break;
+      }
+    }
+  }
+  for (let r = 0; r < cells.length - 1; r++) {
+    const row = cells[r];
+    const hasCol0 = row[0].text.trim() !== "";
+    const hasColLast = numCols > 1 && row[numCols - 1].text.trim() !== "";
+    const midEmpty = row.slice(1, numCols - 1).every((c) => c.text.trim() === "");
+    if (hasCol0 && hasColLast && midEmpty) {
+      const next = cells[r + 1];
+      if (next[0].text.trim() === "" && next.some((c) => c.text.trim())) {
+        for (let c = 1; c < numCols; c++) {
+          const curr = next[c].text.trim();
+          if (curr) row[c].text = row[c].text.trim() ? row[c].text.trim() + " " + curr : curr;
+        }
+        for (let c = 0; c < numCols; c++) next[c].text = "";
+      }
+    }
+  }
+  const filteredCells = cells.filter((row) => row.some((c) => c.text.trim()));
+  const finalRowCount = filteredCells.length;
+  if (finalRowCount < MIN_ROWS) return null;
   const irTable = {
-    rows: numRows,
+    rows: finalRowCount,
     cols: numCols,
-    cells,
-    hasHeader: numRows > 1
+    cells: filteredCells,
+    hasHeader: finalRowCount > 1
   };
   const allItems = rows.flatMap((r) => r.items);
   let minX = Infinity, minY = Infinity, maxX = -Infinity, maxY = -Infinity;
@@ -3671,7 +4223,7 @@ async function parsePdfDocument(buffer, options) {
     const effectivePageCount = Math.min(pageCount, MAX_PAGES);
     const pageFilter = options?.pages ? parsePageRange(options.pages, effectivePageCount) : null;
     const totalTarget = pageFilter ? pageFilter.size : effectivePageCount;
-    const allFontSizes = [];
+    const fontSizeFreq = /* @__PURE__ */ new Map();
     const pageHeights = /* @__PURE__ */ new Map();
     let parsedPages = 0;
     for (let i = 1; i <= effectivePageCount; i++) {
@@ -3688,7 +4240,7 @@ async function parsePdfDocument(buffer, options) {
           warnings.push({ page: i, message: `${hiddenCount}\uAC1C \uC228\uACA8\uC9C4 \uD14D\uC2A4\uD2B8 \uC694\uC18C \uD544\uD130\uB9C1\uB428`, code: "HIDDEN_TEXT_FILTERED" });
         }
         for (const item of visible) {
-          if (item.fontSize > 0) allFontSizes.push(item.fontSize);
+          if (item.fontSize > 0) fontSizeFreq.set(item.fontSize, (fontSizeFreq.get(item.fontSize) || 0) + 1);
         }
         const opList = await page.getOperatorList();
         const pageBlocks = extractPageBlocksWithLines(visible, i, opList, viewport.width, viewport.height);
@@ -3727,7 +4279,7 @@ async function parsePdfDocument(buffer, options) {
         blocks.splice(removed[ri], 1);
       }
     }
-    const medianFontSize = computeMedianFontSize(allFontSizes);
+    const medianFontSize = computeMedianFontSizeFromFreq(fontSizeFreq);
     if (medianFontSize > 0) {
       detectHeadings(blocks, medianFontSize);
     }
@@ -3780,11 +4332,18 @@ function filterHiddenText(items, pageWidth, pageHeight) {
   }
   return { visible, hiddenCount };
 }
-function computeMedianFontSize(sizes) {
-  if (sizes.length === 0) return 0;
-  const sorted = [...sizes].sort((a, b) => a - b);
-  const mid = Math.floor(sorted.length / 2);
-  return sorted.length % 2 === 0 ? (sorted[mid - 1] + sorted[mid]) / 2 : sorted[mid];
+function computeMedianFontSizeFromFreq(freq) {
+  if (freq.size === 0) return 0;
+  let total = 0;
+  for (const count of freq.values()) total += count;
+  const sorted = [...freq.entries()].sort((a, b) => a[0] - b[0]);
+  const mid = Math.floor(total / 2);
+  let cumulative = 0;
+  for (const [size, count] of sorted) {
+    cumulative += count;
+    if (cumulative > mid) return size;
+  }
+  return sorted[sorted.length - 1][0];
 }
 function detectHeadings(blocks, medianFontSize) {
   for (const block of blocks) {
@@ -3810,11 +4369,21 @@ function collapseEvenSpacing(text) {
   if (tokens.length >= 3 && singleCharCount / tokens.length >= 0.7) {
     return tokens.join("");
   }
-  return text;
+  return text.replace(
+    /(?<![가-힣])[가-힣](?: [가-힣\d]){2,}(?![가-힣])/g,
+    (match) => match.replace(/ /g, "")
+  );
 }
 function shouldDemoteTable(table) {
   const allCells = table.cells.flatMap((row) => row.map((c) => c.text.trim())).filter(Boolean);
   const allText = allCells.join(" ");
+  if (table.rows <= 3 && table.cols <= 3) {
+    const totalCells2 = table.rows * table.cols;
+    const emptyCells2 = totalCells2 - allCells.length;
+    if (emptyCells2 >= totalCells2 * 0.3) return true;
+    if (/[□■◆○●▶ㅇ]/.test(allText)) return true;
+    if (/<[^>]+>/.test(allText)) return true;
+  }
   if (allText.length > 200) return false;
   if (/[□■◆○●▶]/.test(allText) && table.rows <= 3) return true;
   const totalCells = table.rows * table.cols;
@@ -3925,6 +4494,7 @@ function extractPageBlocksWithLines(items, pageNum, opList, pageWidth, pageHeigh
   if (items.length === 0) return [];
   let { horizontals, verticals } = extractLines(opList.fnArray, opList.argsArray);
   ({ horizontals, verticals } = filterPageBorderLines(horizontals, verticals, pageWidth, pageHeight));
+  ({ horizontals, verticals } = preprocessLines(horizontals, verticals));
   const grids = buildTableGrids(horizontals, verticals);
   if (grids.length > 0) {
     return extractBlocksWithGrids(items, pageNum, grids, horizontals, verticals);
@@ -3936,14 +4506,19 @@ function extractBlocksWithGrids(items, pageNum, grids, horizontals, verticals) {
   const usedItems = /* @__PURE__ */ new Set();
   const sortedGrids = [...grids].sort((a, b) => b.bbox.y2 - a.bbox.y2);
   for (const grid of sortedGrids) {
+    const numGridRows = grid.rowYs.length - 1;
+    const numGridCols = grid.colXs.length - 1;
+    if (numGridRows === 1 && numGridCols >= 2) continue;
     const tableItems = [];
     const pad = 3;
+    const gridW = grid.bbox.x2 - grid.bbox.x1;
     for (const item of items) {
       if (usedItems.has(item)) continue;
-      if (item.x >= grid.bbox.x1 - pad && item.x + item.w <= grid.bbox.x2 + pad && item.y >= grid.bbox.y1 - pad && item.y <= grid.bbox.y2 + pad) {
-        tableItems.push(item);
-        usedItems.add(item);
-      }
+      if (item.y < grid.bbox.y1 - pad || item.y > grid.bbox.y2 + pad) continue;
+      if (item.x < grid.bbox.x1 - pad || item.x + item.w > grid.bbox.x2 + pad) continue;
+      if (gridW < 120 && item.x + item.w > grid.bbox.x2 - 2) continue;
+      tableItems.push(item);
+      usedItems.add(item);
     }
     const cells = extractCells(grid, horizontals, verticals);
     if (cells.length === 0) continue;
@@ -3967,6 +4542,7 @@ function extractBlocksWithGrids(items, pageNum, grids, horizontals, verticals) {
       const cellItems = cellTextMap.get(cell) || [];
       let text = cellTextToString(cellItems);
       text = text.replace(/^[\s]*[-–—]\s*\d+\s*[-–—][\s]*$/gm, "").trim();
+      text = text.split("\n").map((line) => collapseEvenSpacing(line)).join("\n");
       irGrid[cell.row][cell.col] = {
         text,
         colSpan: cell.colSpan,
@@ -3991,23 +4567,58 @@ function extractBlocksWithGrids(items, pageNum, grids, horizontals, verticals) {
     if (shouldDemoteTable(irTable)) {
       const demoted = demoteTableToText(irTable);
       if (demoted) {
-        blocks.push({ type: "paragraph", text: demoted, pageNumber: pageNum, bbox: tableBbox, style: dominantStyle(tableItems) });
+        const text = numGridRows === 1 ? "\n" + demoted + "\n" : demoted;
+        blocks.push({ type: "paragraph", text, pageNumber: pageNum, bbox: tableBbox, style: dominantStyle(tableItems) });
       }
       continue;
     }
     blocks.push({ type: "table", table: irTable, pageNumber: pageNum, bbox: tableBbox });
   }
-  const remaining = items.filter((i) => !usedItems.has(i));
+  let remaining = items.filter((i) => !usedItems.has(i));
   if (remaining.length > 0) {
     remaining.sort((a, b) => b.y - a.y || a.x - b.x);
-    const textBlocks = detectListBlocks(extractPageBlocksFallback(remaining, pageNum));
-    const allBlocks = [...blocks, ...textBlocks];
-    allBlocks.sort((a, b) => {
+    const clusterItems = remaining.map((i) => ({
+      text: i.text,
+      x: i.x,
+      y: i.y,
+      w: i.w,
+      h: i.h,
+      fontSize: i.fontSize,
+      fontName: i.fontName
+    }));
+    const clusterResults = detectClusterTables(clusterItems, pageNum);
+    if (clusterResults.length > 0) {
+      const ciToIdx = /* @__PURE__ */ new Map();
+      for (let ci = 0; ci < clusterItems.length; ci++) ciToIdx.set(clusterItems[ci], ci);
+      const usedClusterIndices = /* @__PURE__ */ new Set();
+      for (const cr of clusterResults) {
+        for (const ci of cr.usedItems) {
+          const idx = ciToIdx.get(ci);
+          if (idx !== void 0) usedClusterIndices.add(idx);
+        }
+        blocks.push({ type: "table", table: cr.table, pageNumber: pageNum, bbox: cr.bbox });
+      }
+      remaining = remaining.filter((_, idx) => !usedClusterIndices.has(idx));
+    }
+    if (remaining.length > 0) {
+      const allY = remaining.map((i) => i.y);
+      const pageH = safeMax(allY) - safeMin(allY);
+      const groups = xyCutOrder(remaining, Math.max(15, pageH * 0.03));
+      const textBlocks = [];
+      for (const group of groups) {
+        if (group.length === 0) continue;
+        const groupBlocks = extractPageBlocksFallback(group, pageNum);
+        for (const b of groupBlocks) textBlocks.push(b);
+      }
+      const finalTextBlocks = detectListBlocks(textBlocks);
+      for (const b of finalTextBlocks) blocks.push(b);
+    }
+    blocks.sort((a, b) => {
       const ay = a.bbox ? a.bbox.y + a.bbox.height : 0;
       const by = b.bbox ? b.bbox.y + b.bbox.height : 0;
       return by - ay;
     });
-    return mergeAdjacentTableBlocks(allBlocks);
+    return mergeAdjacentTableBlocks(blocks);
   }
   return mergeAdjacentTableBlocks(blocks);
 }
@@ -4034,52 +4645,52 @@ function mergeAdjacentTableBlocks(blocks) {
 function extractPageBlocksFallback(items, pageNum) {
   if (items.length === 0) return [];
   const blocks = [];
-  const allYLines = groupByY(items);
-  const columns = detectColumns(allYLines);
-  if (columns && columns.length >= 3) {
-    const tableText = extractWithColumns(allYLines, columns);
-    const bbox = computeBBox(items, pageNum);
-    blocks.push({ type: "paragraph", text: tableText, pageNumber: pageNum, bbox, style: dominantStyle(items) });
-  } else {
-    const clusterItems = items.map((i) => ({
-      text: i.text,
-      x: i.x,
-      y: i.y,
-      w: i.w,
-      h: i.h,
-      fontSize: i.fontSize,
-      fontName: i.fontName
-    }));
-    const clusterResults = detectClusterTables(clusterItems, pageNum);
-    if (clusterResults.length > 0) {
-      const ciToIdx = /* @__PURE__ */ new Map();
-      for (let ci = 0; ci < clusterItems.length; ci++) ciToIdx.set(clusterItems[ci], ci);
-      const usedIndices = /* @__PURE__ */ new Set();
-      for (const cr of clusterResults) {
-        for (const ci of cr.usedItems) {
-          const idx = ciToIdx.get(ci);
-          if (idx !== void 0) usedIndices.add(idx);
-        }
-        blocks.push({ type: "table", table: cr.table, pageNumber: pageNum, bbox: cr.bbox });
+  const clusterItems = items.map((i) => ({
+    text: i.text,
+    x: i.x,
+    y: i.y,
+    w: i.w,
+    h: i.h,
+    fontSize: i.fontSize,
+    fontName: i.fontName
+  }));
+  const clusterResults = detectClusterTables(clusterItems, pageNum);
+  if (clusterResults.length > 0) {
+    const ciToIdx = /* @__PURE__ */ new Map();
+    for (let ci = 0; ci < clusterItems.length; ci++) ciToIdx.set(clusterItems[ci], ci);
+    const usedIndices = /* @__PURE__ */ new Set();
+    for (const cr of clusterResults) {
+      for (const ci of cr.usedItems) {
+        const idx = ciToIdx.get(ci);
+        if (idx !== void 0) usedIndices.add(idx);
       }
-      const remaining = items.filter((_, idx) => !usedIndices.has(idx));
-      if (remaining.length > 0) {
-        const yLines = groupByY(remaining);
-        for (const line of yLines) {
-          const text = mergeLineSimple(line);
-          if (!text.trim()) continue;
-          const bbox = computeBBox(line, pageNum);
-          blocks.push({ type: "paragraph", text, pageNumber: pageNum, bbox, style: dominantStyle(line) });
-        }
+      blocks.push({ type: "table", table: cr.table, pageNumber: pageNum, bbox: cr.bbox });
+    }
+    const remaining = items.filter((_, idx) => !usedIndices.has(idx));
+    if (remaining.length > 0) {
+      const yLines = groupByY(remaining);
+      for (const line of yLines) {
+        const text = mergeLineSimple(line);
+        if (!text.trim()) continue;
+        const bbox = computeBBox(line, pageNum);
+        blocks.push({ type: "paragraph", text, pageNumber: pageNum, bbox, style: dominantStyle(line) });
       }
-      blocks.sort((a, b) => {
-        const ay = a.bbox ? a.bbox.y + a.bbox.height : 0;
-        const by = b.bbox ? b.bbox.y + b.bbox.height : 0;
-        return by - ay;
-      });
+    }
+    blocks.sort((a, b) => {
+      const ay = a.bbox ? a.bbox.y + a.bbox.height : 0;
+      const by = b.bbox ? b.bbox.y + b.bbox.height : 0;
+      return by - ay;
+    });
+  } else {
+    const allYLines = groupByY(items);
+    const columns = detectColumns(allYLines);
+    if (columns && columns.length >= 3) {
+      const tableText = extractWithColumns(allYLines, columns);
+      const bbox = computeBBox(items, pageNum);
+      blocks.push({ type: "paragraph", text: tableText, pageNumber: pageNum, bbox, style: dominantStyle(items) });
     } else {
       const allY = items.map((i) => i.y);
-      const pageHeight = Math.max(...allY) - Math.min(...allY);
+      const pageHeight = safeMax(allY) - safeMin(allY);
       const gapThreshold = Math.max(15, pageHeight * 0.03);
       const orderedGroups = xyCutOrder(items, gapThreshold);
       for (const group of orderedGroups) {
@@ -4132,22 +4743,76 @@ function dominantStyle(items) {
   return { fontSize: dominantSize, fontName };
 }
 function normalizeItems(rawItems) {
-  return rawItems.filter((i) => typeof i.str === "string" && i.str.trim() !== "").map((i) => {
+  const items = [];
+  const spacePositions = [];
+  for (const i of rawItems) {
+    if (typeof i.str !== "string") continue;
+    const x = Math.round(i.transform[4]);
+    const y = Math.round(i.transform[5]);
+    if (!i.str.trim()) {
+      spacePositions.push({ x, y });
+      continue;
+    }
     const scaleY = Math.abs(i.transform[3]);
     const scaleX = Math.abs(i.transform[0]);
     const fontSize = Math.round(Math.max(scaleY, scaleX));
-    return {
-      text: i.str.trim(),
-      x: Math.round(i.transform[4]),
-      y: Math.round(i.transform[5]),
-      w: Math.round(i.width),
-      h: Math.round(i.height),
-      fontSize,
-      fontName: i.fontName || "",
-      // 0pt 폰트이거나 너비 0 → hidden text (prompt injection 의심)
-      isHidden: fontSize === 0 || i.width === 0 && i.str.trim().length > 0
-    };
-  }).sort((a, b) => b.y - a.y || a.x - b.x);
+    const w = Math.round(i.width);
+    const h = Math.round(i.height);
+    const isHidden = fontSize === 0 || i.width === 0 && i.str.trim().length > 0;
+    let text = i.str.trim();
+    if (/^[\d\s\-().·,☎]+$/.test(text) && /\d/.test(text) && / /.test(text)) {
+      text = text.replace(/ /g, "");
+    }
+    const split = splitEvenSpacedItem(text, x, w, fontSize);
+    if (split) {
+      for (const s of split) {
+        items.push({ text: s.text, x: s.x, y, w: s.w, h, fontSize, fontName: i.fontName || "", isHidden });
+      }
+    } else {
+      items.push({ text, x, y, w, h, fontSize, fontName: i.fontName || "", isHidden });
+    }
+  }
+  const sorted = items.sort((a, b) => b.y - a.y || a.x - b.x);
+  const deduped = [];
+  for (let i = 0; i < sorted.length; i++) {
+    let isDup = false;
+    for (let j = deduped.length - 1; j >= 0; j--) {
+      const prev = deduped[j];
+      if (prev.y - sorted[i].y > 3) break;
+      if (Math.abs(prev.y - sorted[i].y) <= 3 && prev.text === sorted[i].text && Math.abs(prev.x - sorted[i].x) <= 3) {
+        isDup = true;
+        break;
+      }
+    }
+    if (!isDup) deduped.push(sorted[i]);
+  }
+  if (spacePositions.length > 0) {
+    for (const item of deduped) {
+      for (const sp of spacePositions) {
+        if (Math.abs(sp.y - item.y) <= 3) {
+          const dist = item.x - sp.x;
+          if (dist >= 0 && dist <= 20) {
+            item.hasSpaceBefore = true;
+            break;
+          }
+        }
+      }
+    }
+  }
+  return deduped;
+}
+function splitEvenSpacedItem(text, itemX, itemW, fontSize) {
+  if (!/^[가-힣\d](?: [가-힣\d]){2,}$/.test(text)) return null;
+  const chars = text.split(" ");
+  if (chars.length < 3) return null;
+  const charW = itemW / chars.length;
+  if (charW > fontSize * 2) return null;
+  return chars.map((ch, idx) => ({
+    text: ch,
+    x: Math.round(itemX + idx * charW),
+    w: Math.round(charW * 0.8)
+    // 실제 글자 폭은 간격보다 좁음
+  }));
 }
 function groupByY(items) {
   if (items.length === 0) return [];
@@ -4172,14 +4837,14 @@ function isProseSpread(items) {
   for (let i = 1; i < sorted.length; i++) {
     gaps.push(sorted[i].x - (sorted[i - 1].x + sorted[i - 1].w));
   }
-  const maxGap = Math.max(...gaps);
+  const maxGap = safeMax(gaps);
   const avgLen = items.reduce((s, i) => s + i.text.length, 0) / items.length;
   return maxGap < 40 && avgLen < 5;
 }
 function detectColumns(yLines) {
   const allItems = yLines.flat();
   if (allItems.length === 0) return null;
-  const pageWidth = Math.max(...allItems.map((i) => i.x + i.w)) - Math.min(...allItems.map((i) => i.x));
+  const pageWidth = safeMax(allItems.map((i) => i.x + i.w)) - safeMin(allItems.map((i) => i.x));
   if (pageWidth < 100) return null;
   let bigoLineIdx = -1;
   for (let i = 0; i < yLines.length; i++) {
@@ -4211,7 +4876,7 @@ function detectColumns(yLines) {
   }
   const peaks = xClusters.filter((c) => c.count >= 3).sort((a, b) => a.minX - b.minX);
   if (peaks.length < 3) return null;
-  const MERGE_TOL = 30;
+  const MERGE_TOL = 40;
   const merged = [peaks[0]];
   for (let i = 1; i < peaks.length; i++) {
     const prev = merged[merged.length - 1];
@@ -4225,7 +4890,14 @@ function detectColumns(yLines) {
       merged.push({ ...peaks[i] });
     }
   }
-  const columns = merged.filter((c) => c.count >= 3).map((c) => c.minX);
+  const rawColumns = merged.filter((c) => c.count >= 3).map((c) => c.minX);
+  if (rawColumns.length < 3) return null;
+  const MIN_DETECT_COL_WIDTH = 30;
+  const columns = [rawColumns[0]];
+  for (let i = 1; i < rawColumns.length; i++) {
+    if (rawColumns[i] - columns[columns.length - 1] < MIN_DETECT_COL_WIDTH) continue;
+    columns.push(rawColumns[i]);
+  }
   return columns.length >= 3 ? columns : null;
 }
 function findColumn(x, columns) {
@@ -4353,6 +5025,16 @@ function buildGridTable(lines, columns) {
     }
     merged.splice(0, headerEnd, headerRow);
   }
+  for (const row of merged) {
+    for (let c = 0; c < row.length; c++) {
+      if (row[c]) row[c] = collapseEvenSpacing(row[c]);
+    }
+  }
+  const totalCells = merged.length * numCols;
+  const filledCells = merged.reduce((s, row) => s + row.filter((c) => c).length, 0);
+  if (filledCells < totalCells * 0.35 || merged.length < 2 || merged.length <= 3 && numCols >= 7) {
+    return merged.map((r) => r.filter((c) => c).join("	")).join("\n");
+  }
   const md = [];
   md.push("| " + merged[0].join(" | ") + " |");
   md.push("| " + merged[0].map(() => "---").join(" | ") + " |");
@@ -4364,12 +5046,32 @@ function buildGridTable(lines, columns) {
 function mergeLineSimple(items) {
   if (items.length <= 1) return items[0]?.text || "";
   const sorted = [...items].sort((a, b) => a.x - b.x);
+  const isEvenSpaced = detectEvenSpacedItems(sorted);
   let result = sorted[0].text;
   for (let i = 1; i < sorted.length; i++) {
     const gap = sorted[i].x - (sorted[i - 1].x + sorted[i - 1].w);
     const avgFs = (sorted[i].fontSize + sorted[i - 1].fontSize) / 2;
-    if (gap > 15) result += "	";
-    else if (gap < avgFs * 0.15) {
+    const tabThreshold = Math.max(avgFs * 2, 30);
+    if (gap > tabThreshold) {
+      result += "	";
+      result += sorted[i].text;
+      continue;
+    }
+    if (isEvenSpaced[i]) {
+      result += sorted[i].text;
+      continue;
+    }
+    if (sorted[i].hasSpaceBefore && gap >= avgFs * 0.05) {
+      result += " ";
+      result += sorted[i].text;
+      continue;
+    }
+    if (/[□■○●▶◆◇ㅇ]$/.test(sorted[i - 1].text) && /^[가-힣]/.test(sorted[i].text) && gap > 1) {
+      result += " ";
+      result += sorted[i].text;
+      continue;
+    }
+    if (gap < avgFs * 0.15) {
     } else if (gap < avgFs * 0.35 && (/[가-힣]$/.test(result) || /^[가-힣]/.test(sorted[i].text))) {
     } else if (gap > 3) result += " ";
     result += sorted[i].text;
@@ -4378,8 +5080,8 @@ function mergeLineSimple(items) {
 }
 function cleanPdfText(text) {
   return mergeKoreanLines(
-    text.replace(/^[\s]*[-–—]\s*[-–—]?\d+[-–—]?[\s]*[-–—]?[\s]*$/gm, "").replace(/^\s*\d+\s*\/\s*\d+\s*$/gm, "").replace(/\n\d{1,4}\n/g, "\n").replace(/\n\d{1,4}$/, "")
-  ).replace(/^(?!\|).{3,30}$/gm, (line) => collapseEvenSpacing(line)).replace(/\n{3,}/g, "\n\n").trim();
+    text.replace(/^\d{1,4}\n/, "").replace(/^[\s]*[-–—]\s*[-–—]?\d+[-–—]?[\s]*[-–—]?[\s]*$/gm, "").replace(/^\s*\d+\s*\/\s*\d+\s*$/gm, "").replace(/\n\d{1,4}\n/g, "\n").replace(/\n\d{1,4}$/, "").replace(/^#{1,6}\s*\d{1,4}\s*$/gm, "")
+  ).replace(/^(?!\| ---).*$/gm, (line) => collapseEvenSpacing(line)).replace(/([□■◆○●▶ㅇ])\s+([가-힣])\s+([가-힣])/g, "$1 $2$3").replace(/\n{3,}/g, "\n\n").trim();
 }
 function startsWithMarker(line) {
   const t = line.trimStart();
@@ -4571,7 +5273,7 @@ function mergeKoreanLines(text) {
       result[result.length - 1] = prev + " " + currTrimmed;
       continue;
     }
-    if (/[가-힣·,\-]$/.test(prev) && /^[가-힣(]/.test(curr) && !startsWithMarker(curr) && !isStandaloneHeader(prev)) {
+    if (/[가-힣·,\-]$/.test(prev) && /^[가-힣(]/.test(curr) && !startsWithMarker(curr) && !isStandaloneHeader(prev) && !startsWithMarker(prev)) {
       result[result.length - 1] = prev + " " + curr;
     } else {
       result.push(curr);
@@ -4619,7 +5321,7 @@ function getTextContent(el) {
   return el.textContent?.trim() ?? "";
 }
 function parseXml(text) {
-  return new import_xmldom2.DOMParser().parseFromString(text, "text/xml");
+  return new import_xmldom2.DOMParser().parseFromString(stripDtd(text), "text/xml");
 }
 function parseSharedStrings(xml) {
   const doc = parseXml(xml);
@@ -4906,7 +5608,7 @@ function getAttr(el, localName) {
   return null;
 }
 function parseXml2(text) {
-  return new import_xmldom3.DOMParser().parseFromString(text, "text/xml");
+  return new import_xmldom3.DOMParser().parseFromString(stripDtd(text), "text/xml");
 }
 function parseStyles(xml) {
   const doc = parseXml2(xml);
@@ -5306,7 +6008,13 @@ function normalize(s) {
 }
 var MAX_LEVENSHTEIN_LEN = 1e4;
 function levenshtein(a, b) {
-  if (a.length + b.length > MAX_LEVENSHTEIN_LEN) return Math.abs(a.length - b.length);
+  if (a.length + b.length > MAX_LEVENSHTEIN_LEN) {
+    const sampleLen = Math.min(500, a.length, b.length);
+    let diffs = 0;
+    for (let i = 0; i < sampleLen; i++) if (a[i] !== b[i]) diffs++;
+    const sampleRate = sampleLen > 0 ? diffs / sampleLen : 1;
+    return Math.abs(a.length - b.length) + Math.round(Math.min(a.length, b.length) * sampleRate);
+  }
   if (a.length > b.length) [a, b] = [b, a];
   const m = a.length;
   const n = b.length;
@@ -5589,13 +6297,20 @@ function extractInlineFields(text) {
 
 // src/hwpx/generator.ts
 var import_jszip5 = __toESM(require("jszip"), 1);
-var HWPML_NS = "http://www.hancom.co.kr/hwpml/2016/HwpMl";
+var NS_SECTION = "http://www.hancom.co.kr/hwpml/2011/section";
+var NS_PARA = "http://www.hancom.co.kr/hwpml/2011/paragraph";
+var NS_HEAD = "http://www.hancom.co.kr/hwpml/2011/head";
+var NS_OPF = "http://www.idpf.org/2007/opf/";
+var NS_HPF = "http://www.hancom.co.kr/schema/2011/hpf";
+var NS_OCF = "urn:oasis:names:tc:opendocument:xmlns:container";
 async function markdownToHwpx(markdown) {
   const blocks = parseMarkdownToBlocks(markdown);
   const sectionXml = blocksToSectionXml(blocks);
   const zip = new import_jszip5.default();
   zip.file("mimetype", "application/hwp+zip", { compression: "STORE" });
+  zip.file("META-INF/container.xml", generateContainerXml());
   zip.file("Contents/content.hpf", generateManifest());
+  zip.file("Contents/header.xml", generateHeaderXml());
   zip.file("Contents/section0.xml", sectionXml);
   return await zip.generateAsync({ type: "arraybuffer" });
 }
@@ -5640,8 +6355,111 @@ function parseMarkdownToBlocks(md) {
 function escapeXml(text) {
   return text.replace(/&/g, "&amp;").replace(/</g, "&lt;").replace(/>/g, "&gt;").replace(/"/g, "&quot;");
 }
+function generateContainerXml() {
+  return `<?xml version="1.0" encoding="UTF-8" standalone="yes" ?>
+<ocf:container xmlns:ocf="${NS_OCF}" xmlns:hpf="${NS_HPF}">
+  <ocf:rootfiles>
+    <ocf:rootfile full-path="Contents/content.hpf" media-type="application/hwpml-package+xml"/>
+  </ocf:rootfiles>
+</ocf:container>`;
+}
+function generateManifest() {
+  return `<?xml version="1.0" encoding="UTF-8" standalone="yes" ?>
+<opf:package xmlns:opf="${NS_OPF}" xmlns:hpf="${NS_HPF}" xmlns:hh="${NS_HEAD}">
+  <opf:manifest>
+    <opf:item id="header" href="Contents/header.xml" media-type="application/xml"/>
+    <opf:item id="section0" href="Contents/section0.xml" media-type="application/xml"/>
+  </opf:manifest>
+  <opf:spine>
+    <opf:itemref idref="header" linear="no"/>
+    <opf:itemref idref="section0" linear="yes"/>
+  </opf:spine>
+</opf:package>`;
+}
+function generateHeaderXml() {
+  return `<?xml version="1.0" encoding="UTF-8" standalone="yes" ?>
+<hh:head xmlns:hh="${NS_HEAD}" xmlns:hp="${NS_PARA}" version="1.4" secCnt="1">
+  <hh:beginNum page="1" footnote="1" endnote="1" pic="1" tbl="1" equation="1"/>
+  <hh:refList>
+    <hh:fontfaces itemCnt="7">
+      <hh:fontface lang="HANGUL" fontCnt="1">
+        <hh:font id="0" face="\uD568\uCD08\uB86C\uBC14\uD0D5" type="TTF" isEmbedded="0">
+          <hh:typeInfo familyType="FCAT_GOTHIC" weight="6" proportion="4" contrast="0" strokeVariation="1" armStyle="1" letterform="1" midline="1" xHeight="1"/>
+        </hh:font>
+      </hh:fontface>
+      <hh:fontface lang="LATIN" fontCnt="1">
+        <hh:font id="0" face="Times New Roman" type="TTF" isEmbedded="0">
+          <hh:typeInfo familyType="FCAT_OLDSTYLE" weight="5" proportion="4" contrast="2" strokeVariation="0" armStyle="0" letterform="0" midline="0" xHeight="4"/>
+        </hh:font>
+      </hh:fontface>
+      <hh:fontface lang="HANJA" fontCnt="1">
+        <hh:font id="0" face="\uD568\uCD08\uB86C\uBC14\uD0D5" type="TTF" isEmbedded="0">
+          <hh:typeInfo familyType="FCAT_GOTHIC" weight="6" proportion="4" contrast="0" strokeVariation="1" armStyle="1" letterform="1" midline="1" xHeight="1"/>
+        </hh:font>
+      </hh:fontface>
+      <hh:fontface lang="JAPANESE" fontCnt="1">
+        <hh:font id="0" face="\uAD74\uB9BC" type="TTF" isEmbedded="0">
+          <hh:typeInfo familyType="FCAT_GOTHIC" weight="6" proportion="0" contrast="0" strokeVariation="1" armStyle="1" letterform="1" midline="1" xHeight="1"/>
+        </hh:font>
+      </hh:fontface>
+      <hh:fontface lang="OTHER" fontCnt="1">
+        <hh:font id="0" face="\uAD74\uB9BC" type="TTF" isEmbedded="0">
+          <hh:typeInfo familyType="FCAT_GOTHIC" weight="6" proportion="0" contrast="0" strokeVariation="1" armStyle="1" letterform="1" midline="1" xHeight="1"/>
+        </hh:font>
+      </hh:fontface>
+      <hh:fontface lang="SYMBOL" fontCnt="1">
+        <hh:font id="0" face="Symbol" type="TTF" isEmbedded="0">
+          <hh:typeInfo familyType="FCAT_GOTHIC" weight="6" proportion="0" contrast="0" strokeVariation="1" armStyle="1" letterform="1" midline="1" xHeight="1"/>
+        </hh:font>
+      </hh:fontface>
+      <hh:fontface lang="USER" fontCnt="1">
+        <hh:font id="0" face="\uAD74\uB9BC" type="TTF" isEmbedded="0">
+          <hh:typeInfo familyType="FCAT_GOTHIC" weight="6" proportion="0" contrast="0" strokeVariation="1" armStyle="1" letterform="1" midline="1" xHeight="1"/>
+        </hh:font>
+      </hh:fontface>
+    </hh:fontfaces>
+    <hh:borderFills itemCnt="1">
+      <hh:borderFill id="0" threeD="0" shadow="0" centerLine="0" breakCellSeparateLine="0">
+        <hh:slash type="NONE" Crooked="0" isCounter="0"/>
+        <hh:backSlash type="NONE" Crooked="0" isCounter="0"/>
+        <hh:leftBorder type="NONE" width="0.1mm" color="0"/>
+        <hh:rightBorder type="NONE" width="0.1mm" color="0"/>
+        <hh:topBorder type="NONE" width="0.1mm" color="0"/>
+        <hh:bottomBorder type="NONE" width="0.1mm" color="0"/>
+        <hh:diagonal type="NONE" width="0.1mm" color="0"/>
+        <hh:fillInfo/>
+      </hh:borderFill>
+    </hh:borderFills>
+    <hh:charProperties itemCnt="1">
+      <hh:charPr id="0" height="1000" textColor="0" shadeColor="-1" useFontSpace="0" useKerning="0" symMark="NONE" borderFillIDRef="0">
+        <hh:fontRef hangul="0" latin="0" hanja="0" japanese="0" other="0" symbol="0" user="0"/>
+        <hh:ratio hangul="100" latin="100" hanja="100" japanese="100" other="100" symbol="100" user="100"/>
+        <hh:spacing hangul="0" latin="0" hanja="0" japanese="0" other="0" symbol="0" user="0"/>
+        <hh:relSz hangul="100" latin="100" hanja="100" japanese="100" other="100" symbol="100" user="100"/>
+        <hh:offset hangul="0" latin="0" hanja="0" japanese="0" other="0" symbol="0" user="0"/>
+      </hh:charPr>
+    </hh:charProperties>
+    <hh:tabProperties itemCnt="0"/>
+    <hh:numberings itemCnt="0"/>
+    <hh:bullets itemCnt="0"/>
+    <hh:paraProperties itemCnt="1">
+      <hh:paraPr id="0" tabIDRef="0" condense="0" fontLineHeight="0" snapToGrid="0" suppressOverlap="0" checked="0">
+        <hh:parLineBreak lineBreak="BREAK_LINE" wordBreak="BREAK_WORD" breakLatinWord="BREAK_WORD" breakNonLatinWord="BREAK_WORD"/>
+        <hh:parMargin left="0" right="0" prev="0" next="0" indent="0"/>
+        <hh:parBorder borderFillIDRef="0" offsetLeft="0" offsetRight="0" offsetTop="0" offsetBottom="0" connect="0" ignoreMargin="0"/>
+        <hh:parShade borderFillIDRef="0"/>
+        <hh:parTabList/>
+      </hh:paraPr>
+    </hh:paraProperties>
+    <hh:styles itemCnt="1">
+      <hh:style id="0" type="PARA" name="\uBC14\uD0D5\uAE00" engName="Normal" paraPrIDRef="0" charPrIDRef="0" nextStyleIDRef="0" langIDRef="1042" lockForm="0"/>
+    </hh:styles>
+  </hh:refList>
+  <hh:compatibleDocument targetProgram="HWP2018"/>
+</hh:head>`;
+}
 function generateParagraph(text) {
-  return `<hp:p><hp:run><hp:t>${escapeXml(text)}</hp:t></hp:run></hp:p>`;
+  return `<hp:p paraPrIDRef="0" styleIDRef="0"><hp:run charPrIDRef="0"><hp:t>${escapeXml(text)}</hp:t></hp:run></hp:p>`;
 }
 function generateTable(rows) {
   const trElements = rows.map((row) => {
@@ -5665,22 +6483,11 @@ function blocksToSectionXml(blocks) {
         return "";
     }
   }).join("\n  ");
-  return `<?xml version="1.0" encoding="UTF-8"?>
-<hs:sec xmlns:hs="${HWPML_NS}" xmlns:hp="${HWPML_NS}">
+  return `<?xml version="1.0" encoding="UTF-8" standalone="yes" ?>
+<hs:sec xmlns:hs="${NS_SECTION}" xmlns:hp="${NS_PARA}">
   ${body}
 </hs:sec>`;
 }
-function generateManifest() {
-  return `<?xml version="1.0" encoding="UTF-8"?>
-<opf:package xmlns:opf="http://www.idpf.org/2007/opf">
-  <opf:manifest>
-    <opf:item id="s0" href="section0.xml" media-type="application/xml"/>
-  </opf:manifest>
-  <opf:spine>
-    <opf:itemref idref="s0"/>
-  </opf:spine>
-</opf:package>`;
-}
 
 // src/index.ts
 async function parse(input, options) {