kordoc 4.18.1 → 4.18.3

package/dist/{chunk-VDIXNEWV.js → chunk-CAN5P6N6.js}

@@ -1,5 +1,5 @@
 // src/utils.ts
-var VERSION = true ? "4.18.1" : "0.0.0-dev";
+var VERSION = true ? "4.18.3" : "0.0.0-dev";
 function toArrayBuffer(buf) {
   if (buf.byteOffset === 0 && buf.byteLength === buf.buffer.byteLength) {
     return buf.buffer;
@@ -75,8 +75,9 @@ function precheckZipSize(buffer, maxUncompressedSize = 256 * 1024 * 1024, maxEnt
       const extraLen = data.getUint16(pos + 30, true);
       const commentLen = data.getUint16(pos + 32, true);
       const isMedia = !!media && pos + 46 + nameLen <= len && media.re.test(new TextDecoder().decode(new Uint8Array(buffer, pos + 46, nameLen)));
-      if (isMedia) mediaUncompressed += size;
-      if (!(isMedia && media?.skip)) totalUncompressed += size;
+      const never = isMedia && !!media?.never && media.never.test(new TextDecoder().decode(new Uint8Array(buffer, pos + 46, nameLen)));
+      if (isMedia && !never) mediaUncompressed += size;
+      if (!(isMedia && (media?.skip || never))) totalUncompressed += size;
       pos += 46 + nameLen + extraLen + commentLen;
     }
     if (totalUncompressed > maxUncompressedSize) {
@@ -134,6 +135,9 @@ function classifyError(err) {
   if (msg.includes("\uC2DC\uADF8\uB2C8\uCC98") || msg.includes("\uBCF5\uAD6C\uD560 \uC218 \uC5C6")) return "CORRUPTED";
   return "PARSE_ERROR";
 }
+function partExtension(path) {
+  return (/\.([A-Za-z0-9]{1,5})$/.exec(path.slice(path.lastIndexOf("/") + 1))?.[1] ?? "bin").toLowerCase();
+}
 
 export {
   VERSION,
@@ -152,6 +156,7 @@ export {
   safeMin,
   safeMax,
   OPTIONAL_DEP_INSTALL_HINT,
-  classifyError
+  classifyError,
+  partExtension
 };
-//# sourceMappingURL=chunk-VDIXNEWV.js.map
+//# sourceMappingURL=chunk-CAN5P6N6.js.map

package/dist/chunk-CAN5P6N6.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/utils.ts"],"sourcesContent":["/** kordoc 공용 유틸리티 */\n\nimport { format } from \"util\"\n\n/** 빌드 타임에 tsup define으로 주입되는 버전 */\ndeclare const __KORDOC_VERSION__: string\nexport const VERSION: string = typeof __KORDOC_VERSION__ !== \"undefined\" ? __KORDOC_VERSION__ : \"0.0.0-dev\"\n\n/**\n * Node.js Buffer → ArrayBuffer 변환\n * pool Buffer의 공유 ArrayBuffer 문제를 안전하게 처리.\n * offset=0이고 전체 ArrayBuffer를 차지하면 복사 없이 직접 반환.\n */\nexport function toArrayBuffer(buf: Buffer): ArrayBuffer {\n  if (buf.byteOffset === 0 && buf.byteLength === buf.buffer.byteLength) {\n    return buf.buffer as ArrayBuffer\n  }\n  return buf.buffer.slice(buf.byteOffset, buf.byteOffset + buf.byteLength) as ArrayBuffer\n}\n\n/**\n * kordoc 내부 에러 클래스 — 사용자에게 노출해도 안전한 메시지만 포함.\n * MCP 에러 정제에서 instanceof로 판별하여 allowlist 패턴 매칭 없이 안전하게 통과.\n */\nexport class KordocError extends Error {\n  constructor(message: string) {\n    super(message)\n    this.name = \"KordocError\"\n  }\n}\n\n/**\n * ENOENT 판별 — 입력 경로가 존재하지 않음.\n * 래핑된 KordocError 를 뚫고 올라온 경우도 있으므로 메시지까지 함께 본다.\n */\nfunction isFileNotFound(err: unknown): boolean {\n  if (typeof err !== \"object\" || err === null) return false\n  if ((err as NodeJS.ErrnoException).code === \"ENOENT\") return true\n  const msg = (err as Error).message\n  return typeof msg === \"string\" && /ENOENT|no such file or directory/i.test(msg)\n}\n\n/**\n * 에러 메시지 정제 — KordocError는 그대로, ENOENT 만 구분 문구, 나머지는 일반 메시지로 대체.\n * 파일시스템 경로, 스택 트레이스 등 내부 정보 노출 방지.\n */\nexport function sanitizeError(err: unknown): string {\n  if (err instanceof KordocError) return err.message\n  // 없는 파일은 문서 문제가 아니라 고칠 수 있는 입력 문제다 — 사유를 숨기지 않는다.\n  // 경로는 노출하지 않는다(어느 파일인지는 stderr·실패 JSON 의 basename 이 알린다).\n  // 문구는 MCP describeError 의 ENOENT 힌트와 같은 표현을 쓴다.\n  if (isFileNotFound(err)) return \"파일 또는 디렉토리를 찾을 수 없습니다\"\n  return \"문서 처리 중 오류가 발생했습니다\"\n}\n\n/**\n * ZIP 엔트리 경로의 경로 순회 여부 판별.\n * 백슬래시 정규화, .., 절대경로, Windows 드라이브 문자 모두 차단.\n */\nexport function isPathTraversal(name: string): boolean {\n  if (name.includes(\"\\x00\")) return true\n  const normalized = name.replace(/\\\\/g, \"/\")\n  const segments = normalized.split(\"/\")\n  return segments.some(s => s === \"..\") || normalized.startsWith(\"/\") || /^[A-Za-z]:/.test(normalized)\n}\n\n// ─── HWPX 섹션 href 해석 (parser/roundtrip 공용) ────────────────────\n\n/**\n * manifest href를 본문 섹션 경로(`Contents/sectionN.xml`)로 정규화.\n * 비본문 XML(header/script/settings)·path traversal·드라이브 경로는 null로 거른다.\n * 백슬래시 구분자를 슬래시로 통일해 parser와 roundtrip이 동일한 목록을 만든다.\n */\nexport function normalizeSectionHref(href: string): string | null {\n  if (!href) return null\n  let normalized = href.replace(/\\\\/g, \"/\").replace(/^\\/+/, \"\")\n  if (isPathTraversal(normalized)) return null\n  if (/^[Ss]ection\\d+\\.xml$/.test(normalized)) normalized = \"Contents/\" + normalized\n  return /(?:^|\\/)[Ss]ection\\d+\\.xml$/.test(normalized) ? normalized : null\n}\n\n/** sectionN.xml을 N 숫자 순서로 정렬 (section10 > section2). */\nexport function compareSectionPaths(a: string, b: string): number {\n  const ai = Number(a.match(/[Ss]ection(\\d+)\\.xml$/)?.[1] ?? Number.MAX_SAFE_INTEGER)\n  const bi = Number(b.match(/[Ss]ection(\\d+)\\.xml$/)?.[1] ?? Number.MAX_SAFE_INTEGER)\n  return ai === bi ? a.localeCompare(b) : ai - bi\n}\n\n// ─── ZIP 안전 로딩 (ZIP bomb 방지) ────────────────────\n\n/** ZIP 비압축 상한 env 의 최대값 (MB) — 오타로 사실상 무제한이 되지 않게 */\nconst MAX_UNZIP_ENV_MB = 8192\n\n/**\n * ZIP 비압축 크기 상한 — `KORDOC_MAX_UNZIP_MB` 가 양수면 그 값(MB, 최대 8192), 아니면 포맷 기본값.\n * 빈 행이 수백만 개인 XLSX 처럼 내용은 작고 XML 만 큰 실문서를 폐쇄망 대량 적재에서 받으려는 설정 (#91)\n */\nexport function unzipLimitBytes(defaultBytes: number): number {\n  const mb = Number(process.env.KORDOC_MAX_UNZIP_MB)\n  return Number.isFinite(mb) && mb > 0 ? Math.min(mb, MAX_UNZIP_ENV_MB) * 1024 * 1024 : defaultBytes\n}\n\n/**\n * ZIP bomb 사전 검사 — Central Directory에서 비압축 합계와 엔트리 수 확인.\n * HWPX/XLSX/DOCX 등 모든 ZIP 기반 포맷에서 공통 사용.\n */\nexport function precheckZipSize(\n  buffer: ArrayBuffer,\n  maxUncompressedSize = 256 * 1024 * 1024, // parser-shared MAX_DECOMPRESS_SIZE 와 동기\n  maxEntries = 500,\n  /** 그림·개체 파트(엔트리 이름 re) — skip 이면 합계에서 뺀다(파서가 풀지 않는다), 아니면 한도 초과 메시지에 크기를 적는다 (#108).\n   *  never — 파서가 어느 때도 풀지 않는 파트(동영상·음성·OLE 임베드)는 늘 뺀다: 110MB 동영상 하나로 발표 자료가 통째 거부됐다 */\n  media?: { re: RegExp; skip: boolean; never?: RegExp },\n): { totalUncompressed: number; entryCount: number } {\n  try {\n    const data = new DataView(buffer)\n    const len = buffer.byteLength\n    // EOCD 시그니처 역방향 스캔\n    let eocdOffset = -1\n    for (let i = len - 22; i >= Math.max(0, len - 65557); i--) {\n      if (data.getUint32(i, true) === 0x06054b50) { eocdOffset = i; break }\n    }\n    if (eocdOffset < 0) return { totalUncompressed: 0, entryCount: 0 }\n\n    const entryCount = data.getUint16(eocdOffset + 10, true)\n    if (entryCount > maxEntries) {\n      throw new KordocError(`ZIP 엔트리 수 초과: ${entryCount} (최대 ${maxEntries})`)\n    }\n\n    const cdSize = data.getUint32(eocdOffset + 12, true)\n    const cdOffset = data.getUint32(eocdOffset + 16, true)\n    if (cdOffset + cdSize > len) return { totalUncompressed: 0, entryCount }\n\n    let totalUncompressed = 0\n    let mediaUncompressed = 0\n    let pos = cdOffset\n    for (let i = 0; i < entryCount && pos + 46 <= cdOffset + cdSize; i++) {\n      if (data.getUint32(pos, true) !== 0x02014b50) break\n      const size = data.getUint32(pos + 24, true)\n      const nameLen = data.getUint16(pos + 28, true)\n      const extraLen = data.getUint16(pos + 30, true)\n      const commentLen = data.getUint16(pos + 32, true)\n      // 이름이 버퍼 밖이면 그림 파트로 치지 않는다(합계에서 빠지지 않게)\n      const isMedia = !!media && pos + 46 + nameLen <= len &&\n        media.re.test(new TextDecoder().decode(new Uint8Array(buffer, pos + 46, nameLen)))\n      const never = isMedia && !!media?.never && media.never.test(new TextDecoder().decode(new Uint8Array(buffer, pos + 46, nameLen)))\n      if (isMedia && !never) mediaUncompressed += size\n      if (!(isMedia && (media?.skip || never))) totalUncompressed += size\n      pos += 46 + nameLen + extraLen + commentLen\n    }\n\n    if (totalUncompressed > maxUncompressedSize) {\n      const mb = (n: number) => (n / 1024 / 1024).toFixed(1)\n      const mediaNote = media && !media.skip && mediaUncompressed > 0\n        ? ` — 그림·개체 파트가 ${mb(mediaUncompressed)}MB, 이미지 추출을 끄면(images: false·--no-images) 세지 않는다` : \"\"\n      throw new KordocError(`ZIP 비압축 크기 초과: ${mb(totalUncompressed)}MB (최대 ${maxUncompressedSize / 1024 / 1024}MB)${mediaNote}`)\n    }\n\n    return { totalUncompressed, entryCount }\n  } catch (err) {\n    if (err instanceof KordocError) throw err\n    return { totalUncompressed: 0, entryCount: 0 }\n  }\n}\n\n/**\n * XML 파싱 전 프롤로그 정리 — 선두 BOM 제거 + DOCTYPE 제거(XXE/Billion Laughs 방지).\n *\n * BOM: 일부 OpenXML 라이터가 `[Content_Types].xml`·`.rels` 앞에 EF BB BF를 붙인다.\n * XML 스펙상 적법(엔티티는 BOM으로 시작할 수 있음)이고 엑셀·리브레오피스는 여는데,\n * xmldom은 \"선언이 문서 시작이 아님\" fatalError로 파일 전체를 거부했다 (#63).\n */\nexport function stripDtd(xml: string): string {\n  return xml.replace(/^\\uFEFF/, \"\").replace(/<!DOCTYPE\\s[^[>]*(\\[[\\s\\S]*?\\])?\\s*>/gi, \"\")\n}\n\n/** 하이퍼링크 URL 살균 — javascript: 등 XSS 위험 스킴 차단 */\nconst SAFE_HREF_RE = /^(?:https?:|mailto:|tel:|#)/i\nconst HREF_ESCAPE: Record<string, string> = { \"(\": \"%28\", \")\": \"%29\", \"<\": \"%3C\", \">\": \"%3E\", \"\\\"\": \"%22\", \"'\": \"%27\", \"`\": \"%60\" }\nexport function sanitizeHref(href: string): string | null {\n  const trimmed = href.trim()\n  if (!trimmed || !SAFE_HREF_RE.test(trimmed)) return null\n  // 괄호 percent-encoding — 마크다운 링크 목적지에서 불균형 ')'는 링크를 조기 종료시켜\n  // 문법이 깨진다 (원문 하이퍼링크에 ')'가 박힌 실문서 존재). %28/%29는 의미 동일.\n  // 공백·<>\"'` 도 — 스킴만 보고 통과시키면 \"mailto:a <img src=x onerror=…>\" 가 링크 목적지를 깨고 날 태그로\n  // 렌더됐다(pdfjs 는 mailto·tel URI 를 정규화하지 않는다, v4.14.4 리뷰 재현). 인코딩해도 같은 주소다\n  return trimmed.replace(/[()<>\"'`]/g, c => HREF_ESCAPE[c]).replace(/\\s/g, c => encodeURIComponent(c))\n}\n\n/**\n * HTML 엔티티 이스케이프 — 글은 `& < >`, 속성값(attr)은 `\"` 까지. HTML 표 경로(builder tableToHtml)와\n * 인쇄 렌더러가 같이 쓰는 한 벌. 읽는 쪽은 unescapeHtml 로 푼다\n */\nexport function escapeHtml(text: string, attr = false): string {\n  const s = text.replace(/&/g, \"&amp;\").replace(/</g, \"&lt;\").replace(/>/g, \"&gt;\")\n  return attr ? s.replace(/\"/g, \"&quot;\") : s\n}\n\n/** escapeHtml 의 역 (+ 손으로 쓴 HTML 표의 &#39;) — 한 번에 풀어 \"&amp;lt;\" 는 \"&lt;\" 로 남는다 */\nexport function unescapeHtml(text: string): string {\n  return text.replace(/&(lt|gt|quot|#39|amp);/g, (_, e: string) => HTML_ENTITY[e])\n}\nconst HTML_ENTITY: Record<string, string> = { lt: \"<\", gt: \">\", quot: \"\\\"\", \"#39\": \"'\", amp: \"&\" }\n\n/**\n * console.log·info·warn·debug 를 stderr 로 — stdout 이 기계 출력 채널(MCP JSON-RPC·CLI 마크다운/JSON·parse-worker\n * NDJSON)인 진입점용. 라이브러리(pdfjs 등)가 console 로 찍는 경고가 그 채널을 깨지 않게 한다\n */\nexport function routeConsoleToStderr(): void {\n  const toStderr = (...args: unknown[]): void => void process.stderr.write(format(...args) + \"\\n\")\n  console.log = toStderr\n  console.info = toStderr\n  console.warn = toStderr\n  console.debug = toStderr\n}\n\n// ─── 안전한 min/max (스택 오버플로 방지) ─────────────\n\n/** Math.min(...arr) 대체 — 대형 배열에서 스택 오버플로 방지 */\nexport function safeMin(arr: number[]): number {\n  let min = Infinity\n  for (let i = 0; i < arr.length; i++) if (arr[i] < min) min = arr[i]\n  return min\n}\n\n/** Math.max(...arr) 대체 — 대형 배열에서 스택 오버플로 방지 */\nexport function safeMax(arr: number[]): number {\n  let max = -Infinity\n  for (let i = 0; i < arr.length; i++) if (arr[i] > max) max = arr[i]\n  return max\n}\n\n// ─── 에러 분류 ──────────────────────────────────────\n\nimport type { ErrorCode } from \"./types.js\"\n\n/** 에러를 구조화된 ErrorCode로 분류 — KordocError 메시지 패턴 매칭 */\n/** 선택 의존 설치 실패 안내 꼬리 — 네트워크가 막힌 linux/x64 에서 onnxruntime-node 설치 스크립트가 CUDA 바이너리를 받다 실패하면\n *  npm 이 sharp·@huggingface/transformers 까지 선택 의존 설치에서 뺀다(#99). npx 캐시 실행은 npm install 로 못 고친다 */\nexport const OPTIONAL_DEP_INSTALL_HINT =\n  \" 네트워크가 제한된 linux/x64 에서 npx 로 설치했다면 onnxruntime-node 의 CUDA 다운로드 실패로 함께 빠졌을 수 있습니다 — \" +\n  \"`ONNXRUNTIME_NODE_INSTALL=skip npx -y kordoc@^4 …` 로 다시 설치하세요.\"\n\nexport function classifyError(err: unknown): ErrorCode {\n  if (!(err instanceof Error)) return \"PARSE_ERROR\"\n  const msg = err.message\n  // 입력 파일 없음 — 문서 내용 문제가 아니라 호출자의 경로 문제라 PARSE_ERROR 로 덮지 않는다.\n  if (isFileNotFound(err)) return \"FILE_NOT_FOUND\"\n  // DRM 을 암호화보다 먼저 — \"DRM 암호화된 HWPX…\" 류 메시지가 ENCRYPTED 로 떨어지지 않도록\n  if (msg.includes(\"DRM\")) return \"DRM_PROTECTED\"\n  if (msg.includes(\"암호화\") || msg.includes(\"암호로 보호\")) return \"ENCRYPTED\"\n  if (msg.includes(\"optional dependency\")) return \"MISSING_DEPENDENCY\"\n  // V8 문자열 한계 초과 — 이미지 base64 총량이 큰 문서의 JSON 직렬화 (#65)\n  if (msg.includes(\"Invalid string length\") || msg.includes(\"Cannot create a string longer\")) return \"OUTPUT_TOO_LARGE\"\n  if (msg.includes(\"ZIP bomb\") || msg.includes(\"ZIP 비압축 크기 초과\") || msg.includes(\"ZIP 엔트리 수 초과\")) return \"ZIP_BOMB\"\n  if (msg.includes(\"bomb\") || msg.includes(\"크기 초과\") || msg.includes(\"압축 해제\")) return \"DECOMPRESSION_BOMB\"\n  if (msg.includes(\"이미지 기반\")) return \"IMAGE_BASED_PDF\"\n  if (msg.includes(\"섹션\") && (msg.includes(\"찾을 수 없\") || msg.includes(\"없음\"))) return \"NO_SECTIONS\"\n  if (msg.includes(\"시그니처\") || msg.includes(\"복구할 수 없\")) return \"CORRUPTED\"\n  return \"PARSE_ERROR\"\n}\n\n/** ZIP 파트 경로의 확장자 — 마지막 경로 조각에서만, 없으면 \"bin\". \"ppt/media.v2/image1\" 의 \"v2/image1\" 이 그림 파일 이름에 들어가\n *  저장 경로가 깨졌다 */\nexport function partExtension(path: string): string {\n  return (/\\.([A-Za-z0-9]{1,5})$/.exec(path.slice(path.lastIndexOf(\"/\") + 1))?.[1] ?? \"bin\").toLowerCase()\n}\n"],"mappings":";AAMO,IAAM,UAAkB,OAA4C,WAAqB;AAOzF,SAAS,cAAc,KAA0B;AACtD,MAAI,IAAI,eAAe,KAAK,IAAI,eAAe,IAAI,OAAO,YAAY;AACpE,WAAO,IAAI;AAAA,EACb;AACA,SAAO,IAAI,OAAO,MAAM,IAAI,YAAY,IAAI,aAAa,IAAI,UAAU;AACzE;AAMO,IAAM,cAAN,cAA0B,MAAM;AAAA,EACrC,YAAY,SAAiB;AAC3B,UAAM,OAAO;AACb,SAAK,OAAO;AAAA,EACd;AACF;AAMA,SAAS,eAAe,KAAuB;AAC7C,MAAI,OAAO,QAAQ,YAAY,QAAQ,KAAM,QAAO;AACpD,MAAK,IAA8B,SAAS,SAAU,QAAO;AAC7D,QAAM,MAAO,IAAc;AAC3B,SAAO,OAAO,QAAQ,YAAY,oCAAoC,KAAK,GAAG;AAChF;AAMO,SAAS,cAAc,KAAsB;AAClD,MAAI,eAAe,YAAa,QAAO,IAAI;AAI3C,MAAI,eAAe,GAAG,EAAG,QAAO;AAChC,SAAO;AACT;AAMO,SAAS,gBAAgB,MAAuB;AACrD,MAAI,KAAK,SAAS,IAAM,EAAG,QAAO;AAClC,QAAM,aAAa,KAAK,QAAQ,OAAO,GAAG;AAC1C,QAAM,WAAW,WAAW,MAAM,GAAG;AACrC,SAAO,SAAS,KAAK,OAAK,MAAM,IAAI,KAAK,WAAW,WAAW,GAAG,KAAK,aAAa,KAAK,UAAU;AACrG;AASO,SAAS,qBAAqB,MAA6B;AAChE,MAAI,CAAC,KAAM,QAAO;AAClB,MAAI,aAAa,KAAK,QAAQ,OAAO,GAAG,EAAE,QAAQ,QAAQ,EAAE;AAC5D,MAAI,gBAAgB,UAAU,EAAG,QAAO;AACxC,MAAI,uBAAuB,KAAK,UAAU,EAAG,cAAa,cAAc;AACxE,SAAO,8BAA8B,KAAK,UAAU,IAAI,aAAa;AACvE;AAGO,SAAS,oBAAoB,GAAW,GAAmB;AAChE,QAAM,KAAK,OAAO,EAAE,MAAM,uBAAuB,IAAI,CAAC,KAAK,OAAO,gBAAgB;AAClF,QAAM,KAAK,OAAO,EAAE,MAAM,uBAAuB,IAAI,CAAC,KAAK,OAAO,gBAAgB;AAClF,SAAO,OAAO,KAAK,EAAE,cAAc,CAAC,IAAI,KAAK;AAC/C;AAKA,IAAM,mBAAmB;AAMlB,SAAS,gBAAgB,cAA8B;AAC5D,QAAM,KAAK,OAAO,QAAQ,IAAI,mBAAmB;AACjD,SAAO,OAAO,SAAS,EAAE,KAAK,KAAK,IAAI,KAAK,IAAI,IAAI,gBAAgB,IAAI,OAAO,OAAO;AACxF;AAMO,SAAS,gBACd,QACA,sBAAsB,MAAM,OAAO,MACnC,aAAa,KAGb,OACmD;AACnD,MAAI;AACF,UAAM,OAAO,IAAI,SAAS,MAAM;AAChC,UAAM,MAAM,OAAO;AAEnB,QAAI,aAAa;AACjB,aAAS,IAAI,MAAM,IAAI,KAAK,KAAK,IAAI,GAAG,MAAM,KAAK,GAAG,KAAK;AACzD,UAAI,KAAK,UAAU,GAAG,IAAI,MAAM,WAAY;AAAE,qBAAa;AAAG;AAAA,MAAM;AAAA,IACtE;AACA,QAAI,aAAa,EAAG,QAAO,EAAE,mBAAmB,GAAG,YAAY,EAAE;AAEjE,UAAM,aAAa,KAAK,UAAU,aAAa,IAAI,IAAI;AACvD,QAAI,aAAa,YAAY;AAC3B,YAAM,IAAI,YAAY,+CAAiB,UAAU,kBAAQ,UAAU,GAAG;AAAA,IACxE;AAEA,UAAM,SAAS,KAAK,UAAU,aAAa,IAAI,IAAI;AACnD,UAAM,WAAW,KAAK,UAAU,aAAa,IAAI,IAAI;AACrD,QAAI,WAAW,SAAS,IAAK,QAAO,EAAE,mBAAmB,GAAG,WAAW;AAEvE,QAAI,oBAAoB;AACxB,QAAI,oBAAoB;AACxB,QAAI,MAAM;AACV,aAAS,IAAI,GAAG,IAAI,cAAc,MAAM,MAAM,WAAW,QAAQ,KAAK;AACpE,UAAI,KAAK,UAAU,KAAK,IAAI,MAAM,SAAY;AAC9C,YAAM,OAAO,KAAK,UAAU,MAAM,IAAI,IAAI;AAC1C,YAAM,UAAU,KAAK,UAAU,MAAM,IAAI,IAAI;AAC7C,YAAM,WAAW,KAAK,UAAU,MAAM,IAAI,IAAI;AAC9C,YAAM,aAAa,KAAK,UAAU,MAAM,IAAI,IAAI;AAEhD,YAAM,UAAU,CAAC,CAAC,SAAS,MAAM,KAAK,WAAW,OAC/C,MAAM,GAAG,KAAK,IAAI,YAAY,EAAE,OAAO,IAAI,WAAW,QAAQ,MAAM,IAAI,OAAO,CAAC,CAAC;AACnF,YAAM,QAAQ,WAAW,CAAC,CAAC,OAAO,SAAS,MAAM,MAAM,KAAK,IAAI,YAAY,EAAE,OAAO,IAAI,WAAW,QAAQ,MAAM,IAAI,OAAO,CAAC,CAAC;AAC/H,UAAI,WAAW,CAAC,MAAO,sBAAqB;AAC5C,UAAI,EAAE,YAAY,OAAO,QAAQ,QAAS,sBAAqB;AAC/D,aAAO,KAAK,UAAU,WAAW;AAAA,IACnC;AAEA,QAAI,oBAAoB,qBAAqB;AAC3C,YAAM,KAAK,CAAC,OAAe,IAAI,OAAO,MAAM,QAAQ,CAAC;AACrD,YAAM,YAAY,SAAS,CAAC,MAAM,QAAQ,oBAAoB,IAC1D,2DAAgB,GAAG,iBAAiB,CAAC,yHAAqD;AAC9F,YAAM,IAAI,YAAY,qDAAkB,GAAG,iBAAiB,CAAC,oBAAU,sBAAsB,OAAO,IAAI,MAAM,SAAS,EAAE;AAAA,IAC3H;AAEA,WAAO,EAAE,mBAAmB,WAAW;AAAA,EACzC,SAAS,KAAK;AACZ,QAAI,eAAe,YAAa,OAAM;AACtC,WAAO,EAAE,mBAAmB,GAAG,YAAY,EAAE;AAAA,EAC/C;AACF;AASO,SAAS,SAAS,KAAqB;AAC5C,SAAO,IAAI,QAAQ,WAAW,EAAE,EAAE,QAAQ,0CAA0C,EAAE;AACxF;AAGA,IAAM,eAAe;AACrB,IAAM,cAAsC,EAAE,KAAK,OAAO,KAAK,OAAO,KAAK,OAAO,KAAK,OAAO,KAAM,OAAO,KAAK,OAAO,KAAK,MAAM;AAC3H,SAAS,aAAa,MAA6B;AACxD,QAAM,UAAU,KAAK,KAAK;AAC1B,MAAI,CAAC,WAAW,CAAC,aAAa,KAAK,OAAO,EAAG,QAAO;AAKpD,SAAO,QAAQ,QAAQ,cAAc,OAAK,YAAY,CAAC,CAAC,EAAE,QAAQ,OAAO,OAAK,mBAAmB,CAAC,CAAC;AACrG;AAMO,SAAS,WAAW,MAAc,OAAO,OAAe;AAC7D,QAAM,IAAI,KAAK,QAAQ,MAAM,OAAO,EAAE,QAAQ,MAAM,MAAM,EAAE,QAAQ,MAAM,MAAM;AAChF,SAAO,OAAO,EAAE,QAAQ,MAAM,QAAQ,IAAI;AAC5C;AAGO,SAAS,aAAa,MAAsB;AACjD,SAAO,KAAK,QAAQ,2BAA2B,CAAC,GAAG,MAAc,YAAY,CAAC,CAAC;AACjF;AACA,IAAM,cAAsC,EAAE,IAAI,KAAK,IAAI,KAAK,MAAM,KAAM,OAAO,KAAK,KAAK,IAAI;AAiB1F,SAAS,QAAQ,KAAuB;AAC7C,MAAI,MAAM;AACV,WAAS,IAAI,GAAG,IAAI,IAAI,QAAQ,IAAK,KAAI,IAAI,CAAC,IAAI,IAAK,OAAM,IAAI,CAAC;AAClE,SAAO;AACT;AAGO,SAAS,QAAQ,KAAuB;AAC7C,MAAI,MAAM;AACV,WAAS,IAAI,GAAG,IAAI,IAAI,QAAQ,IAAK,KAAI,IAAI,CAAC,IAAI,IAAK,OAAM,IAAI,CAAC;AAClE,SAAO;AACT;AASO,IAAM,4BACX;AAGK,SAAS,cAAc,KAAyB;AACrD,MAAI,EAAE,eAAe,OAAQ,QAAO;AACpC,QAAM,MAAM,IAAI;AAEhB,MAAI,eAAe,GAAG,EAAG,QAAO;AAEhC,MAAI,IAAI,SAAS,KAAK,EAAG,QAAO;AAChC,MAAI,IAAI,SAAS,oBAAK,KAAK,IAAI,SAAS,iCAAQ,EAAG,QAAO;AAC1D,MAAI,IAAI,SAAS,qBAAqB,EAAG,QAAO;AAEhD,MAAI,IAAI,SAAS,uBAAuB,KAAK,IAAI,SAAS,+BAA+B,EAAG,QAAO;AACnG,MAAI,IAAI,SAAS,UAAU,KAAK,IAAI,SAAS,kDAAe,KAAK,IAAI,SAAS,4CAAc,EAAG,QAAO;AACtG,MAAI,IAAI,SAAS,MAAM,KAAK,IAAI,SAAS,2BAAO,KAAK,IAAI,SAAS,2BAAO,EAAG,QAAO;AACnF,MAAI,IAAI,SAAS,iCAAQ,EAAG,QAAO;AACnC,MAAI,IAAI,SAAS,cAAI,MAAM,IAAI,SAAS,4BAAQ,KAAK,IAAI,SAAS,cAAI,GAAI,QAAO;AACjF,MAAI,IAAI,SAAS,0BAAM,KAAK,IAAI,SAAS,kCAAS,EAAG,QAAO;AAC5D,SAAO;AACT;AAIO,SAAS,cAAc,MAAsB;AAClD,UAAQ,wBAAwB,KAAK,KAAK,MAAM,KAAK,YAAY,GAAG,IAAI,CAAC,CAAC,IAAI,CAAC,KAAK,OAAO,YAAY;AACzG;","names":[]}

package/dist/chunk-GS7T56RP.cjs.map

@@ -1 +1 @@
-{"version":3,"sources":["/Users/mong-e/workspace/kordoc/dist/chunk-GS7T56RP.cjs","../node_modules/tsup/assets/cjs_shims.js"],"names":[],"mappings":"AAAA;ACKA,IAAM,iBAAA,EAAmB,CAAA,EAAA,GACvB,OAAO,SAAA,IAAa,YAAA,EAChB,IAAI,GAAA,CAAI,CAAA,KAAA,EAAQ,UAAU,CAAA,CAAA;AAK8B;ADT4B;AACA;AACA;AACA","file":"/Users/mong-e/workspace/kordoc/dist/chunk-GS7T56RP.cjs","sourcesContent":[null,"// Shim globals in cjs bundle\n// There's a weird bug that esbuild will always inject importMetaUrl\n// if we export it as `const importMetaUrl = ... __filename ...`\n// But using a function will not cause this issue\n\nconst getImportMetaUrl = () => \n  typeof document === \"undefined\" \n    ? new URL(`file:${__filename}`).href \n    : (document.currentScript && document.currentScript.tagName.toUpperCase() === 'SCRIPT') \n      ? document.currentScript.src \n      : new URL(\"main.js\", document.baseURI).href;\n\nexport const importMetaUrl = /* @__PURE__ */ getImportMetaUrl()\n"]}
+{"version":3,"sources":["/Users/chris_gomdori/workspace/kordoc/dist/chunk-GS7T56RP.cjs","../node_modules/tsup/assets/cjs_shims.js"],"names":[],"mappings":"AAAA;ACKA,IAAM,iBAAA,EAAmB,CAAA,EAAA,GACvB,OAAO,SAAA,IAAa,YAAA,EAChB,IAAI,GAAA,CAAI,CAAA,KAAA,EAAQ,UAAU,CAAA,CAAA;AAK8B;ADT4B;AACA;AACA;AACA","file":"/Users/chris_gomdori/workspace/kordoc/dist/chunk-GS7T56RP.cjs","sourcesContent":[null,"// Shim globals in cjs bundle\n// There's a weird bug that esbuild will always inject importMetaUrl\n// if we export it as `const importMetaUrl = ... __filename ...`\n// But using a function will not cause this issue\n\nconst getImportMetaUrl = () => \n  typeof document === \"undefined\" \n    ? new URL(`file:${__filename}`).href \n    : (document.currentScript && document.currentScript.tagName.toUpperCase() === 'SCRIPT') \n      ? document.currentScript.src \n      : new URL(\"main.js\", document.baseURI).href;\n\nexport const importMetaUrl = /* @__PURE__ */ getImportMetaUrl()\n"]}

package/dist/{chunk-SIZ63E5G.cjs → chunk-HZV52V2X.cjs}

@@ -1,6 +1,6 @@
 "use strict";Object.defineProperty(exports, "__esModule", {value: true}); function _optionalChain(ops) { let lastAccessLHS = undefined; let value = ops[0]; let i = 1; while (i < ops.length) { const op = ops[i]; const fn = ops[i + 1]; i += 2; if ((op === 'optionalAccess' || op === 'optionalCall') && value == null) { return undefined; } if (op === 'access' || op === 'optionalAccess') { lastAccessLHS = value; value = fn(value); } else if (op === 'call' || op === 'optionalCall') { value = fn((...args) => value.call(lastAccessLHS, ...args)); lastAccessLHS = undefined; } } return value; }
 
-var _chunkPCSOJ523cjs = require('./chunk-PCSOJ523.cjs');
+var _chunkSHRRD2TDcjs = require('./chunk-SHRRD2TD.cjs');
 
 // src/pdf/formula/models.ts
 var _crypto = require('crypto');
@@ -23,7 +23,7 @@ function isOfflineMode() {
 }
 function assertNetworkAllowed(what, hint) {
   if (!isOfflineMode()) return;
-  throw new (0, _chunkPCSOJ523cjs.KordocError)(
+  throw new (0, _chunkSHRRD2TDcjs.KordocError)(
     `\uD3D0\uC1C4\uB9DD \uBAA8\uB4DC(KORDOC_OFFLINE)\uC5D0\uC11C \uCC28\uB2E8\uB428: ${what}` + (hint ? ` \u2014 ${hint}` : "")
   );
 }
@@ -308,4 +308,4 @@ async function sha256OfFile(p) {
 
 
 exports.MFD_MODEL = MFD_MODEL; exports.MFR_ENCODER_MODEL = MFR_ENCODER_MODEL; exports.MFR_DECODER_MODEL = MFR_DECODER_MODEL; exports.MFR_TOKENIZER = MFR_TOKENIZER; exports.ALL_FORMULA_MODELS = ALL_FORMULA_MODELS; exports.getModelsDir = getModelsDir; exports.getFormulaModelsDir = getFormulaModelsDir; exports.getFormulaModelStatus = getFormulaModelStatus; exports.ensureModelsIn = ensureModelsIn; exports.ensureFormulaModels = ensureFormulaModels; exports.ensureSingleModel = ensureSingleModel;
-//# sourceMappingURL=chunk-SIZ63E5G.cjs.map
+//# sourceMappingURL=chunk-HZV52V2X.cjs.map

package/dist/chunk-HZV52V2X.cjs.map

@@ -0,0 +1 @@
+{"version":3,"sources":["/Users/chris_gomdori/workspace/kordoc/dist/chunk-HZV52V2X.cjs","../src/pdf/formula/models.ts","../src/shared/offline.ts","../src/pdf/formula/model-partials.ts"],"names":[],"mappings":"AAAA;AACE;AACF,wDAA6B;AAC7B;AACA;ACMA,gCAA2B;AAC3B,wBAAiC;AACjC,uCAA4C;AAC5C;AACA,wBAAwB;AACxB,4BAA8B;AAC9B,4CAAyB;AACzB,gCAAyB;ADJzB;AACA;AEKA,IAAM,OAAA,kBAAS,IAAI,GAAA,CAAI,CAAC,GAAA,EAAK,MAAA,EAAQ,KAAA,EAAO,IAAI,CAAC,CAAA;AAEjD,SAAS,OAAA,CAAQ,IAAA,EAAuB;AACtC,EAAA,MAAM,EAAA,EAAI,OAAA,CAAQ,GAAA,CAAI,IAAI,CAAA;AAC1B,EAAA,OAAO,CAAC,CAAC,EAAA,GAAK,MAAA,CAAO,GAAA,CAAI,CAAA,CAAE,IAAA,CAAK,CAAA,CAAE,WAAA,CAAY,CAAC,CAAA;AACjD;AAGO,SAAS,aAAA,CAAA,EAAyB;AACvC,EAAA,OAAO,OAAA,CAAQ,gBAAgB,CAAA;AACjC;AAOO,SAAS,oBAAA,CAAqB,IAAA,EAAc,IAAA,EAAqB;AACtE,EAAA,GAAA,CAAI,CAAC,aAAA,CAAc,CAAA,EAAG,MAAA;AACtB,EAAA,MAAM,IAAI,kCAAA;AAAA,IACR,CAAA,gFAAA,EAAiC,IAAI,CAAA,EAAA;AACvC,EAAA;AACF;AFZ4F;AACA;AG7BrD;AACI;AACX;AACP;AACe;AAEpC;AAC0B;AACV,EAAA;AACF,EAAA;AACkB,EAAA;AAC5B,IAAA;AAIwE,MAAA;AACb,MAAA;AACL,MAAA;AAClD,IAAA;AAGiB,MAAA;AACzB,IAAA;AACF,EAAA;AACwF,EAAA;AACjF,EAAA;AACT;AAEoE;AACO,EAAA;AAC3E;AAGqF;AACtD,EAAA;AAC+B,EAAA;AACxD,EAAA;AACA,EAAA;AAAsD,IAAA;AAAU,EAAA;AAAE,IAAA;AAAO,EAAA;AAChD,EAAA;AAG4B,IAAA;AACzC,IAAA;AACF,IAAA;AACe,IAAA;AACyB,IAAA;AAChD,IAAA;AACiB,MAAA;AACnB,MAAA;AACY,IAAA;AACyC,MAAA;AACvD,IAAA;AACI,IAAA;AAAoC,MAAA;AAAU,IAAA;AAAsC,IAAA;AAC1F,EAAA;AACF;AH2B4F;AACA;ACrDxD;AAC5B,EAAA;AACI,EAAA;AACL,EAAA;AACG,EAAA;AACA,EAAA;AACV;AAE4C;AACpC,EAAA;AACI,EAAA;AACL,EAAA;AACG,EAAA;AACA,EAAA;AACV;AAE4C;AACpC,EAAA;AACI,EAAA;AACL,EAAA;AACG,EAAA;AACA,EAAA;AACV;AAEwC;AAChC,EAAA;AACI,EAAA;AACL,EAAA;AACG,EAAA;AACA,EAAA;AACV;AAE4D;AAC1D,EAAA;AACA,EAAA;AACA,EAAA;AACA,EAAA;AACF;AAOqD;AACtB,EAAA;AACI,EAAA;AACH,IAAA;AAC9B,EAAA;AAC2D,EAAA;AAC7D;AAE8C;AACd,EAAA;AAChC;AAa6G;AAC5E,EAAA;AACL,EAAA;AACiB,IAAA;AAC5B,IAAA;AACT,IAAA;AAC4B,MAAA;AACE,MAAA;AAC1B,IAAA;AACG,MAAA;AACX,IAAA;AACa,IAAA;AACoD,MAAA;AAC/D,MAAA;AACF,IAAA;AACI,IAAA;AACyC,MAAA;AACf,MAAA;AACmC,QAAA;AACxD,MAAA;AACO,QAAA;AACV,UAAA;AACA,UAAA;AACQ,UAAA;AACE,UAAA;AAC4D,UAAA;AACvE,QAAA;AACH,MAAA;AACU,IAAA;AACE,MAAA;AACV,QAAA;AACA,QAAA;AACQ,QAAA;AACE,QAAA;AACgD,QAAA;AAC3D,MAAA;AACH,IAAA;AACF,EAAA;AACO,EAAA;AACT;AAGsE;AACH,EAAA;AACnE;AAoBgI;AAC1F,EAAA;AAEV,EAAA;AACiB,IAAA;AACG,IAAA;AAEO,IAAA;AACpC,sBAAA;AACX,QAAA;AACY,QAAA;AACL,QAAA;AACA,QAAA;AACE,QAAA;AACV,MAAA;AACD,MAAA;AACF,IAAA;AAIgD,IAAA;AAClD,EAAA;AACF;AAGuF;AACV,EAAA;AAC7E;AAGsG;AACpE,EAAA;AACI,EAAA;AACK,EAAA;AACG,EAAA;AACO,EAAA;AACe,oBAAA;AAChE,IAAA;AACF,EAAA;AACgD,EAAA;AAClD;AAE4F;AACtF,EAAA;AAC4B,IAAA;AACU,IAAA;AAClC,EAAA;AACC,IAAA;AACT,EAAA;AACI,EAAA;AACyC,IAAA;AACzB,IAAA;AACZ,EAAA;AACC,IAAA;AACT,EAAA;AACF;AAMiB;AAEf,EAAA;AACc,IAAA;AACZ,IAAA;AACF,EAAA;AAKmD,EAAA;AACA,EAAA;AAE/C,EAAA;AACiC,IAAA;AACxB,MAAA;AAAA;AAEO,QAAA;AAChB,MAAA;AACD,IAAA;AAC2B,IAAA;AAChB,MAAA;AAC0C,QAAA;AACpD,MAAA;AACF,IAAA;AAEmD,IAAA;AACQ,IAAA;AAC1C,IAAA;AAEoB,IAAA;AACjC,IAAA;AACyF,MAAA;AACzC,MAAA;AAC5B,QAAA;AACP,wBAAA;AACX,UAAA;AACA,UAAA;AACA,UAAA;AACO,UAAA;AACR,QAAA;AACF,MAAA;AACwB,MAAA;AACf,IAAA;AACqD,MAAA;AACjE,IAAA;AAEa,oBAAA;AACX,MAAA;AACA,MAAA;AACA,MAAA;AACO,MAAA;AACR,IAAA;AAGG,IAAA;AACA,IAAA;AACkC,MAAA;AAC1B,IAAA;AACuD,MAAA;AACnE,IAAA;AAE4B,IAAA;AAChB,MAAA;AAC4D,QAAA;AACtE,MAAA;AACF,IAAA;AAEgC,IAAA;AACnB,oBAAA;AACX,MAAA;AACA,MAAA;AACA,MAAA;AACO,MAAA;AACR,IAAA;AACD,EAAA;AAGI,IAAA;AAAuB,MAAA;AAAU,IAAA;AAAuC,IAAA;AAC9E,EAAA;AACF;AAEwD;AACzB,EAAA;AACI,EAAA;AACY,EAAA;AACZ,IAAA;AACf,MAAA;AAEhB,IAAA;AACD,EAAA;AACoB,EAAA;AACvB;ADlB4F;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA","file":"/Users/chris_gomdori/workspace/kordoc/dist/chunk-HZV52V2X.cjs","sourcesContent":[null,"/**\n * 수식 OCR 모델 자동 다운로드 + SHA-256 검증\n *\n * 캐시 위치: `~/.cache/kordoc/models/pix2text/`\n * 총 ~155MB (MFD 44MB + MFR encoder 87MB + MFR decoder 30MB + tokenizer 40KB)\n *\n * HF URL + SHA 는 Docufinder archive/rust-formula-ocr-wip 에서 검증 완료됨.\n * 모델 버전 갱신 시 URL + SHA 를 함께 갱신해야 한다.\n */\n\nimport { createHash } from \"crypto\"\nimport { createReadStream } from \"fs\"\nimport { mkdir, stat, unlink, rename } from \"fs/promises\"\nimport { createWriteStream } from \"fs\"\nimport { homedir } from \"os\"\nimport { join, dirname } from \"path\"\nimport { pipeline } from \"stream/promises\"\nimport { Readable } from \"stream\"\nimport { assertNetworkAllowed } from \"../../shared/offline.js\"\nimport { cleanupModelDownloadPartials, modelDownloadPartialPath } from \"./model-partials.js\"\n\nexport interface ModelSpec {\n  name: string\n  filename: string\n  url: string\n  sha256: string\n  sizeMb: number\n}\n\n/** 모델 스펙 (SHA-256 은 실제 다운로드로 검증됨 — 변경 금지) */\nexport const MFD_MODEL: ModelSpec = {\n  name: \"Pix2Text MFD\",\n  filename: \"mfd.onnx\",\n  url: \"https://huggingface.co/breezedeus/pix2text-mfd/resolve/main/mfd-v20240618.onnx\",\n  sha256: \"51a8854743b17ae654729af8db82a630c1ccfa06debf4856c8b28055f87d02c1\",\n  sizeMb: 42,\n}\n\nexport const MFR_ENCODER_MODEL: ModelSpec = {\n  name: \"Pix2Text MFR encoder\",\n  filename: \"encoder_model.onnx\",\n  url: \"https://huggingface.co/breezedeus/pix2text-mfr/resolve/main/encoder_model.onnx\",\n  sha256: \"bd8d5c322792e9ec45793af5569e9748f82a3d728a9e00213dbfc56c1486f37d\",\n  sizeMb: 87,\n}\n\nexport const MFR_DECODER_MODEL: ModelSpec = {\n  name: \"Pix2Text MFR decoder\",\n  filename: \"decoder_model.onnx\",\n  url: \"https://huggingface.co/breezedeus/pix2text-mfr/resolve/main/decoder_model.onnx\",\n  sha256: \"fd0f92d7a012f3dae41e1ac79421aea0ea888b5a66cb3f9a004e424f82f3daed\",\n  sizeMb: 30,\n}\n\nexport const MFR_TOKENIZER: ModelSpec = {\n  name: \"Pix2Text MFR tokenizer\",\n  filename: \"tokenizer.json\",\n  url: \"https://huggingface.co/breezedeus/pix2text-mfr/resolve/main/tokenizer.json\",\n  sha256: \"3e2ab757277d22639bec28c9d7972e352d3d1dba223051fa674002dc5ab64df3\",\n  sizeMb: 1,\n}\n\nexport const ALL_FORMULA_MODELS: ReadonlyArray<ModelSpec> = [\n  MFD_MODEL,\n  MFR_ENCODER_MODEL,\n  MFR_DECODER_MODEL,\n  MFR_TOKENIZER,\n]\n\n/**\n * 모델 캐시 하위 디렉토리 경로. 환경 변수 `KORDOC_MODEL_CACHE` 가 설정되면\n * 해당 경로를 우선 사용, 없으면 `~/.cache/kordoc/models/<subdir>/`.\n * (수식 OCR 외 텍스트 OCR 등 다른 모델군도 공용 — subdir 로 격리)\n */\nexport function getModelsDir(subdir: string): string {\n  const override = process.env.KORDOC_MODEL_CACHE\n  if (override && override.trim()) {\n    return join(override, subdir)\n  }\n  return join(homedir(), \".cache\", \"kordoc\", \"models\", subdir)\n}\n\nexport function getFormulaModelsDir(): string {\n  return getModelsDir(\"pix2text\")\n}\n\nexport interface ModelStatus {\n  spec: ModelSpec\n  localPath: string\n  exists: boolean\n  /** SHA 검증까지 끝났는가 (exists && valid) */\n  verified: boolean\n  /** 검증 실패 시 이유 */\n  invalidReason?: string\n}\n\n/** 지정 모델군의 존재/유효성 상태 반환 (다운로드 없이 확인만) */\nexport async function getModelStatusIn(dir: string, specs: ReadonlyArray<ModelSpec>): Promise<ModelStatus[]> {\n  const result: ModelStatus[] = []\n  for (const spec of specs) {\n    const localPath = join(dir, spec.filename)\n    let exists = false\n    try {\n      const s = await stat(localPath)\n      exists = s.isFile() && s.size > 0\n    } catch {\n      exists = false\n    }\n    if (!exists) {\n      result.push({ spec, localPath, exists: false, verified: false })\n      continue\n    }\n    try {\n      const actual = await sha256OfFile(localPath)\n      if (actual === spec.sha256) {\n        result.push({ spec, localPath, exists: true, verified: true })\n      } else {\n        result.push({\n          spec,\n          localPath,\n          exists: true,\n          verified: false,\n          invalidReason: `SHA256 mismatch: expected ${spec.sha256}, got ${actual}`,\n        })\n      }\n    } catch (e) {\n      result.push({\n        spec,\n        localPath,\n        exists: true,\n        verified: false,\n        invalidReason: `SHA compute failed: ${(e as Error).message}`,\n      })\n    }\n  }\n  return result\n}\n\n/** 모든 수식 모델의 존재/유효성 상태 반환 (다운로드 없이 확인만) */\nexport async function getFormulaModelStatus(): Promise<ModelStatus[]> {\n  return getModelStatusIn(getFormulaModelsDir(), ALL_FORMULA_MODELS)\n}\n\nexport interface DownloadProgress {\n  spec: ModelSpec\n  /** 현재까지 다운로드한 바이트 */\n  downloaded: number\n  /** 전체 바이트 (알 수 있으면) */\n  total: number | null\n  /** \"download\" | \"verify\" | \"done\" | \"skip\" */\n  phase: \"download\" | \"verify\" | \"done\" | \"skip\" | \"error\"\n  message?: string\n}\n\nexport type ProgressHandler = (p: DownloadProgress) => void\n\n/**\n * 지정 모델군을 다운로드/검증한다.\n * - 이미 있고 SHA 일치 → skip\n * - 없음 → 다운로드 후 SHA 검증. 실패 시 파일 삭제.\n */\nexport async function ensureModelsIn(dir: string, specs: ReadonlyArray<ModelSpec>, onProgress?: ProgressHandler): Promise<void> {\n  await mkdir(dir, { recursive: true })\n\n  for (const spec of specs) {\n    const localPath = join(dir, spec.filename)\n    await cleanupModelDownloadPartials(localPath)\n\n    if (await isExistingValid(localPath, spec.sha256)) {\n      onProgress?.({\n        spec,\n        downloaded: 0,\n        total: null,\n        phase: \"skip\",\n        message: \"이미 존재 + SHA 일치\",\n      })\n      continue\n    }\n\n    // Keep the destination until atomic replacement; another process may have\n    // finished a valid download since our initial check.\n    await downloadToFile(spec, localPath, onProgress)\n  }\n}\n\n/** 필요한 모든 수식 모델을 다운로드/검증한다. */\nexport async function ensureFormulaModels(onProgress?: ProgressHandler): Promise<void> {\n  return ensureModelsIn(getFormulaModelsDir(), ALL_FORMULA_MODELS, onProgress)\n}\n\n/** 단일 모델만 확인/다운로드 (진단 UI 에서 개별 상태 새로고침용) */\nexport async function ensureSingleModel(spec: ModelSpec, onProgress?: ProgressHandler): Promise<void> {\n  const dir = getFormulaModelsDir()\n  await mkdir(dir, { recursive: true })\n  const localPath = join(dir, spec.filename)\n  await cleanupModelDownloadPartials(localPath)\n  if (await isExistingValid(localPath, spec.sha256)) {\n    onProgress?.({ spec, downloaded: 0, total: null, phase: \"skip\" })\n    return\n  }\n  await downloadToFile(spec, localPath, onProgress)\n}\n\nasync function isExistingValid(localPath: string, sha256Expected: string): Promise<boolean> {\n  try {\n    const s = await stat(localPath)\n    if (!s.isFile() || s.size === 0) return false\n  } catch {\n    return false\n  }\n  try {\n    const actual = await sha256OfFile(localPath)\n    return actual === sha256Expected\n  } catch {\n    return false\n  }\n}\n\nasync function downloadToFile(\n  spec: ModelSpec,\n  localPath: string,\n  onProgress?: ProgressHandler,\n): Promise<void> {\n  // 폐쇄망 모드에서는 요청 자체를 보내지 않는다 (사이드로드 안내로 대체)\n  assertNetworkAllowed(\n    `${spec.name} 모델 다운로드`,\n    \"온라인 PC에서 `kordoc models --export <디렉토리>` 로 내보낸 뒤 이 PC에서 `kordoc models --import <디렉토리>` 하거나, KORDOC_MODEL_CACHE 로 모델 캐시 경로를 지정하세요\",\n  )\n\n  // 먼저 .part 로 받고 검증 후 rename — 중단된 다운로드가 \"정상 파일\"로 오인되는 걸 방지\n  // Batch workers may download the same missing model concurrently. Never share\n  // a partial file: only a fully verified download is atomically published.\n  const partPath = modelDownloadPartialPath(localPath)\n  await mkdir(dirname(localPath), { recursive: true })\n\n  try {\n    const resp = await fetch(spec.url, {\n      headers: {\n        // HF CDN 은 UA 없으면 가끔 403 을 뱉는다\n        \"User-Agent\": \"kordoc-formula-ocr/1.0 (+https://github.com/chrisryugj/kordoc)\",\n      },\n    })\n    if (!resp.ok || !resp.body) {\n      throw new Error(\n        `${spec.name} 다운로드 실패: HTTP ${resp.status} ${resp.statusText} (${spec.url})`,\n      )\n    }\n\n    const lenHeader = resp.headers.get(\"content-length\")\n    const total = lenHeader ? Number.parseInt(lenHeader, 10) : null\n    let downloaded = 0\n\n    const ws = createWriteStream(partPath)\n    try {\n      const reader = Readable.fromWeb(resp.body as unknown as import(\"stream/web\").ReadableStream)\n      reader.on(\"data\", (chunk: Buffer | Uint8Array) => {\n        downloaded += chunk.length\n        onProgress?.({\n          spec,\n          downloaded,\n          total,\n          phase: \"download\",\n        })\n      })\n      await pipeline(reader, ws)\n    } catch (e) {\n      throw new Error(`${spec.name} 스트리밍 실패: ${(e as Error).message}`)\n    }\n\n    onProgress?.({\n      spec,\n      downloaded,\n      total,\n      phase: \"verify\",\n    })\n\n    // SHA 검증\n    let actual: string\n    try {\n      actual = await sha256OfFile(partPath)\n    } catch (e) {\n      throw new Error(`${spec.name} SHA 계산 실패: ${(e as Error).message}`)\n    }\n\n    if (actual !== spec.sha256) {\n      throw new Error(\n        `${spec.name} SHA256 mismatch: expected ${spec.sha256}, got ${actual} — 모델 URL 이 오염되었거나 전송 중 손상되었습니다.`,\n      )\n    }\n\n    await rename(partPath, localPath)\n    onProgress?.({\n      spec,\n      downloaded,\n      total,\n      phase: \"done\",\n    })\n  } finally {\n    // Successful rename already removed this path. Process termination bypasses\n    // finally; the next ensure call reclaims it only after proving the owner dead.\n    try { await unlink(partPath) } catch { /* Preserve the original outcome. */ }\n  }\n}\n\nasync function sha256OfFile(p: string): Promise<string> {\n  const h = createHash(\"sha256\")\n  const stream = createReadStream(p)\n  await pipeline(stream, async function* (src) {\n    for await (const chunk of src) {\n      h.update(chunk)\n      // yield 안 함 (소비만)\n    }\n  })\n  return h.digest(\"hex\")\n}\n","/**\n * 폐쇄망(내부망) 배포용 아웃바운드 통신 킬스위치 + 파일 접근 루트 제한.\n *\n * kordoc 자체는 문서 파서라 상시 외부 통신이 없지만, 두 경로만 나간다:\n *   1. OCR/수식 모델 최초 1회 다운로드 (huggingface.co)\n *   2. `kordoc watch --webhook` 사용자가 명시한 URL\n *\n * `KORDOC_OFFLINE=1` 이면 위 경로를 **시도 전에** 차단한다. 보안성 검토에서\n * \"외부 통신 없음\"을 런타임으로 증명하기 위한 단일 관문 — 새 아웃바운드 경로를\n * 추가하면 반드시 `assertNetworkAllowed()` 를 먼저 통과시킬 것.\n *\n * `KORDOC_ROOT` 는 파일 접근을 특정 디렉토리 하위로 제한한다(미설정 시 제한 없음 —\n * 기존 동작 보존). 내부망 배포에서는 두 변수를 함께 고정하는 것을 권장.\n */\n\nimport { realpathSync } from \"fs\"\nimport { isAbsolute, relative, resolve, sep } from \"path\"\nimport { KordocError } from \"../utils.js\"\n\nconst TRUTHY = new Set([\"1\", \"true\", \"yes\", \"on\"])\n\nfunction envFlag(name: string): boolean {\n  const v = process.env[name]\n  return !!v && TRUTHY.has(v.trim().toLowerCase())\n}\n\n/** 폐쇄망 모드 여부 (`KORDOC_OFFLINE=1|true|yes|on`) */\nexport function isOfflineMode(): boolean {\n  return envFlag(\"KORDOC_OFFLINE\")\n}\n\n/**\n * 아웃바운드 통신 관문. 폐쇄망 모드면 요청을 보내기 전에 예외로 끊는다.\n * @param what 차단 대상 설명 (사용자에게 노출됨 — 호스트/URL 은 넣지 말 것)\n * @param hint 대안 안내 (오프라인 사이드로드 방법 등)\n */\nexport function assertNetworkAllowed(what: string, hint?: string): void {\n  if (!isOfflineMode()) return\n  throw new KordocError(\n    `폐쇄망 모드(KORDOC_OFFLINE)에서 차단됨: ${what}` + (hint ? ` — ${hint}` : \"\"),\n  )\n}\n\n/** 파일 접근 루트 (`KORDOC_ROOT`). 미설정이면 null = 제한 없음 */\nexport function getAccessRoot(): string | null {\n  const raw = process.env.KORDOC_ROOT\n  if (!raw || !raw.trim()) return null\n  const resolved = resolve(raw.trim())\n  try {\n    return realpathSync(resolved)\n  } catch {\n    // 루트가 실재하지 않으면 모든 접근이 막히는 게 안전한 기본값\n    return resolved\n  }\n}\n\n/** 경로가 루트 하위인지 (루트 자신 포함). 경계 문자열 매칭이 아닌 세그먼트 단위 비교 */\nexport function isWithinRoot(realPath: string, root: string): boolean {\n  if (realPath === root) return true\n  const rel = relative(root, realPath)\n  return rel !== \"\" && !rel.startsWith(\"..\") && !isAbsolute(rel) && !rel.startsWith(`..${sep}`)\n}\n\n/**\n * `KORDOC_ROOT` 가 설정된 경우 경로가 루트 하위인지 검증한다.\n * 인자는 이미 realpath 로 해석된 절대경로여야 한다 (심볼릭 링크 탈출 방지).\n */\nexport function assertWithinRoot(realPath: string): void {\n  const root = getAccessRoot()\n  if (!root) return\n  if (!isWithinRoot(realPath, root)) {\n    throw new KordocError(`허용된 작업 디렉토리(KORDOC_ROOT) 밖의 경로입니다: ${realPath}`)\n  }\n}\n","/** Process-owned temporary model downloads. Never infer abandonment from age alone. */\nimport { createHash, randomUUID } from \"node:crypto\"\nimport { readFileSync, readlinkSync } from \"node:fs\"\nimport { readdir, unlink } from \"node:fs/promises\"\nimport { hostname } from \"node:os\"\nimport { basename, dirname, join } from \"node:path\"\n\nlet scope: string | undefined\nfunction ownerScope(): string {\n  if (scope) return scope\n  let namespace = \"\"\n  if (process.platform === \"linux\") {\n    try {\n      // A cache may be shared by containers whose PID spaces differ. A missing\n      // PID in this namespace says nothing about a download in another one.\n      // Boot identity also separates machines with matching hostnames/ns inodes.\n      const boot = readFileSync(\"/proc/sys/kernel/random/boot_id\", \"utf8\").trim()\n      if (!boot) throw new Error(\"Kernel boot identity unavailable\")\n      namespace = `${boot}:${readlinkSync(\"/proc/self/ns/pid\")}`\n    } catch {\n      // Fail closed if namespace identity is unavailable: only this process can\n      // recognize its own paths; it will not scavenge another process's files.\n      namespace = randomUUID()\n    }\n  }\n  scope = createHash(\"sha256\").update(`${process.platform}\\0${hostname()}\\0${namespace}`).digest(\"hex\").slice(0, 32)\n  return scope\n}\n\nexport function modelDownloadPartialPath(localPath: string): string {\n  return `${localPath}.kordoc-${ownerScope()}-${process.pid}-${randomUUID()}.part`\n}\n\n/** Best effort: cache maintenance must not prevent use of a valid model. */\nexport async function cleanupModelDownloadPartials(localPath: string): Promise<void> {\n  const dir = dirname(localPath)\n  const prefix = `${basename(localPath)}.kordoc-${ownerScope()}-`\n  let entries\n  try { entries = await readdir(dir, { withFileTypes: true }) } catch { return }\n  for (const entry of entries) {\n    // Do not follow symlinks, touch other models, or guess ownership for old\n    // UUID-only partials/foreign hosts/namespaces. Their owner could still be live.\n    if (!entry.isFile() || !entry.name.startsWith(prefix)) continue\n    const match = /^([1-9]\\d*)-([0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12})\\.part$/.exec(entry.name.slice(prefix.length))\n    if (!match) continue\n    const pid = Number(match[1])\n    if (!Number.isSafeInteger(pid) || pid > 0x7fffffff) continue\n    try {\n      process.kill(pid, 0)\n      continue // Includes a reused PID: retaining an orphan beats deleting live data.\n    } catch (err) {\n      if ((err as NodeJS.ErrnoException).code !== \"ESRCH\") continue // EPERM/unknown = potentially live.\n    }\n    try { await unlink(join(dir, entry.name)) } catch { /* Another cleaner may have won. */ }\n  }\n}\n"]}

package/dist/{chunk-X625PCZM.js → chunk-IXX3WHIE.js}

@@ -6,10 +6,10 @@ import {
   extractPageBlocksWithLines,
   getOcrModelsDir,
   parseCharacterDict
-} from "./chunk-6V7QRQY3.js";
+} from "./chunk-BJ6RYYXV.js";
 import {
   OPTIONAL_DEP_INSTALL_HINT
-} from "./chunk-VDIXNEWV.js";
+} from "./chunk-CAN5P6N6.js";
 
 // src/ocr/ruling-lines.ts
 var INK_LUMA_MAX = 205;
@@ -2058,4 +2058,4 @@ export {
   runPdfOcr,
   ocrItemsToBlocks
 };
-//# sourceMappingURL=chunk-X625PCZM.js.map
+//# sourceMappingURL=chunk-IXX3WHIE.js.map

package/dist/{chunk-6V7QRQY3.js → chunk-J2A46VCA.js}

@@ -1,88 +1,15 @@
+#!/usr/bin/env node
 import {
   CELL_EDGES,
   HEADING_RATIO_H1,
   HEADING_RATIO_H2,
   HEADING_RATIO_H3,
   tidyScriptTags
-} from "./chunk-HUGBJVAA.js";
-import {
-  ensureModelsIn,
-  getModelsDir
-} from "./chunk-KXR4CTFT.js";
+} from "./chunk-UVUKT3IQ.js";
 import {
   safeMax,
   safeMin
-} from "./chunk-VDIXNEWV.js";
-
-// src/ocr/models.ts
-import { join } from "path";
-import { stat } from "fs/promises";
-var OCR_DET_MODEL = {
-  name: "PP-OCRv5 mobile det",
-  filename: "det.onnx",
-  url: "https://huggingface.co/PaddlePaddle/PP-OCRv5_mobile_det_onnx/resolve/main/inference.onnx",
-  sha256: "a431985659dc921974177a95adcfbb90fd9e51989a5e04d70d0b75f597b6e61d",
-  sizeMb: 5
-};
-var OCR_REC_MODEL = {
-  name: "PP-OCRv5 korean rec",
-  filename: "rec_korean.onnx",
-  url: "https://huggingface.co/PaddlePaddle/korean_PP-OCRv5_mobile_rec_onnx/resolve/main/inference.onnx",
-  sha256: "92f0b7785e64fc9090106a241cf4c1eb97472824558272751b88a2a4476d3a08",
-  sizeMb: 13
-};
-var OCR_REC_DICT = {
-  name: "PP-OCRv5 korean dict",
-  filename: "rec_korean.yml",
-  url: "https://huggingface.co/PaddlePaddle/korean_PP-OCRv5_mobile_rec_onnx/resolve/main/inference.yml",
-  sha256: "f757fa1c40e99edcf27e9cce879b93eb2a51fa46f5ef39095689b8c37dd75998",
-  sizeMb: 1
-};
-var ALL_OCR_MODELS = [OCR_DET_MODEL, OCR_REC_MODEL, OCR_REC_DICT];
-function getOcrModelsDir() {
-  return getModelsDir("ppocr");
-}
-async function ensureOcrModels(onProgress) {
-  return ensureModelsIn(getOcrModelsDir(), ALL_OCR_MODELS, onProgress);
-}
-async function ocrModelsCached() {
-  const dir = getOcrModelsDir();
-  for (const spec of ALL_OCR_MODELS) {
-    try {
-      if (!(await stat(join(dir, spec.filename))).size) return false;
-    } catch {
-      return false;
-    }
-  }
-  return true;
-}
-function parseCharacterDict(yml) {
-  const lines = yml.split("\n");
-  const chars = [];
-  let inDict = false;
-  let dictIndent = -1;
-  for (const line of lines) {
-    if (!inDict) {
-      const m2 = /^(\s*)character_dict:\s*$/.exec(line);
-      if (m2) {
-        inDict = true;
-        dictIndent = m2[1].length;
-      }
-      continue;
-    }
-    const m = /^(\s*)- (.*)$/.exec(line);
-    if (m && m[1].length >= dictIndent) {
-      let v = m[2];
-      if (v.length >= 2 && (v.startsWith("'") && v.endsWith("'") || v.startsWith('"') && v.endsWith('"'))) {
-        v = v.slice(1, -1).replace(/''/g, "'");
-      }
-      chars.push(v);
-      continue;
-    }
-    if (line.trim() !== "") break;
-  }
-  return chars;
-}
+} from "./chunk-VGDTONTB.js";
 
 // src/pdf/line-extract.ts
 import { OPS } from "pdfjs-dist/legacy/build/pdf.mjs";
@@ -3477,7 +3404,9 @@ function findVerticalCutWithOutlierFilter(items, minGap) {
     const filtered = items.filter((i) => i.w >= narrowThreshold);
     if (filtered.length >= 2 && filtered.length < items.length && filtered.length >= items.length * 0.7) {
       const filteredCut = findVerticalCut(filtered);
-      if (filteredCut.gap > edgeCut.gap && filteredCut.gap >= minGap) {
+      const sameRow = (a, b) => Math.abs(a.y - b.y) <= Math.max(a.fontSize, b.fontSize) * 0.5;
+      const inLine = items.some((r) => r.w < narrowThreshold && r.x < filteredCut.position && r.x + r.w > filteredCut.position && filtered.some((o) => sameRow(o, r) && o.x + o.w <= r.x + 1) && filtered.some((o) => sameRow(o, r) && o.x >= r.x + r.w - 1));
+      if (filteredCut.gap > edgeCut.gap && filteredCut.gap >= minGap && !inLine) {
         return filteredCut;
       }
     }
@@ -4197,7 +4126,7 @@ function detectSpecialKoreanTables(blocks) {
 }
 function removeHeaderFooterBlocks(blocks, pageHeights, warnings, notes, tables = false) {
   const ZONE_RATIO = 0.12;
-  const MIN_REPEAT = 3;
+  const MIN_REPEAT2 = 3;
   const topEntries = [];
   const bottomEntries = [];
   for (let bi = 0; bi < blocks.length; bi++) {
@@ -4234,7 +4163,7 @@ function removeHeaderFooterBlocks(blocks, pageHeights, warnings, notes, tables =
       const pages = [...patternPages.get(p) ?? []];
       const span = pages.length ? Math.max(...pages) - Math.min(...pages) + 1 : 0;
       const pageNumbered = !tables && (patternNumbers.get(p)?.size ?? 0) > 1;
-      if (count >= MIN_REPEAT && pages.length >= MIN_REPEAT && (pages.length >= span * 0.4 || pageNumbered)) {
+      if (count >= MIN_REPEAT2 && pages.length >= MIN_REPEAT2 && (pages.length >= span * 0.4 || pageNumbered)) {
         repeatedPatterns.add(p);
       }
     }
@@ -4382,6 +4311,78 @@ function mergeSliverColumns(grid, colXs, maxWidth) {
   return merged;
 }
 
+// src/pdf/side-tabs.ts
+var BAND = 0.12;
+var MAX_CHARS = 10;
+var MIN_REPEAT = 3;
+var RUN_GAP = 5;
+var TAB_CELL_CHARS = 24;
+var TAB_MIN_HEIGHT = 0.3;
+var SIDE_TAB_TABLES = /* @__PURE__ */ new WeakSet();
+function isSideTabTable(box, table, others, pageWidth, pageHeight) {
+  if (table.cols !== 1 || box.x2 - box.x1 > pageWidth * BAND || box.y2 - box.y1 < pageHeight * TAB_MIN_HEIGHT) return false;
+  const side = box.x1 >= pageWidth * (1 - BAND) ? "R" : box.x2 <= pageWidth * BAND ? "L" : null;
+  if (!side) return false;
+  const texts = table.cells.flat().map((c) => c.text.replace(/\s+/g, ""));
+  if (!texts.some(Boolean) || texts.some((t) => [...t].length > TAB_CELL_CHARS)) return false;
+  return others.every((it) => it.y < box.y1 || it.y > box.y2 || (side === "R" ? it.x + it.w <= box.x1 + 1 : it.x >= box.x2 - 1));
+}
+function removeSideTabs(blocks, pageWidths) {
+  const edges = /* @__PURE__ */ new Map();
+  for (const b of blocks) {
+    if (!b.bbox || b.pageNumber === void 0 || [...b.text?.trim() ?? ""].length <= MAX_CHARS) continue;
+    const e = edges.get(b.pageNumber) ?? { lefts: [], rights: [] };
+    e.lefts.push(b.bbox.x);
+    e.rights.push(b.bbox.x + b.bbox.width);
+    edges.set(b.pageNumber, e);
+  }
+  const q = (xs, p) => {
+    const v = [...xs].sort((a, b) => a - b);
+    return v[Math.min(v.length - 1, Math.floor(p * v.length))];
+  };
+  const body = /* @__PURE__ */ new Map();
+  for (const [page, e] of edges) body.set(page, { left: q(e.lefts, 0.2), right: q(e.rights, 0.8) });
+  const all = [...body.values()];
+  const fallback = all.length ? { left: q(all.map((e) => e.left), 0.5), right: q(all.map((e) => e.right), 0.5) } : void 0;
+  const key = (b) => {
+    const w = b.pageNumber !== void 0 ? pageWidths.get(b.pageNumber) : void 0;
+    const text = b.text?.trim();
+    const area = (b.pageNumber !== void 0 ? body.get(b.pageNumber) : void 0) ?? fallback;
+    if (b.type === "table") {
+      if (!w || !b.bbox || !b.table || !SIDE_TAB_TABLES.has(b.table)) return null;
+      return `
${b.bbox.x >= w * (1 - BAND) ? "R" : "L"}${Math.round(b.bbox.x / 8)}`;
+    }
+    if (!w || !area || !b.bbox || !text || [...text].length > MAX_CHARS) return null;
+    const side = b.bbox.x + b.bbox.width <= Math.min(w * BAND, area.left - 1) ? "L" : b.bbox.x >= Math.max(w * (1 - BAND), area.right + 1) ? "R" : null;
+    return side && text.replace(/\s+/g, "");
+  };
+  const pages = /* @__PURE__ */ new Map();
+  for (const b of blocks) {
+    const k = key(b);
+    if (k) pages.set(k, (pages.get(k) ?? /* @__PURE__ */ new Set()).add(b.pageNumber));
+  }
+  const running = /* @__PURE__ */ new Set();
+  for (const [k, ps] of pages) {
+    const list = [...ps].sort((a, b) => a - b);
+    const steps = list.slice(1).map((p, i) => p - list[i]).sort((a, b) => a - b);
+    const gap = Math.max(RUN_GAP, (steps[steps.length >> 1] ?? 1) * 3);
+    let run = [];
+    const flush = () => {
+      if (run.length >= MIN_REPEAT && run.length >= (run[run.length - 1] - run[0] + 1) * 0.4) for (const p of run) running.add(`${p}\0${k}`);
+      run = [];
+    };
+    for (const p of list) {
+      if (run.length && p - run[run.length - 1] > gap) flush();
+      run.push(p);
+    }
+    flush();
+  }
+  return running.size ? blocks.filter((b) => {
+    const k = key(b);
+    return !k || !running.has(`${b.pageNumber}\0${k}`);
+  }) : blocks;
+}
+
 // src/pdf/table-roles.ts
 var TOC_BLOCKS = /* @__PURE__ */ new WeakSet();
 var PAGE_LABEL = /^(?:\d{1,4}|[ivxlc]{1,7})$/i;
@@ -4488,6 +4489,8 @@ var MIN_GAP_FACTOR = 2;
 var MIN_GAP_ABSOLUTE = 20;
 var MIN_COL_FILL_RATIO = 0.4;
 function detectClusterTables(items, pageNum, rejected) {
+  const tab = sideTabGlyphs(items);
+  if (tab.size) items = items.filter((i) => !tab.has(i));
   if (items.length < MIN_ROWS * MIN_COLS) return [];
   const { merged, originMap } = mergeEvenSpacedClusters(items);
   const rows = mergeOverlappingRows(groupByBaseline(merged));
@@ -4547,6 +4550,36 @@ function detectClusterTables(items, pageNum, rejected) {
     return !prose;
   });
 }
+var SIDE_TAB_MIN = 6;
+function sideTabGlyphs(items) {
+  const found = /* @__PURE__ */ new Set();
+  const singles = items.filter((i) => i.fontSize > 0 && [...i.text.trim()].length === 1);
+  if (singles.length < SIDE_TAB_MIN) return found;
+  const cols = [];
+  for (const it of [...singles].sort((a, b) => a.x - b.x)) {
+    const c = cols.find((col) => Math.abs(col[0].x - it.x) <= Math.max(1.5, it.fontSize * 0.25));
+    if (c) c.push(it);
+    else cols.push([it]);
+  }
+  for (const col of cols) {
+    if (col.length < SIDE_TAB_MIN) continue;
+    if (col.filter((g) => /[\p{L}\p{N}]/u.test(g.text)).length < col.length * 0.8) continue;
+    col.sort((a, b) => b.y - a.y);
+    const fs = [...col.map((i) => i.fontSize)].sort((a, b) => a - b)[col.length >> 1];
+    if (col.slice(1).some((it, k) => col[k].y - it.y > fs * 3)) continue;
+    const members = new Set(col);
+    const top = col[0].y + fs, bottom = col[col.length - 1].y - fs;
+    const others = items.filter((i) => !members.has(i) && i.y <= top && i.y >= bottom);
+    const left = Math.min(...col.map((i) => i.x)), right = Math.max(...col.map((i) => i.x + i.w));
+    const rest = items.filter((i) => !members.has(i) && i.text.trim());
+    const outside = rest.every((i) => i.x >= right + fs) || rest.every((i) => i.x + i.w <= left - fs);
+    if (!outside) continue;
+    const alone = col.filter((g) => !others.some((i) => Math.abs(i.y - g.y) <= Y_TOL)).length;
+    if (alone < col.length * 0.4) continue;
+    for (const g of col) found.add(g);
+  }
+  return found;
+}
 function detectCompactStatisticalTables(rows, pageNum) {
   const found = [];
   const parts = /* @__PURE__ */ new Map();
@@ -6734,6 +6767,11 @@ function extractPageBlocksWithLines(items, pageNum, opList, pageWidth, pageHeigh
     if (carry) carry.clip = void 0;
     return [];
   }
+  const tab = sideTabGlyphs(items);
+  if (tab.size) {
+    const glyphs = [...tab].sort((a, b) => b.y - a.y).map((g) => ({ type: "paragraph", text: g.text.trim(), pageNumber: pageNum, bbox: computeBBox([g], pageNum), style: dominantStyle([g]) }));
+    return [...glyphs, ...extractPageBlocksWithLines(items.filter((i) => !tab.has(i)), pageNum, opList, pageWidth, pageHeight, extraLines, detectTables, carry, lexicon)];
+  }
   const lex = lexicon ?? new WrapLexicon();
   for (const line of streamLines(items)) lex.addLine(mergeLineSimple(line));
   const extracted = extractLines(opList.fnArray, opList.argsArray);
@@ -7095,6 +7133,12 @@ function extractBlocksWithGrids(items, pageNum, pageWidth, pageHeight, grids, ho
     const emptyPart = !hasContent && !!grid.cells && !grid.clipParent && !nestedAttached;
     if (!hasContent && !emptyPart) continue;
     if (emptyPart) EMPTY_PARTS.add(irTable);
+    const ownItems = numCols === 1 ? new Set(tableItems) : null;
+    if (ownItems && isSideTabTable(grid.bbox, irTable, items.filter((it) => !ownItems.has(it)), pageWidth, pageHeight)) {
+      SIDE_TAB_TABLES.add(irTable);
+      blocks.push({ type: "table", table: irTable, pageNumber: pageNum, bbox: { page: pageNum, x: grid.bbox.x1, y: grid.bbox.y1, width: gridW, height: grid.bbox.y2 - grid.bbox.y1 } });
+      continue;
+    }
     if (grid.clipParent) {
       const nb = { page: pageNum, x: grid.bbox.x1, y: grid.bbox.y1, width: grid.bbox.x2 - grid.bbox.x1, height: grid.bbox.y2 - grid.bbox.y1 };
       pendingNested.push({ parent: grid.clipParent, block: { type: "table", table: irTable, pageNumber: pageNum, bbox: nb } });
@@ -7564,13 +7608,6 @@ function extractPageBlocksFallback(items, pageNum, fullPage = false, detectTable
 }
 
 export {
-  OCR_DET_MODEL,
-  OCR_REC_MODEL,
-  OCR_REC_DICT,
-  getOcrModelsDir,
-  ensureOcrModels,
-  ocrModelsCached,
-  parseCharacterDict,
   extractLines,
   extractImageRegions,
   filterHiddenText,
@@ -7609,6 +7646,7 @@ export {
   trimTrailingEmptyTableCols,
   NO_EDGES,
   joinCellEdges,
+  removeSideTabs,
   extractPageBlocksWithLines
 };
-//# sourceMappingURL=chunk-6V7QRQY3.js.map
+//# sourceMappingURL=chunk-J2A46VCA.js.map