kordoc 4.24.36 → 4.24.37

package/README.md

@@ -125,6 +125,7 @@ await writeFile("보고서.hwpx", Buffer.from(hwpx))
 | 법령 별표 | 272건 · 346표 | HWP **346/346**, PDF **343/346 · 99.13%** |
 | PDF 전체 | 1,911문서 중 1,729건 채점 | 글 커버리지 **99.80%** |
 | OCR | 53문서 · 102쪽 · 66표 | CER **0.0239** · 글자 재현율 **99.32%** · 표 구조 일치 **34/66** |
+| KoOCR-Bench v2.0c (외부) | 공개 877쪽 (표·글·손글씨·순서) | 종합 기본값 **0.807** · `ocrRec: "auto"` **0.828** (4.24.37) |
 | opendataloader-bench | PDF 200문서 | 종합 기본값 **0.964** · OCR 끔 **0.939** |
 | HWPX 렌더 배치 | 한컴 PDF 짝 1,565문서(쪽 수 일치 1,508) · 글 조각 111,480 · 그림 4,300 | 글 위치(±1.5pt) **106,904 · 95.9%** · 그림 위치(±2pt) **4,213 · 98.0%** |
 
@@ -132,15 +133,15 @@ await writeFile("보고서.hwpx", Buffer.from(hwpx))
 
 ## 최근 업데이트
 
-최신 배포: **[4.24.36](https://github.com/chrisryugj/kordoc/releases/tag/v4.24.36)** · 2026-11-01
+최신 배포: **[4.24.37](https://github.com/chrisryugj/kordoc/releases/tag/v4.24.37)** · 2026-11-02
 
 | 버전 | 주요 변경 | 수치 변화 |
 | --- | --- | --- |
+| 4.24.37 | OCR 표: 통계표 묶음 안 빈칸 열·괘선 없는 본문 행 펼치기·행 합침 아래 행 가르기·이웃 칸 값 몰림·음영 머리 띠와 칠한 첫 열 경계 · 머리 행 있는 표를 차트로 풀던 것 · 시험지 문항 상자 풀기 · `ocrRec: "auto"`(손글씨 쪽만 손글씨 모델) | KoOCR-Bench v2.0c 공개 877쪽 종합 0.742 → **0.807**(`ocrRec: "auto"` **0.828**) · 자체 벤치 문서별 그대로 |
 | 4.24.36 | PDF 띄어쓰기(한컴 PDF): 자간 줄인 글의 낱말 공백 되살림 · 공백 글리프 위에 덧찍은 쉼표("55만원,") · 날짜 "2026. 8. 28." 공백 유지 · 줄 끝에서 갈린 연도 "20⏎25년" | PDF 글 어절 F1 99.076 → **99.103%** · 그 밖 지표 그대로 |
 | 4.24.35 | PDF: 자간 줄인 수치의 같은 숫자 둘("129.9") 지키기 · 글자마다 찍은 첨자 줄을 본문에 붙이기 · 수 가운데서 꺾인 줄("1,2⏎96억"·"(4.1⏎6.~5.16.)")과 칸 소수 잇기 · 표지 책등 외기둥 세로 글 | PDF 표 칸 완전 일치 0.99264 → **0.99267** · 칸 내용 0.98902 → **0.98905** · PDF 글 순서 99.359 → **99.360%** · 어절 F1 99.073 → **99.076%** · 그 밖 지표 그대로 |
 | 4.24.34 | to-hwpx: 파일 여러 개·폴더 일괄 변환(하위 폴더, 이미 있는 .hwpx 건너뜀) · PDF: 기호 글꼴 작은 네모로 찍고 공백으로 적은 "·" 글머리 되살림 · 글자 단위로 나눈 줄 머리 어미("는데"·"졌다")·겹조사 잇기 | PDF 표 칸 완전 일치 0.99221 → **0.99264** · 칸 내용 0.98896 → **0.98902** · PDF 글 재현율 99.883 → **99.885%** · 순서 99.356 → **99.359%** · 어절 F1 99.071 → **99.073%** · 그 밖 지표 그대로 |
 | 4.24.33 | PDF 표: 모아찍기 쪽 넘김에서 오른쪽 쪽 머리 상자로 넘어간 칸 글을 앞 칸에 잇기 · 칸을 한 겹 더 깐 1×1 클립(통계표 숫자 칸)을 중첩 표 대신 칸 글로 | PDF 표 칸 내용 일치 0.98794 → **0.98896** · 칸 완전 일치 0.99153 → **0.99221** · 그 밖 벤치 지표 그대로 |
-| 4.24.32 | PDF·OCR 두 단: 같은 줄 격자 두 단(속기록·시험지)을 단으로 · 옆 단 머리 문항이 앞 선지 하위 목록이 되던 것 · 짝·홀 쪽에서 칸 순서가 뒤집히는 머리말 상자 걷기 | PDF 글 읽기 순서 99.348% → **99.356%** · 정확도 99.680% → **99.682%** · 어절 F1 99.069% → **99.071%** · OCR 정답(텍스트층) 회의록 두 단이 바로잡혀 CER 0.0237 → 0.0239(정답 변화, 채점 그대로) |
 
 이전 버전과 상세 내용은 [변경 이력](CHANGELOG.md)에 있습니다.
 

package/dist/{-P5YOBEUM.js → -HX3RWDQF.js}

@@ -53,26 +53,26 @@ import {
   usesGaejosikMunche,
   usesLogicStructure,
   validateHwpx
-} from "./chunk-AZMS6M6S.js";
+} from "./chunk-GCFG67SF.js";
 import {
   blocksToPages
-} from "./chunk-IJT4GDQX.js";
+} from "./chunk-J26TNCL7.js";
 import {
   extractClickHereFields,
   patchHwp,
   patchHwpx
-} from "./chunk-EBYDC56L.js";
+} from "./chunk-7GN7O3K2.js";
 import {
   addMemoHwpx
-} from "./chunk-JYYA6APD.js";
+} from "./chunk-YNIZLSM4.js";
 import {
   placeSealHwpx
-} from "./chunk-U6NRQARP.js";
-import "./chunk-UX6WP55C.js";
+} from "./chunk-P3SZJJRK.js";
+import "./chunk-VRCXXZHT.js";
 import {
   convertToHwpx,
   hwpToHwpx
-} from "./chunk-F4O2VM2W.js";
+} from "./chunk-B2VRTOQZ.js";
 import {
   DEFAULT_REDACT_RULES,
   redactMarkdown,
@@ -80,7 +80,7 @@ import {
 } from "./chunk-3WSNLH4B.js";
 import {
   blocksToChunks
-} from "./chunk-ZPGXCXSX.js";
+} from "./chunk-MTSWWXSB.js";
 import {
   blocksToPdf,
   extractRenderedRegions,
@@ -91,13 +91,13 @@ import {
   renderHwp5Pages,
   renderHwpxToSvg,
   renderSceneToHtml
-} from "./chunk-MFBSYY52.js";
-import "./chunk-46M3WU3M.js";
-import "./chunk-B3WLIR4K.js";
-import "./chunk-N5GRO67A.js";
+} from "./chunk-XCMDIXWE.js";
+import "./chunk-ONCXLQV4.js";
+import "./chunk-T7K3CLXM.js";
+import "./chunk-DRDW5ULI.js";
 import "./chunk-VKVDMH7X.js";
-import "./chunk-FUAWTU6H.js";
-import "./chunk-KDUI534B.js";
+import "./chunk-LM7LYH54.js";
+import "./chunk-J42O5KBX.js";
 import "./chunk-VQFA55IC.js";
 import {
   detectFormat,
@@ -109,7 +109,7 @@ import {
   isZipFile
 } from "./chunk-GHHVT4Q5.js";
 import "./chunk-H7K5YHWQ.js";
-import "./chunk-IQJKLIJV.js";
+import "./chunk-JIMG76FW.js";
 import {
   AGENCY_COMMON_NAME,
   PRESET_ALIAS,
@@ -119,7 +119,7 @@ import {
   markdownToHwpx,
   normalizeGongmunPreset,
   placeTotalRow
-} from "./chunk-WH33GYVU.js";
+} from "./chunk-BT5RPLV4.js";
 import {
   ValueCursor,
   applySplices,
@@ -141,20 +141,20 @@ import {
   measureTextWidth,
   simulateWrap,
   simulateWrapKeepWord
-} from "./chunk-PZOF2YT5.js";
-import "./chunk-TPGYFICT.js";
+} from "./chunk-LWSMZJLR.js";
+import "./chunk-TQBNSAAS.js";
 import "./chunk-OFVHQ2TZ.js";
 import "./chunk-WFKFG2RT.js";
-import "./chunk-TBHWTT3M.js";
+import "./chunk-7MGZWDO4.js";
 import {
   blocksToMarkdown,
   flattenLayoutTables,
   hasStructuredCellContent
-} from "./chunk-LMJNZEGE.js";
+} from "./chunk-33BKUIIC.js";
 import "./chunk-KFF47XZ7.js";
 import {
   VERSION
-} from "./chunk-5LOELQQQ.js";
+} from "./chunk-BYOAA442.js";
 export {
   AGENCY_COMMON_NAME,
   BUILTIN_TEMPLATES,
@@ -268,4 +268,4 @@ export {
   usesLogicStructure,
   validateHwpx
 };
-//# sourceMappingURL=-P5YOBEUM.js.map
+//# sourceMappingURL=-HX3RWDQF.js.map

package/dist/{chunk-VX3TCX6T.js → chunk-2LDFUVVF.js}

@@ -39,12 +39,12 @@ import {
   sanitizeText,
   tidyScriptTags,
   unframeLayoutTables
-} from "./chunk-LI7GH25U.js";
+} from "./chunk-RR2FKPFN.js";
 import {
   KordocError,
   VERSION,
   sanitizeHref
-} from "./chunk-NDKS7IDO.js";
+} from "./chunk-64MVXWS5.js";
 
 // src/hwp5/to-hwpx/index.ts
 import JSZip from "jszip";
@@ -5270,7 +5270,7 @@ async function convertToHwpx(input, options) {
   if (format === "hwpx") return { data: bytes, mode: "preserve", warnings: ["\uC774\uBBF8 HWPX \uBB38\uC11C\uC785\uB2C8\uB2E4 \u2014 \uADF8\uB300\uB85C \uB461\uB2C8\uB2E4"] };
   if (format !== "hwp3" && format !== "hwpml") throw new KordocError(`HWP \uBB38\uC11C\uAC00 \uC544\uB2D9\uB2C8\uB2E4 (${format})`);
   const { parse } = await import("./index.js");
-  const { markdownToHwpx } = await import("./generator-OGOQGTEG.js");
+  const { markdownToHwpx } = await import("./generator-2WFWLJFG.js");
   const parsed = await parse(ab, { password: options?.password });
   if (!parsed.success) throw new KordocError(parsed.error ?? "\uBB38\uC11C\uB97C \uC77D\uC9C0 \uBABB\uD588\uC2B5\uB2C8\uB2E4");
   return {
@@ -5332,4 +5332,4 @@ export {
   hwpToHwpx,
   convertToHwpx
 };
-//# sourceMappingURL=chunk-VX3TCX6T.js.map
+//# sourceMappingURL=chunk-2LDFUVVF.js.map

package/dist/{chunk-LMJNZEGE.js → chunk-33BKUIIC.js}

@@ -5,7 +5,7 @@ import {
 import {
   escapeHtml,
   sanitizeHref
-} from "./chunk-5LOELQQQ.js";
+} from "./chunk-BYOAA442.js";
 
 // src/table/builder.ts
 var MAX_COLS = 200;
@@ -673,4 +673,4 @@ export {
   hasInlineCellBlocks,
   cellToMarkdown
 };
-//# sourceMappingURL=chunk-LMJNZEGE.js.map
+//# sourceMappingURL=chunk-33BKUIIC.js.map

package/dist/{chunk-KBU3KKZF.js → chunk-4MWGWS3B.js}

@@ -1,7 +1,7 @@
 #!/usr/bin/env node
 import {
   KordocError
-} from "./chunk-5LOELQQQ.js";
+} from "./chunk-BYOAA442.js";
 
 // src/hwpx/profile-io.ts
 import { z } from "zod";
@@ -89,4 +89,4 @@ export {
   formatProfileSchema,
   parseFormatProfileJson
 };
-//# sourceMappingURL=chunk-KBU3KKZF.js.map
+//# sourceMappingURL=chunk-4MWGWS3B.js.map

package/dist/{chunk-BPBY6762.js → chunk-4OWD6WRS.js}

@@ -14,6 +14,7 @@ import {
   findOcrTextGaps,
   gapGlyphs,
   grayCrop,
+  handwritingFallbackModels,
   inkBounds,
   inkStats,
   leaderRuns,
@@ -33,17 +34,125 @@ import {
   splitRowBands,
   starRun,
   tallInkCount
-} from "./chunk-DNDMC7WW.js";
+} from "./chunk-JRVZBYQ2.js";
 import {
   validateOcrOptions
-} from "./chunk-G5JSF2AY.js";
+} from "./chunk-IFWGHRCQ.js";
 import {
   disableOrtTelemetry
 } from "./chunk-DC6CTQTK.js";
 import {
   KordocError,
   OPTIONAL_DEP_INSTALL_HINT
-} from "./chunk-NDKS7IDO.js";
+} from "./chunk-64MVXWS5.js";
+
+// src/ocr/ocr-stacked-rows.ts
+var NUMERIC = /^[-+−△▲▽▼]?[\d.,]+[%p]?$|^[-–—·∼~]$|^\([-+−△▲▽▼]?[\d.,]+\)$/;
+var blank = () => ({ text: "", colSpan: 1, rowSpan: 1 });
+function rowAnchors(t, r, coveredFromAbove) {
+  const out = [];
+  const row = t.cells[r] ?? [];
+  for (let c = 0; c < row.length; ) {
+    if (coveredFromAbove[c]) {
+      c++;
+      continue;
+    }
+    const cell = row[c];
+    out.push({ c, cell });
+    c += Math.max(1, cell.colSpan ?? 1);
+  }
+  return out;
+}
+function splitTable(t) {
+  let changed = false;
+  const cover = Array.from({ length: t.rows }, () => Array(t.cols).fill(null));
+  for (let r = 0; r < t.rows; r++) for (let c = 0; c < (t.cells[r] ?? []).length; c++) {
+    const cell = t.cells[r][c], rs = Math.max(1, cell.rowSpan ?? 1), cs = Math.max(1, cell.colSpan ?? 1);
+    if (cover[r][c]) continue;
+    for (let i = r + 1; i < Math.min(t.rows, r + rs); i++) for (let j = c; j < Math.min(t.cols, c + cs); j++) cover[i][j] = cell;
+  }
+  const newRows = [];
+  for (let r = 0; r < t.rows; r++) {
+    const row = t.cells[r] ?? [];
+    const anchors = rowAnchors(t, r, cover[r]);
+    const lines = anchors.map((a) => a.cell.text.split("\n").map((s) => s.trim()).filter(Boolean));
+    const counts = /* @__PURE__ */ new Map();
+    for (const ls of lines) if (ls.length >= 2) counts.set(ls.length, (counts.get(ls.length) ?? 0) + 1);
+    const [k, n] = [...counts].sort((a, b) => b[1] - a[1])[0] ?? [0, 0];
+    const nonEmpty = lines.filter((ls) => ls.length).length;
+    const near = (ls) => ls.length >= 2 && ls.length <= k && ls.length >= Math.ceil(k * 0.8);
+    const stacked = lines.filter(near);
+    const flat = stacked.flat();
+    const numericShare = flat.length ? flat.filter((s) => NUMERIC.test(s.replace(/\s+/g, ""))).length / flat.length : 0;
+    const spanOk = anchors.every((a, i) => (a.cell.rowSpan ?? 1) === 1 || lines[i].length <= 1);
+    const ok = spanOk && k >= 2 && n >= 2 && stacked.length * 2 >= nonEmpty && lines.every((ls) => ls.length <= 1 || near(ls)) && numericShare >= 0.6;
+    if (!ok) {
+      newRows.push(row);
+      continue;
+    }
+    const sub = Array.from({ length: k }, () => Array.from({ length: row.length }, blank));
+    anchors.forEach(({ c, cell }, idx) => {
+      const ls = lines[idx], cs = Math.max(1, cell.colSpan ?? 1);
+      if (ls.length >= 2) for (let i = 0; i < k; i++) sub[i][c] = { ...cell, text: ls[i] ?? "", colSpan: cs, rowSpan: 1 };
+      else {
+        cell.rowSpan = (cell.rowSpan ?? 1) + k - 1;
+        sub[0][c] = cell;
+      }
+    });
+    for (const owner of new Set(cover[r].filter((x) => !!x))) owner.rowSpan = (owner.rowSpan ?? 1) + k - 1;
+    newRows.push(...sub);
+    changed = true;
+  }
+  if (changed) {
+    t.cells = newRows;
+    t.rows = newRows.length;
+  }
+  return changed;
+}
+function splitStackedOcrRows(blocks) {
+  for (const b of blocks) if (b.type === "table" && b.table) splitTable(b.table);
+  return blocks;
+}
+function unframeOcrLayoutTables(blocks, pageHeight) {
+  const out = [];
+  for (const b of blocks) {
+    const t = b.table;
+    if (b.type !== "table" || !t || !b.bbox || b.bbox.height < pageHeight * 0.5 || t.cols > 4) {
+      out.push(b);
+      continue;
+    }
+    const anchors = [];
+    const covered = Array.from({ length: t.rows }, () => /* @__PURE__ */ new Set());
+    for (let r = 0; r < t.rows; r++) for (let c = 0; c < (t.cells[r] ?? []).length; c++) {
+      if (covered[r].has(c)) continue;
+      const cell = t.cells[r][c];
+      const rs = Math.max(1, cell.rowSpan ?? 1), cs = Math.max(1, cell.colSpan ?? 1);
+      for (let i = r; i < Math.min(t.rows, r + rs); i++) for (let j = c; j < c + cs; j++) if (i !== r || j !== c) covered[i].add(j);
+      anchors.push({ r, c, cell });
+    }
+    const texts = anchors.map((a) => a.cell.text.trim()).filter(Boolean);
+    if (!texts.length) {
+      out.push(b);
+      continue;
+    }
+    const numeric = texts.filter((s) => NUMERIC.test(s.replace(/\s+/g, ""))).length;
+    const avg = texts.reduce((n, s) => n + s.length, 0) / texts.length;
+    if (numeric >= texts.length * 0.3 || avg < 15 || texts.filter((s) => /[①-⑤]/.test(s)).length < 2) {
+      out.push(b);
+      continue;
+    }
+    anchors.sort((a, b2) => a.c - b2.c || a.r - b2.r);
+    for (const { cell } of anchors) {
+      if (cell.blocks?.length) {
+        out.push(...cell.blocks);
+        continue;
+      }
+      const text = cell.text.trim();
+      if (text) out.push({ type: "paragraph", text, pageNumber: b.pageNumber, bbox: b.bbox });
+    }
+  }
+  return out;
+}
 
 // src/ocr/ruling-lines.ts
 var INK_LUMA_MAX = 205;
@@ -122,6 +231,14 @@ function detectRulingLines(rgba, width, height, scale) {
       if (lo - p.x1 <= chainEnd && p.x2 - hi <= chainEnd && cover >= (p.x2 - p.x1) * 0.5) fillH.push({ ...inside[0], x1: p.x1, x2: p.x2 });
     }
   }
+  for (const a of fill.horizontals) {
+    if (fillH.some((f) => f === a || Math.abs(f.y1 - a.y1) <= tol && f.x1 <= a.x1 + tol && f.x2 >= a.x2 - tol)) continue;
+    const p = ruleH.find((p2) => (Math.abs(p2.x1 - a.x1) <= tol || Math.abs(p2.x2 - a.x2) <= tol) && a.x1 >= p2.x1 - tol && a.x2 <= p2.x2 + tol && a.x2 - a.x1 >= (p2.x2 - p2.x1) * 0.7);
+    if (!p || ruleH.some((q) => Math.abs(q.y1 - a.y1) <= tol)) continue;
+    const loose = Math.abs(p.x1 - a.x1) <= tol ? a.x2 : a.x1;
+    const depthPx = Math.round(FILL_MIN_DEPTH_PT * scale);
+    if (fill.verticals.some((v) => Math.abs(v.x1 - loose) <= tol * 2 && Math.abs(v.y1 - a.y1) <= tol * 2 && v.y2 - v.y1 >= depthPx * 4)) fillH.push({ ...a, x1: p.x1, x2: p.x2 });
+  }
   horizontals.push(...fillH);
   const edgeH = [...ruleH, ...fillH];
   const depth = Math.round(FILL_MIN_DEPTH_PT * scale);
@@ -131,6 +248,7 @@ function detectRulingLines(rgba, width, height, scale) {
       const top = edgeH.filter((h) => h.x1 <= a.x1 + tol && h.x2 >= a.x1 - tol && h.y1 < a.y1 && h.y1 >= a.y1 - depth);
       verticals.push({ ...a, y1: top.length ? Math.min(...top.map((h) => h.y1)) : a.y1 });
     } else if (edgeH.some((p) => (Math.abs(p.x1 - a.x1) <= tol || Math.abs(p.x2 - a.x1) <= tol) && p.y1 >= a.y1 - tol && p.y1 <= a.y2 + tol)) verticals.push(a);
+    else if ([a.y1, a.y2].every((y) => edgeH.some((p) => Math.abs(p.y1 - y) <= tol * 2 && p.x1 < a.x1 - tol && p.x2 > a.x1 + tol)) && ruleH.some((p) => p.y1 > a.y1 + tol * 2 && p.y1 < a.y2 - tol * 2 && p.x1 < a.x1 - tol && p.x2 > a.x1 + tol)) verticals.push(a);
   }
   for (const v of shortV) {
     if (!endsOn(v.x1, v.y1).length || !endsOn(v.x1, v.y2).length) continue;
@@ -428,11 +546,11 @@ function trimInklessEnds(rgba, pageW, box, text, steps, stepPx) {
     for (let x = Math.max(0, Math.floor(cx - half)); x <= Math.min(w - 1, Math.ceil(cx + half)); x++) if (cols[x] && cols[x] < h * 0.9) return false;
     return true;
   };
-  const blank = (i) => !chars[i].trim();
+  const blank2 = (i) => !chars[i].trim();
   const mark = (i) => /\p{P}/u.test(chars[i]);
   let a = 0, b = chars.length;
-  while (a < b && (blank(a) || !mark(a) && inkless(a))) a++;
-  while (b > a && (blank(b - 1) || !mark(b - 1) && inkless(b - 1))) b--;
+  while (a < b && (blank2(a) || !mark(a) && inkless(a))) a++;
+  while (b > a && (blank2(b - 1) || !mark(b - 1) && inkless(b - 1))) b--;
   const cut = (from, to) => chars.slice(from, to).some((c) => c.trim());
   if (a === b || !cut(0, a) && !cut(b, chars.length)) return { text, steps };
   return { text: chars.slice(a, b).join(""), steps: steps.slice(a, b) };
@@ -1126,17 +1244,17 @@ function midBlankRow(gray, w, h, ink) {
   const isInk = (v) => ink.darkInk ? v <= ink.threshold : v > ink.threshold;
   const colInk = new Uint32Array(w);
   for (let y = 0; y < h; y++) for (let x = 0; x < w; x++) if (isInk(gray[y * w + x])) colInk[x]++;
-  const blank = (y) => {
+  const blank2 = (y) => {
     let n = 0;
     for (let x = 0; x < w; x++) if (colInk[x] < h * 0.85 && isInk(gray[y * w + x])) n++;
     return n <= Math.max(1, w * 0.02);
   };
   let best = null;
   for (let y = Math.floor(h * 0.3); y < Math.ceil(h * 0.7); y++) {
-    if (!blank(y)) continue;
+    if (!blank2(y)) continue;
     let a = y, b = y + 1;
-    while (a > 0 && blank(a - 1)) a--;
-    while (b < h && blank(b)) b++;
+    while (a > 0 && blank2(a - 1)) a--;
+    while (b < h && blank2(b)) b++;
     if (!best || Math.abs((a + b) / 2 - h / 2) < Math.abs((best[0] + best[1]) / 2 - h / 2)) best = [a, b];
     y = b;
   }
@@ -1938,12 +2056,27 @@ function builtinOcrConfig(o) {
   if (!o || o.ocrDet === void 0 && o.ocrDetLongSide === void 0 && o.ocrRec === void 0 && o.ocrRecDict === void 0 && !o.ocrLines) return DEFAULT_BUILTIN_OCR;
   const bad = validateOcrOptions(o);
   if (bad) throw new KordocError(bad);
+  const models = resolveOcrModels(o);
   return {
-    models: resolveOcrModels(o),
+    models,
+    ...o.ocrRec === "auto" ? { handwriting: handwritingFallbackModels(models) } : {},
     tuning: o.ocrDetLongSide === void 0 ? DEFAULT_OCR_TUNING : { ...DEFAULT_OCR_TUNING, detLongSide: o.ocrDetLongSide },
     lines: o.ocrLines === "all" ? "all" : o.ocrLines ? "kept" : void 0
   };
 }
+var AUTO_REC_PAGE_CONF = 0.9;
+var AUTO_REC_GAIN = 0.06;
+async function recognizeOcrPage(engine, config, tuning, rgba, width, height, stats, cellRules, rowRules) {
+  const items = await engine.recognizePage(rgba, width, height, stats, tuning, cellRules, rowRules);
+  if (!config.handwriting) return items;
+  const base = meanConfidence(items);
+  if (base >= AUTO_REC_PAGE_CONF) return items;
+  const altStats = { droppedLowConf: 0, ...stats.dropped ? { dropped: [] } : {} };
+  const alt = await (await getOcrEngine(config.handwriting)).recognizePage(rgba, width, height, altStats, tuning, cellRules, rowRules);
+  if (meanConfidence(alt) - base < AUTO_REC_GAIN) return items;
+  Object.assign(stats, altStats);
+  return alt;
+}
 var announced = /* @__PURE__ */ new Set();
 function reportModelDownload(p) {
   if (p.phase !== "download" || announced.has(p.spec.filename)) return;
@@ -1974,6 +2107,7 @@ async function runPdfOcr(buffer, targets, mode, warnings, onProgress, detectTabl
     () => import("@hyzyla/pdfium")
   );
   if (mode === "builtin") await ensureOcrModelsFor(config.models, reportModelDownload);
+  if (mode === "builtin" && config.handwriting) await ensureOcrModelsFor(config.handwriting, reportModelDownload);
   const engine = mode === "builtin" ? await getOcrEngine(config.models) : null;
   const pdfium = await pdfiumMod.PDFiumLibrary.init();
   const doc = await pdfium.loadDocument(new Uint8Array(buffer));
@@ -2023,12 +2157,14 @@ async function ocrOnePage(page, pageNo, mode, engine, warnings, detectTables, ve
     const stats = { droppedLowConf: 0, ...config.lines === "all" ? { dropped: [] } : {} };
     const ruling = vectorOps ? void 0 : detectRulingLines(upright, rw, rh, rh / pdfH);
     const tuning = regions ? regionTuning(config.tuning) : config.tuning;
-    const items = await engine.recognizePage(
+    const items = await recognizeOcrPage(
+      engine,
+      config,
+      tuning,
       upright,
       rw,
       rh,
       stats,
-      tuning,
       ruling ? [...ruling.cellDividers ?? [], ...cellEdgeVerticals(ruling)] : void 0,
       ruling ? cellEdgeHorizontals(ruling) : void 0
     );
@@ -2207,7 +2343,7 @@ function ocrItemsToBlocks(items, pageNumber, pdfW, pdfH, scale, extraLines, dete
     copyOcrTextGaps(it, normalized, 1 / scale);
     return normalized;
   });
-  return extractPageBlocksWithLines(norm, pageNumber, opList, pdfW, pdfH, extraLines, detectTables);
+  return unframeOcrLayoutTables(splitStackedOcrRows(extractPageBlocksWithLines(norm, pageNumber, opList, pdfW, pdfH, extraLines, detectTables)), pdfH);
 }
 async function tryImport2(name, loader) {
   try {
@@ -2252,6 +2388,7 @@ export {
   MAX_OCR_PIXELS,
   DEFAULT_BUILTIN_OCR,
   builtinOcrConfig,
+  recognizeOcrPage,
   reportModelDownload,
   toOcrLine,
   pageLines,
@@ -2262,4 +2399,4 @@ export {
   groupOcrLines,
   ocrItemsToBlocks
 };
-//# sourceMappingURL=chunk-BPBY6762.js.map
+//# sourceMappingURL=chunk-4OWD6WRS.js.map