kordoc 4.24.14 → 4.24.16

package/README.md

@@ -108,19 +108,19 @@ await writeFile("보고서.hwpx", Buffer.from(hwpx))
 
 ## 검증 결과
 
-**4.22.14 · 2026-10-19 검증**(렌더 배치·PDF 글·표·OCR 은 4.24.14 · 2026-10-28). 고정 코퍼스와 참조 기준으로 측정한 결과입니다.
+**4.22.14 · 2026-10-19 검증**(렌더 배치는 4.24.14, PDF 글·표·OCR 은 4.24.16 · 2026-10-28). 고정 코퍼스와 참조 기준으로 측정한 결과입니다.
 
 | 대상 | 규모 | 결과 |
 | --- | --- | --- |
 | HWPX | 2,424문서 · 10,342표 | 표 구조 일치 **10,342/10,342 · 100%** |
 | HWP 5.x ↔ HWPX | 1,130쌍 · 4,315표 | 짝 문서의 표 구조 일치 **100%** |
-| PDF 글 | 744쌍 | 글자 재현율 **99.883%** · 정확도 **99.678%** · 읽기 순서 **99.312%** · 어절 F1 **99.066%** |
+| PDF 글 | 744쌍 | 글자 재현율 **99.883%** · 정확도 **99.678%** · 읽기 순서 **99.312%** · 어절 F1 **99.067%** |
 | PDF 표 | 708쌍 · 2,331표 | 표 탐지 **99.83%** · 구조 일치 **98.24%** · 칸 F1 **0.9919** |
 | HWPX 재조판 | 다중줄 25,187문단 (12개 폴더) | 모든 줄 시작 일치 **24,941/25,187 · 99.02%** |
 | PDF 중첩표 (텍스트층) | 29문서 · 101표 | 구조·칸 내용 **101/101 · 100%** |
 | 법령 별표 | 272건 · 346표 | HWP **346/346**, PDF **342/346 · 98.84%** |
 | PDF 전체 | 1,911문서 중 1,729건 채점 | 글 커버리지 **99.80%** |
-| OCR | 53문서 · 102쪽 · 67표 | CER **0.0260** · 글자 재현율 **99.19%** · 표 구조 일치 **26/67** |
+| OCR | 53문서 · 102쪽 · 67표 | CER **0.0250** · 글자 재현율 **99.27%** · 표 구조 일치 **26/67** |
 | opendataloader-bench | PDF 200문서 | 종합 기본값 **0.964** · OCR 끔 **0.939** |
 | HWPX 렌더 배치 (4.24.14) | 한컴 PDF 짝 · 글 조각 90,840 · 그림 4,107 | 글 위치(±1.5pt) **87,643 · 96.5%** · 그림 위치(±2pt) **4,019 · 97.9%** |
 
@@ -128,15 +128,15 @@ await writeFile("보고서.hwpx", Buffer.from(hwpx))
 
 ## 최근 업데이트
 
-최신 배포: **[4.24.14](https://github.com/chrisryugj/kordoc/releases/tag/v4.24.14)** · 2026-10-28
+최신 배포: **[4.24.16](https://github.com/chrisryugj/kordoc/releases/tag/v4.24.16)** · 2026-10-28
 
 | 버전 | 주요 변경 | 수치 변화 |
 | --- | --- | --- |
+| 4.24.16 | OCR: 짧은 목차 리더 ··, 과제 번호의 Ⅲ(파·피로 읽음)·네모 번호 앞 [ ·붙임표가 붙은 원문자 ①, 한글 뒤 로마 숫자 획 수(Ⅰ↔Ⅱ), 칸 테두리 앞·빈칸 긴 서식 칸 끝 닫는 괄호 | OCR CER 0.0255 → **0.0250** · 글자 재현율 99.23% → **99.27%** · 정밀도 99.53% → **99.54%** · 열화 입력 8종 개선(원본 0.0397 → **0.0394**, 스캔만 0.0692 → 0.0693) |
+| 4.24.15 | OCR: 검출기가 놓친 서식 빈칸 네모 □□-□□□·목차 줄 머리 대시·줄 끝 홑 쪽 번호, 큰 글머리 ■, 리터 ℓ, 작은 고리 。, 윗줄에 묻힌 줄 머리 가운뎃점, 빗금 칸 환각 · PDF: 줄 간격 넓은 중·일문 두 줄 묶음 문단 잇기 | OCR CER 0.0260 → **0.0255** · 글자 재현율 99.19% → **99.23%** · 열화 입력 8종 개선(원본 0.0405 → **0.0397**, 잡음만 0.0485 → 0.0486) · PDF 어절 F1 99.066% → **99.067%** |
 | 4.24.14 | 렌더: 칸 간격(`cellSpacing`) 표의 칸 자리 · OCR: 공백·따옴표로 읽은 홑낫표 「, 검출기가 놓친 쪽 번호, 라틴 약어 앞·C 로 읽은 줄 머리 ○, 고리만 든 칸 ○, 마름모 글머리 ◆·◇·◈, 서식 칸 끝 닫는 괄호 | 렌더 글 위치 87,577 → **87,643** · OCR CER 0.0263 → **0.0260** · 열화 입력 9종 모두 개선(원본 0.0409 → **0.0405**) |
 | 4.24.13 | 기관팩(`.kpack`): 기관 문서 몇 건으로 서식 묶음을 만들어 그 서식대로 새 문서 · OCR: 선 없는 표 칸 순서·넓게 벌린 균등배분, 글자 상자 안 획 조각, 줄 머리 □, 기간 날짜 끝점, 전화 기호 ☎ · PDF·HWP: 한컴 소프트 하이픈 붙임표(증감 칸 음수 부호) · patch 추가 문단 강조 서식·렌더 자란 표 겹침(#175) | 기관팩 처음 보는 문서 재현 서울 68.3% → **74.3%** · 광양 77.1% → **94.5%** · OCR CER 0.0281 → **0.0263** · 열화 입력 9종 모두 개선(원본 0.0429 → **0.0409**) |
 | 4.24.12 | OCR: 조직도 상자 연결선을 괘선으로(과·팀 상자 열 단위 읽기), 요일 앞 날짜 마침표, 목차 리더 잡티, 화살표 모양 ⇨·➔·▶, 쉼표로 읽은 」 | OCR CER 0.0306 → **0.0281** · 글자 재현율 99.12% → **99.14%** · 열화 입력 9종 모두 개선(원본 0.0475 → **0.0429**) |
-| 4.24.11 | OCR: 표 칸·행 경계를 가로지른 검출 박스 가르기, 날짜 쉼표, 직함 앞 전(前)·현(現) · 렌더: 쪽 번호 글자 크기 · 생성: 서울 방침서 제목 줄간격·요약 정렬 | OCR CER 0.0310 → **0.0306** · 글자 재현율 99.10% → **99.12%** · 렌더 글 위치 87,546 → **87,577** · 생성 재현 0.806 → **0.807** |
-| 4.24.10 | 보안: 데모 lockfile 의 zod·fast-uri, 정부 GitLab 미러 복구(Java SDK `kordoc-pom.xml`) · OCR: 줄 머리 ▣·❐·‣, `\|` 로 읽은 로마 숫자, 날짜 끝 마침표 · PDF: 세로쓰기 、。, 그러데이션 배경 괘선, 중·일문 자동 간격 | OCR CER 0.0313 → **0.0310** · 글자 재현율 99.06% → **99.10%** · 열화 입력 9종 모두 개선(스캔 0.0734 → **0.0730**) · PDF 어절 F1 99.064% → **99.066%** |
 
 이전 버전과 상세 내용은 [변경 이력](CHANGELOG.md)에 있습니다.
 

package/dist/{-MG25ST6Y.js → -ZKOPHIGY.js}

@@ -64,10 +64,10 @@ import {
   usesGaejosikMunche,
   usesLogicStructure,
   validateHwpx
-} from "./chunk-UUNCLJB7.js";
+} from "./chunk-BHY2QAPM.js";
 import {
   blocksToChunks
-} from "./chunk-NVD24SLL.js";
+} from "./chunk-2UXGDAWQ.js";
 import {
   SPACE_EM_FIXED,
   SPACE_EM_FONT,
@@ -92,20 +92,20 @@ import {
   renderSceneToHtml,
   simulateWrap,
   simulateWrapKeepWord
-} from "./chunk-3LFNTOL7.js";
-import "./chunk-UTLHDVC5.js";
+} from "./chunk-UCEGWDWX.js";
+import "./chunk-F3VIVCIF.js";
 import {
   blocksToPages
-} from "./chunk-LSFVBS2I.js";
+} from "./chunk-FK7KGS5D.js";
 import "./chunk-VKVDMH7X.js";
 import "./chunk-VQFA55IC.js";
 import "./chunk-BYMPPVT3.js";
-import "./chunk-AM6TECXD.js";
+import "./chunk-DH564VPR.js";
 import {
   blocksToMarkdown,
   flattenLayoutTables,
   hasStructuredCellContent
-} from "./chunk-NIPFEMJJ.js";
+} from "./chunk-QOMLSZLM.js";
 import {
   ValueCursor,
   applySplices,
@@ -119,11 +119,11 @@ import {
   isLabelCell,
   placeSealHwpx,
   scanSectionXml
-} from "./chunk-DIFJPJGN.js";
-import "./chunk-ERYRF3EA.js";
+} from "./chunk-S2VWNC2W.js";
+import "./chunk-TSCQDCAR.js";
 import {
   VERSION
-} from "./chunk-ZIGWQ3IS.js";
+} from "./chunk-6PPVAQ2N.js";
 import {
   DEFAULT_REDACT_RULES,
   redactMarkdown,
@@ -239,4 +239,4 @@ export {
   usesLogicStructure,
   validateHwpx
 };
-//# sourceMappingURL=-MG25ST6Y.js.map
+//# sourceMappingURL=-ZKOPHIGY.js.map

package/dist/{chunk-NVD24SLL.js → chunk-2UXGDAWQ.js}

@@ -2,7 +2,7 @@
 import {
   assignTableIds,
   blocksToMarkdown
-} from "./chunk-NIPFEMJJ.js";
+} from "./chunk-QOMLSZLM.js";
 
 // src/chunks.ts
 var LIST_MARKER_RE = /^(?:[□■◇◆○◎●◦ㅇ•▪▸▶※-]|\d{1,3}[.)]|[가나다라마바사아자차카타파하][.)]|\([가나다라마바사아자차카타파하0-9]{1,3}\)|[①-⑳㉮-㉻㈎-㈛])\s/;
@@ -82,4 +82,4 @@ function blocksToChunks(blocks, options) {
 export {
   blocksToChunks
 };
-//# sourceMappingURL=chunk-NVD24SLL.js.map
+//# sourceMappingURL=chunk-2UXGDAWQ.js.map

package/dist/{chunk-MIJYCKJA.js → chunk-3TXAIEXL.js}

@@ -30,7 +30,7 @@ import {
   splitRowBands,
   starRun,
   tallInkCount
-} from "./chunk-LMQT2JFS.js";
+} from "./chunk-6K2SYTNJ.js";
 import {
   validateOcrOptions
 } from "./chunk-BYMPPVT3.js";
@@ -41,11 +41,11 @@ import {
   ocrSourcePath,
   parseCharacterDict,
   resolveOcrModels
-} from "./chunk-VBOFP27O.js";
+} from "./chunk-J5M5OW4S.js";
 import {
   KordocError,
   OPTIONAL_DEP_INSTALL_HINT
-} from "./chunk-ZIGWQ3IS.js";
+} from "./chunk-6PPVAQ2N.js";
 
 // src/ocr/ruling-lines.ts
 var INK_LUMA_MAX = 205;
@@ -535,6 +535,14 @@ function gaps({ chars, cx, gray, w, h, ink }) {
   const pitch = d.length >= 2 ? d.sort((a, b) => a - b)[d.length >> 1] : bandH;
   const nearTo = (g) => (i) => Math.abs(cx(i) - (g.x0 + g.x1) / 2) <= Math.max(pitch * 0.45, (g.x1 - g.x0) / 2 + 1);
   const loose = glyphs.filter((g) => /^[\u00b7.]/.test(g.mark) && !live.some(nearTo(g)));
+  const tail = [...live].reverse().findIndex((i) => !/^\d$/.test(chars[i]));
+  if (tail >= 1 && tail <= 3 && tail < live.length) {
+    const prev = live[live.length - 1 - tail], num = live[live.length - tail];
+    const run = loose.filter((g) => (g.x0 + g.x1) / 2 > cx(prev) && (g.x0 + g.x1) / 2 < cx(num));
+    const mark = run[0]?.mark[0];
+    if (run.length >= 2 && run.length <= 3 && run.every((g) => g.mark[0] === mark) && /^[가-힣?)]$/.test(chars[prev]) && !glyphs.some((g) => !run.includes(g) && (g.x0 + g.x1) / 2 > cx(prev) && (g.x0 + g.x1) / 2 < cx(num)))
+      chars[prev] += " " + mark.repeat(run.length) + " ";
+  }
   for (const g of glyphs.sort((a, b) => a.x0 - b.x0)) {
     const mid = (g.x0 + g.x1) / 2;
     const near = nearTo(g);
@@ -547,6 +555,7 @@ function gaps({ chars, cx, gray, w, h, ink }) {
     }
     if (cover.length) {
       if (g.mark.startsWith("\u25A1") && cover.length === 1 && /^[0Oo\u25cb]$/.test(chars[cover[0]])) chars[cover[0]] = "\u25A1";
+      if (g.mark === "\u3002" && cover.length === 1 && /^[\u00b7\u2022\u318d]$/.test(chars[cover[0]])) chars[cover[0]] = "\u3002";
       continue;
     }
     if (g.mark === "\u25A1\0") continue;
@@ -611,13 +620,14 @@ function roman({ chars, cx, gray, w, h, ink }) {
     if (serifOne(gray, w, h, ink, cx(i))) chars[i] = "\u2160";
   }
   for (let i = 0; i < chars.length; i++) {
-    if (chars[i] !== "|" || chars[i - 1] === "|" || chars[i + 1] === "|") continue;
+    const hangul = /^[파피]$/.test(chars[i]);
+    if (chars[i] !== "|" && !hangul || chars[i - 1] === "|" || chars[i + 1] === "|") continue;
     let p = i - 1, q = i + 1;
     while (p >= 0 && !chars[p].trim()) p--;
     while (q < chars.length && !chars[q].trim()) q++;
-    const glued = /[가-힣]$/.test(chars[i - 1] ?? "") || (chars[i + 1] ?? "") === "." || chars[p] === "-" && chars[q] === "-";
+    const glued = hangul ? chars[p] === "(" && chars[q] === "-" : /[가-힣]$/.test(chars[i - 1] ?? "") || (chars[i + 1] ?? "") === "." || chars[p] === "-" && chars[q] === "-";
     if (!glued) continue;
-    const n = romanStems(gray, w, h, ink, p >= 0 ? cx(p) : -Infinity, q < chars.length ? cx(q) : Infinity);
+    const n = romanStems(gray, w, h, ink, p >= 0 ? cx(p) : -Infinity, q < chars.length ? cx(q) : Infinity, /[가-힣]$/.test(chars[p] ?? ""));
     if (n) chars[i] = String.fromCharCode(8543 + n);
   }
   const inRoman = (i) => ROMAN.test(chars[i]) || chars[i] === "1" && (ROMAN.test(chars[i - 1] ?? "") || ROMAN.test(chars[i + 1] ?? ""));
@@ -632,7 +642,7 @@ function roman({ chars, cx, gray, w, h, ink }) {
       continue;
     }
     let e = s;
-    while (e < chars.length && inRoman(e)) e++;
+    while (e < chars.length && (inRoman(e) || e > s && chars[e] === "|" && chars[e + 1] !== "|")) e++;
     const tok = chars.slice(s, e).join("");
     if (e < chars.length && /^[A-Za-z]/.test(chars[e])) {
       s = e;
@@ -641,7 +651,7 @@ function roman({ chars, cx, gray, w, h, ink }) {
     let p = s - 1, q = e;
     while (p >= 0 && !chars[p].trim()) p--;
     while (q < chars.length && !chars[q].trim()) q++;
-    const n = romanStems(gray, w, h, ink, p >= 0 ? cx(p) : -Infinity, q < chars.length ? cx(q) : Infinity);
+    const n = romanStems(gray, w, h, ink, p >= 0 ? cx(p) : -Infinity, q < chars.length ? cx(q) : Infinity, /[가-힣]$/.test(chars[p] ?? ""));
     const canon = n ? [["\u2160", "I"], ["\u2161", "II"], ["\u2162", "III"]][n - 1] : null;
     if (canon && !canon.includes(tok)) {
       chars[s] = canon[0];
@@ -660,7 +670,9 @@ function restoreSymbols(text, ringDecided = false) {
   s = s.replace(/[\u2206\u0394]/g, "\u25B3");
   s = s.replace(/\(cid:\d*\)/g, "");
   s = s.replace(/(^|[^\uac00-\ud7a3])(전|현)\(\s?\)(?=\S)/g, (_, a, b) => `${a}${b}(${b === "\uC804" ? "\u524D" : "\u73FE"})`);
+  if (/[가-힣]/.test(s)) s = s.replace(/(\d)l(?![A-Za-z])/g, "$1\u2113");
   s = s.replace(/([(（])\s?[A-Za-z@&]\s?(?=(?:0\d{1,2}-)?\d{3,4}-\d{4}\b)/g, "$1\u260E");
+  if (!s.includes("]")) s = s.replace(/^\[(?=[\u2460-\u2473])/, "");
   return smartQuotes(dateDots(dateCommas(joinDigitGroups(s))));
 }
 function dateCommas(s) {
@@ -907,6 +919,182 @@ var MIN_INK_RATIO = 0.01;
 var MAX_LOCAL_MODEL_BYTES = 1024 * 1024 * 1024;
 var DET_MEAN = [0.485, 0.456, 0.406];
 var DET_STD = [0.229, 0.224, 0.225];
+function rowHeadDashes(rgba, width, items) {
+  const out = [];
+  const dark = (x, y) => {
+    const p = (y * width + x) * 4;
+    return rgba[p] * 77 + rgba[p + 1] * 150 + rgba[p + 2] * 29 >> 8 < 160;
+  };
+  const inBox = (x, y) => items.some((b) => x >= b.x && x < b.x + b.w && y >= b.y && y < b.y + b.h);
+  const done = /* @__PURE__ */ new Set();
+  for (const it of [...items]) {
+    if (done.has(it) || it.h < 12) continue;
+    const row = items.filter((o) => Math.min(o.y + o.h, it.y + it.h) - Math.max(o.y, it.y) >= Math.min(o.h, it.h) * 0.6);
+    for (const o of row) done.add(o);
+    if (row.reduce((k, o) => k + (o.text.match(/[가-힣]/g)?.length ?? 0), 0) < 2) continue;
+    const first = row.reduce((m, o) => o.x < m.x ? o : m);
+    if (!/^[가-힣]/.test(first.text)) continue;
+    if (row.filter((o) => /\d/.test(o.text)).length >= 2) continue;
+    const y0 = Math.min(...row.map((o) => o.y)), y1 = Math.max(...row.map((o) => o.y + o.h)), h = y1 - y0;
+    const x0 = Math.max(0, Math.round(first.x - 2.5 * h)), x1 = first.x - 2;
+    let minX = -1, maxX = -1, minY = Infinity, maxY = -1, groups = 0;
+    for (let x = x0; x < x1; x++) {
+      let on = false;
+      for (let y = y0; y < y1; y++) if (dark(x, y) && !inBox(x, y)) {
+        on = true;
+        if (y < minY) minY = y;
+        if (y > maxY) maxY = y;
+      }
+      if (!on) continue;
+      if (maxX < 0 || x - maxX > 2) {
+        groups++;
+        if (groups === 1) minX = x;
+      }
+      maxX = x;
+    }
+    const bw = maxX - minX + 1, bh = maxY - minY + 1, cy = ((minY + maxY) / 2 - y0) / h;
+    if (groups !== 1 || bh > h * 0.2 || bw < h * 0.3 || bw > h * 1.2 || cy < 0.3 || cy > 0.7) continue;
+    const faint = (x, y) => {
+      const p = (y * width + x) * 4;
+      return rgba[p] * 77 + rgba[p + 1] * 150 + rgba[p + 2] * 29 >> 8 < 230;
+    };
+    let touch = minX <= x0;
+    for (let y = minY; y <= maxY && !touch; y++) for (let d = 2; d <= 4 && !touch; d++) touch = minX - d >= 0 && faint(minX - d, y) || faint(maxX + d, y);
+    if (touch) continue;
+    out.push({ text: "-", x: minX, y: y0, w: bw, h, confidence: first.confidence });
+  }
+  return out;
+}
+function rowTailBoxes(rgba, width, items) {
+  const out = [];
+  const dark = (x, y) => {
+    const p = (y * width + x) * 4;
+    return rgba[p] * 77 + rgba[p + 1] * 150 + rgba[p + 2] * 29 >> 8 < 160;
+  };
+  const done = /* @__PURE__ */ new Set();
+  for (const it of items) {
+    if (done.has(it) || it.h < 12) continue;
+    const row = items.filter((o) => Math.min(o.y + o.h, it.y + it.h) - Math.max(o.y, it.y) >= Math.min(o.h, it.h) * 0.6);
+    for (const o of row) done.add(o);
+    if (row.reduce((k, o) => k + (o.text.match(/[가-힣]/g)?.length ?? 0), 0) < 2) continue;
+    const last = row.reduce((m, o) => o.x + o.w > m.x + m.w ? o : m);
+    if (/\d$/.test(last.text.trim())) continue;
+    const y0 = Math.min(...row.map((o) => o.y)), y1 = Math.max(...row.map((o) => o.y + o.h)), h = y1 - y0;
+    const x0 = last.x + last.w + 4, x1 = Math.floor(width * 0.95);
+    if (x1 - x0 < h) continue;
+    let groups = 0, minX = -1, maxX = -1, minY = Infinity, maxY = -1;
+    for (let x = x0; x < x1; x++) {
+      let on = false;
+      for (let y = y0; y < y1; y++) if (dark(x, y)) {
+        on = true;
+        if (y < minY) minY = y;
+        if (y > maxY) maxY = y;
+      }
+      if (!on) continue;
+      if (maxX < 0 || x - maxX > h) {
+        groups++;
+        if (groups === 1) minX = x;
+      }
+      maxX = x;
+    }
+    if (groups !== 1 || maxY - minY + 1 < h * 0.5 || maxX - minX + 1 > h * 4) continue;
+    const height = Math.floor(rgba.length / 4 / width);
+    let rule = false;
+    for (let x = minX; x <= maxX && !rule; x++) rule = minY > 1 && dark(x, minY - 2) || maxY + 2 < height && dark(x, maxY + 2);
+    if (rule) continue;
+    const pad = 4;
+    out.push({ x: Math.max(0, minX - pad), y: Math.max(0, minY - pad), w: maxX - minX + 1 + 2 * pad, h: maxY - minY + 1 + 2 * pad });
+  }
+  return out;
+}
+function placeholderSquares(rgba, width, height, items) {
+  const dark = new Uint8Array(width * height);
+  for (let p = 0, i = 0; p < dark.length; p++, i += 4) dark[p] = rgba[i] * 77 + rgba[i + 1] * 150 + rgba[i + 2] * 29 >> 8 < 128 ? 1 : 0;
+  const covered = (x, y) => items.some((b) => x >= b.x - 2 && x < b.x + b.w + 2 && y >= b.y - 2 && y < b.y + b.h + 2);
+  const seen = new Uint8Array(width * height);
+  const squares = [], bars = [];
+  const stack = [];
+  for (let start = 0; start < dark.length; start++) {
+    if (!dark[start] || seen[start]) continue;
+    const c = { x0: width, y0: height, x1: -1, y1: -1, n: 0, px: [] };
+    stack.push(start);
+    seen[start] = 1;
+    while (stack.length) {
+      const p = stack.pop(), x = p % width, y = (p - x) / width;
+      c.n++;
+      if (c.px.length < 4e3) c.px.push(p);
+      if (x < c.x0) c.x0 = x;
+      if (x > c.x1) c.x1 = x;
+      if (y < c.y0) c.y0 = y;
+      if (y > c.y1) c.y1 = y;
+      for (const q of [p - 1, p + 1, p - width, p + width]) if (q >= 0 && q < dark.length && !seen[q] && dark[q] && Math.abs(q % width - x) <= 1) {
+        seen[q] = 1;
+        stack.push(q);
+      }
+      if (c.n > 4e3) break;
+    }
+    while (stack.length) {
+      const p = stack.pop();
+      seen[p] = 1;
+    }
+    const w = c.x1 - c.x0 + 1, h = c.y1 - c.y0 + 1;
+    if (c.n > 4e3 || covered(c.x0 + c.x1 >> 1, c.y0 + c.y1 >> 1)) continue;
+    if (h >= 2 && h <= 6 && w >= 4 && w <= 30 && c.n >= w * h * 0.7) {
+      bars.push(c);
+      continue;
+    }
+    if (w < 12 || h < 12 || w > 60 || h > 60 || w > h * 1.33 || h > w * 1.33) continue;
+    const at = (x, y) => dark[y * width + x] === 1;
+    const row = (y) => {
+      let k = 0;
+      for (let x = c.x0; x <= c.x1; x++) if (at(x, y)) k++;
+      return k / w;
+    };
+    const col = (x) => {
+      let k = 0;
+      for (let y = c.y0; y <= c.y1; y++) if (at(x, y)) k++;
+      return k / h;
+    };
+    const edge = (vals) => Math.max(...vals) >= 0.8;
+    const top = edge([row(c.y0), row(c.y0 + 1)]), bottom = edge([row(c.y1), row(c.y1 - 1)]), left = edge([col(c.x0), col(c.x0 + 1)]), right = edge([col(c.x1), col(c.x1 - 1)]);
+    let inner = 0;
+    for (let y = c.y0 + Math.ceil(h * 0.25); y <= c.y1 - Math.ceil(h * 0.25); y++) for (let x = c.x0 + Math.ceil(w * 0.25); x <= c.x1 - Math.ceil(w * 0.25); x++) if (at(x, y)) inner++;
+    if (top && bottom && left && right && inner === 0) squares.push(c);
+  }
+  squares.sort((a, b) => a.y0 - b.y0 || a.x0 - b.x0);
+  const out = [];
+  const used = /* @__PURE__ */ new Set();
+  for (const s0 of squares) {
+    if (used.has(s0)) continue;
+    const run = [s0];
+    used.add(s0);
+    const side = s0.y1 - s0.y0 + 1, cy = (s0.y0 + s0.y1) / 2;
+    let text = "\u25A1";
+    for (; ; ) {
+      const last = run[run.length - 1];
+      const next = squares.find((q) => !used.has(q) && Math.abs((q.y0 + q.y1) / 2 - cy) <= side * 0.3 && q.x0 > last.x1 && q.x0 - last.x1 <= side * 2);
+      if (!next) break;
+      const gap = next.x0 - last.x1;
+      const bar = bars.find((b) => b.x0 > last.x1 && b.x1 < next.x0 && Math.abs((b.y0 + b.y1) / 2 - cy) <= side * 0.3);
+      let stroke = 0;
+      if (!bar) for (let y = Math.round(cy - side * 0.15); y <= Math.round(cy + side * 0.15); y++) {
+        let k = 0;
+        for (let x = last.x1 + 1; x < next.x0; x++) {
+          k = dark[y * width + x] ? k + 1 : 0;
+          if (k > stroke) stroke = k;
+        }
+      }
+      if (bar || stroke >= Math.max(4, side * 0.25)) text += "-";
+      else if (gap > side) break;
+      text += "\u25A1";
+      used.add(next);
+      run.push(next);
+    }
+    const x0 = run[0].x0, x1 = run[run.length - 1].x1;
+    out.push({ text, x: x0, y: s0.y0, w: x1 - x0 + 1, h: side, confidence: 0.9 });
+  }
+  return out;
+}
 function footerInkBox(rgba, width, height, boxes) {
   const y0 = Math.floor(height * 0.88), x0 = Math.floor(width * 0.3), x1 = Math.ceil(width * 0.7);
   let minX = Infinity, maxX = -1, minY = Infinity, maxY = -1;
@@ -936,11 +1124,28 @@ function closingParens(rgba, width, items) {
     if ((t.match(/\(/g)?.length ?? 0) <= (t.match(/\)/g)?.length ?? 0) || !/\([^()]*$/.test(t)) continue;
     const x0 = it.x + it.w + 2;
     const next = items.filter((o) => o !== it && o.x >= x0 && o.y < it.y + it.h && o.y + o.h > it.y).reduce((m, o) => Math.min(m, o.x), Infinity);
-    const x1 = Math.min(width, next, x0 + it.h * 8);
+    let x1 = Math.min(width, next, x0 + it.h * 16);
+    const pageH = Math.floor(rgba.length / 4 / width);
+    const colInk = (x) => {
+      let k = 0;
+      for (let y = it.y; y < it.y + it.h; y++) if (dark(x, y)) k++;
+      return k;
+    };
+    for (let x = x0; x < x1; x++) if (colInk(x) >= it.h * 0.9) {
+      x1 = x;
+      while (x1 > x0 && colInk(x1 - 1) >= it.h * 0.5) x1--;
+      break;
+    }
+    const ruleRow = /* @__PURE__ */ new Set();
+    for (let y = it.y; y < it.y + it.h; y++) {
+      let k = 0;
+      for (let x = x0; x < x1; x++) if (dark(x, y)) k++;
+      if (k > (x1 - x0) * 0.8) for (const d of [-1, 0, 1]) ruleRow.add(y + d);
+    }
     let minX = Infinity, maxX = -1, minY = Infinity, maxY = -1, cols = 0, prevOn = false, runs = 0;
     for (let x = x0; x < x1; x++) {
       let on = false;
-      for (let y = it.y; y < it.y + it.h; y++) if (dark(x, y)) {
+      for (let y = it.y; y < it.y + it.h; y++) if (!ruleRow.has(y) && dark(x, y)) {
         on = true;
         if (y < minY) minY = y;
         if (y > maxY) maxY = y;
@@ -954,9 +1159,14 @@ function closingParens(rgba, width, items) {
       prevOn = on;
     }
     if (runs !== 1 || maxX - minX + 1 > it.h * 0.5 || maxY - minY + 1 < it.h * 0.6) continue;
-    let rule = false;
-    for (let x = minX; x <= maxX && !rule; x++) rule = minY > 1 && dark(x, minY - 2) || dark(x, Math.min(maxY + 2, Math.floor(rgba.length / 4 / width) - 1));
-    if (rule) continue;
+    const inkRow = (y) => {
+      for (let x = minX; x <= maxX; x++) if (dark(x, y)) return true;
+      return false;
+    };
+    let over = 0;
+    for (let y = minY - 1; y >= 0 && !ruleRow.has(y) && inkRow(y) && over <= it.h; y--) over++;
+    for (let y = maxY + 1; y < pageH && !ruleRow.has(y) && inkRow(y) && over <= it.h; y++) over++;
+    if (over > it.h * 0.6) continue;
     const meanX = (ya, yb) => {
       let sx = 0, k = 0;
       for (let y = ya; y < yb; y++) for (let x = minX; x <= maxX; x++) if (dark(x, y)) {
@@ -1115,6 +1325,12 @@ var OcrEngine = class _OcrEngine {
       const cur = best.get(job.group);
       if (!cur || score(job, r.confidence) > score(cur.job, cur.confidence)) best.set(job.group, { job, ...r });
     });
+    jobs.forEach((job, i) => {
+      const r = results[i], cur = best.get(job.group);
+      if (!job.keep || !r || !cur || cur.job === job) return;
+      const b = cur.text.replace(/\s/g, "");
+      if (r.text.replace(/\s/g, "") === "\xB7" + b && /^[가-힣]/.test(b)) best.set(job.group, { ...cur, text: "\xB7" + cur.text, steps: [r.steps[0], ...cur.steps] });
+    });
     let items = [];
     for (const r of best.values()) {
       const { job, confidence, stepPx } = r;
@@ -1123,6 +1339,10 @@ var OcrEngine = class _OcrEngine {
       const raw = tuning.postprocess && job.rot === 0 ? restoreGlyphs(rgba, width, job.box, read, steps, stepPx, note) : read;
       let text = tuning.postprocess ? restoreSymbols(raw.trim(), note.ringLead) : raw;
       if (!text.trim()) continue;
+      if (tuning.postprocess && job.rot === 0 && text !== "/" && /^[(:;/)]+$/.test(text) && text.includes("/")) {
+        const g = grayCrop(rgba, width, job.box);
+        if (tallInkCount(g, job.box.w, job.box.h, inkStats(g)) === 1) text = "/";
+      }
       if (tuning.postprocess && job.rot === 0 && !/^[◎●▪□■▣❐❏○ㅇ◆◇◈]/.test(text)) {
         const chars = [...read], k = chars.findIndex((c) => c.trim());
         if (k >= 0 && steps.length === chars.length) {
@@ -1195,6 +1415,18 @@ var OcrEngine = class _OcrEngine {
     if (tuning.postprocess) items = dropNestedFragments(items);
     if (tuning.postprocess) for (const it of items) this.outerSquare(rgba, width, it, items);
     if (tuning.postprocess) items.push(...closingParens(rgba, width, items));
+    if (tuning.postprocess) items.push(...placeholderSquares(rgba, width, height, items));
+    if (tuning.postprocess) items.push(...rowHeadDashes(rgba, width, items));
+    if (tuning.postprocess) {
+      const tails = rowTailBoxes(rgba, width, items);
+      if (tails.length) {
+        const res = await this.recognizeJobs(rgba, width, tails.map((box, k) => ({ box, rot: 0, group: k })), tuning.recBatch);
+        tails.forEach((box, k) => {
+          const t = res[k]?.text.trim() ?? "";
+          if (/^\d{1,4}$/.test(t) && (res[k]?.confidence ?? 0) >= tuning.textScore) items.push({ text: t, ...this.itemBox(rgba, width, box, tuning), confidence: res[k].confidence });
+        });
+      }
+    }
     if (tuning.postprocess) restoreBulletItems(items);
     items.sort((a, b) => a.y - b.y || a.x - b.x);
     return items;
@@ -1864,4 +2096,4 @@ export {
   groupOcrLines,
   ocrItemsToBlocks
 };
-//# sourceMappingURL=chunk-MIJYCKJA.js.map
+//# sourceMappingURL=chunk-3TXAIEXL.js.map