kordoc 4.14.2 → 4.14.3
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +38 -5
- package/dist/{-AQKXLFIM.js → -7XBPXCXN.js} +13 -11
- package/dist/{chunk-FOJZNMMK.js → chunk-2LQGSDZ6.js} +1 -1
- package/dist/chunk-2LQGSDZ6.js.map +1 -0
- package/dist/chunk-4R67INKT.js +1002 -0
- package/dist/chunk-4R67INKT.js.map +1 -0
- package/dist/{chunk-ILU76UF3.js → chunk-4RJ5D7MT.js} +2 -2
- package/dist/chunk-4RJ5D7MT.js.map +1 -0
- package/dist/{chunk-EQZI77PP.cjs → chunk-5VVS2QIJ.cjs} +3 -3
- package/dist/{chunk-EQZI77PP.cjs.map → chunk-5VVS2QIJ.cjs.map} +1 -1
- package/dist/{chunk-3DUADAY2.js → chunk-A275ZQLX.js} +209 -87
- package/dist/chunk-A275ZQLX.js.map +1 -0
- package/dist/{chunk-EISO3POT.js → chunk-CVEERPMP.js} +788 -194
- package/dist/chunk-CVEERPMP.js.map +1 -0
- package/dist/{chunk-ZWD734EY.js → chunk-D6KQKMU3.js} +2 -2
- package/dist/{chunk-QOMNEWCQ.cjs → chunk-EBZUZC6F.cjs} +305 -31
- package/dist/chunk-EBZUZC6F.cjs.map +1 -0
- package/dist/{chunk-HR2XNERB.cjs → chunk-EMGF6EEV.cjs} +2 -2
- package/dist/{chunk-HR2XNERB.cjs.map → chunk-EMGF6EEV.cjs.map} +1 -1
- package/dist/{chunk-GDW4YP7K.js → chunk-FAZR34CA.js} +2 -2
- package/dist/{chunk-TNCKOL2O.js → chunk-GTINVM5K.js} +2 -2
- package/dist/{chunk-L7DMSMAG.cjs → chunk-HX4UUF7Y.cjs} +1 -1
- package/dist/chunk-HX4UUF7Y.cjs.map +1 -0
- package/dist/{chunk-UET462HB.js → chunk-I4NNYWIR.js} +2 -2
- package/dist/chunk-IH56ZISC.js +620 -0
- package/dist/chunk-IH56ZISC.js.map +1 -0
- package/dist/{chunk-EDNJLUSE.js → chunk-IHZEFW4L.js} +2 -2
- package/dist/chunk-IW2CB755.cjs +1002 -0
- package/dist/chunk-IW2CB755.cjs.map +1 -0
- package/dist/{chunk-ZWOFQRIB.js → chunk-JCVIOLBF.js} +3 -3
- package/dist/{chunk-JQEFP45L.js → chunk-L5EH2NVO.js} +2 -2
- package/dist/{chunk-247ANM72.cjs → chunk-NOAQRCV6.cjs} +219 -97
- package/dist/chunk-NOAQRCV6.cjs.map +1 -0
- package/dist/{chunk-7IAQ4HOK.js → chunk-NQCUAGLR.js} +1 -1
- package/dist/chunk-NQCUAGLR.js.map +1 -0
- package/dist/{chunk-NS55SM7A.js → chunk-O2PGQNYE.js} +2 -2
- package/dist/{chunk-QC5T6D4J.js → chunk-QGW7NMTX.js} +299 -25
- package/dist/chunk-QGW7NMTX.js.map +1 -0
- package/dist/{chunk-A7QC2ZVP.js → chunk-UVVEBI22.js} +209 -87
- package/dist/chunk-UVVEBI22.js.map +1 -0
- package/dist/chunk-VJHAUMCF.js +949 -0
- package/dist/chunk-VJHAUMCF.js.map +1 -0
- package/dist/{chunk-C7FQBSON.js → chunk-VNFJPTU4.js} +2 -2
- package/dist/{chunk-VQ5W4MYO.js → chunk-VYOGXNQ3.js} +298 -24
- package/dist/chunk-VYOGXNQ3.js.map +1 -0
- package/dist/{chunk-DOGPOY4B.js → chunk-WB5LW52Z.js} +2 -2
- package/dist/{chunk-TSFS5VKD.js → chunk-YGBZABI4.js} +332 -141
- package/dist/chunk-YGBZABI4.js.map +1 -0
- package/dist/chunks-I7TEMJXH.js +10 -0
- package/dist/cli.js +228 -97
- package/dist/cli.js.map +1 -1
- package/dist/{formula-PJJ6S3Z2.cjs → formula-CPO2TZGB.cjs} +9 -9
- package/dist/{formula-PJJ6S3Z2.cjs.map → formula-CPO2TZGB.cjs.map} +1 -1
- package/dist/{formula-5S7WWXGH.js → formula-NOIDZEBA.js} +4 -4
- package/dist/{formula-JGFTXWYT.js → formula-X44GV7TN.js} +3 -3
- package/dist/{image-ocr-WCAYXLOA.js → image-ocr-72RAR4WR.js} +34 -16
- package/dist/image-ocr-72RAR4WR.js.map +1 -0
- package/dist/{image-ocr-FLIL7MGN.js → image-ocr-R6W6A75P.js} +32 -14
- package/dist/image-ocr-R6W6A75P.js.map +1 -0
- package/dist/{image-ocr-ZTYAGV3W.cjs → image-ocr-SUGT64ZM.cjs} +36 -18
- package/dist/image-ocr-SUGT64ZM.cjs.map +1 -0
- package/dist/index.cjs +2416 -1184
- package/dist/index.cjs.map +1 -1
- package/dist/index.d.cts +62 -25
- package/dist/index.d.ts +62 -25
- package/dist/index.js +1853 -621
- package/dist/index.js.map +1 -1
- package/dist/mcp.js +74 -53
- package/dist/mcp.js.map +1 -1
- package/dist/{model-bundle-FSPZZ2EM.js → model-bundle-57CLEBCO.js} +5 -5
- package/dist/{models-CJPAQK5R.js → models-7ELIFTD7.js} +5 -5
- package/dist/{parser-SCUTZAIJ.js → parser-4NKAYWAX.js} +261 -241
- package/dist/parser-4NKAYWAX.js.map +1 -0
- package/dist/{parser-A4TMPJAM.js → parser-HWFQRSUQ.js} +261 -241
- package/dist/parser-HWFQRSUQ.js.map +1 -0
- package/dist/{parser-VCXE4I5A.cjs → parser-RANGXGBD.cjs} +289 -269
- package/dist/parser-RANGXGBD.cjs.map +1 -0
- package/dist/pdf-ocr-BNC34MNU.js +13 -0
- package/dist/pdf-ocr-QX2GB33Q.js +16 -0
- package/dist/pdf-ocr-VHSFACH7.cjs +14 -0
- package/dist/{pdf-ocr-MBZZ57LD.cjs.map → pdf-ocr-VHSFACH7.cjs.map} +1 -1
- package/dist/{profile-io-LBM263K7.js → profile-io-V52PL5ZO.js} +3 -3
- package/dist/{rasterize-46RG4ZF6.js → rasterize-ZIGLJDOG.js} +3 -3
- package/dist/redact-TIQEQI3M.js +18 -0
- package/dist/redact-doc-KXZ3TZAG.js +1216 -0
- package/dist/redact-doc-KXZ3TZAG.js.map +1 -0
- package/dist/{render-NGP5T24S.js → render-U5PE6V2E.js} +7 -7
- package/dist/seal-DNXJBILR.js +10 -0
- package/dist/{setup-AU4OCHNH.js → setup-YRDM4ODP.js} +3 -3
- package/dist/visual-5XWMA4EV.js +19 -0
- package/dist/{watch-MKPUCPXM.js → watch-5IOLBATT.js} +12 -12
- package/package.json +2 -2
- package/dist/chunk-247ANM72.cjs.map +0 -1
- package/dist/chunk-3DUADAY2.js.map +0 -1
- package/dist/chunk-422427XJ.cjs +0 -571
- package/dist/chunk-422427XJ.cjs.map +0 -1
- package/dist/chunk-7IAQ4HOK.js.map +0 -1
- package/dist/chunk-A7QC2ZVP.js.map +0 -1
- package/dist/chunk-DZIXKL7E.js +0 -145
- package/dist/chunk-DZIXKL7E.js.map +0 -1
- package/dist/chunk-EISO3POT.js.map +0 -1
- package/dist/chunk-EP7KOI2V.js +0 -571
- package/dist/chunk-EP7KOI2V.js.map +0 -1
- package/dist/chunk-FOJZNMMK.js.map +0 -1
- package/dist/chunk-ILU76UF3.js.map +0 -1
- package/dist/chunk-KVWV4ULP.js +0 -516
- package/dist/chunk-KVWV4ULP.js.map +0 -1
- package/dist/chunk-L7DMSMAG.cjs.map +0 -1
- package/dist/chunk-QC5T6D4J.js.map +0 -1
- package/dist/chunk-QOMNEWCQ.cjs.map +0 -1
- package/dist/chunk-TSFS5VKD.js.map +0 -1
- package/dist/chunk-VQ5W4MYO.js.map +0 -1
- package/dist/chunks-SX77P5CV.js +0 -10
- package/dist/image-ocr-FLIL7MGN.js.map +0 -1
- package/dist/image-ocr-WCAYXLOA.js.map +0 -1
- package/dist/image-ocr-ZTYAGV3W.cjs.map +0 -1
- package/dist/parser-A4TMPJAM.js.map +0 -1
- package/dist/parser-SCUTZAIJ.js.map +0 -1
- package/dist/parser-VCXE4I5A.cjs.map +0 -1
- package/dist/pdf-ocr-76SAUXMZ.js +0 -13
- package/dist/pdf-ocr-CPU55CVF.js +0 -16
- package/dist/pdf-ocr-MBZZ57LD.cjs +0 -14
- package/dist/redact-7ILEAUTS.js +0 -12
- package/dist/seal-RGUPOLXD.js +0 -10
- package/dist/visual-NZLYIYEM.js +0 -19
- /package/dist/{-AQKXLFIM.js.map → -7XBPXCXN.js.map} +0 -0
- /package/dist/{chunk-ZWD734EY.js.map → chunk-D6KQKMU3.js.map} +0 -0
- /package/dist/{chunk-GDW4YP7K.js.map → chunk-FAZR34CA.js.map} +0 -0
- /package/dist/{chunk-TNCKOL2O.js.map → chunk-GTINVM5K.js.map} +0 -0
- /package/dist/{chunk-UET462HB.js.map → chunk-I4NNYWIR.js.map} +0 -0
- /package/dist/{chunk-EDNJLUSE.js.map → chunk-IHZEFW4L.js.map} +0 -0
- /package/dist/{chunk-ZWOFQRIB.js.map → chunk-JCVIOLBF.js.map} +0 -0
- /package/dist/{chunk-JQEFP45L.js.map → chunk-L5EH2NVO.js.map} +0 -0
- /package/dist/{chunk-NS55SM7A.js.map → chunk-O2PGQNYE.js.map} +0 -0
- /package/dist/{chunk-C7FQBSON.js.map → chunk-VNFJPTU4.js.map} +0 -0
- /package/dist/{chunk-DOGPOY4B.js.map → chunk-WB5LW52Z.js.map} +0 -0
- /package/dist/{chunks-SX77P5CV.js.map → chunks-I7TEMJXH.js.map} +0 -0
- /package/dist/{formula-5S7WWXGH.js.map → formula-NOIDZEBA.js.map} +0 -0
- /package/dist/{formula-JGFTXWYT.js.map → formula-X44GV7TN.js.map} +0 -0
- /package/dist/{model-bundle-FSPZZ2EM.js.map → model-bundle-57CLEBCO.js.map} +0 -0
- /package/dist/{models-CJPAQK5R.js.map → models-7ELIFTD7.js.map} +0 -0
- /package/dist/{pdf-ocr-76SAUXMZ.js.map → pdf-ocr-BNC34MNU.js.map} +0 -0
- /package/dist/{pdf-ocr-CPU55CVF.js.map → pdf-ocr-QX2GB33Q.js.map} +0 -0
- /package/dist/{profile-io-LBM263K7.js.map → profile-io-V52PL5ZO.js.map} +0 -0
- /package/dist/{rasterize-46RG4ZF6.js.map → rasterize-ZIGLJDOG.js.map} +0 -0
- /package/dist/{redact-7ILEAUTS.js.map → redact-TIQEQI3M.js.map} +0 -0
- /package/dist/{render-NGP5T24S.js.map → render-U5PE6V2E.js.map} +0 -0
- /package/dist/{seal-RGUPOLXD.js.map → seal-DNXJBILR.js.map} +0 -0
- /package/dist/{setup-AU4OCHNH.js.map → setup-YRDM4ODP.js.map} +0 -0
- /package/dist/{visual-NZLYIYEM.js.map → visual-5XWMA4EV.js.map} +0 -0
- /package/dist/{watch-MKPUCPXM.js.map → watch-5IOLBATT.js.map} +0 -0
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"sources":["../src/ocr/ruling-lines.ts","../src/ocr/engine.ts","../src/ocr/models.ts","../src/ocr/line-split.ts","../src/ocr/postprocess.ts","../src/ocr/crop.ts","../src/ocr/deskew.ts","../src/ocr/pdf-ocr.ts"],"sourcesContent":["/**\n * 래스터 괘선 감지 — 스캔/이미지 페이지의 픽셀에서 표 괘선(수평/수직 선)을\n * 찾아 선 기반 표 파이프라인(line-detector → table-grid)에 공급한다.\n *\n * PDF 그래픽 ops 가 없는 OCR 경로는 종전에 클러스터 감지기만 탔는데,\n * 병합 라벨 셀 + 다중줄 서술형 서식(정부 제출 서식류)은 라인 클러스터로\n * 행 경계를 잡을 수 없다. 괘선은 이미지에 실존하므로 이진화 + 런렝스로\n * 직접 감지한다 — ML 없음, 축 정렬 전제(스캔 스큐 보정은 스코프 밖).\n *\n * 오탐 방어 3겹 (실측: 우수사례 제출 서식 스캔 — 진짜 괘선은 길고 얇음):\n * - 최소 길이 20pt: 글리프 획(가로 ≤한 글자폭·세로 ≤글자높이)은 걸러짐\n * - 두께 상한 2.5pt: 굵은 제목 획(6~13px)·텍스트 줄 밴드·채운 색상바 탈락\n * - 양측 잉크 포위 제외: 색상바(붙임 박스 등) 안 흰 글자 틈새 슬리버 탈락\n */\n\nimport type { LineSegment } from \"../pdf/line-types.js\"\n\n/** 잉크 판정 휘도 상한 — 셀 배경 음영(회색 ~220)은 통과, 실선(진회~검정·남색)만 잉크 */\nconst INK_LUMA_MAX = 190\n/** 최소 선 길이 (pt) — 최소 셀 변(≈7mm)과 대형 글리프 획(≤1em) 사이 */\nconst MIN_LINE_LENGTH_PT = 20\n/** 선 두께 상한 (pt) — 진짜 괘선(0.12~0.7mm ≈ 0.3~2pt)과 굵은 글리프 획(2pt+) 분리 */\nconst MAX_LINE_WIDTH_PT = 2.5\n/** 런 내 허용 끊김 (px) — 스캔 노이즈/JPEG 아티팩트 */\nconst GAP_TOL_PX = 2\n/** 밴드 병합 시 요구하는 x(또는 y) 범위 겹침 비율 (짧은 쪽 기준) */\nconst BAND_OVERLAP_RATIO = 0.5\n/** 밴드 양측(±2px)이 이 비율 이상 잉크면 채운 영역 내부 슬리버로 판정 */\nconst SURROUND_INK_RATIO = 0.35\n\n/** 픽셀 공간(top-left origin) 선분 — 두께는 밴드 픽셀 수 */\nexport interface PxSegment {\n x1: number; y1: number\n x2: number; y2: number\n thicknessPx: number\n}\n\nexport interface RulingLines {\n horizontals: PxSegment[]\n verticals: PxSegment[]\n}\n\n/**\n * RGBA 래스터에서 수평/수직 괘선 감지.\n * @param scale 렌더 스케일 (px/pt) — 길이·두께 임계를 pt 기준으로 환산\n */\nexport function detectRulingLines(\n rgba: Uint8Array,\n width: number,\n height: number,\n scale: number,\n): RulingLines {\n const minLenPx = Math.round(MIN_LINE_LENGTH_PT * scale)\n const maxThickPx = Math.max(1, Math.floor(MAX_LINE_WIDTH_PT * scale))\n\n // 잉크 마스크 1회 구축 — 수평/수직 두 패스가 공유\n const ink = new Uint8Array(width * height)\n for (let p = 0, i = 0; p < ink.length; p++, i += 4) {\n // ITU-R BT.601 휘도 근사 (정수 시프트)\n const luma = (rgba[i] * 77 + rgba[i + 1] * 150 + rgba[i + 2] * 29) >> 8\n if (luma <= INK_LUMA_MAX && rgba[i + 3] >= 128) ink[p] = 1\n }\n\n const horizontals = detectBands(ink, width, height, minLenPx, maxThickPx, false)\n const verticals = detectBands(ink, width, height, minLenPx, maxThickPx, true)\n return { horizontals, verticals }\n}\n\n/** 열린 밴드 — 인접 스캔라인의 런을 누적 */\ninterface Band {\n lo: number // 런 진행축 시작 (h: x1, v: y1)\n hi: number // 런 진행축 끝\n first: number // 직교축 시작 스캔라인 (h: yTop, v: xLeft)\n last: number // 직교축 마지막 스캔라인\n}\n\n/**\n * 스캔라인 런렝스 → 인접 라인 밴드 병합 → 두께 필터 → 선분 방출.\n * transpose=false 면 수평선(행 스캔), true 면 수직선(열 스캔).\n */\nfunction detectBands(\n ink: Uint8Array,\n width: number,\n height: number,\n minLenPx: number,\n maxThickPx: number,\n transpose: boolean,\n): PxSegment[] {\n const lines = transpose ? width : height // 스캔라인 개수 (직교축)\n const span = transpose ? height : width // 런 진행축 길이\n const at = transpose\n ? (line: number, pos: number) => ink[pos * width + line]\n : (line: number, pos: number) => ink[line * width + pos]\n\n const out: PxSegment[] = []\n let open: Band[] = []\n\n /** 밴드 양측(±2px 스캔라인)의 잉크 비율 — 채운 색상바 안 흰 글자 틈새 슬리버 판별 */\n const surroundInkRatio = (b: Band, side: number): number => {\n if (side < 0 || side >= lines) return 0\n let dark = 0, total = 0\n for (let pos = b.lo; pos <= b.hi; pos += 3) {\n total++\n if (at(side, pos) === 1) dark++\n }\n return total > 0 ? dark / total : 0\n }\n\n const emit = (b: Band) => {\n const thick = b.last - b.first + 1\n if (thick > maxThickPx) return // 색상바·텍스트 밴드\n if (\n surroundInkRatio(b, b.first - 2) >= SURROUND_INK_RATIO &&\n surroundInkRatio(b, b.last + 2) >= SURROUND_INK_RATIO\n ) return // 채운 영역 내부 슬리버\n const center = (b.first + b.last) / 2\n out.push(\n transpose\n ? { x1: center, y1: b.lo, x2: center, y2: b.hi, thicknessPx: thick }\n : { x1: b.lo, y1: center, x2: b.hi, y2: center, thicknessPx: thick },\n )\n }\n\n for (let line = 0; line < lines; line++) {\n // 1) 이 스캔라인의 잉크 런 수집 (GAP_TOL_PX 이하 끊김 허용)\n const runs: Array<{ lo: number; hi: number }> = []\n let runStart = -1\n let gap = 0\n for (let pos = 0; pos <= span; pos++) {\n const on = pos < span && at(line, pos) === 1\n if (on) {\n if (runStart < 0) runStart = pos\n gap = 0\n } else if (runStart >= 0) {\n if (++gap > GAP_TOL_PX || pos >= span) {\n const hi = pos - gap\n if (hi - runStart + 1 >= minLenPx) runs.push({ lo: runStart, hi })\n runStart = -1\n gap = 0\n }\n }\n }\n\n // 2) 열린 밴드와 병합 — 직전 스캔라인까지 이어졌고 범위가 겹치면 확장\n const next: Band[] = []\n const used = new Set<number>()\n for (const band of open) {\n if (band.last !== line - 1) { emit(band); continue } // 연속 끊김 → 확정\n let merged = false\n for (let r = 0; r < runs.length; r++) {\n if (used.has(r)) continue\n const run = runs[r]\n const overlap = Math.min(band.hi, run.hi) - Math.max(band.lo, run.lo) + 1\n const shorter = Math.min(band.hi - band.lo, run.hi - run.lo) + 1\n if (overlap >= shorter * BAND_OVERLAP_RATIO) {\n band.lo = Math.min(band.lo, run.lo)\n band.hi = Math.max(band.hi, run.hi)\n band.last = line\n next.push(band)\n used.add(r)\n merged = true\n break\n }\n }\n if (!merged) emit(band)\n }\n for (let r = 0; r < runs.length; r++) {\n if (!used.has(r)) next.push({ lo: runs[r].lo, hi: runs[r].hi, first: line, last: line })\n }\n open = next\n }\n for (const band of open) emit(band)\n return out\n}\n\n/**\n * 픽셀 선분 → PDF pt LineSegment (bottom-up, classifyAndAdd 규약:\n * 수평선 y1==y2·x1<x2, 수직선 x1==x2·y1<y2).\n */\nexport function rulingToPdfLines(\n ruling: RulingLines,\n scale: number,\n pdfHeight: number,\n): { horizontals: LineSegment[]; verticals: LineSegment[] } {\n const horizontals: LineSegment[] = ruling.horizontals.map(s => {\n const y = pdfHeight - (s.y1 + s.y2) / 2 / scale\n return {\n x1: Math.min(s.x1, s.x2) / scale,\n y1: y,\n x2: Math.max(s.x1, s.x2) / scale,\n y2: y,\n lineWidth: s.thicknessPx / scale,\n }\n })\n const verticals: LineSegment[] = ruling.verticals.map(s => {\n const x = (s.x1 + s.x2) / 2 / scale\n const yA = pdfHeight - Math.max(s.y1, s.y2) / scale\n const yB = pdfHeight - Math.min(s.y1, s.y2) / scale\n return { x1: x, y1: yA, x2: x, y2: yB, lineWidth: s.thicknessPx / scale }\n })\n return { horizontals, verticals }\n}\n","/**\n * 내장 텍스트 OCR 엔진 — PP-OCRv5 korean (det DBNet + rec SVTR/CTC) ONNX 추론.\n *\n * 파이프라인: 페이지 RGBA → det(선 검출) → 박스 픽셀 분석(대비·행 밴드) → 라인 crop\n * → rec 배치(CTC 인식) → 표기 후처리 → OcrItem[]\n * (좌표는 입력 픽셀 기준 top-left origin — 호출자가 PDF 좌표계로 변환).\n *\n * 전·후처리는 공식 inference.yml 스펙 그대로:\n * - det: BGR, 긴 변 960 리사이즈(32 배수), mean/std [0.485,0.456,0.406]/[0.229,0.224,0.225],\n * DBPostProcess thresh 0.3 / box_thresh 0.6 / unclip_ratio 1.5\n * - rec: BGR, 높이 48 고정 비율 리사이즈 + 우측 zero-pad(최소 폭 320), (x/255-0.5)/0.5,\n * CTC 디코드 (blank=0, 1..N=사전, N+1=공백), text_score 0.5.\n * 배치(recBatch>1)는 공식처럼 폭 비율로 정렬해 묶고 배치 최대 폭까지 zero-pad — 기본은 1\n * DB 후처리의 contour+minAreaRect 는 축정렬 connected-component bbox 로 근사\n * (공문서 스캔은 수평 텍스트가 지배적).\n *\n * 공식 파이프라인 밖의 보강 (근거는 bench/ocr-accuracy.mjs 코퍼스 실측, 각 모듈 주석):\n * - 키 큰 박스(세로로 쌓인 글자)는 행 밴드로 갈라 밴드마다 인식 (line-split.ts)\n * - 잉크 대비가 낮은 박스(배경 도안)는 인식하지 않음 — 환각 방지\n * - 사전 밖 공문서 기호(○ △)·둥근 따옴표·천 단위 숫자 공백 복원, 점류 조각 폐기 (postprocess.ts)\n * - 결과 좌표는 det 박스(unclip 여백) 대신 박스 안 잉크 외곽 — 텍스트층 아이템과 같은 기하\n *\n * 의존성(onnxruntime-node, sharp)은 optional — 미설치 시 create()가 명확한 에러.\n * 모델 미다운로드 시에도 즉시 실패 — 호출자가 ensureOcrModels() 먼저.\n */\n\nimport type { InferenceSession } from \"onnxruntime-node\"\nimport { readFile } from \"fs/promises\"\nimport { join } from \"path\"\nimport {\n OCR_DET_MODEL,\n OCR_REC_MODEL,\n OCR_REC_DICT,\n getOcrModelsDir,\n parseCharacterDict,\n} from \"./models.js\"\nimport { grayCrop, inkBounds, inkStats, splitRowBands } from \"./line-split.js\"\nimport { isDotFragment, restoreBulletItems, restoreSymbols } from \"./postprocess.js\"\nimport { bandBoxes, lineCrop, type Box, REC_HEIGHT } from \"./crop.js\"\n\n/** OCR 인식 결과 한 줄 — 좌표는 입력 이미지 픽셀 (top-left origin, y down) */\nexport interface OcrItem {\n text: string\n x: number\n y: number\n w: number\n h: number\n /** CTC 평균 신뢰도 0~1 */\n confidence: number\n}\n\n/**\n * 엔진 튜닝 — 기본값이 제품 동작. 벤치·실험만 덮어쓴다 (공개 API 아님).\n */\nexport interface OcrTuning {\n /** det 입력 긴 변 (공식 resize_long). 1920 보존 실험은 recall Δ median +0.04pp에\n * 속도 -25%로 기각 (2026-08-02) */\n detLongSide: number\n detThresh: number\n detBoxThresh: number\n detUnclip: number\n /** 인식 신뢰도 하한 (공식 drop_score) */\n textScore: number\n /** rec 배치 크기. 공식 rec_batch_num 은 6 이지만 onnxruntime CPU 에선 배치(최대 폭까지\n * zero-pad)가 오히려 느리다 — 코퍼스 16쪽 교대 실측 1: 0.57 · 4: 0.93 · 8: 0.87 s/page,\n * 정확도 차는 박스 내 CER 0.08pp (2026-09-23). 기본 1 */\n recBatch: number\n /** 키 큰 박스 행 밴드 분할 */\n splitTall: boolean\n /** 이 대비(전경/배경 평균 휘도 차) 미만 박스는 글자가 아님 — 0 이면 끔 */\n minInkContrast: number\n /** 기호·따옴표 복원 + 점류 조각 폐기 */\n postprocess: boolean\n /** 결과 좌표를 det 박스(unclip 여백 포함) 대신 박스 안 잉크 외곽으로 */\n tightBoxes: boolean\n}\n\nexport const DEFAULT_OCR_TUNING: Readonly<OcrTuning> = Object.freeze({\n detLongSide: 960,\n detThresh: 0.3,\n detBoxThresh: 0.6,\n detUnclip: 1.5,\n textScore: 0.5,\n recBatch: 1,\n splitTall: true,\n minInkContrast: 35,\n postprocess: true,\n tightBoxes: true,\n})\n\nconst DET_MIN_SIZE = 3\nconst DET_MAX_BOXES = 1000\nconst REC_MIN_WIDTH = 320\n/** 배치 텐서 폭 합 상한 — 긴 줄 여러 개를 한 텐서로 묶어 메모리가 튀지 않게 */\nconst REC_BATCH_MAX_PIXELS = 48 * 16000\n/** 키 큰 박스 판정 (공식 파이프라인의 세로 판정 h/w ≥ 1.5 와 같은 문턱) */\nconst TALL_RATIO = 1.5\n/** 밴드로 갈라지지 않는 키 큰 박스 중 이 비율 이상은 90° 회전 글자 후보 */\nconst ROTATE_RATIO = 3\n\n// det: BGR 채널 순서에 yml 기재 순서 그대로 적용 (mean[0]→B)\nconst DET_MEAN = [0.485, 0.456, 0.406]\nconst DET_STD = [0.229, 0.224, 0.225]\n\ntype SharpFactory = (\n input: Uint8Array | Buffer,\n options?: { raw?: { width: number; height: number; channels: number } },\n) => SharpChain\ninterface SharpChain {\n resize(w: number, h: number, opts?: { fit?: string }): SharpChain\n removeAlpha(): SharpChain\n raw(): { toBuffer(): Promise<Buffer> }\n}\n\n/** 인식 대상 한 줄 — rot 는 crop 회전(90=반시계, 270=시계) */\ninterface LineJob { box: Box; rot: 0 | 90 | 270; group: number }\n\nexport class OcrEngine {\n private det: InferenceSession\n private rec: InferenceSession\n private dict: string[]\n private ort: typeof import(\"onnxruntime-node\")\n private sharp: SharpFactory\n\n private constructor(parts: {\n det: InferenceSession\n rec: InferenceSession\n dict: string[]\n ort: typeof import(\"onnxruntime-node\")\n sharp: SharpFactory\n }) {\n this.det = parts.det\n this.rec = parts.rec\n this.dict = parts.dict\n this.ort = parts.ort\n this.sharp = parts.sharp\n }\n\n static async create(): Promise<OcrEngine> {\n const [ortMod, sharpModRaw] = await Promise.all([\n tryImport<typeof import(\"onnxruntime-node\")>(\"onnxruntime-node\", () => import(\"onnxruntime-node\")),\n tryImport<{ default?: SharpFactory } & SharpFactory>(\n \"sharp\",\n () => import(\"sharp\") as unknown as Promise<{ default?: SharpFactory } & SharpFactory>,\n ),\n ])\n const sharpAny = sharpModRaw as { default?: SharpFactory } | SharpFactory\n const sharpMod: SharpFactory =\n typeof sharpAny === \"function\" ? sharpAny : (sharpAny.default ?? (sharpAny as unknown as SharpFactory))\n\n const dir = getOcrModelsDir()\n const sessionOpts: import(\"onnxruntime-node\").InferenceSession.SessionOptions = {\n graphOptimizationLevel: \"all\",\n executionProviders: [\"cpu\"],\n logSeverityLevel: 3, // paddle2onnx 변환 잔여물 W 로그 폭주 억제\n }\n const [det, rec, dictYml] = await Promise.all([\n ortMod.InferenceSession.create(join(dir, OCR_DET_MODEL.filename), sessionOpts),\n ortMod.InferenceSession.create(join(dir, OCR_REC_MODEL.filename), sessionOpts),\n readFile(join(dir, OCR_REC_DICT.filename), \"utf-8\"),\n ])\n const dict = parseCharacterDict(dictYml)\n if (dict.length === 0) throw new Error(\"OCR 사전 파싱 실패 — 모델 캐시를 삭제 후 재다운로드하세요\")\n\n return new OcrEngine({ det, rec, dict, ort: ortMod, sharp: sharpMod })\n }\n\n /** onnxruntime-node 1.14+ InferenceSession.release() — 구버전은 무시 */\n async destroy(): Promise<void> {\n for (const s of [this.det, this.rec]) {\n const rel = (s as unknown as { release?: () => Promise<void> }).release\n if (typeof rel === \"function\") {\n try { await rel.call(s) } catch { /* ignore */ }\n }\n }\n }\n\n /**\n * 페이지 RGBA 픽셀 → 텍스트 라인 인식.\n * 반환 좌표는 입력 픽셀 기준. 라인은 위→아래, 좌→우 정렬.\n * @param stats 저신뢰(conf<0.5) 폐기 라인 카운트 출력 — 종전엔 무음 폐기라 관측 불가\n */\n async recognizePage(\n rgba: Uint8Array,\n width: number,\n height: number,\n stats?: { droppedLowConf: number },\n tuning: Readonly<OcrTuning> = DEFAULT_OCR_TUNING,\n ): Promise<OcrItem[]> {\n if (width < DET_MIN_SIZE || height < DET_MIN_SIZE) return []\n const boxes = await this.detect(rgba, width, height, tuning)\n\n // 박스 픽셀 분석 → 인식 작업(라인) 목록. group = 한 결과로 합칠 후보 묶음(회전 후보)\n const jobs: LineJob[] = []\n let group = 0\n for (const b of boxes) {\n const gray = grayCrop(rgba, width, b)\n const ink = inkStats(gray)\n if (tuning.minInkContrast > 0 && ink.contrast < tuning.minInkContrast) continue\n if (tuning.splitTall && b.h >= b.w * TALL_RATIO) {\n const bands = splitRowBands(gray, b.w, b.h, ink, 0.45)\n if (bands.length >= 2) {\n for (const sub of bandBoxes(b, bands, height)) jobs.push({ box: sub, rot: 0, group: group++ })\n continue\n }\n if (b.h >= b.w * ROTATE_RATIO) {\n for (const rot of [0, 90, 270] as const) jobs.push({ box: b, rot, group })\n group++\n continue\n }\n }\n jobs.push({ box: b, rot: 0, group: group++ })\n }\n\n const results = await this.recognizeJobs(rgba, width, jobs, tuning.recBatch)\n\n // 회전 후보 그룹은 최고 신뢰도 하나만\n const best = new Map<number, { job: LineJob; text: string; confidence: number }>()\n jobs.forEach((job, i) => {\n const r = results[i]\n if (!r) return\n const cur = best.get(job.group)\n if (!cur || r.confidence > cur.confidence) best.set(job.group, { job, ...r })\n })\n\n const items: OcrItem[] = []\n for (const { job, text: raw, confidence } of best.values()) {\n const text = tuning.postprocess ? restoreSymbols(raw.trim()) : raw\n if (!text.trim()) continue\n if (tuning.postprocess && isDotFragment(text)) continue\n if (confidence < tuning.textScore) { if (stats) stats.droppedLowConf++; continue }\n let b = job.box\n if (tuning.tightBoxes) {\n const gray = grayCrop(rgba, width, b)\n const t = inkBounds(gray, b.w, b.h, inkStats(gray))\n b = { x: b.x + t.x0, y: b.y + t.y0, w: t.x1 - t.x0, h: t.y1 - t.y0 }\n }\n items.push({ text, x: b.x, y: b.y, w: b.w, h: b.h, confidence })\n }\n if (tuning.postprocess) restoreBulletItems(items)\n items.sort((a, b) => (a.y - b.y) || (a.x - b.x))\n return items\n }\n\n // ─── det ─────────────────────────────────────────────\n\n private async detect(\n rgba: Uint8Array,\n width: number,\n height: number,\n tuning: Readonly<OcrTuning>,\n ): Promise<Box[]> {\n const ratio = tuning.detLongSide / Math.max(width, height)\n const dw = Math.max(32, Math.round((width * ratio) / 32) * 32)\n const dh = Math.max(32, Math.round((height * ratio) / 32) * 32)\n\n const rgb = await this.sharp(rgba, { raw: { width, height, channels: 4 } })\n .resize(dw, dh, { fit: \"fill\" })\n .removeAlpha()\n .raw()\n .toBuffer()\n\n // HWC RGB → CHW BGR float32 정규화\n const plane = dw * dh\n const input = new Float32Array(3 * plane)\n for (let i = 0; i < plane; i++) {\n const r = rgb[i * 3] / 255\n const g = rgb[i * 3 + 1] / 255\n const b = rgb[i * 3 + 2] / 255\n input[i] = (b - DET_MEAN[0]) / DET_STD[0]\n input[plane + i] = (g - DET_MEAN[1]) / DET_STD[1]\n input[2 * plane + i] = (r - DET_MEAN[2]) / DET_STD[2]\n }\n\n const tensor = new this.ort.Tensor(\"float32\", input, [1, 3, dh, dw])\n const out = await this.det.run({ [this.det.inputNames[0]]: tensor })\n const probMap = out[this.det.outputNames[0]].data as Float32Array\n\n const rawBoxes = componentBoxes(probMap, dw, dh, tuning.detThresh, tuning.detBoxThresh)\n const sx = width / dw\n const sy = height / dh\n const boxes: Box[] = []\n for (const rb of rawBoxes.slice(0, DET_MAX_BOXES)) {\n // unclip: DB 는 학습 시 텍스트 영역을 수축시키므로 검출 박스를 되팽창\n const bw = rb.x2 - rb.x1 + 1\n const bh = rb.y2 - rb.y1 + 1\n const delta = (bw * bh * tuning.detUnclip) / (2 * (bw + bh))\n const x1 = Math.max(0, Math.floor((rb.x1 - delta) * sx))\n const y1 = Math.max(0, Math.floor((rb.y1 - delta) * sy))\n const x2 = Math.min(width, Math.ceil((rb.x2 + 1 + delta) * sx))\n const y2 = Math.min(height, Math.ceil((rb.y2 + 1 + delta) * sy))\n if (x2 - x1 < DET_MIN_SIZE || y2 - y1 < DET_MIN_SIZE) continue\n boxes.push({ x: x1, y: y1, w: x2 - x1, h: y2 - y1 })\n }\n return boxes\n }\n\n // ─── rec ─────────────────────────────────────────────\n\n /** 라인 작업들을 폭 비율 순으로 배치 인식 — 결과는 jobs 순서 */\n private async recognizeJobs(\n rgba: Uint8Array,\n pageW: number,\n jobs: LineJob[],\n batchSize: number,\n ): Promise<Array<{ text: string; confidence: number } | null>> {\n const crops = jobs.map(j => lineCrop(rgba, pageW, j.box, j.rot))\n const order = crops.map((_, i) => i).sort((a, b) => crops[a].w - crops[b].w)\n const results: Array<{ text: string; confidence: number } | null> = new Array(jobs.length).fill(null)\n const plane = REC_HEIGHT\n for (let s = 0; s < order.length;) {\n // 폭 오름차순이라 배치 마지막 원소가 최대 폭\n let e = s + 1\n while (e < order.length && e - s < Math.max(1, batchSize)\n && (e - s + 1) * Math.max(REC_MIN_WIDTH, crops[order[e]].w) * plane <= REC_BATCH_MAX_PIXELS) e++\n const idx = order.slice(s, e)\n const bw = Math.max(REC_MIN_WIDTH, crops[idx[idx.length - 1]].w)\n const n = idx.length\n const chw = 3 * REC_HEIGHT * bw\n const input = new Float32Array(n * chw) // pad 영역 0 (공식 zero-pad 와 동일)\n idx.forEach((ci, k) => {\n const c = crops[ci]\n const base = k * chw\n const p = REC_HEIGHT * bw\n for (let y = 0; y < REC_HEIGHT; y++) {\n for (let x = 0; x < c.w; x++) {\n const src = (y * c.w + x) * 3\n const dst = base + y * bw + x\n // HWC RGB → CHW BGR, (x/255-0.5)/0.5\n input[dst] = c.rgb[src + 2] / 127.5 - 1\n input[dst + p] = c.rgb[src + 1] / 127.5 - 1\n input[dst + 2 * p] = c.rgb[src] / 127.5 - 1\n }\n }\n })\n const tensor = new this.ort.Tensor(\"float32\", input, [n, 3, REC_HEIGHT, bw])\n const out = await this.rec.run({ [this.rec.inputNames[0]]: tensor })\n const logits = out[this.rec.outputNames[0]]\n const [, T, C] = logits.dims as number[]\n const data = logits.data as Float32Array\n idx.forEach((ci, k) => {\n results[ci] = ctcDecode(data.subarray(k * T * C, (k + 1) * T * C), T, C, this.dict)\n })\n s = e\n }\n return results\n }\n}\n\n/** CTC greedy 디코드 — 연속 중복 붕괴 → blank(0) 제거 → 사전 매핑 (테스트용 export) */\nexport function ctcDecode(\n data: Float32Array,\n T: number,\n C: number,\n dict: string[],\n): { text: string; confidence: number } | null {\n let text = \"\"\n let confSum = 0\n let confCount = 0\n let prev = -1\n for (let t = 0; t < T; t++) {\n const off = t * C\n let best = 0\n let bestV = data[off]\n for (let c = 1; c < C; c++) {\n const v = data[off + c]\n if (v > bestV) { bestV = v; best = c }\n }\n const repeat = best === prev\n prev = best\n if (best === 0 || repeat) continue\n // 모델 출력이 softmax 확률이 아니면 (>1) 해당 스텝만 정규화\n let p = bestV\n if (p > 1.0001 || p < 0) {\n let denom = 0\n for (let c = 0; c < C; c++) denom += Math.exp(data[off + c] - bestV)\n p = 1 / denom\n }\n confSum += p\n confCount++\n if (best >= 1 && best <= dict.length) text += dict[best - 1]\n else if (best === dict.length + 1) text += \" \"\n }\n if (!text) return null\n return { text, confidence: confCount > 0 ? confSum / confCount : 0 }\n}\n\n/** 이진화 확률맵의 4-연결 성분 bbox (score = 성분 평균 확률, 테스트용 export) */\nexport function componentBoxes(\n prob: Float32Array,\n w: number,\n h: number,\n thresh: number = DEFAULT_OCR_TUNING.detThresh,\n boxThresh: number = DEFAULT_OCR_TUNING.detBoxThresh,\n): Array<{ x1: number; y1: number; x2: number; y2: number }> {\n const visited = new Uint8Array(w * h)\n const boxes: Array<{ x1: number; y1: number; x2: number; y2: number; score: number }> = []\n const stack: number[] = []\n\n for (let start = 0; start < w * h; start++) {\n if (visited[start] || prob[start] <= thresh) continue\n let x1 = start % w, x2 = x1, y1 = (start / w) | 0, y2 = y1\n let sum = 0\n let count = 0\n stack.length = 0\n stack.push(start)\n visited[start] = 1\n while (stack.length) {\n const p = stack.pop()!\n const px = p % w\n const py = (p / w) | 0\n sum += prob[p]\n count++\n if (px < x1) x1 = px\n if (px > x2) x2 = px\n if (py < y1) y1 = py\n if (py > y2) y2 = py\n // 4-이웃\n if (px > 0 && !visited[p - 1] && prob[p - 1] > thresh) { visited[p - 1] = 1; stack.push(p - 1) }\n if (px < w - 1 && !visited[p + 1] && prob[p + 1] > thresh) { visited[p + 1] = 1; stack.push(p + 1) }\n if (py > 0 && !visited[p - w] && prob[p - w] > thresh) { visited[p - w] = 1; stack.push(p - w) }\n if (py < h - 1 && !visited[p + w] && prob[p + w] > thresh) { visited[p + w] = 1; stack.push(p + w) }\n }\n if (x2 - x1 + 1 < DET_MIN_SIZE && y2 - y1 + 1 < DET_MIN_SIZE) continue\n boxes.push({ x1, y1, x2, y2, score: sum / count })\n }\n\n return boxes\n .filter(b => b.score >= boxThresh)\n .sort((a, b) => (a.y1 - b.y1) || (a.x1 - b.x1))\n}\n\nasync function tryImport<T>(name: string, loader: () => Promise<T>): Promise<T> {\n try {\n return await loader()\n } catch (e) {\n throw new Error(\n `내장 OCR 을 사용하려면 optional dependency '${name}' 이 필요합니다. ` +\n `\\`npm install ${name}\\` 후 다시 실행하세요. 원인: ${(e as Error).message}`,\n )\n }\n}\n\n// ─── 엔진 싱글턴 (watch/서버 장기 실행에서 세션 재사용) ───\nlet enginePromise: Promise<OcrEngine> | null = null\n\nexport function getOcrEngine(): Promise<OcrEngine> {\n if (!enginePromise) {\n enginePromise = OcrEngine.create().catch(err => {\n enginePromise = null // 실패는 캐시하지 않음 — 모델 설치 후 재시도 가능\n throw err\n })\n }\n return enginePromise\n}\n","/**\n * 내장 텍스트 OCR 모델 (PP-OCRv5 korean) 스펙 + 다운로드/검증.\n *\n * 캐시 위치: `~/.cache/kordoc/models/ppocr/` — 총 ~18MB (det 4.6 + rec 12.8 + dict 0.1)\n * 다운로드 인프라는 수식 OCR 과 공용 (formula/models.ts ensureModelsIn).\n *\n * 모델 출처: PaddlePaddle 공식 HuggingFace org 의 ONNX 변환본 (Apache-2.0).\n * SHA-256 은 실제 다운로드 + HF API lfs.oid 대조로 검증됨 — 변경 금지.\n * 한국어 사전은 rec 리포의 inference.yml `PostProcess.character_dict` 에\n * 내장 (11,945자 — 완성형 한글 11,172 음절 전량 + 자모/라틴/기호).\n */\n\nimport { join } from \"path\"\nimport {\n type ModelSpec,\n type ModelStatus,\n type ProgressHandler,\n ensureModelsIn,\n getModelStatusIn,\n getModelsDir,\n} from \"../pdf/formula/models.js\"\n\nexport const OCR_DET_MODEL: ModelSpec = {\n name: \"PP-OCRv5 mobile det\",\n filename: \"det.onnx\",\n url: \"https://huggingface.co/PaddlePaddle/PP-OCRv5_mobile_det_onnx/resolve/main/inference.onnx\",\n sha256: \"a431985659dc921974177a95adcfbb90fd9e51989a5e04d70d0b75f597b6e61d\",\n sizeMb: 5,\n}\n\nexport const OCR_REC_MODEL: ModelSpec = {\n name: \"PP-OCRv5 korean rec\",\n filename: \"rec_korean.onnx\",\n url: \"https://huggingface.co/PaddlePaddle/korean_PP-OCRv5_mobile_rec_onnx/resolve/main/inference.onnx\",\n sha256: \"92f0b7785e64fc9090106a241cf4c1eb97472824558272751b88a2a4476d3a08\",\n sizeMb: 13,\n}\n\nexport const OCR_REC_DICT: ModelSpec = {\n name: \"PP-OCRv5 korean dict\",\n filename: \"rec_korean.yml\",\n url: \"https://huggingface.co/PaddlePaddle/korean_PP-OCRv5_mobile_rec_onnx/resolve/main/inference.yml\",\n sha256: \"f757fa1c40e99edcf27e9cce879b93eb2a51fa46f5ef39095689b8c37dd75998\",\n sizeMb: 1,\n}\n\nexport const ALL_OCR_MODELS: ReadonlyArray<ModelSpec> = [OCR_DET_MODEL, OCR_REC_MODEL, OCR_REC_DICT]\n\nexport function getOcrModelsDir(): string {\n return getModelsDir(\"ppocr\")\n}\n\n/** 모든 텍스트 OCR 모델 다운로드/검증 (있으면 skip) */\nexport async function ensureOcrModels(onProgress?: ProgressHandler): Promise<void> {\n return ensureModelsIn(getOcrModelsDir(), ALL_OCR_MODELS, onProgress)\n}\n\n/** 텍스트 OCR 모델 상태 (다운로드 없이 확인만) */\nexport async function getOcrModelStatus(): Promise<ModelStatus[]> {\n return getModelStatusIn(getOcrModelsDir(), ALL_OCR_MODELS)\n}\n\n/**\n * inference.yml 에서 `PostProcess.character_dict` 리스트를 추출.\n * 전체 YAML 파서 없이 해당 블록의 `- <char>` 라인만 순서대로 읽는다\n * (공식 yml 구조 고정 — 들여쓰기 2칸, 한 줄 한 글자).\n *\n * CTC 클래스 배열: index 0 = blank, 1..N = 사전 순서, N+1(마지막) = space.\n */\nexport function parseCharacterDict(yml: string): string[] {\n const lines = yml.split(\"\\n\")\n const chars: string[] = []\n let inDict = false\n let dictIndent = -1\n for (const line of lines) {\n if (!inDict) {\n const m = /^(\\s*)character_dict:\\s*$/.exec(line)\n if (m) {\n inDict = true\n dictIndent = m[1].length\n }\n continue\n }\n // 리스트 항목 — YAML 은 키와 같은 들여쓰기의 \"- x\" 도 허용 (공식 yml 실측: 둘 다 2칸)\n const m = /^(\\s*)- (.*)$/.exec(line)\n if (m && m[1].length >= dictIndent) {\n let v = m[2]\n // YAML 인용 문자 처리 ('#' 같은 예약 문자는 인용됨)\n if (v.length >= 2 && ((v.startsWith(\"'\") && v.endsWith(\"'\")) || (v.startsWith('\"') && v.endsWith('\"')))) {\n v = v.slice(1, -1).replace(/''/g, \"'\")\n }\n chars.push(v)\n continue\n }\n // 들여쓰기가 얕아지면 블록 종료 (빈 줄은 통과)\n if (line.trim() !== \"\") break\n }\n return chars\n}\n\nexport function ocrModelPath(spec: ModelSpec): string {\n return join(getOcrModelsDir(), spec.filename)\n}\n","/**\n * 검출 박스 픽셀 분석 — 행 밴드 분할 + 잉크 대비.\n *\n * DB 검출기는 줄 간격보다 글자 간격이 넓으면 글자를 세로로 잇는다: 표 머리의 세로쓰기\n * (\"국/균/도/시\", \"사회복지과\"), 균등배분 목차의 같은 열 글자들이 키 큰 박스 하나가 되고,\n * 인식기(높이 48 고정)는 이를 짓눌러 빈 문자열이나 저신뢰 쓰레기를 낸다(실측: 코퍼스\n * 80쪽에서 GT 글자 0.55% 가 이런 박스와 함께 폐기). 박스 안 행 투영의 빈 띠로 밴드를\n * 갈라 밴드마다 따로 인식하면 글자가 자기 위치(행)를 되찾는다.\n *\n * 잉크 대비는 배경 무늬(연한 바탕의 흰 도안)를 글자로 읽는 환각을 거른다 — 진짜 글자는\n * 전경/배경 평균 휘도 차가 크다.\n */\n\n/** RGBA 페이지에서 박스 영역 휘도(BT.601 정수 근사) 추출 */\nexport function grayCrop(\n rgba: Uint8Array,\n pageW: number,\n box: { x: number; y: number; w: number; h: number },\n): Uint8Array {\n const out = new Uint8Array(box.w * box.h)\n for (let y = 0; y < box.h; y++) {\n let si = ((box.y + y) * pageW + box.x) * 4\n let di = y * box.w\n for (let x = 0; x < box.w; x++, si += 4, di++) {\n out[di] = (rgba[si] * 77 + rgba[si + 1] * 150 + rgba[si + 2] * 29) >> 8\n }\n }\n return out\n}\n\nexport interface InkStats {\n /** Otsu 임계값 */\n threshold: number\n /** 잉크 = 소수 클래스. true 면 어두운 쪽이 잉크(일반 문서) */\n darkInk: boolean\n /** 두 클래스 평균 휘도 차 (0~255) — 글자 대비 */\n contrast: number\n /** 잉크 픽셀 비율 */\n inkRatio: number\n}\n\n/** Otsu 이진화 — 잉크 극성(소수 클래스)과 대비 */\nexport function inkStats(gray: Uint8Array): InkStats {\n const hist = new Uint32Array(256)\n for (let i = 0; i < gray.length; i++) hist[gray[i]]++\n const total = gray.length\n let sumAll = 0\n for (let v = 0; v < 256; v++) sumAll += v * hist[v]\n let wB = 0, sumB = 0, best = -1, thr = 127\n for (let t = 0; t < 256; t++) {\n wB += hist[t]\n if (wB === 0) continue\n const wF = total - wB\n if (wF === 0) break\n sumB += t * hist[t]\n const mB = sumB / wB\n const mF = (sumAll - sumB) / wF\n const between = wB * wF * (mB - mF) * (mB - mF)\n if (between > best) { best = between; thr = t }\n }\n let nDark = 0, sDark = 0\n for (let v = 0; v <= thr; v++) { nDark += hist[v]; sDark += v * hist[v] }\n const nLight = total - nDark\n const sLight = sumAll - sDark\n if (nDark === 0 || nLight === 0) return { threshold: thr, darkInk: true, contrast: 0, inkRatio: 0 }\n const darkInk = nDark <= nLight\n return {\n threshold: thr,\n darkInk,\n contrast: sLight / nLight - sDark / nDark,\n inkRatio: (darkInk ? nDark : nLight) / total,\n }\n}\n\n/**\n * 행 투영의 빈 띠로 밴드 분할. 반환: 밴드별 [y0, y1) 행 구간과 괘선을 뺀 잉크 [x0, x1) 열 구간\n * (박스 로컬). 분할 불가면 길이 1.\n * @param minBandRatio 밴드 최소 높이 = 이 비율 × 글자 크기. 이보다 얇은 조각(한 글자 안의\n * 획 띠 \"을\"의 ㅇ/ㅡ/ㄹ, 밑줄, 박스 끝에 걸린 이웃 글자 조각)은 간격이 좁은 쪽 이웃 밴드에\n * 흡수된다. 글자 크기는 밴드별 잉크 가로 폭의 중앙값 — 박스 폭은 unclip 여백만큼 글자보다\n * 넓고(◎ 40px 에 박스 92px) 박스 전체 잉크 폭은 옆 열 글자 조각에 끌려가서, 둘 다 쓰면\n * 제대로 된 글자 밴드까지 합쳐진다(함평 목차 실측). 획 조각도 가로 폭은 글자 폭(ㅡ)이라\n * 중앙값이 흔들리지 않는다\n */\nexport function splitRowBands(\n gray: Uint8Array,\n w: number,\n h: number,\n ink: InkStats,\n minBandRatio: number,\n): Array<{ y0: number; y1: number; x0: number; x1: number }> {\n const isInk = (v: number) => (ink.darkInk ? v <= ink.threshold : v > ink.threshold)\n // 박스를 관통하는 괘선 — 세로선(행 85%+ 에 잉크인 열)은 투영에서 빼고, 가로선(남은 폭\n // 85%+ 가 잉크인 행)은 빈 행(구분자)으로 본다. 표 칸 사이 세로쓰기 라벨(\"국/균/도/시\")\n // 박스는 칸 경계 세로선을 끼고 잡혀 빈 행이 하나도 없다(부천 예산서 실측)\n const colInk = new Uint32Array(w)\n for (let y = 0; y < h; y++) {\n const off = y * w\n for (let x = 0; x < w; x++) if (isInk(gray[off + x])) colInk[x]++\n }\n const ruleCol = new Uint8Array(w)\n let liveW = 0\n for (let x = 0; x < w; x++) { if (colInk[x] >= h * 0.85) ruleCol[x] = 1; else liveW++ }\n const rowInk = new Uint32Array(h)\n for (let y = 0; y < h; y++) {\n let n = 0\n const off = y * w\n for (let x = 0; x < w; x++) if (!ruleCol[x] && isInk(gray[off + x])) n++\n rowInk[y] = n >= liveW * 0.85 ? 0 : n\n }\n // 안티앨리어싱·잡티 허용: 폭의 2% (최소 1px) 이하면 빈 행\n const blank = Math.max(1, Math.floor(w * 0.02))\n let bands: Array<[number, number]> = []\n let start = -1\n for (let y = 0; y <= h; y++) {\n const on = y < h && rowInk[y] > blank\n if (on && start < 0) start = y\n else if (!on && start >= 0) { bands.push([start, y]); start = -1 }\n }\n const inkCols = (y0: number, y1: number): [number, number] => {\n let x0 = w, x1 = -1\n for (let y = y0; y < y1; y++) {\n const off = y * w\n for (let x = 0; x < w; x++) if (!ruleCol[x] && isInk(gray[off + x])) { if (x < x0) x0 = x; if (x > x1) x1 = x }\n }\n return x1 >= x0 ? [x0, x1 + 1] : [0, w]\n }\n if (bands.length <= 1) return [{ y0: bands[0]?.[0] ?? 0, y1: bands[0]?.[1] ?? h, x0: 0, x1: w }]\n const widths = bands.map(([y0, y1]) => { const [a, b] = inkCols(y0, y1); return b - a }).sort((a, b) => a - b)\n const charSize = widths[widths.length >> 1]\n const minBand = Math.max(3, minBandRatio * charSize)\n // 글자 안 빈 행(“업”의 어/ㅂ 사이 2px) — 간격이 글자 크기 15% 이하이고 합쳐도 한 글자\n // 높이(1.3배) 안이면 한 밴드. 세로쓰기의 글자 사이 간격은 이보다 넓고, 두 글자를 합치면\n // 높이 상한에 걸린다\n for (let i = 0; i + 1 < bands.length;) {\n const gap = bands[i + 1][0] - bands[i][1]\n if (gap <= Math.max(2, charSize * 0.15) && bands[i + 1][1] - bands[i][0] <= charSize * 1.3) {\n bands.splice(i, 2, [bands[i][0], bands[i + 1][1]])\n } else i++\n }\n // 박스 위아래 끝에 걸린 얇은 조각 = 이웃 줄 글자의 끝자락 — 합치지 않고 버린다\n while (bands.length > 1 && bands[0][0] <= 1 && bands[0][1] - bands[0][0] < minBand) bands.shift()\n while (bands.length > 1 && bands[bands.length - 1][1] >= h - 1 && bands[bands.length - 1][1] - bands[bands.length - 1][0] < minBand) bands.pop()\n // 얇은 밴드 흡수 — 간격이 더 좁은 쪽 이웃과 합침 (가장 얇은 것부터 반복)\n for (;;) {\n let idx = -1, minH = Infinity\n for (let i = 0; i < bands.length; i++) {\n const bh = bands[i][1] - bands[i][0]\n if (bh < minBand && bh < minH) { minH = bh; idx = i }\n }\n if (idx < 0 || bands.length === 1) break\n const gapPrev = idx > 0 ? bands[idx][0] - bands[idx - 1][1] : Infinity\n const gapNext = idx < bands.length - 1 ? bands[idx + 1][0] - bands[idx][1] : Infinity\n const j = gapPrev <= gapNext ? idx - 1 : idx + 1\n const a = Math.min(idx, j), b = Math.max(idx, j)\n bands = [...bands.slice(0, a), [bands[a][0], bands[b][1]], ...bands.slice(b + 1)]\n }\n if (bands.length === 1) return [{ y0: bands[0][0], y1: bands[0][1], x0: 0, x1: w }]\n return bands.map(([y0, y1]) => { const [x0, x1] = inkCols(y0, y1); return { y0, y1, x0, x1 } })\n}\n\n/**\n * 박스 안 잉크 외곽 (박스 로컬, [x0,x1)·[y0,y1)) — 관통 괘선(행/열 85%+ 잉크)의 픽셀은 뺀다.\n * det 박스는 unclip 여백만큼 글자보다 커서(본문 10pt 에 박스 높이 ≈ 1.5em) 이를 그대로\n * 좌표로 넘기면 글자 크기·줄 기준선·칸 배정이 텍스트층 아이템과 어긋난다. 표 칸 숫자 박스는\n * 칸 경계 괘선을 물고 잡히므로(고산 예산서 실측) 괘선 픽셀을 빼지 않으면 외곽이 박스 전체가\n * 된다. 잉크가 없으면 박스 전체.\n */\nexport function inkBounds(\n gray: Uint8Array,\n w: number,\n h: number,\n ink: InkStats,\n): { x0: number; y0: number; x1: number; y1: number } {\n const isInk = (v: number) => (ink.darkInk ? v <= ink.threshold : v > ink.threshold)\n const colInk = new Uint32Array(w)\n const rowInk = new Uint32Array(h)\n for (let y = 0; y < h; y++) {\n const off = y * w\n for (let x = 0; x < w; x++) if (isInk(gray[off + x])) { colInk[x]++; rowInk[y]++ }\n }\n const ruleCol = new Uint8Array(w), ruleRow = new Uint8Array(h)\n for (let x = 0; x < w; x++) if (colInk[x] >= h * 0.85) ruleCol[x] = 1\n for (let y = 0; y < h; y++) if (rowInk[y] >= w * 0.85) ruleRow[y] = 1\n let x0 = w, x1 = -1, y0 = h, y1 = -1\n for (let y = 0; y < h; y++) {\n if (ruleRow[y]) continue\n const off = y * w\n for (let x = 0; x < w; x++) {\n if (ruleCol[x] || !isInk(gray[off + x])) continue\n if (x < x0) x0 = x\n if (x > x1) x1 = x\n if (y < y0) y0 = y\n if (y > y1) y1 = y\n }\n }\n if (x1 < x0 || y1 < y0) return { x0: 0, y0: 0, x1: w, y1: h }\n return { x0, y0, x1: x1 + 1, y1: y1 + 1 }\n}\n","/**\n * OCR 인식 문자열 후처리 — 한국 공문서 표기 관행 복원.\n *\n * PP-OCRv5 korean 인식기의 사전(11,945자)에는 공문서 글머리·표기 기호 일부가 없다\n * (○ U+25CB·△ U+25B3·◎·■·「」·【】 — 사전 실측). 모델은 사전 안의 닮은 글자로\n * 읽는다: ○ → 라틴 O(코퍼스 박스 단위 실측 145회)·o, △ → 증가 기호 ∆(29회).\n * 공문서에서 쓰이지 않는 쪽만 되돌린다 — 모호한 것(ㅇ 글머리는 실제 자모, 0 은\n * \"0원\" 같은 숫자)은 건드리지 않는다.\n *\n * 따옴표: 모델은 ‘ ’ 를 곧은 ' 로 읽는다(114회). 공문서 텍스트층은 둥근 따옴표가 압도적\n * (코퍼스 GT ‘’ 129 : ' 7 — 한컴 자동 고침) → 여닫음 문맥으로 둥근 따옴표를 복원한다.\n */\n\nconst HANGUL_START = /^[\\uac00-\\ud7a3]/\n\n/** 박스 텍스트 한 줄 후처리 (위치 독립 규칙) */\nexport function restoreSymbols(text: string): string {\n let s = text\n // ○ 글머리: 줄 머리 O/o + (공백) + 한글\n s = s.replace(/^([Oo])(\\s?)(?=[\\uac00-\\ud7a3])/, \"\\u25cb$2\")\n // ○○ 자리표시(○○시·○○○ 과장): 라틴 글자와 붙지 않은 O 2개 이상이 한글과 (공백 하나 사이로) 이웃\n s = s.replace(/(?<![A-Za-z])O{2,}(?![A-Za-z])/g, (m, i: number) =>\n /[\\uac00-\\ud7a3]$/.test(s.slice(Math.max(0, i - 2), i).trimEnd()) || /^\\s?[\\uac00-\\ud7a3]/.test(s.slice(i + m.length))\n ? \"\\u25cb\".repeat(m.length) : m)\n // △ 감액 표시: 사전 밖 → ∆(U+2206)·그리스 Δ(U+0394) 로 읽힘\n s = s.replace(/[\\u2206\\u0394]/g, \"\\u25b3\")\n return smartQuotes(joinDigitGroups(s))\n}\n\n/**\n * 천 단위 숫자 안의 끼어든 공백 제거 — 모델이 쉼표 뒤에 공백을 넣는다(\"385, 426\"·\"4,802, 164\",\n * 예산서 표에서 수십 회 실측). 쉼표 앞이 1~3자리, 뒤가 정확히 3자리인 무리만 잇는다\n * (\"1, 2, 3\" 같은 목록·\"10, 20명\"은 뒤가 3자리가 아니라 그대로).\n */\nexport function joinDigitGroups(s: string): string {\n return s.replace(/(?<![\\d,])\\d{1,3}(?:, ?\\d{3})+(?![\\d])/g, m => m.replace(/, /g, \",\"))\n}\n\n/**\n * 곧은 따옴표 → 둥근 따옴표. 한 줄 안의 따옴표를 짝으로 보고 여닫음을 번갈아 준다 —\n * OCR 띄어쓰기가 불안정해(\"내에서'기록관리시스템'\") 앞 글자가 공백인지로는 여는 자리를\n * 가를 수 없다. 개수가 홀수면(줄바꿈에 걸친 인용) 첫 따옴표가 여는 자리(줄 머리·공백·\n * 여는 괄호 뒤)가 아닐 때 닫는 것부터. 연도 생략 '24 는 짝에서 빼고 ’.\n */\nexport function smartQuotes(s: string): string {\n if (!/['\"]/.test(s)) return s\n const chars = [...s]\n for (const q of [\"'\", '\"']) {\n const [open, close] = q === \"'\" ? [\"\\u2018\", \"\\u2019\"] : [\"\\u201c\", \"\\u201d\"]\n const pos: number[] = []\n for (let i = 0; i < chars.length; i++) {\n if (chars[i] !== q) continue\n const prev = i > 0 ? chars[i - 1] : \"\"\n if (q === \"'\" && !/[\\p{L}\\p{N}]/u.test(prev) && /^\\d{2}(?!\\d)/.test(chars.slice(i + 1, i + 4).join(\"\"))) {\n chars[i] = close // 연도 생략\n continue\n }\n pos.push(i)\n }\n if (pos.length === 0) continue\n const first = pos[0]\n const prev = first > 0 ? chars[first - 1] : \"\"\n const atOpen = prev === \"\" || /[\\s(\\[{<\\u300c\\u300e\\u3010\\u3008\\u300a\\u2018\\u201c·,:]/.test(prev)\n let isOpen = pos.length % 2 === 0 || atOpen\n for (const i of pos) { chars[i] = isOpen ? open : close; isOpen = !isOpen }\n }\n return chars.join(\"\")\n}\n\n/** 점류만으로 된 조각 — 목차 리더·잡티(\".\", \"..\", \"…\", \"·\"). 단독으로는 뜻이 없다 */\nexport function isDotFragment(text: string): boolean {\n return /^[\\s.\\u00b7\\u2024\\u2025\\u2026\\u2027\\u2219\\u22c5\\u318d]+$/.test(text)\n}\n\n/**\n * 이웃 인지 ○ 복원 — 글머리 ○ 가 제 박스로 떨어져 \"O\" 한 글자로 읽힌 경우,\n * 같은 줄(세로 중심 차 < 높이 절반) 바로 오른쪽(글자 높이 3배 이내) 박스가 한글로 시작하면 ○.\n */\nexport function restoreBulletItems<T extends { text: string; x: number; y: number; w: number; h: number }>(items: T[]): void {\n for (const it of items) {\n if (it.text !== \"O\" && it.text !== \"o\") continue\n const cy = it.y + it.h / 2\n let next: T | null = null\n for (const o of items) {\n if (o === it || o.x < it.x + it.w * 0.5) continue\n if (Math.abs(o.y + o.h / 2 - cy) > Math.max(it.h, o.h) / 2) continue\n if (o.x - (it.x + it.w) > it.h * 3) continue\n if (!next || o.x < next.x) next = o\n }\n if (next && HANGUL_START.test(next.text.trimStart())) it.text = \"\\u25cb\"\n }\n}\n","/**\n * 인식 입력 준비 — 밴드 서브 박스 좌표, 라인 crop 리사이즈(회전 포함).\n * (engine.ts 에서 분리 — 엔진은 세션·배치·후처리 흐름만)\n */\n\nexport interface Box { x: number; y: number; w: number; h: number }\n\n/** rec 입력 높이 (공식 RecResizeImg image_shape [3, 48, 320]) */\nexport const REC_HEIGHT = 48\nconst REC_MAX_WIDTH = 3200\nconst MIN_SIZE = 3\n\n/**\n * 밴드 → 페이지 좌표 서브 박스. 위아래는 밴드 높이 20% 여유(이웃 밴드와의 간격 절반 이내),\n * 좌우는 괘선을 뺀 잉크 폭 + 밴드 높이 20% 여유(원 박스 안) — 칸 경계 세로선이 crop 에\n * 들어가 \"|국\" 으로 읽히지 않게\n */\nexport function bandBoxes(b: Box, bands: Array<{ y0: number; y1: number; x0: number; x1: number }>, pageH: number): Box[] {\n const out: Box[] = []\n for (let i = 0; i < bands.length; i++) {\n const { y0, y1, x0, x1 } = bands[i]\n const bh = y1 - y0\n const upGap = i > 0 ? (y0 - bands[i - 1].y1) / 2 : y0\n const downGap = i < bands.length - 1 ? (bands[i + 1].y0 - y1) / 2 : b.h - y1\n const top = Math.max(0, Math.round(b.y + y0 - Math.min(bh * 0.2, upGap)))\n const bot = Math.min(pageH, Math.round(b.y + y1 + Math.min(bh * 0.2, downGap)))\n const left = Math.max(b.x, Math.round(b.x + x0 - bh * 0.2))\n const right = Math.min(b.x + b.w, Math.round(b.x + x1 + bh * 0.2))\n if (bot - top >= MIN_SIZE && right - left >= MIN_SIZE) out.push({ x: left, y: top, w: right - left, h: bot - top })\n }\n return out\n}\n\n/**\n * 라인 crop → rec 입력 크기(높이 48, 비율 유지 폭) RGB. 회전은 crop 좌표계에서\n * (90 = 반시계 np.rot90, 270 = 시계). 리사이즈는 bilinear(반픽셀 중심, cv2 INTER_LINEAR 등가)\n * — 공식 전처리(cv2.resize 기본)와 같은 보간.\n */\nexport function lineCrop(rgba: Uint8Array, pageW: number, box: Box, rot: 0 | 90 | 270): { rgb: Uint8Array; w: number } {\n const srcW = rot === 0 ? box.w : box.h\n const srcH = rot === 0 ? box.h : box.w\n const rw = Math.min(REC_MAX_WIDTH, Math.max(16, Math.round((srcW * REC_HEIGHT) / srcH)))\n const rgb = new Uint8Array(rw * REC_HEIGHT * 3)\n // 회전 좌표 → 페이지 픽셀\n const at = (u: number, v: number): number => {\n let px: number, py: number\n if (rot === 0) { px = u; py = v }\n else if (rot === 90) { px = box.w - 1 - v; py = u }\n else { px = v; py = box.h - 1 - u }\n return ((box.y + py) * pageW + (box.x + px)) * 4\n }\n const fx = srcW / rw\n const fy = srcH / REC_HEIGHT\n for (let dy = 0; dy < REC_HEIGHT; dy++) {\n let sy = (dy + 0.5) * fy - 0.5\n if (sy < 0) sy = 0\n const y0 = Math.min(srcH - 1, Math.floor(sy))\n const y1 = Math.min(srcH - 1, y0 + 1)\n const wy = sy - y0\n for (let dx = 0; dx < rw; dx++) {\n let sx = (dx + 0.5) * fx - 0.5\n if (sx < 0) sx = 0\n const x0 = Math.min(srcW - 1, Math.floor(sx))\n const x1 = Math.min(srcW - 1, x0 + 1)\n const wx = sx - x0\n const i00 = at(x0, y0), i01 = at(x1, y0), i10 = at(x0, y1), i11 = at(x1, y1)\n const o = (dy * rw + dx) * 3\n for (let c = 0; c < 3; c++) {\n const top = rgba[i00 + c] * (1 - wx) + rgba[i01 + c] * wx\n const bottom = rgba[i10 + c] * (1 - wx) + rgba[i11 + c] * wx\n rgb[o + c] = Math.round(top * (1 - wy) + bottom * wy)\n }\n }\n }\n return { rgb, w: rw }\n}\n\n","/**\n * 스캔 기울기 보정 — 투영 프로파일 방식.\n *\n * 스캔·사진 입력은 0.5~3° 기울기가 흔하다. det 은 축정렬 박스(connected-component bbox)\n * 라 기울어진 줄은 박스가 위아래 줄을 물고, 괘선 감지(ruling-lines.ts)는 축정렬 런렝스라\n * 기울면 표 괘선이 통째로 사라진다. 인식 전에 페이지를 바로 세운다.\n *\n * 추정: 페이지를 긴 변 ~1000px 로 줄여 잉크 픽셀을 모으고, 후보 각 θ 마다\n * y' = y·cosθ + x·sinθ 의 행 히스토그램 제곱합(줄이 수평일 때 가장 뾰족)을 비교 —\n * ±MAX 를 0.2° 간격으로 훑고 최댓값 주변을 0.02° 간격으로 다듬는다.\n * 적용 조건: |θ| ≥ MIN_APPLY_DEG 이고 0° 대비 뾰족함이 뚜렷할 때만 (클린 렌더는 무보정).\n */\n\n/** 추정 범위 (°) — 이보다 큰 회전은 기울기가 아니라 방향 문제 (범위 밖) */\nconst MAX_SKEW_DEG = 5\n/** 이보다 작은 기울기는 보정하지 않음 — 리샘플 흐림이 이득보다 크다 */\nconst MIN_APPLY_DEG = 0.25\n/** 0° 대비 투영 뾰족함 비율 하한 — 글자 적은 페이지·도형의 우연한 봉우리 거부 */\nconst MIN_GAIN = 1.03\nconst TARGET_LONG = 1000\n\nexport interface SkewEstimate {\n /** 페이지 기울기 (°, 양수 = 시계 방향으로 기움 → 같은 각만큼 반시계로 돌리면 바로 섬) */\n angle: number\n /** 최적각 투영 점수 / 0° 점수 */\n gain: number\n}\n\n/** 기울기 추정 (적용 여부와 무관한 원값) */\nexport function estimateSkew(rgba: Uint8Array, width: number, height: number): SkewEstimate {\n const step = Math.max(1, Math.ceil(Math.max(width, height) / TARGET_LONG))\n const sw = Math.floor(width / step), sh = Math.floor(height / step)\n // 축소 휘도 + 전역 문턱 (잉크 = 어두운 픽셀)\n const xs: number[] = [], ys: number[] = []\n for (let y = 0; y < sh; y++) {\n for (let x = 0; x < sw; x++) {\n const i = ((y * step) * width + x * step) * 4\n const l = (rgba[i] * 77 + rgba[i + 1] * 150 + rgba[i + 2] * 29) >> 8\n if (l < 128) { xs.push(x); ys.push(y) }\n }\n }\n if (xs.length < 200) return { angle: 0, gain: 1 }\n const diag = Math.ceil(Math.hypot(sw, sh))\n const hist = new Float64Array(2 * diag + 2)\n const score = (deg: number): number => {\n const t = (deg * Math.PI) / 180\n const c = Math.cos(t), s = Math.sin(t)\n hist.fill(0)\n for (let k = 0; k < xs.length; k++) hist[Math.round(ys[k] * c + xs[k] * s) + diag]++\n let sum = 0\n for (let i = 0; i < hist.length; i++) sum += hist[i] * hist[i]\n return sum\n }\n const s0 = score(0)\n let best = 0, bestS = s0\n for (let d = -MAX_SKEW_DEG; d <= MAX_SKEW_DEG + 1e-9; d += 0.2) {\n const v = score(d)\n if (v > bestS) { bestS = v; best = d }\n }\n const coarse = best\n for (let d = coarse - 0.2; d <= coarse + 0.2 + 1e-9; d += 0.02) {\n const v = score(d)\n if (v > bestS) { bestS = v; best = d }\n }\n // 투영 회전각 θ 에서 줄이 수평 → 페이지는 −θ 만큼 기운 것\n return { angle: +(-best).toFixed(2), gain: s0 > 0 ? bestS / s0 : 1 }\n}\n\n/**\n * 필요하면 페이지를 바로 세운 RGBA 를 돌려준다 (같은 크기, 중심 회전, 바깥은 흰색).\n * 보정하지 않으면 입력 그대로 + angle 0.\n */\nexport function deskewPage(rgba: Uint8Array, width: number, height: number): { rgba: Uint8Array; angle: number } {\n const est = estimateSkew(rgba, width, height)\n if (Math.abs(est.angle) < MIN_APPLY_DEG || est.gain < MIN_GAIN) return { rgba, angle: 0 }\n return { rgba: rotateRgba(rgba, width, height, est.angle), angle: est.angle }\n}\n\n/**\n * 중심 기준 회전 (deg, 양수 = 화면상 반시계), 같은 캔버스 크기, bilinear, 바깥 흰색.\n * 출력 픽셀 (x,y) 는 입력의 역회전 위치에서 표본.\n */\nexport function rotateRgba(rgba: Uint8Array, width: number, height: number, deg: number): Uint8Array {\n const out = new Uint8Array(rgba.length)\n const t = (deg * Math.PI) / 180\n const c = Math.cos(t), s = Math.sin(t)\n const cx = (width - 1) / 2, cy = (height - 1) / 2\n for (let y = 0; y < height; y++) {\n const dy = y - cy\n for (let x = 0; x < width; x++) {\n const dx = x - cx\n // 화면 좌표(y 아래)에서 반시계 deg 회전의 역변환\n const sx = c * dx - s * dy + cx\n const sy = s * dx + c * dy + cy\n const o = (y * width + x) * 4\n const x0 = Math.floor(sx), y0 = Math.floor(sy)\n if (x0 < 0 || y0 < 0 || x0 >= width - 1 || y0 >= height - 1) {\n out[o] = 255; out[o + 1] = 255; out[o + 2] = 255; out[o + 3] = 255\n continue\n }\n const wx = sx - x0, wy = sy - y0\n const i00 = (y0 * width + x0) * 4, i01 = i00 + 4, i10 = i00 + width * 4, i11 = i10 + 4\n for (let ch = 0; ch < 4; ch++) {\n const top = rgba[i00 + ch] * (1 - wx) + rgba[i01 + ch] * wx\n const bot = rgba[i10 + ch] * (1 - wx) + rgba[i11 + ch] * wx\n out[o + ch] = Math.round(top * (1 - wy) + bot * wy)\n }\n }\n }\n return out\n}\n","/**\n * PDF OCR 브릿지 — 대상 페이지를 pdfium 으로 래스터해 내장 엔진 또는\n * 사용자 OcrProvider 로 인식하고, 페이지별 IRBlock[] 을 돌려준다.\n *\n * - 내장 엔진: OcrItem 좌표를 PDF 포인트로 환산해 기존 블록 파이프라인\n * (extractPageBlocksWithLines — xy-cut 읽기순서·클러스터 표 감지)에 태운다.\n * 스캔 문서도 표 구조 복원이 가능한 이유.\n * - 사용자 프로바이더: 페이지 PNG → 텍스트 → 페이지당 paragraph 블록 (종전 계약).\n *\n * 에러 계약 (v4.1.0 리뷰 F6/F7):\n * - 환경 오류(의존성·모델 미설치, PDF 오픈 실패)는 throw — 호출자가 NEEDS_OCR 폴백.\n * - 페이지 단위 실패는 본문 오염 대신 warnings(OCR_FAILED) 로 기록하고 계속.\n */\n\nimport type { IRBlock, OcrProvider, ParseWarning } from \"../types.js\"\nimport type { NormItem } from \"../pdf/text-line.js\"\nimport type { LineSegment } from \"../pdf/line-types.js\"\nimport { extractPageBlocksWithLines } from \"../pdf/page-blocks.js\"\nimport { detectRulingLines, rulingToPdfLines } from \"./ruling-lines.js\"\nimport { getOcrEngine, type OcrItem } from \"./engine.js\"\nimport { deskewPage } from \"./deskew.js\"\nimport { ensureOcrModels } from \"./models.js\"\n\n/** OCR 렌더 스케일 (72dpi × 3 = 216dpi) — 10pt 본문이 rec 입력 높이(48px)에 근접 */\nconst OCR_RENDER_SCALE = 3\n/** 페이지 하나당 OCR 타임아웃 — det+rec 수십 라인 기준 넉넉히 */\nconst PAGE_TIMEOUT_MS = 120_000\n\nexport type OcrMode = \"builtin\" | OcrProvider\n\n/**\n * 대상 페이지들을 OCR 해 페이지별 블록 맵 반환.\n * @param buffer 원본 PDF (pdfjs 가 detach 하기 전에 clone 해 둔 것)\n * @param targets 1-based 페이지 번호 집합\n */\nexport async function runPdfOcr(\n buffer: ArrayBuffer,\n targets: Set<number>,\n mode: OcrMode,\n warnings: ParseWarning[],\n onProgress?: (current: number, total: number) => void,\n detectTables = true,\n): Promise<Map<number, IRBlock[]>> {\n const result = new Map<number, IRBlock[]>()\n if (targets.size === 0) return result\n\n // 환경 준비 — 실패는 그대로 throw (호출자가 NEEDS_OCR 폴백)\n const pdfiumMod = await tryImport<typeof import(\"@hyzyla/pdfium\")>(\n \"@hyzyla/pdfium\",\n () => import(\"@hyzyla/pdfium\"),\n )\n if (mode === \"builtin\") {\n await ensureOcrModels(p => {\n if (p.phase === \"download\" && p.downloaded === 0) {\n process.stderr.write(`[kordoc-ocr] ${p.spec.name} 다운로드 중 (~${p.spec.sizeMb}MB)...\\n`)\n }\n })\n }\n const engine = mode === \"builtin\" ? await getOcrEngine() : null\n\n const pdfium = await pdfiumMod.PDFiumLibrary.init()\n const doc = await pdfium.loadDocument(new Uint8Array(buffer))\n try {\n let done = 0\n for (const page of doc.pages()) {\n // pdfium page.number 는 0-based pageIndex — 대외 계약(1-based)으로 환산\n const pageNo = page.number + 1\n if (!targets.has(pageNo)) continue\n onProgress?.(++done, targets.size)\n try {\n const blocks = await withTimeout(\n ocrOnePage(page, pageNo, mode, engine, warnings, detectTables),\n PAGE_TIMEOUT_MS,\n `OCR 페이지 ${pageNo} 타임아웃 (${PAGE_TIMEOUT_MS / 1000}초)`,\n )\n result.set(pageNo, blocks)\n } catch (e) {\n warnings.push({\n page: pageNo,\n message: `페이지 ${pageNo} OCR 실패: ${e instanceof Error ? e.message : String(e)}`,\n code: \"OCR_FAILED\",\n })\n }\n }\n } finally {\n doc.destroy()\n pdfium.destroy()\n }\n return result\n}\n\nasync function ocrOnePage(\n page: import(\"@hyzyla/pdfium\").PDFiumPage,\n pageNo: number,\n mode: OcrMode,\n engine: Awaited<ReturnType<typeof getOcrEngine>> | null,\n warnings: ParseWarning[],\n detectTables: boolean,\n): Promise<IRBlock[]> {\n const { originalWidth: pdfW, originalHeight: pdfH } = page.getOriginalSize()\n const rendered = await page.render({\n scale: OCR_RENDER_SCALE,\n render: async ({ data }) => data,\n })\n const { data: bgra, width: rw, height: rh } = rendered\n const rgba = bgraToRgba(bgra)\n\n if (mode === \"builtin\") {\n // 스캔 기울기 보정 — 인식과 괘선 감지가 같은(바로 선) 래스터를 본다. 클린 렌더는 무보정\n const { rgba: upright } = deskewPage(rgba, rw, rh)\n const stats = { droppedLowConf: 0 }\n const items = await engine!.recognizePage(upright, rw, rh, stats)\n if (stats.droppedLowConf > 0) {\n warnings.push({\n page: pageNo,\n message: `페이지 ${pageNo}: 저신뢰 OCR 라인 ${stats.droppedLowConf}개 폐기 (인식 결손 가능)`,\n code: \"OCR_LOW_CONF\",\n })\n }\n // 래스터에서 표 괘선 감지 — 스캔본 병합셀 서식도 선 기반 표 파이프라인을 탄다\n const scale = rh / pdfH\n const ruling = detectRulingLines(upright, rw, rh, scale)\n const extraLines = rulingToPdfLines(ruling, scale, pdfH)\n return ocrItemsToBlocks(items, pageNo, pdfW, pdfH, scale, extraLines, detectTables)\n }\n\n // 사용자 프로바이더 — PNG 인코딩 후 호출, 페이지당 paragraph (종전 계약)\n type SharpPngFactory = (\n input: Uint8Array,\n opts: { raw: { width: number; height: number; channels: number } },\n ) => { png(): { toBuffer(): Promise<Buffer> } }\n const sharpModRaw = await tryImport<{ default?: SharpPngFactory } & SharpPngFactory>(\n \"sharp\",\n () => import(\"sharp\") as unknown as Promise<{ default?: SharpPngFactory } & SharpPngFactory>,\n )\n const sharpAny = sharpModRaw as { default?: SharpPngFactory } | SharpPngFactory\n const sharp: SharpPngFactory =\n typeof sharpAny === \"function\" ? sharpAny : (sharpAny.default ?? (sharpAny as unknown as SharpPngFactory))\n const png = await sharp(rgba, { raw: { width: rw, height: rh, channels: 4 } })\n .png()\n .toBuffer()\n\n const text = await mode(new Uint8Array(png), pageNo, \"image/png\")\n if (!text.trim()) {\n warnings.push({ page: pageNo, message: `페이지 ${pageNo} OCR 결과 없음`, code: \"OCR_FAILED\" })\n return []\n }\n return [{ type: \"paragraph\", text: text.trim(), pageNumber: pageNo }]\n}\n\n/**\n * OcrItem(렌더 픽셀, top-left origin) → NormItem(PDF pt, bottom-up baseline)\n * 으로 환산해 기존 블록 파이프라인에 태운다. 괘선은 그래픽 ops 대신\n * 래스터 감지 결과(extraLines)로 공급 — 없으면 클러스터 감지기 몫이다.\n * (이미지 직접 입력 경로 image-ocr.ts 와 공유)\n */\nexport function ocrItemsToBlocks(\n items: OcrItem[],\n pageNumber: number,\n pdfW: number,\n pdfH: number,\n scale: number,\n extraLines?: { horizontals: LineSegment[]; verticals: LineSegment[] },\n detectTables = true,\n): IRBlock[] {\n const norm: NormItem[] = items.map(it => {\n const h = it.h / scale\n return {\n text: it.text,\n x: Math.round(it.x / scale),\n // NormItem.y 는 pdfjs transform[5] = 베이스라인 (bottom-up) — 잉크 하단으로 근사\n y: Math.round(pdfH - (it.y + it.h) / scale),\n w: Math.round(it.w / scale),\n h: Math.round(h),\n // 박스는 잉크 외곽(engine tightBoxes) — 한글 줄 잉크 높이 ≈ 0.9em\n fontSize: Math.max(1, Math.round(h / 0.9)),\n fontName: \"ocr\",\n isHidden: false,\n }\n })\n return extractPageBlocksWithLines(norm, pageNumber, { fnArray: [], argsArray: [] }, pdfW, pdfH, extraLines, detectTables)\n}\n\nasync function tryImport<T>(name: string, loader: () => Promise<T>): Promise<T> {\n try {\n return await loader()\n } catch (e) {\n throw new Error(\n `OCR 을 사용하려면 optional dependency '${name}' 이 필요합니다. ` +\n `\\`npm install ${name}\\` 후 다시 실행하세요. 원인: ${(e as Error).message}`,\n )\n }\n}\n\nasync function withTimeout<T>(promise: Promise<T>, ms: number, msg: string): Promise<T> {\n // 타임아웃 패배 후 원 promise 의 사후 reject 가 unhandled rejection 이 되지 않게 흡수\n promise.catch(() => {})\n let timer: NodeJS.Timeout | undefined\n try {\n return await Promise.race([\n promise,\n new Promise<never>((_, reject) => {\n timer = setTimeout(() => reject(new Error(msg)), ms)\n }),\n ])\n } finally {\n if (timer) clearTimeout(timer)\n }\n}\n\n/** pdfium BGRA → RGBA */\nfunction bgraToRgba(bgra: Uint8Array): Uint8Array {\n const out = new Uint8Array(bgra.length)\n for (let i = 0; i < bgra.length; i += 4) {\n out[i] = bgra[i + 2]\n out[i + 1] = bgra[i + 1]\n out[i + 2] = bgra[i]\n out[i + 3] = bgra[i + 3]\n }\n return out\n}\n"],"mappings":";;;;;;;;;AAkBA,IAAM,eAAe;AAErB,IAAM,qBAAqB;AAE3B,IAAM,oBAAoB;AAE1B,IAAM,aAAa;AAEnB,IAAM,qBAAqB;AAE3B,IAAM,qBAAqB;AAkBpB,SAAS,kBACd,MACA,OACA,QACA,OACa;AACb,QAAM,WAAW,KAAK,MAAM,qBAAqB,KAAK;AACtD,QAAM,aAAa,KAAK,IAAI,GAAG,KAAK,MAAM,oBAAoB,KAAK,CAAC;AAGpE,QAAM,MAAM,IAAI,WAAW,QAAQ,MAAM;AACzC,WAAS,IAAI,GAAG,IAAI,GAAG,IAAI,IAAI,QAAQ,KAAK,KAAK,GAAG;AAElD,UAAM,OAAQ,KAAK,CAAC,IAAI,KAAK,KAAK,IAAI,CAAC,IAAI,MAAM,KAAK,IAAI,CAAC,IAAI,MAAO;AACtE,QAAI,QAAQ,gBAAgB,KAAK,IAAI,CAAC,KAAK,IAAK,KAAI,CAAC,IAAI;AAAA,EAC3D;AAEA,QAAM,cAAc,YAAY,KAAK,OAAO,QAAQ,UAAU,YAAY,KAAK;AAC/E,QAAM,YAAY,YAAY,KAAK,OAAO,QAAQ,UAAU,YAAY,IAAI;AAC5E,SAAO,EAAE,aAAa,UAAU;AAClC;AAcA,SAAS,YACP,KACA,OACA,QACA,UACA,YACA,WACa;AACb,QAAM,QAAQ,YAAY,QAAQ;AAClC,QAAM,OAAO,YAAY,SAAS;AAClC,QAAM,KAAK,YACP,CAAC,MAAc,QAAgB,IAAI,MAAM,QAAQ,IAAI,IACrD,CAAC,MAAc,QAAgB,IAAI,OAAO,QAAQ,GAAG;AAEzD,QAAM,MAAmB,CAAC;AAC1B,MAAI,OAAe,CAAC;AAGpB,QAAM,mBAAmB,CAAC,GAAS,SAAyB;AAC1D,QAAI,OAAO,KAAK,QAAQ,MAAO,QAAO;AACtC,QAAI,OAAO,GAAG,QAAQ;AACtB,aAAS,MAAM,EAAE,IAAI,OAAO,EAAE,IAAI,OAAO,GAAG;AAC1C;AACA,UAAI,GAAG,MAAM,GAAG,MAAM,EAAG;AAAA,IAC3B;AACA,WAAO,QAAQ,IAAI,OAAO,QAAQ;AAAA,EACpC;AAEA,QAAM,OAAO,CAAC,MAAY;AACxB,UAAM,QAAQ,EAAE,OAAO,EAAE,QAAQ;AACjC,QAAI,QAAQ,WAAY;AACxB,QACE,iBAAiB,GAAG,EAAE,QAAQ,CAAC,KAAK,sBACpC,iBAAiB,GAAG,EAAE,OAAO,CAAC,KAAK,mBACnC;AACF,UAAM,UAAU,EAAE,QAAQ,EAAE,QAAQ;AACpC,QAAI;AAAA,MACF,YACI,EAAE,IAAI,QAAQ,IAAI,EAAE,IAAI,IAAI,QAAQ,IAAI,EAAE,IAAI,aAAa,MAAM,IACjE,EAAE,IAAI,EAAE,IAAI,IAAI,QAAQ,IAAI,EAAE,IAAI,IAAI,QAAQ,aAAa,MAAM;AAAA,IACvE;AAAA,EACF;AAEA,WAAS,OAAO,GAAG,OAAO,OAAO,QAAQ;AAEvC,UAAM,OAA0C,CAAC;AACjD,QAAI,WAAW;AACf,QAAI,MAAM;AACV,aAAS,MAAM,GAAG,OAAO,MAAM,OAAO;AACpC,YAAM,KAAK,MAAM,QAAQ,GAAG,MAAM,GAAG,MAAM;AAC3C,UAAI,IAAI;AACN,YAAI,WAAW,EAAG,YAAW;AAC7B,cAAM;AAAA,MACR,WAAW,YAAY,GAAG;AACxB,YAAI,EAAE,MAAM,cAAc,OAAO,MAAM;AACrC,gBAAM,KAAK,MAAM;AACjB,cAAI,KAAK,WAAW,KAAK,SAAU,MAAK,KAAK,EAAE,IAAI,UAAU,GAAG,CAAC;AACjE,qBAAW;AACX,gBAAM;AAAA,QACR;AAAA,MACF;AAAA,IACF;AAGA,UAAM,OAAe,CAAC;AACtB,UAAM,OAAO,oBAAI,IAAY;AAC7B,eAAW,QAAQ,MAAM;AACvB,UAAI,KAAK,SAAS,OAAO,GAAG;AAAE,aAAK,IAAI;AAAG;AAAA,MAAS;AACnD,UAAI,SAAS;AACb,eAAS,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;AACpC,YAAI,KAAK,IAAI,CAAC,EAAG;AACjB,cAAM,MAAM,KAAK,CAAC;AAClB,cAAM,UAAU,KAAK,IAAI,KAAK,IAAI,IAAI,EAAE,IAAI,KAAK,IAAI,KAAK,IAAI,IAAI,EAAE,IAAI;AACxE,cAAM,UAAU,KAAK,IAAI,KAAK,KAAK,KAAK,IAAI,IAAI,KAAK,IAAI,EAAE,IAAI;AAC/D,YAAI,WAAW,UAAU,oBAAoB;AAC3C,eAAK,KAAK,KAAK,IAAI,KAAK,IAAI,IAAI,EAAE;AAClC,eAAK,KAAK,KAAK,IAAI,KAAK,IAAI,IAAI,EAAE;AAClC,eAAK,OAAO;AACZ,eAAK,KAAK,IAAI;AACd,eAAK,IAAI,CAAC;AACV,mBAAS;AACT;AAAA,QACF;AAAA,MACF;AACA,UAAI,CAAC,OAAQ,MAAK,IAAI;AAAA,IACxB;AACA,aAAS,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;AACpC,UAAI,CAAC,KAAK,IAAI,CAAC,EAAG,MAAK,KAAK,EAAE,IAAI,KAAK,CAAC,EAAE,IAAI,IAAI,KAAK,CAAC,EAAE,IAAI,OAAO,MAAM,MAAM,KAAK,CAAC;AAAA,IACzF;AACA,WAAO;AAAA,EACT;AACA,aAAW,QAAQ,KAAM,MAAK,IAAI;AAClC,SAAO;AACT;AAMO,SAAS,iBACd,QACA,OACA,WAC0D;AAC1D,QAAM,cAA6B,OAAO,YAAY,IAAI,OAAK;AAC7D,UAAM,IAAI,aAAa,EAAE,KAAK,EAAE,MAAM,IAAI;AAC1C,WAAO;AAAA,MACL,IAAI,KAAK,IAAI,EAAE,IAAI,EAAE,EAAE,IAAI;AAAA,MAC3B,IAAI;AAAA,MACJ,IAAI,KAAK,IAAI,EAAE,IAAI,EAAE,EAAE,IAAI;AAAA,MAC3B,IAAI;AAAA,MACJ,WAAW,EAAE,cAAc;AAAA,IAC7B;AAAA,EACF,CAAC;AACD,QAAM,YAA2B,OAAO,UAAU,IAAI,OAAK;AACzD,UAAM,KAAK,EAAE,KAAK,EAAE,MAAM,IAAI;AAC9B,UAAM,KAAK,YAAY,KAAK,IAAI,EAAE,IAAI,EAAE,EAAE,IAAI;AAC9C,UAAM,KAAK,YAAY,KAAK,IAAI,EAAE,IAAI,EAAE,EAAE,IAAI;AAC9C,WAAO,EAAE,IAAI,GAAG,IAAI,IAAI,IAAI,GAAG,IAAI,IAAI,WAAW,EAAE,cAAc,MAAM;AAAA,EAC1E,CAAC;AACD,SAAO,EAAE,aAAa,UAAU;AAClC;;;AC9KA,SAAS,gBAAgB;AACzB,SAAS,YAAY;;;ACNd,IAAM,gBAA2B;AAAA,EACtC,MAAM;AAAA,EACN,UAAU;AAAA,EACV,KAAK;AAAA,EACL,QAAQ;AAAA,EACR,QAAQ;AACV;AAEO,IAAM,gBAA2B;AAAA,EACtC,MAAM;AAAA,EACN,UAAU;AAAA,EACV,KAAK;AAAA,EACL,QAAQ;AAAA,EACR,QAAQ;AACV;AAEO,IAAM,eAA0B;AAAA,EACrC,MAAM;AAAA,EACN,UAAU;AAAA,EACV,KAAK;AAAA,EACL,QAAQ;AAAA,EACR,QAAQ;AACV;AAEO,IAAM,iBAA2C,CAAC,eAAe,eAAe,YAAY;AAE5F,SAAS,kBAA0B;AACxC,SAAO,aAAa,OAAO;AAC7B;AAGA,eAAsB,gBAAgB,YAA6C;AACjF,SAAO,eAAe,gBAAgB,GAAG,gBAAgB,UAAU;AACrE;AAcO,SAAS,mBAAmB,KAAuB;AACxD,QAAM,QAAQ,IAAI,MAAM,IAAI;AAC5B,QAAM,QAAkB,CAAC;AACzB,MAAI,SAAS;AACb,MAAI,aAAa;AACjB,aAAW,QAAQ,OAAO;AACxB,QAAI,CAAC,QAAQ;AACX,YAAMA,KAAI,4BAA4B,KAAK,IAAI;AAC/C,UAAIA,IAAG;AACL,iBAAS;AACT,qBAAaA,GAAE,CAAC,EAAE;AAAA,MACpB;AACA;AAAA,IACF;AAEA,UAAM,IAAI,gBAAgB,KAAK,IAAI;AACnC,QAAI,KAAK,EAAE,CAAC,EAAE,UAAU,YAAY;AAClC,UAAI,IAAI,EAAE,CAAC;AAEX,UAAI,EAAE,UAAU,MAAO,EAAE,WAAW,GAAG,KAAK,EAAE,SAAS,GAAG,KAAO,EAAE,WAAW,GAAG,KAAK,EAAE,SAAS,GAAG,IAAK;AACvG,YAAI,EAAE,MAAM,GAAG,EAAE,EAAE,QAAQ,OAAO,GAAG;AAAA,MACvC;AACA,YAAM,KAAK,CAAC;AACZ;AAAA,IACF;AAEA,QAAI,KAAK,KAAK,MAAM,GAAI;AAAA,EAC1B;AACA,SAAO;AACT;;;ACpFO,SAAS,SACd,MACA,OACA,KACY;AACZ,QAAM,MAAM,IAAI,WAAW,IAAI,IAAI,IAAI,CAAC;AACxC,WAAS,IAAI,GAAG,IAAI,IAAI,GAAG,KAAK;AAC9B,QAAI,OAAO,IAAI,IAAI,KAAK,QAAQ,IAAI,KAAK;AACzC,QAAI,KAAK,IAAI,IAAI;AACjB,aAAS,IAAI,GAAG,IAAI,IAAI,GAAG,KAAK,MAAM,GAAG,MAAM;AAC7C,UAAI,EAAE,IAAK,KAAK,EAAE,IAAI,KAAK,KAAK,KAAK,CAAC,IAAI,MAAM,KAAK,KAAK,CAAC,IAAI,MAAO;AAAA,IACxE;AAAA,EACF;AACA,SAAO;AACT;AAcO,SAAS,SAAS,MAA4B;AACnD,QAAM,OAAO,IAAI,YAAY,GAAG;AAChC,WAAS,IAAI,GAAG,IAAI,KAAK,QAAQ,IAAK,MAAK,KAAK,CAAC,CAAC;AAClD,QAAM,QAAQ,KAAK;AACnB,MAAI,SAAS;AACb,WAAS,IAAI,GAAG,IAAI,KAAK,IAAK,WAAU,IAAI,KAAK,CAAC;AAClD,MAAI,KAAK,GAAG,OAAO,GAAG,OAAO,IAAI,MAAM;AACvC,WAAS,IAAI,GAAG,IAAI,KAAK,KAAK;AAC5B,UAAM,KAAK,CAAC;AACZ,QAAI,OAAO,EAAG;AACd,UAAM,KAAK,QAAQ;AACnB,QAAI,OAAO,EAAG;AACd,YAAQ,IAAI,KAAK,CAAC;AAClB,UAAM,KAAK,OAAO;AAClB,UAAM,MAAM,SAAS,QAAQ;AAC7B,UAAM,UAAU,KAAK,MAAM,KAAK,OAAO,KAAK;AAC5C,QAAI,UAAU,MAAM;AAAE,aAAO;AAAS,YAAM;AAAA,IAAE;AAAA,EAChD;AACA,MAAI,QAAQ,GAAG,QAAQ;AACvB,WAAS,IAAI,GAAG,KAAK,KAAK,KAAK;AAAE,aAAS,KAAK,CAAC;AAAG,aAAS,IAAI,KAAK,CAAC;AAAA,EAAE;AACxE,QAAM,SAAS,QAAQ;AACvB,QAAM,SAAS,SAAS;AACxB,MAAI,UAAU,KAAK,WAAW,EAAG,QAAO,EAAE,WAAW,KAAK,SAAS,MAAM,UAAU,GAAG,UAAU,EAAE;AAClG,QAAM,UAAU,SAAS;AACzB,SAAO;AAAA,IACL,WAAW;AAAA,IACX;AAAA,IACA,UAAU,SAAS,SAAS,QAAQ;AAAA,IACpC,WAAW,UAAU,QAAQ,UAAU;AAAA,EACzC;AACF;AAYO,SAAS,cACd,MACA,GACA,GACA,KACA,cAC2D;AAC3D,QAAM,QAAQ,CAAC,MAAe,IAAI,UAAU,KAAK,IAAI,YAAY,IAAI,IAAI;AAIzE,QAAM,SAAS,IAAI,YAAY,CAAC;AAChC,WAAS,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,UAAM,MAAM,IAAI;AAChB,aAAS,IAAI,GAAG,IAAI,GAAG,IAAK,KAAI,MAAM,KAAK,MAAM,CAAC,CAAC,EAAG,QAAO,CAAC;AAAA,EAChE;AACA,QAAM,UAAU,IAAI,WAAW,CAAC;AAChC,MAAI,QAAQ;AACZ,WAAS,IAAI,GAAG,IAAI,GAAG,KAAK;AAAE,QAAI,OAAO,CAAC,KAAK,IAAI,KAAM,SAAQ,CAAC,IAAI;AAAA,QAAQ;AAAA,EAAQ;AACtF,QAAM,SAAS,IAAI,YAAY,CAAC;AAChC,WAAS,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,QAAI,IAAI;AACR,UAAM,MAAM,IAAI;AAChB,aAAS,IAAI,GAAG,IAAI,GAAG,IAAK,KAAI,CAAC,QAAQ,CAAC,KAAK,MAAM,KAAK,MAAM,CAAC,CAAC,EAAG;AACrE,WAAO,CAAC,IAAI,KAAK,QAAQ,OAAO,IAAI;AAAA,EACtC;AAEA,QAAM,QAAQ,KAAK,IAAI,GAAG,KAAK,MAAM,IAAI,IAAI,CAAC;AAC9C,MAAI,QAAiC,CAAC;AACtC,MAAI,QAAQ;AACZ,WAAS,IAAI,GAAG,KAAK,GAAG,KAAK;AAC3B,UAAM,KAAK,IAAI,KAAK,OAAO,CAAC,IAAI;AAChC,QAAI,MAAM,QAAQ,EAAG,SAAQ;AAAA,aACpB,CAAC,MAAM,SAAS,GAAG;AAAE,YAAM,KAAK,CAAC,OAAO,CAAC,CAAC;AAAG,cAAQ;AAAA,IAAG;AAAA,EACnE;AACA,QAAM,UAAU,CAAC,IAAY,OAAiC;AAC5D,QAAI,KAAK,GAAG,KAAK;AACjB,aAAS,IAAI,IAAI,IAAI,IAAI,KAAK;AAC5B,YAAM,MAAM,IAAI;AAChB,eAAS,IAAI,GAAG,IAAI,GAAG,IAAK,KAAI,CAAC,QAAQ,CAAC,KAAK,MAAM,KAAK,MAAM,CAAC,CAAC,GAAG;AAAE,YAAI,IAAI,GAAI,MAAK;AAAG,YAAI,IAAI,GAAI,MAAK;AAAA,MAAE;AAAA,IAChH;AACA,WAAO,MAAM,KAAK,CAAC,IAAI,KAAK,CAAC,IAAI,CAAC,GAAG,CAAC;AAAA,EACxC;AACA,MAAI,MAAM,UAAU,EAAG,QAAO,CAAC,EAAE,IAAI,MAAM,CAAC,IAAI,CAAC,KAAK,GAAG,IAAI,MAAM,CAAC,IAAI,CAAC,KAAK,GAAG,IAAI,GAAG,IAAI,EAAE,CAAC;AAC/F,QAAM,SAAS,MAAM,IAAI,CAAC,CAAC,IAAI,EAAE,MAAM;AAAE,UAAM,CAAC,GAAG,CAAC,IAAI,QAAQ,IAAI,EAAE;AAAG,WAAO,IAAI;AAAA,EAAE,CAAC,EAAE,KAAK,CAAC,GAAG,MAAM,IAAI,CAAC;AAC7G,QAAM,WAAW,OAAO,OAAO,UAAU,CAAC;AAC1C,QAAM,UAAU,KAAK,IAAI,GAAG,eAAe,QAAQ;AAInD,WAAS,IAAI,GAAG,IAAI,IAAI,MAAM,UAAS;AACrC,UAAM,MAAM,MAAM,IAAI,CAAC,EAAE,CAAC,IAAI,MAAM,CAAC,EAAE,CAAC;AACxC,QAAI,OAAO,KAAK,IAAI,GAAG,WAAW,IAAI,KAAK,MAAM,IAAI,CAAC,EAAE,CAAC,IAAI,MAAM,CAAC,EAAE,CAAC,KAAK,WAAW,KAAK;AAC1F,YAAM,OAAO,GAAG,GAAG,CAAC,MAAM,CAAC,EAAE,CAAC,GAAG,MAAM,IAAI,CAAC,EAAE,CAAC,CAAC,CAAC;AAAA,IACnD,MAAO;AAAA,EACT;AAEA,SAAO,MAAM,SAAS,KAAK,MAAM,CAAC,EAAE,CAAC,KAAK,KAAK,MAAM,CAAC,EAAE,CAAC,IAAI,MAAM,CAAC,EAAE,CAAC,IAAI,QAAS,OAAM,MAAM;AAChG,SAAO,MAAM,SAAS,KAAK,MAAM,MAAM,SAAS,CAAC,EAAE,CAAC,KAAK,IAAI,KAAK,MAAM,MAAM,SAAS,CAAC,EAAE,CAAC,IAAI,MAAM,MAAM,SAAS,CAAC,EAAE,CAAC,IAAI,QAAS,OAAM,IAAI;AAE/I,aAAS;AACP,QAAI,MAAM,IAAI,OAAO;AACrB,aAAS,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK;AACrC,YAAM,KAAK,MAAM,CAAC,EAAE,CAAC,IAAI,MAAM,CAAC,EAAE,CAAC;AACnC,UAAI,KAAK,WAAW,KAAK,MAAM;AAAE,eAAO;AAAI,cAAM;AAAA,MAAE;AAAA,IACtD;AACA,QAAI,MAAM,KAAK,MAAM,WAAW,EAAG;AACnC,UAAM,UAAU,MAAM,IAAI,MAAM,GAAG,EAAE,CAAC,IAAI,MAAM,MAAM,CAAC,EAAE,CAAC,IAAI;AAC9D,UAAM,UAAU,MAAM,MAAM,SAAS,IAAI,MAAM,MAAM,CAAC,EAAE,CAAC,IAAI,MAAM,GAAG,EAAE,CAAC,IAAI;AAC7E,UAAM,IAAI,WAAW,UAAU,MAAM,IAAI,MAAM;AAC/C,UAAM,IAAI,KAAK,IAAI,KAAK,CAAC,GAAG,IAAI,KAAK,IAAI,KAAK,CAAC;AAC/C,YAAQ,CAAC,GAAG,MAAM,MAAM,GAAG,CAAC,GAAG,CAAC,MAAM,CAAC,EAAE,CAAC,GAAG,MAAM,CAAC,EAAE,CAAC,CAAC,GAAG,GAAG,MAAM,MAAM,IAAI,CAAC,CAAC;AAAA,EAClF;AACA,MAAI,MAAM,WAAW,EAAG,QAAO,CAAC,EAAE,IAAI,MAAM,CAAC,EAAE,CAAC,GAAG,IAAI,MAAM,CAAC,EAAE,CAAC,GAAG,IAAI,GAAG,IAAI,EAAE,CAAC;AAClF,SAAO,MAAM,IAAI,CAAC,CAAC,IAAI,EAAE,MAAM;AAAE,UAAM,CAAC,IAAI,EAAE,IAAI,QAAQ,IAAI,EAAE;AAAG,WAAO,EAAE,IAAI,IAAI,IAAI,GAAG;AAAA,EAAE,CAAC;AAChG;AASO,SAAS,UACd,MACA,GACA,GACA,KACoD;AACpD,QAAM,QAAQ,CAAC,MAAe,IAAI,UAAU,KAAK,IAAI,YAAY,IAAI,IAAI;AACzE,QAAM,SAAS,IAAI,YAAY,CAAC;AAChC,QAAM,SAAS,IAAI,YAAY,CAAC;AAChC,WAAS,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,UAAM,MAAM,IAAI;AAChB,aAAS,IAAI,GAAG,IAAI,GAAG,IAAK,KAAI,MAAM,KAAK,MAAM,CAAC,CAAC,GAAG;AAAE,aAAO,CAAC;AAAK,aAAO,CAAC;AAAA,IAAI;AAAA,EACnF;AACA,QAAM,UAAU,IAAI,WAAW,CAAC,GAAG,UAAU,IAAI,WAAW,CAAC;AAC7D,WAAS,IAAI,GAAG,IAAI,GAAG,IAAK,KAAI,OAAO,CAAC,KAAK,IAAI,KAAM,SAAQ,CAAC,IAAI;AACpE,WAAS,IAAI,GAAG,IAAI,GAAG,IAAK,KAAI,OAAO,CAAC,KAAK,IAAI,KAAM,SAAQ,CAAC,IAAI;AACpE,MAAI,KAAK,GAAG,KAAK,IAAI,KAAK,GAAG,KAAK;AAClC,WAAS,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,QAAI,QAAQ,CAAC,EAAG;AAChB,UAAM,MAAM,IAAI;AAChB,aAAS,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,UAAI,QAAQ,CAAC,KAAK,CAAC,MAAM,KAAK,MAAM,CAAC,CAAC,EAAG;AACzC,UAAI,IAAI,GAAI,MAAK;AACjB,UAAI,IAAI,GAAI,MAAK;AACjB,UAAI,IAAI,GAAI,MAAK;AACjB,UAAI,IAAI,GAAI,MAAK;AAAA,IACnB;AAAA,EACF;AACA,MAAI,KAAK,MAAM,KAAK,GAAI,QAAO,EAAE,IAAI,GAAG,IAAI,GAAG,IAAI,GAAG,IAAI,EAAE;AAC5D,SAAO,EAAE,IAAI,IAAI,IAAI,KAAK,GAAG,IAAI,KAAK,EAAE;AAC1C;;;ACzLA,IAAM,eAAe;AAGd,SAAS,eAAe,MAAsB;AACnD,MAAI,IAAI;AAER,MAAI,EAAE,QAAQ,mCAAmC,UAAU;AAE3D,MAAI,EAAE,QAAQ,mCAAmC,CAAC,GAAG,MACnD,mBAAmB,KAAK,EAAE,MAAM,KAAK,IAAI,GAAG,IAAI,CAAC,GAAG,CAAC,EAAE,QAAQ,CAAC,KAAK,sBAAsB,KAAK,EAAE,MAAM,IAAI,EAAE,MAAM,CAAC,IACjH,SAAS,OAAO,EAAE,MAAM,IAAI,CAAC;AAEnC,MAAI,EAAE,QAAQ,mBAAmB,QAAQ;AACzC,SAAO,YAAY,gBAAgB,CAAC,CAAC;AACvC;AAOO,SAAS,gBAAgB,GAAmB;AACjD,SAAO,EAAE,QAAQ,2CAA2C,OAAK,EAAE,QAAQ,OAAO,GAAG,CAAC;AACxF;AAQO,SAAS,YAAY,GAAmB;AAC7C,MAAI,CAAC,OAAO,KAAK,CAAC,EAAG,QAAO;AAC5B,QAAM,QAAQ,CAAC,GAAG,CAAC;AACnB,aAAW,KAAK,CAAC,KAAK,GAAG,GAAG;AAC1B,UAAM,CAAC,MAAM,KAAK,IAAI,MAAM,MAAM,CAAC,UAAU,QAAQ,IAAI,CAAC,UAAU,QAAQ;AAC5E,UAAM,MAAgB,CAAC;AACvB,aAAS,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK;AACrC,UAAI,MAAM,CAAC,MAAM,EAAG;AACpB,YAAMC,QAAO,IAAI,IAAI,MAAM,IAAI,CAAC,IAAI;AACpC,UAAI,MAAM,OAAO,CAAC,gBAAgB,KAAKA,KAAI,KAAK,eAAe,KAAK,MAAM,MAAM,IAAI,GAAG,IAAI,CAAC,EAAE,KAAK,EAAE,CAAC,GAAG;AACvG,cAAM,CAAC,IAAI;AACX;AAAA,MACF;AACA,UAAI,KAAK,CAAC;AAAA,IACZ;AACA,QAAI,IAAI,WAAW,EAAG;AACtB,UAAM,QAAQ,IAAI,CAAC;AACnB,UAAM,OAAO,QAAQ,IAAI,MAAM,QAAQ,CAAC,IAAI;AAC5C,UAAM,SAAS,SAAS,MAAM,yDAAyD,KAAK,IAAI;AAChG,QAAI,SAAS,IAAI,SAAS,MAAM,KAAK;AACrC,eAAW,KAAK,KAAK;AAAE,YAAM,CAAC,IAAI,SAAS,OAAO;AAAO,eAAS,CAAC;AAAA,IAAO;AAAA,EAC5E;AACA,SAAO,MAAM,KAAK,EAAE;AACtB;AAGO,SAAS,cAAc,MAAuB;AACnD,SAAO,2DAA2D,KAAK,IAAI;AAC7E;AAMO,SAAS,mBAA2F,OAAkB;AAC3H,aAAW,MAAM,OAAO;AACtB,QAAI,GAAG,SAAS,OAAO,GAAG,SAAS,IAAK;AACxC,UAAM,KAAK,GAAG,IAAI,GAAG,IAAI;AACzB,QAAI,OAAiB;AACrB,eAAW,KAAK,OAAO;AACrB,UAAI,MAAM,MAAM,EAAE,IAAI,GAAG,IAAI,GAAG,IAAI,IAAK;AACzC,UAAI,KAAK,IAAI,EAAE,IAAI,EAAE,IAAI,IAAI,EAAE,IAAI,KAAK,IAAI,GAAG,GAAG,EAAE,CAAC,IAAI,EAAG;AAC5D,UAAI,EAAE,KAAK,GAAG,IAAI,GAAG,KAAK,GAAG,IAAI,EAAG;AACpC,UAAI,CAAC,QAAQ,EAAE,IAAI,KAAK,EAAG,QAAO;AAAA,IACpC;AACA,QAAI,QAAQ,aAAa,KAAK,KAAK,KAAK,UAAU,CAAC,EAAG,IAAG,OAAO;AAAA,EAClE;AACF;;;ACnFO,IAAM,aAAa;AAC1B,IAAM,gBAAgB;AACtB,IAAM,WAAW;AAOV,SAAS,UAAU,GAAQ,OAAkE,OAAsB;AACxH,QAAM,MAAa,CAAC;AACpB,WAAS,IAAI,GAAG,IAAI,MAAM,QAAQ,KAAK;AACrC,UAAM,EAAE,IAAI,IAAI,IAAI,GAAG,IAAI,MAAM,CAAC;AAClC,UAAM,KAAK,KAAK;AAChB,UAAM,QAAQ,IAAI,KAAK,KAAK,MAAM,IAAI,CAAC,EAAE,MAAM,IAAI;AACnD,UAAM,UAAU,IAAI,MAAM,SAAS,KAAK,MAAM,IAAI,CAAC,EAAE,KAAK,MAAM,IAAI,EAAE,IAAI;AAC1E,UAAM,MAAM,KAAK,IAAI,GAAG,KAAK,MAAM,EAAE,IAAI,KAAK,KAAK,IAAI,KAAK,KAAK,KAAK,CAAC,CAAC;AACxE,UAAM,MAAM,KAAK,IAAI,OAAO,KAAK,MAAM,EAAE,IAAI,KAAK,KAAK,IAAI,KAAK,KAAK,OAAO,CAAC,CAAC;AAC9E,UAAM,OAAO,KAAK,IAAI,EAAE,GAAG,KAAK,MAAM,EAAE,IAAI,KAAK,KAAK,GAAG,CAAC;AAC1D,UAAM,QAAQ,KAAK,IAAI,EAAE,IAAI,EAAE,GAAG,KAAK,MAAM,EAAE,IAAI,KAAK,KAAK,GAAG,CAAC;AACjE,QAAI,MAAM,OAAO,YAAY,QAAQ,QAAQ,SAAU,KAAI,KAAK,EAAE,GAAG,MAAM,GAAG,KAAK,GAAG,QAAQ,MAAM,GAAG,MAAM,IAAI,CAAC;AAAA,EACpH;AACA,SAAO;AACT;AAOO,SAAS,SAAS,MAAkB,OAAe,KAAU,KAAmD;AACrH,QAAM,OAAO,QAAQ,IAAI,IAAI,IAAI,IAAI;AACrC,QAAM,OAAO,QAAQ,IAAI,IAAI,IAAI,IAAI;AACrC,QAAM,KAAK,KAAK,IAAI,eAAe,KAAK,IAAI,IAAI,KAAK,MAAO,OAAO,aAAc,IAAI,CAAC,CAAC;AACvF,QAAM,MAAM,IAAI,WAAW,KAAK,aAAa,CAAC;AAE9C,QAAM,KAAK,CAAC,GAAW,MAAsB;AAC3C,QAAI,IAAY;AAChB,QAAI,QAAQ,GAAG;AAAE,WAAK;AAAG,WAAK;AAAA,IAAE,WACvB,QAAQ,IAAI;AAAE,WAAK,IAAI,IAAI,IAAI;AAAG,WAAK;AAAA,IAAE,OAC7C;AAAE,WAAK;AAAG,WAAK,IAAI,IAAI,IAAI;AAAA,IAAE;AAClC,aAAS,IAAI,IAAI,MAAM,SAAS,IAAI,IAAI,OAAO;AAAA,EACjD;AACA,QAAM,KAAK,OAAO;AAClB,QAAM,KAAK,OAAO;AAClB,WAAS,KAAK,GAAG,KAAK,YAAY,MAAM;AACtC,QAAI,MAAM,KAAK,OAAO,KAAK;AAC3B,QAAI,KAAK,EAAG,MAAK;AACjB,UAAM,KAAK,KAAK,IAAI,OAAO,GAAG,KAAK,MAAM,EAAE,CAAC;AAC5C,UAAM,KAAK,KAAK,IAAI,OAAO,GAAG,KAAK,CAAC;AACpC,UAAM,KAAK,KAAK;AAChB,aAAS,KAAK,GAAG,KAAK,IAAI,MAAM;AAC9B,UAAI,MAAM,KAAK,OAAO,KAAK;AAC3B,UAAI,KAAK,EAAG,MAAK;AACjB,YAAM,KAAK,KAAK,IAAI,OAAO,GAAG,KAAK,MAAM,EAAE,CAAC;AAC5C,YAAM,KAAK,KAAK,IAAI,OAAO,GAAG,KAAK,CAAC;AACpC,YAAM,KAAK,KAAK;AAChB,YAAM,MAAM,GAAG,IAAI,EAAE,GAAG,MAAM,GAAG,IAAI,EAAE,GAAG,MAAM,GAAG,IAAI,EAAE,GAAG,MAAM,GAAG,IAAI,EAAE;AAC3E,YAAM,KAAK,KAAK,KAAK,MAAM;AAC3B,eAAS,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,cAAM,MAAM,KAAK,MAAM,CAAC,KAAK,IAAI,MAAM,KAAK,MAAM,CAAC,IAAI;AACvD,cAAM,SAAS,KAAK,MAAM,CAAC,KAAK,IAAI,MAAM,KAAK,MAAM,CAAC,IAAI;AAC1D,YAAI,IAAI,CAAC,IAAI,KAAK,MAAM,OAAO,IAAI,MAAM,SAAS,EAAE;AAAA,MACtD;AAAA,IACF;AAAA,EACF;AACA,SAAO,EAAE,KAAK,GAAG,GAAG;AACtB;;;AJEO,IAAM,qBAA0C,OAAO,OAAO;AAAA,EACnE,aAAa;AAAA,EACb,WAAW;AAAA,EACX,cAAc;AAAA,EACd,WAAW;AAAA,EACX,WAAW;AAAA,EACX,UAAU;AAAA,EACV,WAAW;AAAA,EACX,gBAAgB;AAAA,EAChB,aAAa;AAAA,EACb,YAAY;AACd,CAAC;AAED,IAAM,eAAe;AACrB,IAAM,gBAAgB;AACtB,IAAM,gBAAgB;AAEtB,IAAM,uBAAuB,KAAK;AAElC,IAAM,aAAa;AAEnB,IAAM,eAAe;AAGrB,IAAM,WAAW,CAAC,OAAO,OAAO,KAAK;AACrC,IAAM,UAAU,CAAC,OAAO,OAAO,KAAK;AAe7B,IAAM,YAAN,MAAM,WAAU;AAAA,EACb;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EAEA,YAAY,OAMjB;AACD,SAAK,MAAM,MAAM;AACjB,SAAK,MAAM,MAAM;AACjB,SAAK,OAAO,MAAM;AAClB,SAAK,MAAM,MAAM;AACjB,SAAK,QAAQ,MAAM;AAAA,EACrB;AAAA,EAEA,aAAa,SAA6B;AACxC,UAAM,CAAC,QAAQ,WAAW,IAAI,MAAM,QAAQ,IAAI;AAAA,MAC9C,UAA6C,oBAAoB,MAAM,OAAO,kBAAkB,CAAC;AAAA,MACjG;AAAA,QACE;AAAA,QACA,MAAM,OAAO,OAAO;AAAA,MACtB;AAAA,IACF,CAAC;AACD,UAAM,WAAW;AACjB,UAAM,WACJ,OAAO,aAAa,aAAa,WAAY,SAAS,WAAY;AAEpE,UAAM,MAAM,gBAAgB;AAC5B,UAAM,cAA0E;AAAA,MAC9E,wBAAwB;AAAA,MACxB,oBAAoB,CAAC,KAAK;AAAA,MAC1B,kBAAkB;AAAA;AAAA,IACpB;AACA,UAAM,CAAC,KAAK,KAAK,OAAO,IAAI,MAAM,QAAQ,IAAI;AAAA,MAC5C,OAAO,iBAAiB,OAAO,KAAK,KAAK,cAAc,QAAQ,GAAG,WAAW;AAAA,MAC7E,OAAO,iBAAiB,OAAO,KAAK,KAAK,cAAc,QAAQ,GAAG,WAAW;AAAA,MAC7E,SAAS,KAAK,KAAK,aAAa,QAAQ,GAAG,OAAO;AAAA,IACpD,CAAC;AACD,UAAM,OAAO,mBAAmB,OAAO;AACvC,QAAI,KAAK,WAAW,EAAG,OAAM,IAAI,MAAM,wJAAqC;AAE5E,WAAO,IAAI,WAAU,EAAE,KAAK,KAAK,MAAM,KAAK,QAAQ,OAAO,SAAS,CAAC;AAAA,EACvE;AAAA;AAAA,EAGA,MAAM,UAAyB;AAC7B,eAAW,KAAK,CAAC,KAAK,KAAK,KAAK,GAAG,GAAG;AACpC,YAAM,MAAO,EAAmD;AAChE,UAAI,OAAO,QAAQ,YAAY;AAC7B,YAAI;AAAE,gBAAM,IAAI,KAAK,CAAC;AAAA,QAAE,QAAQ;AAAA,QAAe;AAAA,MACjD;AAAA,IACF;AAAA,EACF;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAOA,MAAM,cACJ,MACA,OACA,QACA,OACA,SAA8B,oBACV;AACpB,QAAI,QAAQ,gBAAgB,SAAS,aAAc,QAAO,CAAC;AAC3D,UAAM,QAAQ,MAAM,KAAK,OAAO,MAAM,OAAO,QAAQ,MAAM;AAG3D,UAAM,OAAkB,CAAC;AACzB,QAAI,QAAQ;AACZ,eAAW,KAAK,OAAO;AACrB,YAAM,OAAO,SAAS,MAAM,OAAO,CAAC;AACpC,YAAM,MAAM,SAAS,IAAI;AACzB,UAAI,OAAO,iBAAiB,KAAK,IAAI,WAAW,OAAO,eAAgB;AACvE,UAAI,OAAO,aAAa,EAAE,KAAK,EAAE,IAAI,YAAY;AAC/C,cAAM,QAAQ,cAAc,MAAM,EAAE,GAAG,EAAE,GAAG,KAAK,IAAI;AACrD,YAAI,MAAM,UAAU,GAAG;AACrB,qBAAW,OAAO,UAAU,GAAG,OAAO,MAAM,EAAG,MAAK,KAAK,EAAE,KAAK,KAAK,KAAK,GAAG,OAAO,QAAQ,CAAC;AAC7F;AAAA,QACF;AACA,YAAI,EAAE,KAAK,EAAE,IAAI,cAAc;AAC7B,qBAAW,OAAO,CAAC,GAAG,IAAI,GAAG,EAAY,MAAK,KAAK,EAAE,KAAK,GAAG,KAAK,MAAM,CAAC;AACzE;AACA;AAAA,QACF;AAAA,MACF;AACA,WAAK,KAAK,EAAE,KAAK,GAAG,KAAK,GAAG,OAAO,QAAQ,CAAC;AAAA,IAC9C;AAEA,UAAM,UAAU,MAAM,KAAK,cAAc,MAAM,OAAO,MAAM,OAAO,QAAQ;AAG3E,UAAM,OAAO,oBAAI,IAAgE;AACjF,SAAK,QAAQ,CAAC,KAAK,MAAM;AACvB,YAAM,IAAI,QAAQ,CAAC;AACnB,UAAI,CAAC,EAAG;AACR,YAAM,MAAM,KAAK,IAAI,IAAI,KAAK;AAC9B,UAAI,CAAC,OAAO,EAAE,aAAa,IAAI,WAAY,MAAK,IAAI,IAAI,OAAO,EAAE,KAAK,GAAG,EAAE,CAAC;AAAA,IAC9E,CAAC;AAED,UAAM,QAAmB,CAAC;AAC1B,eAAW,EAAE,KAAK,MAAM,KAAK,WAAW,KAAK,KAAK,OAAO,GAAG;AAC1D,YAAM,OAAO,OAAO,cAAc,eAAe,IAAI,KAAK,CAAC,IAAI;AAC/D,UAAI,CAAC,KAAK,KAAK,EAAG;AAClB,UAAI,OAAO,eAAe,cAAc,IAAI,EAAG;AAC/C,UAAI,aAAa,OAAO,WAAW;AAAE,YAAI,MAAO,OAAM;AAAkB;AAAA,MAAS;AACjF,UAAI,IAAI,IAAI;AACZ,UAAI,OAAO,YAAY;AACrB,cAAM,OAAO,SAAS,MAAM,OAAO,CAAC;AACpC,cAAM,IAAI,UAAU,MAAM,EAAE,GAAG,EAAE,GAAG,SAAS,IAAI,CAAC;AAClD,YAAI,EAAE,GAAG,EAAE,IAAI,EAAE,IAAI,GAAG,EAAE,IAAI,EAAE,IAAI,GAAG,EAAE,KAAK,EAAE,IAAI,GAAG,EAAE,KAAK,EAAE,GAAG;AAAA,MACrE;AACA,YAAM,KAAK,EAAE,MAAM,GAAG,EAAE,GAAG,GAAG,EAAE,GAAG,GAAG,EAAE,GAAG,GAAG,EAAE,GAAG,WAAW,CAAC;AAAA,IACjE;AACA,QAAI,OAAO,YAAa,oBAAmB,KAAK;AAChD,UAAM,KAAK,CAAC,GAAG,MAAO,EAAE,IAAI,EAAE,KAAO,EAAE,IAAI,EAAE,CAAE;AAC/C,WAAO;AAAA,EACT;AAAA;AAAA,EAIA,MAAc,OACZ,MACA,OACA,QACA,QACgB;AAChB,UAAM,QAAQ,OAAO,cAAc,KAAK,IAAI,OAAO,MAAM;AACzD,UAAM,KAAK,KAAK,IAAI,IAAI,KAAK,MAAO,QAAQ,QAAS,EAAE,IAAI,EAAE;AAC7D,UAAM,KAAK,KAAK,IAAI,IAAI,KAAK,MAAO,SAAS,QAAS,EAAE,IAAI,EAAE;AAE9D,UAAM,MAAM,MAAM,KAAK,MAAM,MAAM,EAAE,KAAK,EAAE,OAAO,QAAQ,UAAU,EAAE,EAAE,CAAC,EACvE,OAAO,IAAI,IAAI,EAAE,KAAK,OAAO,CAAC,EAC9B,YAAY,EACZ,IAAI,EACJ,SAAS;AAGZ,UAAM,QAAQ,KAAK;AACnB,UAAM,QAAQ,IAAI,aAAa,IAAI,KAAK;AACxC,aAAS,IAAI,GAAG,IAAI,OAAO,KAAK;AAC9B,YAAM,IAAI,IAAI,IAAI,CAAC,IAAI;AACvB,YAAM,IAAI,IAAI,IAAI,IAAI,CAAC,IAAI;AAC3B,YAAM,IAAI,IAAI,IAAI,IAAI,CAAC,IAAI;AAC3B,YAAM,CAAC,KAAK,IAAI,SAAS,CAAC,KAAK,QAAQ,CAAC;AACxC,YAAM,QAAQ,CAAC,KAAK,IAAI,SAAS,CAAC,KAAK,QAAQ,CAAC;AAChD,YAAM,IAAI,QAAQ,CAAC,KAAK,IAAI,SAAS,CAAC,KAAK,QAAQ,CAAC;AAAA,IACtD;AAEA,UAAM,SAAS,IAAI,KAAK,IAAI,OAAO,WAAW,OAAO,CAAC,GAAG,GAAG,IAAI,EAAE,CAAC;AACnE,UAAM,MAAM,MAAM,KAAK,IAAI,IAAI,EAAE,CAAC,KAAK,IAAI,WAAW,CAAC,CAAC,GAAG,OAAO,CAAC;AACnE,UAAM,UAAU,IAAI,KAAK,IAAI,YAAY,CAAC,CAAC,EAAE;AAE7C,UAAM,WAAW,eAAe,SAAS,IAAI,IAAI,OAAO,WAAW,OAAO,YAAY;AACtF,UAAM,KAAK,QAAQ;AACnB,UAAM,KAAK,SAAS;AACpB,UAAM,QAAe,CAAC;AACtB,eAAW,MAAM,SAAS,MAAM,GAAG,aAAa,GAAG;AAEjD,YAAM,KAAK,GAAG,KAAK,GAAG,KAAK;AAC3B,YAAM,KAAK,GAAG,KAAK,GAAG,KAAK;AAC3B,YAAM,QAAS,KAAK,KAAK,OAAO,aAAc,KAAK,KAAK;AACxD,YAAM,KAAK,KAAK,IAAI,GAAG,KAAK,OAAO,GAAG,KAAK,SAAS,EAAE,CAAC;AACvD,YAAM,KAAK,KAAK,IAAI,GAAG,KAAK,OAAO,GAAG,KAAK,SAAS,EAAE,CAAC;AACvD,YAAM,KAAK,KAAK,IAAI,OAAO,KAAK,MAAM,GAAG,KAAK,IAAI,SAAS,EAAE,CAAC;AAC9D,YAAM,KAAK,KAAK,IAAI,QAAQ,KAAK,MAAM,GAAG,KAAK,IAAI,SAAS,EAAE,CAAC;AAC/D,UAAI,KAAK,KAAK,gBAAgB,KAAK,KAAK,aAAc;AACtD,YAAM,KAAK,EAAE,GAAG,IAAI,GAAG,IAAI,GAAG,KAAK,IAAI,GAAG,KAAK,GAAG,CAAC;AAAA,IACrD;AACA,WAAO;AAAA,EACT;AAAA;AAAA;AAAA,EAKA,MAAc,cACZ,MACA,OACA,MACA,WAC6D;AAC7D,UAAM,QAAQ,KAAK,IAAI,OAAK,SAAS,MAAM,OAAO,EAAE,KAAK,EAAE,GAAG,CAAC;AAC/D,UAAM,QAAQ,MAAM,IAAI,CAAC,GAAG,MAAM,CAAC,EAAE,KAAK,CAAC,GAAG,MAAM,MAAM,CAAC,EAAE,IAAI,MAAM,CAAC,EAAE,CAAC;AAC3E,UAAM,UAA8D,IAAI,MAAM,KAAK,MAAM,EAAE,KAAK,IAAI;AACpG,UAAM,QAAQ;AACd,aAAS,IAAI,GAAG,IAAI,MAAM,UAAS;AAEjC,UAAI,IAAI,IAAI;AACZ,aAAO,IAAI,MAAM,UAAU,IAAI,IAAI,KAAK,IAAI,GAAG,SAAS,MAClD,IAAI,IAAI,KAAK,KAAK,IAAI,eAAe,MAAM,MAAM,CAAC,CAAC,EAAE,CAAC,IAAI,SAAS,qBAAsB;AAC/F,YAAM,MAAM,MAAM,MAAM,GAAG,CAAC;AAC5B,YAAM,KAAK,KAAK,IAAI,eAAe,MAAM,IAAI,IAAI,SAAS,CAAC,CAAC,EAAE,CAAC;AAC/D,YAAM,IAAI,IAAI;AACd,YAAM,MAAM,IAAI,aAAa;AAC7B,YAAM,QAAQ,IAAI,aAAa,IAAI,GAAG;AACtC,UAAI,QAAQ,CAAC,IAAI,MAAM;AACrB,cAAM,IAAI,MAAM,EAAE;AAClB,cAAM,OAAO,IAAI;AACjB,cAAM,IAAI,aAAa;AACvB,iBAAS,IAAI,GAAG,IAAI,YAAY,KAAK;AACnC,mBAAS,IAAI,GAAG,IAAI,EAAE,GAAG,KAAK;AAC5B,kBAAM,OAAO,IAAI,EAAE,IAAI,KAAK;AAC5B,kBAAM,MAAM,OAAO,IAAI,KAAK;AAE5B,kBAAM,GAAG,IAAI,EAAE,IAAI,MAAM,CAAC,IAAI,QAAQ;AACtC,kBAAM,MAAM,CAAC,IAAI,EAAE,IAAI,MAAM,CAAC,IAAI,QAAQ;AAC1C,kBAAM,MAAM,IAAI,CAAC,IAAI,EAAE,IAAI,GAAG,IAAI,QAAQ;AAAA,UAC5C;AAAA,QACF;AAAA,MACF,CAAC;AACD,YAAM,SAAS,IAAI,KAAK,IAAI,OAAO,WAAW,OAAO,CAAC,GAAG,GAAG,YAAY,EAAE,CAAC;AAC3E,YAAM,MAAM,MAAM,KAAK,IAAI,IAAI,EAAE,CAAC,KAAK,IAAI,WAAW,CAAC,CAAC,GAAG,OAAO,CAAC;AACnE,YAAM,SAAS,IAAI,KAAK,IAAI,YAAY,CAAC,CAAC;AAC1C,YAAM,CAAC,EAAE,GAAG,CAAC,IAAI,OAAO;AACxB,YAAM,OAAO,OAAO;AACpB,UAAI,QAAQ,CAAC,IAAI,MAAM;AACrB,gBAAQ,EAAE,IAAI,UAAU,KAAK,SAAS,IAAI,IAAI,IAAI,IAAI,KAAK,IAAI,CAAC,GAAG,GAAG,GAAG,KAAK,IAAI;AAAA,MACpF,CAAC;AACD,UAAI;AAAA,IACN;AACA,WAAO;AAAA,EACT;AACF;AAGO,SAAS,UACd,MACA,GACA,GACA,MAC6C;AAC7C,MAAI,OAAO;AACX,MAAI,UAAU;AACd,MAAI,YAAY;AAChB,MAAI,OAAO;AACX,WAAS,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,UAAM,MAAM,IAAI;AAChB,QAAI,OAAO;AACX,QAAI,QAAQ,KAAK,GAAG;AACpB,aAAS,IAAI,GAAG,IAAI,GAAG,KAAK;AAC1B,YAAM,IAAI,KAAK,MAAM,CAAC;AACtB,UAAI,IAAI,OAAO;AAAE,gBAAQ;AAAG,eAAO;AAAA,MAAE;AAAA,IACvC;AACA,UAAM,SAAS,SAAS;AACxB,WAAO;AACP,QAAI,SAAS,KAAK,OAAQ;AAE1B,QAAI,IAAI;AACR,QAAI,IAAI,UAAU,IAAI,GAAG;AACvB,UAAI,QAAQ;AACZ,eAAS,IAAI,GAAG,IAAI,GAAG,IAAK,UAAS,KAAK,IAAI,KAAK,MAAM,CAAC,IAAI,KAAK;AACnE,UAAI,IAAI;AAAA,IACV;AACA,eAAW;AACX;AACA,QAAI,QAAQ,KAAK,QAAQ,KAAK,OAAQ,SAAQ,KAAK,OAAO,CAAC;AAAA,aAClD,SAAS,KAAK,SAAS,EAAG,SAAQ;AAAA,EAC7C;AACA,MAAI,CAAC,KAAM,QAAO;AAClB,SAAO,EAAE,MAAM,YAAY,YAAY,IAAI,UAAU,YAAY,EAAE;AACrE;AAGO,SAAS,eACd,MACA,GACA,GACA,SAAiB,mBAAmB,WACpC,YAAoB,mBAAmB,cACoB;AAC3D,QAAM,UAAU,IAAI,WAAW,IAAI,CAAC;AACpC,QAAM,QAAkF,CAAC;AACzF,QAAM,QAAkB,CAAC;AAEzB,WAAS,QAAQ,GAAG,QAAQ,IAAI,GAAG,SAAS;AAC1C,QAAI,QAAQ,KAAK,KAAK,KAAK,KAAK,KAAK,OAAQ;AAC7C,QAAI,KAAK,QAAQ,GAAG,KAAK,IAAI,KAAM,QAAQ,IAAK,GAAG,KAAK;AACxD,QAAI,MAAM;AACV,QAAI,QAAQ;AACZ,UAAM,SAAS;AACf,UAAM,KAAK,KAAK;AAChB,YAAQ,KAAK,IAAI;AACjB,WAAO,MAAM,QAAQ;AACnB,YAAM,IAAI,MAAM,IAAI;AACpB,YAAM,KAAK,IAAI;AACf,YAAM,KAAM,IAAI,IAAK;AACrB,aAAO,KAAK,CAAC;AACb;AACA,UAAI,KAAK,GAAI,MAAK;AAClB,UAAI,KAAK,GAAI,MAAK;AAClB,UAAI,KAAK,GAAI,MAAK;AAClB,UAAI,KAAK,GAAI,MAAK;AAElB,UAAI,KAAK,KAAK,CAAC,QAAQ,IAAI,CAAC,KAAK,KAAK,IAAI,CAAC,IAAI,QAAQ;AAAE,gBAAQ,IAAI,CAAC,IAAI;AAAG,cAAM,KAAK,IAAI,CAAC;AAAA,MAAE;AAC/F,UAAI,KAAK,IAAI,KAAK,CAAC,QAAQ,IAAI,CAAC,KAAK,KAAK,IAAI,CAAC,IAAI,QAAQ;AAAE,gBAAQ,IAAI,CAAC,IAAI;AAAG,cAAM,KAAK,IAAI,CAAC;AAAA,MAAE;AACnG,UAAI,KAAK,KAAK,CAAC,QAAQ,IAAI,CAAC,KAAK,KAAK,IAAI,CAAC,IAAI,QAAQ;AAAE,gBAAQ,IAAI,CAAC,IAAI;AAAG,cAAM,KAAK,IAAI,CAAC;AAAA,MAAE;AAC/F,UAAI,KAAK,IAAI,KAAK,CAAC,QAAQ,IAAI,CAAC,KAAK,KAAK,IAAI,CAAC,IAAI,QAAQ;AAAE,gBAAQ,IAAI,CAAC,IAAI;AAAG,cAAM,KAAK,IAAI,CAAC;AAAA,MAAE;AAAA,IACrG;AACA,QAAI,KAAK,KAAK,IAAI,gBAAgB,KAAK,KAAK,IAAI,aAAc;AAC9D,UAAM,KAAK,EAAE,IAAI,IAAI,IAAI,IAAI,OAAO,MAAM,MAAM,CAAC;AAAA,EACnD;AAEA,SAAO,MACJ,OAAO,OAAK,EAAE,SAAS,SAAS,EAChC,KAAK,CAAC,GAAG,MAAO,EAAE,KAAK,EAAE,MAAQ,EAAE,KAAK,EAAE,EAAG;AAClD;AAEA,eAAe,UAAa,MAAc,QAAsC;AAC9E,MAAI;AACF,WAAO,MAAM,OAAO;AAAA,EACtB,SAAS,GAAG;AACV,UAAM,IAAI;AAAA,MACR,+EAAuC,IAAI,0DACxB,IAAI,wEAAuB,EAAY,OAAO;AAAA,IACnE;AAAA,EACF;AACF;AAGA,IAAI,gBAA2C;AAExC,SAAS,eAAmC;AACjD,MAAI,CAAC,eAAe;AAClB,oBAAgB,UAAU,OAAO,EAAE,MAAM,SAAO;AAC9C,sBAAgB;AAChB,YAAM;AAAA,IACR,CAAC;AAAA,EACH;AACA,SAAO;AACT;;;AKxbA,IAAM,eAAe;AAErB,IAAM,gBAAgB;AAEtB,IAAM,WAAW;AACjB,IAAM,cAAc;AAUb,SAAS,aAAa,MAAkB,OAAe,QAA8B;AAC1F,QAAM,OAAO,KAAK,IAAI,GAAG,KAAK,KAAK,KAAK,IAAI,OAAO,MAAM,IAAI,WAAW,CAAC;AACzE,QAAM,KAAK,KAAK,MAAM,QAAQ,IAAI,GAAG,KAAK,KAAK,MAAM,SAAS,IAAI;AAElE,QAAM,KAAe,CAAC,GAAG,KAAe,CAAC;AACzC,WAAS,IAAI,GAAG,IAAI,IAAI,KAAK;AAC3B,aAAS,IAAI,GAAG,IAAI,IAAI,KAAK;AAC3B,YAAM,KAAM,IAAI,OAAQ,QAAQ,IAAI,QAAQ;AAC5C,YAAM,IAAK,KAAK,CAAC,IAAI,KAAK,KAAK,IAAI,CAAC,IAAI,MAAM,KAAK,IAAI,CAAC,IAAI,MAAO;AACnE,UAAI,IAAI,KAAK;AAAE,WAAG,KAAK,CAAC;AAAG,WAAG,KAAK,CAAC;AAAA,MAAE;AAAA,IACxC;AAAA,EACF;AACA,MAAI,GAAG,SAAS,IAAK,QAAO,EAAE,OAAO,GAAG,MAAM,EAAE;AAChD,QAAM,OAAO,KAAK,KAAK,KAAK,MAAM,IAAI,EAAE,CAAC;AACzC,QAAM,OAAO,IAAI,aAAa,IAAI,OAAO,CAAC;AAC1C,QAAM,QAAQ,CAAC,QAAwB;AACrC,UAAM,IAAK,MAAM,KAAK,KAAM;AAC5B,UAAM,IAAI,KAAK,IAAI,CAAC,GAAG,IAAI,KAAK,IAAI,CAAC;AACrC,SAAK,KAAK,CAAC;AACX,aAAS,IAAI,GAAG,IAAI,GAAG,QAAQ,IAAK,MAAK,KAAK,MAAM,GAAG,CAAC,IAAI,IAAI,GAAG,CAAC,IAAI,CAAC,IAAI,IAAI;AACjF,QAAI,MAAM;AACV,aAAS,IAAI,GAAG,IAAI,KAAK,QAAQ,IAAK,QAAO,KAAK,CAAC,IAAI,KAAK,CAAC;AAC7D,WAAO;AAAA,EACT;AACA,QAAM,KAAK,MAAM,CAAC;AAClB,MAAI,OAAO,GAAG,QAAQ;AACtB,WAAS,IAAI,CAAC,cAAc,KAAK,eAAe,MAAM,KAAK,KAAK;AAC9D,UAAM,IAAI,MAAM,CAAC;AACjB,QAAI,IAAI,OAAO;AAAE,cAAQ;AAAG,aAAO;AAAA,IAAE;AAAA,EACvC;AACA,QAAM,SAAS;AACf,WAAS,IAAI,SAAS,KAAK,KAAK,SAAS,MAAM,MAAM,KAAK,MAAM;AAC9D,UAAM,IAAI,MAAM,CAAC;AACjB,QAAI,IAAI,OAAO;AAAE,cAAQ;AAAG,aAAO;AAAA,IAAE;AAAA,EACvC;AAEA,SAAO,EAAE,OAAO,EAAE,CAAC,MAAM,QAAQ,CAAC,GAAG,MAAM,KAAK,IAAI,QAAQ,KAAK,EAAE;AACrE;AAMO,SAAS,WAAW,MAAkB,OAAe,QAAqD;AAC/G,QAAM,MAAM,aAAa,MAAM,OAAO,MAAM;AAC5C,MAAI,KAAK,IAAI,IAAI,KAAK,IAAI,iBAAiB,IAAI,OAAO,SAAU,QAAO,EAAE,MAAM,OAAO,EAAE;AACxF,SAAO,EAAE,MAAM,WAAW,MAAM,OAAO,QAAQ,IAAI,KAAK,GAAG,OAAO,IAAI,MAAM;AAC9E;AAMO,SAAS,WAAW,MAAkB,OAAe,QAAgB,KAAyB;AACnG,QAAM,MAAM,IAAI,WAAW,KAAK,MAAM;AACtC,QAAM,IAAK,MAAM,KAAK,KAAM;AAC5B,QAAM,IAAI,KAAK,IAAI,CAAC,GAAG,IAAI,KAAK,IAAI,CAAC;AACrC,QAAM,MAAM,QAAQ,KAAK,GAAG,MAAM,SAAS,KAAK;AAChD,WAAS,IAAI,GAAG,IAAI,QAAQ,KAAK;AAC/B,UAAM,KAAK,IAAI;AACf,aAAS,IAAI,GAAG,IAAI,OAAO,KAAK;AAC9B,YAAM,KAAK,IAAI;AAEf,YAAM,KAAK,IAAI,KAAK,IAAI,KAAK;AAC7B,YAAM,KAAK,IAAI,KAAK,IAAI,KAAK;AAC7B,YAAM,KAAK,IAAI,QAAQ,KAAK;AAC5B,YAAM,KAAK,KAAK,MAAM,EAAE,GAAG,KAAK,KAAK,MAAM,EAAE;AAC7C,UAAI,KAAK,KAAK,KAAK,KAAK,MAAM,QAAQ,KAAK,MAAM,SAAS,GAAG;AAC3D,YAAI,CAAC,IAAI;AAAK,YAAI,IAAI,CAAC,IAAI;AAAK,YAAI,IAAI,CAAC,IAAI;AAAK,YAAI,IAAI,CAAC,IAAI;AAC/D;AAAA,MACF;AACA,YAAM,KAAK,KAAK,IAAI,KAAK,KAAK;AAC9B,YAAM,OAAO,KAAK,QAAQ,MAAM,GAAG,MAAM,MAAM,GAAG,MAAM,MAAM,QAAQ,GAAG,MAAM,MAAM;AACrF,eAAS,KAAK,GAAG,KAAK,GAAG,MAAM;AAC7B,cAAM,MAAM,KAAK,MAAM,EAAE,KAAK,IAAI,MAAM,KAAK,MAAM,EAAE,IAAI;AACzD,cAAM,MAAM,KAAK,MAAM,EAAE,KAAK,IAAI,MAAM,KAAK,MAAM,EAAE,IAAI;AACzD,YAAI,IAAI,EAAE,IAAI,KAAK,MAAM,OAAO,IAAI,MAAM,MAAM,EAAE;AAAA,MACpD;AAAA,IACF;AAAA,EACF;AACA,SAAO;AACT;;;ACtFA,IAAM,mBAAmB;AAEzB,IAAM,kBAAkB;AASxB,eAAsB,UACpB,QACA,SACA,MACA,UACA,YACA,eAAe,MACkB;AACjC,QAAM,SAAS,oBAAI,IAAuB;AAC1C,MAAI,QAAQ,SAAS,EAAG,QAAO;AAG/B,QAAM,YAAY,MAAMC;AAAA,IACtB;AAAA,IACA,MAAM,OAAO,gBAAgB;AAAA,EAC/B;AACA,MAAI,SAAS,WAAW;AACtB,UAAM,gBAAgB,OAAK;AACzB,UAAI,EAAE,UAAU,cAAc,EAAE,eAAe,GAAG;AAChD,gBAAQ,OAAO,MAAM,gBAAgB,EAAE,KAAK,IAAI,sCAAa,EAAE,KAAK,MAAM;AAAA,CAAU;AAAA,MACtF;AAAA,IACF,CAAC;AAAA,EACH;AACA,QAAM,SAAS,SAAS,YAAY,MAAM,aAAa,IAAI;AAE3D,QAAM,SAAS,MAAM,UAAU,cAAc,KAAK;AAClD,QAAM,MAAM,MAAM,OAAO,aAAa,IAAI,WAAW,MAAM,CAAC;AAC5D,MAAI;AACF,QAAI,OAAO;AACX,eAAW,QAAQ,IAAI,MAAM,GAAG;AAE9B,YAAM,SAAS,KAAK,SAAS;AAC7B,UAAI,CAAC,QAAQ,IAAI,MAAM,EAAG;AAC1B,mBAAa,EAAE,MAAM,QAAQ,IAAI;AACjC,UAAI;AACF,cAAM,SAAS,MAAM;AAAA,UACnB,WAAW,MAAM,QAAQ,MAAM,QAAQ,UAAU,YAAY;AAAA,UAC7D;AAAA,UACA,0BAAW,MAAM,8BAAU,kBAAkB,GAAI;AAAA,QACnD;AACA,eAAO,IAAI,QAAQ,MAAM;AAAA,MAC3B,SAAS,GAAG;AACV,iBAAS,KAAK;AAAA,UACZ,MAAM;AAAA,UACN,SAAS,sBAAO,MAAM,sBAAY,aAAa,QAAQ,EAAE,UAAU,OAAO,CAAC,CAAC;AAAA,UAC5E,MAAM;AAAA,QACR,CAAC;AAAA,MACH;AAAA,IACF;AAAA,EACF,UAAE;AACA,QAAI,QAAQ;AACZ,WAAO,QAAQ;AAAA,EACjB;AACA,SAAO;AACT;AAEA,eAAe,WACb,MACA,QACA,MACA,QACA,UACA,cACoB;AACpB,QAAM,EAAE,eAAe,MAAM,gBAAgB,KAAK,IAAI,KAAK,gBAAgB;AAC3E,QAAM,WAAW,MAAM,KAAK,OAAO;AAAA,IACjC,OAAO;AAAA,IACP,QAAQ,OAAO,EAAE,KAAK,MAAM;AAAA,EAC9B,CAAC;AACD,QAAM,EAAE,MAAM,MAAM,OAAO,IAAI,QAAQ,GAAG,IAAI;AAC9C,QAAM,OAAO,WAAW,IAAI;AAE5B,MAAI,SAAS,WAAW;AAEtB,UAAM,EAAE,MAAM,QAAQ,IAAI,WAAW,MAAM,IAAI,EAAE;AACjD,UAAM,QAAQ,EAAE,gBAAgB,EAAE;AAClC,UAAM,QAAQ,MAAM,OAAQ,cAAc,SAAS,IAAI,IAAI,KAAK;AAChE,QAAI,MAAM,iBAAiB,GAAG;AAC5B,eAAS,KAAK;AAAA,QACZ,MAAM;AAAA,QACN,SAAS,sBAAO,MAAM,yCAAgB,MAAM,cAAc;AAAA,QAC1D,MAAM;AAAA,MACR,CAAC;AAAA,IACH;AAEA,UAAM,QAAQ,KAAK;AACnB,UAAM,SAAS,kBAAkB,SAAS,IAAI,IAAI,KAAK;AACvD,UAAM,aAAa,iBAAiB,QAAQ,OAAO,IAAI;AACvD,WAAO,iBAAiB,OAAO,QAAQ,MAAM,MAAM,OAAO,YAAY,YAAY;AAAA,EACpF;AAOA,QAAM,cAAc,MAAMA;AAAA,IACxB;AAAA,IACA,MAAM,OAAO,OAAO;AAAA,EACtB;AACA,QAAM,WAAW;AACjB,QAAM,QACJ,OAAO,aAAa,aAAa,WAAY,SAAS,WAAY;AACpE,QAAM,MAAM,MAAM,MAAM,MAAM,EAAE,KAAK,EAAE,OAAO,IAAI,QAAQ,IAAI,UAAU,EAAE,EAAE,CAAC,EAC1E,IAAI,EACJ,SAAS;AAEZ,QAAM,OAAO,MAAM,KAAK,IAAI,WAAW,GAAG,GAAG,QAAQ,WAAW;AAChE,MAAI,CAAC,KAAK,KAAK,GAAG;AAChB,aAAS,KAAK,EAAE,MAAM,QAAQ,SAAS,sBAAO,MAAM,kCAAc,MAAM,aAAa,CAAC;AACtF,WAAO,CAAC;AAAA,EACV;AACA,SAAO,CAAC,EAAE,MAAM,aAAa,MAAM,KAAK,KAAK,GAAG,YAAY,OAAO,CAAC;AACtE;AAQO,SAAS,iBACd,OACA,YACA,MACA,MACA,OACA,YACA,eAAe,MACJ;AACX,QAAM,OAAmB,MAAM,IAAI,QAAM;AACvC,UAAM,IAAI,GAAG,IAAI;AACjB,WAAO;AAAA,MACL,MAAM,GAAG;AAAA,MACT,GAAG,KAAK,MAAM,GAAG,IAAI,KAAK;AAAA;AAAA,MAE1B,GAAG,KAAK,MAAM,QAAQ,GAAG,IAAI,GAAG,KAAK,KAAK;AAAA,MAC1C,GAAG,KAAK,MAAM,GAAG,IAAI,KAAK;AAAA,MAC1B,GAAG,KAAK,MAAM,CAAC;AAAA;AAAA,MAEf,UAAU,KAAK,IAAI,GAAG,KAAK,MAAM,IAAI,GAAG,CAAC;AAAA,MACzC,UAAU;AAAA,MACV,UAAU;AAAA,IACZ;AAAA,EACF,CAAC;AACD,SAAO,2BAA2B,MAAM,YAAY,EAAE,SAAS,CAAC,GAAG,WAAW,CAAC,EAAE,GAAG,MAAM,MAAM,YAAY,YAAY;AAC1H;AAEA,eAAeA,WAAa,MAAc,QAAsC;AAC9E,MAAI;AACF,WAAO,MAAM,OAAO;AAAA,EACtB,SAAS,GAAG;AACV,UAAM,IAAI;AAAA,MACR,kEAAoC,IAAI,0DACrB,IAAI,wEAAuB,EAAY,OAAO;AAAA,IACnE;AAAA,EACF;AACF;AAEA,eAAe,YAAe,SAAqB,IAAY,KAAyB;AAEtF,UAAQ,MAAM,MAAM;AAAA,EAAC,CAAC;AACtB,MAAI;AACJ,MAAI;AACF,WAAO,MAAM,QAAQ,KAAK;AAAA,MACxB;AAAA,MACA,IAAI,QAAe,CAAC,GAAG,WAAW;AAChC,gBAAQ,WAAW,MAAM,OAAO,IAAI,MAAM,GAAG,CAAC,GAAG,EAAE;AAAA,MACrD,CAAC;AAAA,IACH,CAAC;AAAA,EACH,UAAE;AACA,QAAI,MAAO,cAAa,KAAK;AAAA,EAC/B;AACF;AAGA,SAAS,WAAW,MAA8B;AAChD,QAAM,MAAM,IAAI,WAAW,KAAK,MAAM;AACtC,WAAS,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK,GAAG;AACvC,QAAI,CAAC,IAAI,KAAK,IAAI,CAAC;AACnB,QAAI,IAAI,CAAC,IAAI,KAAK,IAAI,CAAC;AACvB,QAAI,IAAI,CAAC,IAAI,KAAK,CAAC;AACnB,QAAI,IAAI,CAAC,IAAI,KAAK,IAAI,CAAC;AAAA,EACzB;AACA,SAAO;AACT;","names":["m","prev","tryImport"]}
|
|
@@ -1,7 +1,7 @@
|
|
|
1
1
|
#!/usr/bin/env node
|
|
2
2
|
import {
|
|
3
3
|
KordocError
|
|
4
|
-
} from "./chunk-
|
|
4
|
+
} from "./chunk-I4NNYWIR.js";
|
|
5
5
|
|
|
6
6
|
// src/hwpx/parser-shared.ts
|
|
7
7
|
import { DOMParser } from "@xmldom/xmldom";
|
|
@@ -335,4 +335,4 @@ export {
|
|
|
335
335
|
hwpFaceToCssStack,
|
|
336
336
|
parseRenderStyles
|
|
337
337
|
};
|
|
338
|
-
//# sourceMappingURL=chunk-
|
|
338
|
+
//# sourceMappingURL=chunk-4RJ5D7MT.js.map
|
|
@@ -0,0 +1 @@
|
|
|
1
|
+
{"version":3,"sources":["../src/hwpx/parser-shared.ts","../src/shared/xml.ts","../src/render/head-styles.ts"],"sourcesContent":["/**\n * HWPX 파서 공유 상수/타입/유틸 (parser.ts에서 분리).\n * ZIP 한도, 섹션 공유 상태(SectionShared), walk 컨텍스트(WalkCtx), XML 유틸.\n */\n\nimport { DOMParser } from \"@xmldom/xmldom\"\nimport { KordocError } from \"../utils.js\"\nimport type { CellContext, IRBlock, ParseWarning } from \"../types.js\"\n// WalkCtx.styleMap 타입 참조 — 타입 전용이라 styles.ts와의 순환은 컴파일 시 소거됨\nimport type { HwpxStyleMap } from \"./styles.js\"\nimport type { NoteNumberFormat } from \"./notes.js\"\n\n// 256MB — rhwp 1만 건 실문서 서베이에서 section1.xml 단독 75.2MB(압축비 35:1) 정상\n// 문서가 확인됨 (rhwp #1917). 종전 100MB 총합 컷은 대형 실문서를 ZIP bomb 으로 오인 거부.\nexport const MAX_DECOMPRESS_SIZE = 256 * 1024 * 1024\n/** 손상 ZIP 복구 시 최대 엔트리 수 */\nexport const MAX_ZIP_ENTRIES = 500\n\n/** ZIP bomb 가드 전용 에러 — per-section catch가 XML fatalError(PARTIAL_PARSE로 강등)와\n * 구분해 이것만 재던진다. KordocError 서브클래스라 sanitizeError allowlist는 그대로 통과 */\nexport class ZipBombError extends KordocError {\n constructor(message: string) {\n super(message)\n this.name = \"ZipBombError\"\n }\n}\n\n/** colSpan/rowSpan을 안전한 범위로 클램핑 */\nexport function clampSpan(val: number, max: number): number {\n return Math.max(1, Math.min(val, max))\n}\n\n// XML DOM 헬퍼 정본은 src/shared/xml.ts — 하위호환 재수출 (section-walker, render/* 등이\n// 이 모듈에서 import). MAX_XML_DEPTH 의미(요소 깊이 ≠ 표 중첩 단계)는 정본 주석 참조.\nexport { MAX_XML_DEPTH, findChildByLocalName, extractTextFromNode } from \"../shared/xml.js\"\n\n/** 셀 컨텍스트 확장 — 중첩표/이미지/다중문단 블록과 제목셀 여부를 IRCell로 전달 (v3.0) */\nexport interface CellCtxEx extends CellContext {\n blocks?: IRBlock[]\n /** 중첩표/이미지 등 구조 콘텐츠 존재 — true일 때만 IRCell.blocks로 attach */\n hasStructure?: boolean\n isHeader?: boolean\n /**\n * cell.text 평탄화 조립용 임시 상태 (#52 후속) — 인라인 흐름(글자취급 표·같은 문단\n * 텍스트)이 \"열려\" 있어 다음 인라인 항목을 `\\n`이 아니라 공백으로 이어야 하는지.\n * 문단 경계·블록/float 표에서 닫힌다(false). 최종 IRCell로는 나가지 않는 워크 전용 필드.\n */\n lineOpen?: boolean\n /**\n * keepEmptyParagraphs 조립용 임시 상태 (#57) — 이 셀에서 문단(빈 문단 포함)을 하나라도\n * 봤는지. 선두 빈 문단은 cell.text가 계속 \"\"라 truthiness로 구분할 수 없어 별도 추적한다.\n * 옵션 off일 땐 미사용. 최종 IRCell로는 나가지 않는 워크 전용 필드.\n */\n paraSeen?: boolean\n}\n\nexport interface TableState {\n rows: CellContext[][]\n currentRow: CellContext[]\n cell: CellCtxEx | null\n /** hp:caption 텍스트 — IRTable.caption으로 전달 (v3.0) */\n caption?: string\n /** hp:caption 내부 블록(중첩표 포함) — IRTable.captionBlocks로 전달 (#55) */\n captionBlocks?: IRBlock[]\n /**\n * 글자취급(treatAsChar=\"1\") 표 여부 — 부모 셀 텍스트 평탄화 시 앞뒤 텍스트와 같은 줄로\n * 공백 연결할지 판단 (#52 후속). 블록/float 표(기본 undefined)는 종전대로 `\\n`.\n */\n inline?: boolean\n /** hp:tbl id — IRTable.sourceId (렌더 region 조인 키, #76) */\n sourceId?: string\n}\n\n/** 섹션 간 공유 상태 — 자동번호 카운터, 머리말/꼬리말, 변경추적 */\nexport interface SectionShared {\n /** numbering id → 레벨별(1..10) 카운터. -1 = 미사용(start값으로 초기화 — 0은 start=\"0\"의 유효값) */\n numState: Map<string, number[]>\n pageText: { headers: string[]; footers: string[] }\n track: { deleteDepth: number; warned: boolean }\n /** content.hpf kordoc-layout 메타 (\"default\"|\"gongmun\") — 자사 생성 파일 왕복 채널\n * 게이트. null/미설정 = 외래 파일 (id 기반 인라인 강조·인용 복원 꺼짐) */\n kordocLayout?: string | null\n /** 표 후행 빈 열(앵커 있는 입력란) 보존 — ParseOptions.keepTrailingEmptyCols (#47) */\n keepTrailingEmptyCols?: boolean\n /** 빈 문단 보존 — ParseOptions.keepEmptyParagraphs (#57) */\n keepEmptyParagraphs?: boolean\n /** 실제 페이지 경계 상태 (#66) — base: 이전 섹션까지 누적 페이지 수,\n * allUsable: 지금까지 파싱한 전 섹션이 조판 캐시(linesegarray)를 신뢰 가능 */\n pageState: { base: number; allUsable: boolean }\n}\n\nexport function createSectionShared(): SectionShared {\n return {\n numState: new Map(),\n pageText: { headers: [], footers: [] },\n track: { deleteDepth: 0, warned: false },\n pageState: { base: 0, allUsable: true },\n }\n}\n\n/** walk 함수들이 공유하는 파싱 컨텍스트 — 개별 optional 파라미터를 하나로 묶어 시그니처 안정화 */\nexport interface WalkCtx {\n styleMap?: HwpxStyleMap\n warnings?: ParseWarning[]\n sectionNum?: number\n shared: SectionShared\n /** secPr outlineShapeIDRef — 개요(OUTLINE) 문단이 사용하는 numbering id */\n outlineNumId?: string\n /** secPr footNotePr/endNotePr 번호 모양 — 각주·미주 본문 참조 부호 재구성 (notes.ts) */\n noteFormats?: { footnote?: NoteNumberFormat; endnote?: NoteNumberFormat }\n /** 현재 페이지 (#66) — top-level 문단 진입 시 paraPage로 갱신, 셀/중첩은 호스트 상속 */\n page?: number\n /** 섹션 프리패스 결과: top-level <hp:p> → 섹션 내 0-based 페이지 */\n paraPage?: Map<Element, number>\n /** 이전 섹션까지 누적 페이지 수 (전역 페이지 = pageBase + 섹션 내 페이지 + 1) */\n pageBase?: number\n}\n\n/** xmldom DOMParser 생성 — onError 콜백으로 malformed XML 경고 수집 */\nexport function createXmlParser(warnings?: ParseWarning[]): DOMParser {\n return new DOMParser({\n onError(level: \"warning\" | \"error\" | \"fatalError\", msg: string) {\n if (level === \"fatalError\") throw new KordocError(`XML 파싱 실패: ${msg}`)\n warnings?.push({ code: \"MALFORMED_XML\", message: `XML ${level === \"warning\" ? \"경고\" : \"오류\"}: ${msg}` })\n },\n })\n}\n\n/** 수집된 머리말/꼬리말을 본문 앞/뒤 문단으로 배치 */\nexport function applyPageText(blocks: IRBlock[], shared: SectionShared): void {\n const { headers, footers } = shared.pageText\n if (headers.length > 0) {\n blocks.unshift(...headers.map(t => ({ type: \"paragraph\" as const, text: t, pageNumber: 1 })))\n }\n if (footers.length > 0) {\n blocks.push(...footers.map(t => ({ type: \"paragraph\" as const, text: t })))\n }\n}\n\n","/**\n * XML DOM 공용 헬퍼 — hwpx(parser-shared)·hwpml(parser)·docx(equation)가 공유.\n *\n * xmldom Element를 네임스페이스 접두사 무시(localName 기준)로 다루는 최소 유틸.\n * 정본은 이 모듈 — hwpx/parser-shared.ts는 하위호환 재수출만 유지한다.\n */\n\n/** XML DOM 재귀 최대 깊이 — 악성 파일의 스택 오버플로 방지.\n * 좌표계가 다른 hwp5 MAX_NEST_DEPTH(8, 표 중첩 단계)·filler/소스맵 16\n * (표 중첩 단계)과 달리 이건 \"XML 요소\" 깊이라 표 1단이 여러 depth를\n * 소모한다 — 상수 통일 금지 (의미가 다름) */\nexport const MAX_XML_DEPTH = 200\n\n/** 요소의 localName (네임스페이스 접두사 제거) */\nexport function localName(el: Element): string {\n return (el.tagName || el.localName || \"\").replace(/^[^:]+:/, \"\")\n}\n\n/** 자식 중 지정된 localName(접두사 제거)을 가진 첫 번째 Element 반환 */\nexport function findChildByLocalName(parent: Element, name: string): Element | null {\n const children = parent.childNodes\n if (!children) return null\n for (let i = 0; i < children.length; i++) {\n const ch = children[i] as Element\n if (ch.nodeType !== 1) continue\n if (localName(ch) === name) return ch\n }\n return null\n}\n\n/** 자식 중 지정된 localName을 가진 모든 Element (직계만) */\nexport function childrenByLocalName(parent: Element, name: string): Element[] {\n const out: Element[] = []\n const children = parent.childNodes\n if (!children) return out\n for (let i = 0; i < children.length; i++) {\n const ch = children[i] as Element\n if (ch.nodeType !== 1) continue\n if (localName(ch) === name) out.push(ch)\n }\n return out\n}\n\n/** 직계 자식 Element 전부 (텍스트/주석 노드 제외) */\nexport function elementChildren(parent: Element): Element[] {\n const out: Element[] = []\n const children = parent.childNodes\n if (!children) return out\n for (let i = 0; i < children.length; i++) {\n if (children[i].nodeType === 1) out.push(children[i] as Element)\n }\n return out\n}\n\n/** 텍스트를 담는 노드 타입 — TEXT_NODE(3)과 CDATA_SECTION_NODE(4).\n * CDATA 를 빠뜨리면 `<hp:t><![CDATA[본문]]></hp:t>` 로 저장된 문서의 텍스트가\n * 경고 없이 통째로 사라진다 (rhwp f4c7de75 와 같은 결함). 한컴 저장본은 CDATA 를\n * 쓰지 않지만 제3자 도구 산출물·특수문자 다수 문단에서는 유효한 XML 표현이다. */\nfunction isTextNode(node: Node): boolean {\n return node.nodeType === 3 || node.nodeType === 4\n}\n\n/** 노드 내 모든 텍스트를 재귀적으로 추출 (요소 단위 trim — hwpx 기존 시맨틱).\n * MAX_XML_DEPTH 가드 — 악성 심층 XML 스택 오버플로 방지 */\nexport function extractTextFromNode(node: Node, depth: number = 0): string {\n let result = \"\"\n if (depth > MAX_XML_DEPTH) return result\n const children = node.childNodes\n if (!children) return result\n for (let i = 0; i < children.length; i++) {\n const child = children[i]\n if (isTextNode(child)) result += child.textContent || \"\"\n else if (child.nodeType === 1) result += extractTextFromNode(child, depth + 1)\n }\n return result.trim()\n}\n\n/** 노드 내 모든 텍스트를 공백 원형 그대로 추출 (trim 없음 — hwpml CHAR 텍스트 시맨틱) */\nexport function rawTextContent(node: Node, depth: number = 0): string {\n if (depth > MAX_XML_DEPTH) return \"\"\n const children = node.childNodes\n if (!children) return \"\"\n const parts: string[] = []\n for (let i = 0; i < children.length; i++) {\n const child = children[i]\n if (isTextNode(child)) parts.push(child.nodeValue || \"\")\n else if (child.nodeType === 1) parts.push(rawTextContent(child, depth + 1))\n }\n return parts.join(\"\")\n}\n","/**\n * 레이아웃 보존 렌더 — header.xml 스타일 테이블 (charPr/paraPr/borderFill).\n * 파서 IR과 무관하게 렌더에 필요한 속성만 추출한다.\n */\n\nimport { createXmlParser, findChildByLocalName } from \"../hwpx/parser-shared.js\"\n\nexport interface RenderCharStyle {\n /** 글자 크기 (1/100pt — 1000 = 10pt) */\n height: number\n bold: boolean\n italic: boolean\n underline: boolean\n /** #RRGGBB — 검정이면 undefined */\n color?: string\n /** 장평 % */\n ratio: number\n /** 자간 % */\n spacing: number\n /** CSS font-family 스택 — charPr fontRef(hangul)에서 해석. 없으면 root 기본 폴백 */\n fontFamily?: string\n /** HWP 원본 글꼴명 (fontfaces face) — reflow 폭 테이블 선택용 */\n face?: string\n}\n\nexport type ParaAlign = \"JUSTIFY\" | \"LEFT\" | \"RIGHT\" | \"CENTER\" | \"DISTRIBUTE\" | \"DISTRIBUTE_SPACE\"\n\n/** reflow(Tier-2)용 문단 기하 — 줄간격·여백. paraPr `<hh:margin>`·`<hh:lineSpacing>`은\n * `<hp:switch>/<hp:case>` 안(손자)이라 재귀 탐색으로 뽑는다. 단위 HWPUNIT. */\nexport interface RenderParaGeom {\n /** PERCENT(기본 160=160%) / FIXED / BETWEEN_LINES / AT_LEAST */\n lineSpacingType: string\n lineSpacingValue: number\n /** 왼쪽 들여쓰기(HWPUNIT) */\n marginLeft: number\n marginRight: number\n /** 첫 줄 들여쓰기/내어쓰기 — 음수=둘째 줄부터 더 들어감(hanging) */\n marginIntent: number\n /** 문단 위 간격(prev) */\n spaceBefore: number\n /** 문단 아래 간격(next) */\n spaceAfter: number\n /** breakSetting의 한글 줄나눔 — 'keep'(어절)/'charAll'(글자). 없으면 호출자 모드.\n * 주의: 저장 속성 의미는 이름과 반대 — breakNonLatinWord \"BREAK_WORD\"=어절,\n * \"KEEP_WORD\"=글자 (한글 COM 실렌더 실측) */\n wrapMode?: \"keep\" | \"charAll\"\n /** 탭 정의(tabPrIDRef)가 내어쓰기용 자동 탭(autoTabLeft) — 첫 줄 탭이 내어쓰기 위치에 선다 */\n autoTabLeft?: boolean\n}\n\nexport const DEFAULT_PARA_GEOM: RenderParaGeom = {\n lineSpacingType: \"PERCENT\", lineSpacingValue: 160,\n marginLeft: 0, marginRight: 0, marginIntent: 0, spaceBefore: 0, spaceAfter: 0,\n}\n\nexport interface RenderBorderEdge {\n /** SOLID/DASH/DOT… — NONE은 edge 자체를 생략 */\n type: string\n /** pt 단위 굵기 */\n widthPt: number\n color: string\n}\n\nexport interface RenderBorderFill {\n left?: RenderBorderEdge\n right?: RenderBorderEdge\n top?: RenderBorderEdge\n bottom?: RenderBorderEdge\n /** 배경색 #RRGGBB (없으면 undefined) */\n fill?: string\n}\n\nexport interface RenderStyles {\n charPr: Map<string, RenderCharStyle>\n paraAlign: Map<string, ParaAlign>\n /** reflow용 문단 기하 (줄간격·여백) — id별 */\n paraGeom: Map<string, RenderParaGeom>\n borderFill: Map<string, RenderBorderFill>\n}\n\nexport const DEFAULT_CHAR: RenderCharStyle = { height: 1000, bold: false, italic: false, underline: false, ratio: 100, spacing: 0 }\n\n// ─── 글꼴 매핑 (HWP fontfaces face → CSS font-family 스택) ──────────────\n//\n// HWPX charPr 는 <hh:fontRef hangul=\"N\">으로 fontfaces 테이블의 실제 글꼴명을 가리킨다.\n// 렌더는 이 글꼴명을 SVG <text font-family>로 내보내 뷰어(WebView2/브라우저)가 원본과\n// 같은 글꼴로 그리게 한다. 미설치 글꼴은 계열(명조=serif / 고딕=sans) 폴백으로 수렴시켜\n// 최소한 획 계열(바탕↔돋움)은 원본과 일치시킨다 — 공문서 제목 고딕이 바탕체로 나오던 회귀 해소.\n\n/** 자주 쓰는 글꼴의 별칭 스택 (한글명↔영문 시스템명 병기 — 뷰어 OS별 등록명 차이 흡수) */\nconst FONT_ALIASES: Record<string, string> = {\n \"함초롬바탕\": \"'HCR Batang','함초롬바탕','한컴바탕'\",\n \"한컴바탕\": \"'HCR Batang','함초롬바탕','한컴바탕'\",\n \"함초롬돋움\": \"'HCR Dotum','함초롬돋움','한컴돋움'\",\n \"한컴돋움\": \"'HCR Dotum','함초롬돋움','한컴돋움'\",\n \"맑은 고딕\": \"'Malgun Gothic','맑은 고딕'\",\n \"맑은고딕\": \"'Malgun Gothic','맑은 고딕'\",\n \"굴림\": \"'Gulim','굴림'\",\n \"굴림체\": \"'GulimChe','굴림체','Gulim'\",\n \"돋움\": \"'Dotum','돋움'\",\n \"돋움체\": \"'DotumChe','돋움체','Dotum'\",\n \"바탕\": \"'Batang','바탕'\",\n \"바탕체\": \"'BatangChe','바탕체','Batang'\",\n \"궁서\": \"'Gungsuh','궁서'\",\n \"궁서체\": \"'GungsuhChe','궁서체','Gungsuh'\",\n \"나눔고딕\": \"'NanumGothic','나눔고딕'\",\n \"나눔명조\": \"'NanumMyeongjo','나눔명조'\",\n \"맑은 고딕 Semilight\": \"'Malgun Gothic Semilight','맑은 고딕'\",\n}\n\n/** 명조/바탕 계열(serif) 여부 — 아니면 고딕/돋움(sans)으로 본다 */\nfunction isSerifFace(face: string): boolean {\n return /바탕|명조|Batang|Myeong|Mincho|궁서|Gungsuh|Serif|신명|순명|Song|송/i.test(face)\n}\n\n/** CSS font-family 토큰 인용 — 영숫자·하이픈만이면 무인용, 그 외(공백·한글)는 작은따옴표 */\nfunction cssQuote(name: string): string {\n return /^[A-Za-z0-9_-]+$/.test(name) ? name : `'${name.replace(/['\"\\\\]/g, \"\")}'`\n}\n\n/**\n * HWP 글꼴명 → CSS font-family 스택. 정확한 글꼴 먼저, 이어서 계열 폴백(명조=serif /\n * 고딕=sans, 각 한국어 렌더 가능 글꼴 포함)을 붙여 미설치 시에도 획 계열을 보존한다.\n */\nexport function hwpFaceToCssStack(face: string | null | undefined): string {\n const trimmed = (face ?? \"\").trim()\n if (!trimmed) return \"\"\n const generic = isSerifFace(trimmed)\n ? \"'HCR Batang','Batang','Noto Serif KR',serif\"\n : \"'Malgun Gothic','HCR Dotum','Noto Sans KR',sans-serif\"\n const head = FONT_ALIASES[trimmed] ?? cssQuote(trimmed)\n return `${head},${generic}`\n}\n\n/** fontfaces 테이블에서 HANGUL 그룹의 font id → 글꼴명 맵 (없으면 첫 그룹) */\nfunction collectHangulFonts(root: Element): Map<string, string> {\n const map = new Map<string, string>()\n const findFaces = (el: Element, depth: number): Element | null => {\n if (depth > 24) return null\n for (const ch of Array.from(el.childNodes)) {\n if (ch.nodeType !== 1) continue\n const e = ch as Element\n if ((e.tagName || \"\").replace(/^[^:]+:/, \"\") === \"fontfaces\") return e\n const f = findFaces(e, depth + 1)\n if (f) return f\n }\n return null\n }\n const faces = findFaces(root, 0)\n if (!faces) return map\n let group: Element | null = null\n let firstGroup: Element | null = null\n for (const ch of Array.from(faces.childNodes)) {\n if (ch.nodeType !== 1) continue\n const e = ch as Element\n if ((e.tagName || \"\").replace(/^[^:]+:/, \"\") !== \"fontface\") continue\n if (!firstGroup) firstGroup = e\n if ((e.getAttribute(\"lang\") ?? \"\").toUpperCase() === \"HANGUL\") { group = e; break }\n }\n group = group ?? firstGroup\n if (!group) return map\n for (const ch of Array.from(group.childNodes)) {\n if (ch.nodeType !== 1) continue\n const e = ch as Element\n if ((e.tagName || \"\").replace(/^[^:]+:/, \"\") !== \"font\") continue\n const id = e.getAttribute(\"id\")\n const face = e.getAttribute(\"face\")\n if (id != null && face) map.set(id, face)\n }\n return map\n}\n\n/** \"0.12 mm\" / \"0.5 mm\" → pt */\nfunction borderWidthPt(v: string | null | undefined): number {\n const n = parseFloat(v ?? \"\")\n if (!Number.isFinite(n)) return 0.34\n return n * 2.834645 // mm → pt\n}\n\nfunction parseEdge(el: Element | null): RenderBorderEdge | undefined {\n if (!el) return undefined\n const type = el.getAttribute(\"type\") ?? \"NONE\"\n if (type === \"NONE\") return undefined\n return { type, widthPt: borderWidthPt(el.getAttribute(\"width\")), color: el.getAttribute(\"color\") ?? \"#000000\" }\n}\n\n/** 서브트리에서 localName이 일치하는 첫 요소를 재귀 탐색 (switch/case 래핑 대응) */\nfunction findDeep(el: Element, name: string, depth = 0): Element | null {\n if (depth > 32) return null\n const children = el.childNodes\n if (!children) return null\n for (let i = 0; i < children.length; i++) {\n const ch = children[i]\n if (ch.nodeType !== 1) continue\n const e = ch as Element\n if ((e.tagName || \"\").replace(/^[^:]+:/, \"\") === name) return e\n const found = findDeep(e, name, depth + 1)\n if (found) return found\n }\n return null\n}\n\n/** paraPr 요소 → 문단 기하 (margin·lineSpacing은 hp:switch/case 안이라 재귀 탐색) */\nfunction parseParaGeom(el: Element): RenderParaGeom {\n const g: RenderParaGeom = { ...DEFAULT_PARA_GEOM }\n const ls = findDeep(el, \"lineSpacing\")\n if (ls) {\n g.lineSpacingType = ls.getAttribute(\"type\") ?? \"PERCENT\"\n g.lineSpacingValue = Number(ls.getAttribute(\"value\")) || 160\n }\n const margin = findDeep(el, \"margin\")\n if (margin) {\n const v = (n: string): number => {\n const c = findDeep(margin, n)\n return c ? Number(c.getAttribute(\"value\")) || 0 : 0\n }\n g.marginLeft = v(\"left\")\n g.marginRight = v(\"right\")\n g.marginIntent = v(\"intent\")\n g.spaceBefore = v(\"prev\")\n g.spaceAfter = v(\"next\")\n }\n const bs = findDeep(el, \"breakSetting\")\n if (bs) {\n // 속성 의미 역전 주의(위 RenderParaGeom.wrapMode 주석): BREAK_WORD=어절, KEEP_WORD=글자\n const nl = bs.getAttribute(\"breakNonLatinWord\")\n if (nl === \"BREAK_WORD\") g.wrapMode = \"keep\"\n else if (nl === \"KEEP_WORD\") g.wrapMode = \"charAll\"\n }\n return g\n}\n\n/** header.xml(구버전 head.xml) → 렌더용 스타일 맵 */\nexport function parseRenderStyles(headXml: string): RenderStyles {\n const styles: RenderStyles = { charPr: new Map(), paraAlign: new Map(), paraGeom: new Map(), borderFill: new Map() }\n const doc = createXmlParser().parseFromString(headXml, \"text/xml\")\n const root = doc.documentElement as unknown as Element | null\n if (!root) return styles\n\n // fontfaces(HANGUL) 선파싱 — charPr fontRef 해석에 쓴다\n const hangulFonts = collectHangulFonts(root)\n // 탭 정의 — 문단의 tabPrIDRef 는 워크가 끝난 뒤 해석(순서 무관)\n const autoTabPr = new Set<string>()\n const paraTabRef = new Map<string, string>()\n\n const walk = (el: Element): void => {\n const tag = (el.tagName || \"\").replace(/^[^:]+:/, \"\")\n if (tag === \"charPr\") {\n const id = el.getAttribute(\"id\")\n if (id != null) {\n const ratioEl = findChildByLocalName(el, \"ratio\")\n const spacingEl = findChildByLocalName(el, \"spacing\")\n const underlineEl = findChildByLocalName(el, \"underline\")\n const textColor = el.getAttribute(\"textColor\")\n const fontRef = findChildByLocalName(el, \"fontRef\")\n const fontId = fontRef?.getAttribute(\"hangul\") ?? fontRef?.getAttribute(\"latin\")\n const face = fontId != null ? hangulFonts.get(fontId) : undefined\n styles.charPr.set(id, {\n height: Number(el.getAttribute(\"height\")) || 1000,\n bold: findChildByLocalName(el, \"bold\") != null,\n italic: findChildByLocalName(el, \"italic\") != null,\n underline: underlineEl != null && (underlineEl.getAttribute(\"type\") ?? \"NONE\") !== \"NONE\",\n color: textColor && textColor !== \"#000000\" && textColor.toLowerCase() !== \"none\" ? textColor : undefined,\n ratio: Number(ratioEl?.getAttribute(\"hangul\")) || 100,\n spacing: Number(spacingEl?.getAttribute(\"hangul\")) || 0,\n fontFamily: face ? hwpFaceToCssStack(face) : undefined,\n face,\n })\n }\n } else if (tag === \"tabPr\") {\n const id = el.getAttribute(\"id\")\n if (id != null && el.getAttribute(\"autoTabLeft\") === \"1\") autoTabPr.add(id)\n } else if (tag === \"paraPr\") {\n const id = el.getAttribute(\"id\")\n if (id != null) {\n const align = findChildByLocalName(el, \"align\")\n styles.paraAlign.set(id, (align?.getAttribute(\"horizontal\") as ParaAlign) || \"JUSTIFY\")\n styles.paraGeom.set(id, parseParaGeom(el))\n const tab = el.getAttribute(\"tabPrIDRef\")\n if (tab != null) paraTabRef.set(id, tab)\n }\n } else if (tag === \"borderFill\") {\n const id = el.getAttribute(\"id\")\n if (id != null) {\n const bf: RenderBorderFill = {\n left: parseEdge(findChildByLocalName(el, \"leftBorder\")),\n right: parseEdge(findChildByLocalName(el, \"rightBorder\")),\n top: parseEdge(findChildByLocalName(el, \"topBorder\")),\n bottom: parseEdge(findChildByLocalName(el, \"bottomBorder\")),\n }\n const fillBrush = findChildByLocalName(el, \"fillBrush\")\n const winBrush = fillBrush ? findChildByLocalName(fillBrush, \"winBrush\") : null\n const face = winBrush?.getAttribute(\"faceColor\")\n if (face && face.toLowerCase() !== \"none\") bf.fill = face\n styles.borderFill.set(id, bf)\n }\n }\n const children = el.childNodes\n for (let i = 0; i < children.length; i++) {\n const ch = children[i]\n if (ch.nodeType === 1) walk(ch as Element)\n }\n }\n walk(root)\n for (const [id, tab] of paraTabRef) {\n const g = styles.paraGeom.get(id)\n if (g && autoTabPr.has(tab)) g.autoTabLeft = true\n }\n return styles\n}\n"],"mappings":";;;;;;AAKA,SAAS,iBAAiB;;;ACMnB,IAAM,gBAAgB;AAGtB,SAAS,UAAU,IAAqB;AAC7C,UAAQ,GAAG,WAAW,GAAG,aAAa,IAAI,QAAQ,WAAW,EAAE;AACjE;AAGO,SAAS,qBAAqB,QAAiB,MAA8B;AAClF,QAAM,WAAW,OAAO;AACxB,MAAI,CAAC,SAAU,QAAO;AACtB,WAAS,IAAI,GAAG,IAAI,SAAS,QAAQ,KAAK;AACxC,UAAM,KAAK,SAAS,CAAC;AACrB,QAAI,GAAG,aAAa,EAAG;AACvB,QAAI,UAAU,EAAE,MAAM,KAAM,QAAO;AAAA,EACrC;AACA,SAAO;AACT;AAGO,SAAS,oBAAoB,QAAiB,MAAyB;AAC5E,QAAM,MAAiB,CAAC;AACxB,QAAM,WAAW,OAAO;AACxB,MAAI,CAAC,SAAU,QAAO;AACtB,WAAS,IAAI,GAAG,IAAI,SAAS,QAAQ,KAAK;AACxC,UAAM,KAAK,SAAS,CAAC;AACrB,QAAI,GAAG,aAAa,EAAG;AACvB,QAAI,UAAU,EAAE,MAAM,KAAM,KAAI,KAAK,EAAE;AAAA,EACzC;AACA,SAAO;AACT;AAGO,SAAS,gBAAgB,QAA4B;AAC1D,QAAM,MAAiB,CAAC;AACxB,QAAM,WAAW,OAAO;AACxB,MAAI,CAAC,SAAU,QAAO;AACtB,WAAS,IAAI,GAAG,IAAI,SAAS,QAAQ,KAAK;AACxC,QAAI,SAAS,CAAC,EAAE,aAAa,EAAG,KAAI,KAAK,SAAS,CAAC,CAAY;AAAA,EACjE;AACA,SAAO;AACT;AAMA,SAAS,WAAW,MAAqB;AACvC,SAAO,KAAK,aAAa,KAAK,KAAK,aAAa;AAClD;AAIO,SAAS,oBAAoB,MAAY,QAAgB,GAAW;AACzE,MAAI,SAAS;AACb,MAAI,QAAQ,cAAe,QAAO;AAClC,QAAM,WAAW,KAAK;AACtB,MAAI,CAAC,SAAU,QAAO;AACtB,WAAS,IAAI,GAAG,IAAI,SAAS,QAAQ,KAAK;AACxC,UAAM,QAAQ,SAAS,CAAC;AACxB,QAAI,WAAW,KAAK,EAAG,WAAU,MAAM,eAAe;AAAA,aAC7C,MAAM,aAAa,EAAG,WAAU,oBAAoB,OAAO,QAAQ,CAAC;AAAA,EAC/E;AACA,SAAO,OAAO,KAAK;AACrB;AAGO,SAAS,eAAe,MAAY,QAAgB,GAAW;AACpE,MAAI,QAAQ,cAAe,QAAO;AAClC,QAAM,WAAW,KAAK;AACtB,MAAI,CAAC,SAAU,QAAO;AACtB,QAAM,QAAkB,CAAC;AACzB,WAAS,IAAI,GAAG,IAAI,SAAS,QAAQ,KAAK;AACxC,UAAM,QAAQ,SAAS,CAAC;AACxB,QAAI,WAAW,KAAK,EAAG,OAAM,KAAK,MAAM,aAAa,EAAE;AAAA,aAC9C,MAAM,aAAa,EAAG,OAAM,KAAK,eAAe,OAAO,QAAQ,CAAC,CAAC;AAAA,EAC5E;AACA,SAAO,MAAM,KAAK,EAAE;AACtB;;;AD3EO,IAAM,sBAAsB,MAAM,OAAO;AAEzC,IAAM,kBAAkB;AAIxB,IAAM,eAAN,cAA2B,YAAY;AAAA,EAC5C,YAAY,SAAiB;AAC3B,UAAM,OAAO;AACb,SAAK,OAAO;AAAA,EACd;AACF;AAGO,SAAS,UAAU,KAAa,KAAqB;AAC1D,SAAO,KAAK,IAAI,GAAG,KAAK,IAAI,KAAK,GAAG,CAAC;AACvC;AA6DO,SAAS,sBAAqC;AACnD,SAAO;AAAA,IACL,UAAU,oBAAI,IAAI;AAAA,IAClB,UAAU,EAAE,SAAS,CAAC,GAAG,SAAS,CAAC,EAAE;AAAA,IACrC,OAAO,EAAE,aAAa,GAAG,QAAQ,MAAM;AAAA,IACvC,WAAW,EAAE,MAAM,GAAG,WAAW,KAAK;AAAA,EACxC;AACF;AAqBO,SAAS,gBAAgB,UAAsC;AACpE,SAAO,IAAI,UAAU;AAAA,IACnB,QAAQ,OAA2C,KAAa;AAC9D,UAAI,UAAU,aAAc,OAAM,IAAI,YAAY,kCAAc,GAAG,EAAE;AACrE,gBAAU,KAAK,EAAE,MAAM,iBAAiB,SAAS,OAAO,UAAU,YAAY,iBAAO,cAAI,KAAK,GAAG,GAAG,CAAC;AAAA,IACvG;AAAA,EACF,CAAC;AACH;AAGO,SAAS,cAAc,QAAmB,QAA6B;AAC5E,QAAM,EAAE,SAAS,QAAQ,IAAI,OAAO;AACpC,MAAI,QAAQ,SAAS,GAAG;AACtB,WAAO,QAAQ,GAAG,QAAQ,IAAI,QAAM,EAAE,MAAM,aAAsB,MAAM,GAAG,YAAY,EAAE,EAAE,CAAC;AAAA,EAC9F;AACA,MAAI,QAAQ,SAAS,GAAG;AACtB,WAAO,KAAK,GAAG,QAAQ,IAAI,QAAM,EAAE,MAAM,aAAsB,MAAM,EAAE,EAAE,CAAC;AAAA,EAC5E;AACF;;;AEvFO,IAAM,oBAAoC;AAAA,EAC/C,iBAAiB;AAAA,EAAW,kBAAkB;AAAA,EAC9C,YAAY;AAAA,EAAG,aAAa;AAAA,EAAG,cAAc;AAAA,EAAG,aAAa;AAAA,EAAG,YAAY;AAC9E;AA2BO,IAAM,eAAgC,EAAE,QAAQ,KAAM,MAAM,OAAO,QAAQ,OAAO,WAAW,OAAO,OAAO,KAAK,SAAS,EAAE;AAUlI,IAAM,eAAuC;AAAA,EAC3C,kCAAS;AAAA,EACT,4BAAQ;AAAA,EACR,kCAAS;AAAA,EACT,4BAAQ;AAAA,EACR,6BAAS;AAAA,EACT,4BAAQ;AAAA,EACR,gBAAM;AAAA,EACN,sBAAO;AAAA,EACP,gBAAM;AAAA,EACN,sBAAO;AAAA,EACP,gBAAM;AAAA,EACN,sBAAO;AAAA,EACP,gBAAM;AAAA,EACN,sBAAO;AAAA,EACP,4BAAQ;AAAA,EACR,4BAAQ;AAAA,EACR,uCAAmB;AACrB;AAGA,SAAS,YAAY,MAAuB;AAC1C,SAAO,4DAA4D,KAAK,IAAI;AAC9E;AAGA,SAAS,SAAS,MAAsB;AACtC,SAAO,mBAAmB,KAAK,IAAI,IAAI,OAAO,IAAI,KAAK,QAAQ,WAAW,EAAE,CAAC;AAC/E;AAMO,SAAS,kBAAkB,MAAyC;AACzE,QAAM,WAAW,QAAQ,IAAI,KAAK;AAClC,MAAI,CAAC,QAAS,QAAO;AACrB,QAAM,UAAU,YAAY,OAAO,IAC/B,gDACA;AACJ,QAAM,OAAO,aAAa,OAAO,KAAK,SAAS,OAAO;AACtD,SAAO,GAAG,IAAI,IAAI,OAAO;AAC3B;AAGA,SAAS,mBAAmB,MAAoC;AAC9D,QAAM,MAAM,oBAAI,IAAoB;AACpC,QAAM,YAAY,CAAC,IAAa,UAAkC;AAChE,QAAI,QAAQ,GAAI,QAAO;AACvB,eAAW,MAAM,MAAM,KAAK,GAAG,UAAU,GAAG;AAC1C,UAAI,GAAG,aAAa,EAAG;AACvB,YAAM,IAAI;AACV,WAAK,EAAE,WAAW,IAAI,QAAQ,WAAW,EAAE,MAAM,YAAa,QAAO;AACrE,YAAM,IAAI,UAAU,GAAG,QAAQ,CAAC;AAChC,UAAI,EAAG,QAAO;AAAA,IAChB;AACA,WAAO;AAAA,EACT;AACA,QAAM,QAAQ,UAAU,MAAM,CAAC;AAC/B,MAAI,CAAC,MAAO,QAAO;AACnB,MAAI,QAAwB;AAC5B,MAAI,aAA6B;AACjC,aAAW,MAAM,MAAM,KAAK,MAAM,UAAU,GAAG;AAC7C,QAAI,GAAG,aAAa,EAAG;AACvB,UAAM,IAAI;AACV,SAAK,EAAE,WAAW,IAAI,QAAQ,WAAW,EAAE,MAAM,WAAY;AAC7D,QAAI,CAAC,WAAY,cAAa;AAC9B,SAAK,EAAE,aAAa,MAAM,KAAK,IAAI,YAAY,MAAM,UAAU;AAAE,cAAQ;AAAG;AAAA,IAAM;AAAA,EACpF;AACA,UAAQ,SAAS;AACjB,MAAI,CAAC,MAAO,QAAO;AACnB,aAAW,MAAM,MAAM,KAAK,MAAM,UAAU,GAAG;AAC7C,QAAI,GAAG,aAAa,EAAG;AACvB,UAAM,IAAI;AACV,SAAK,EAAE,WAAW,IAAI,QAAQ,WAAW,EAAE,MAAM,OAAQ;AACzD,UAAM,KAAK,EAAE,aAAa,IAAI;AAC9B,UAAM,OAAO,EAAE,aAAa,MAAM;AAClC,QAAI,MAAM,QAAQ,KAAM,KAAI,IAAI,IAAI,IAAI;AAAA,EAC1C;AACA,SAAO;AACT;AAGA,SAAS,cAAc,GAAsC;AAC3D,QAAM,IAAI,WAAW,KAAK,EAAE;AAC5B,MAAI,CAAC,OAAO,SAAS,CAAC,EAAG,QAAO;AAChC,SAAO,IAAI;AACb;AAEA,SAAS,UAAU,IAAkD;AACnE,MAAI,CAAC,GAAI,QAAO;AAChB,QAAM,OAAO,GAAG,aAAa,MAAM,KAAK;AACxC,MAAI,SAAS,OAAQ,QAAO;AAC5B,SAAO,EAAE,MAAM,SAAS,cAAc,GAAG,aAAa,OAAO,CAAC,GAAG,OAAO,GAAG,aAAa,OAAO,KAAK,UAAU;AAChH;AAGA,SAAS,SAAS,IAAa,MAAc,QAAQ,GAAmB;AACtE,MAAI,QAAQ,GAAI,QAAO;AACvB,QAAM,WAAW,GAAG;AACpB,MAAI,CAAC,SAAU,QAAO;AACtB,WAAS,IAAI,GAAG,IAAI,SAAS,QAAQ,KAAK;AACxC,UAAM,KAAK,SAAS,CAAC;AACrB,QAAI,GAAG,aAAa,EAAG;AACvB,UAAM,IAAI;AACV,SAAK,EAAE,WAAW,IAAI,QAAQ,WAAW,EAAE,MAAM,KAAM,QAAO;AAC9D,UAAM,QAAQ,SAAS,GAAG,MAAM,QAAQ,CAAC;AACzC,QAAI,MAAO,QAAO;AAAA,EACpB;AACA,SAAO;AACT;AAGA,SAAS,cAAc,IAA6B;AAClD,QAAM,IAAoB,EAAE,GAAG,kBAAkB;AACjD,QAAM,KAAK,SAAS,IAAI,aAAa;AACrC,MAAI,IAAI;AACN,MAAE,kBAAkB,GAAG,aAAa,MAAM,KAAK;AAC/C,MAAE,mBAAmB,OAAO,GAAG,aAAa,OAAO,CAAC,KAAK;AAAA,EAC3D;AACA,QAAM,SAAS,SAAS,IAAI,QAAQ;AACpC,MAAI,QAAQ;AACV,UAAM,IAAI,CAAC,MAAsB;AAC/B,YAAM,IAAI,SAAS,QAAQ,CAAC;AAC5B,aAAO,IAAI,OAAO,EAAE,aAAa,OAAO,CAAC,KAAK,IAAI;AAAA,IACpD;AACA,MAAE,aAAa,EAAE,MAAM;AACvB,MAAE,cAAc,EAAE,OAAO;AACzB,MAAE,eAAe,EAAE,QAAQ;AAC3B,MAAE,cAAc,EAAE,MAAM;AACxB,MAAE,aAAa,EAAE,MAAM;AAAA,EACzB;AACA,QAAM,KAAK,SAAS,IAAI,cAAc;AACtC,MAAI,IAAI;AAEN,UAAM,KAAK,GAAG,aAAa,mBAAmB;AAC9C,QAAI,OAAO,aAAc,GAAE,WAAW;AAAA,aAC7B,OAAO,YAAa,GAAE,WAAW;AAAA,EAC5C;AACA,SAAO;AACT;AAGO,SAAS,kBAAkB,SAA+B;AAC/D,QAAM,SAAuB,EAAE,QAAQ,oBAAI,IAAI,GAAG,WAAW,oBAAI,IAAI,GAAG,UAAU,oBAAI,IAAI,GAAG,YAAY,oBAAI,IAAI,EAAE;AACnH,QAAM,MAAM,gBAAgB,EAAE,gBAAgB,SAAS,UAAU;AACjE,QAAM,OAAO,IAAI;AACjB,MAAI,CAAC,KAAM,QAAO;AAGlB,QAAM,cAAc,mBAAmB,IAAI;AAE3C,QAAM,YAAY,oBAAI,IAAY;AAClC,QAAM,aAAa,oBAAI,IAAoB;AAE3C,QAAM,OAAO,CAAC,OAAsB;AAClC,UAAM,OAAO,GAAG,WAAW,IAAI,QAAQ,WAAW,EAAE;AACpD,QAAI,QAAQ,UAAU;AACpB,YAAM,KAAK,GAAG,aAAa,IAAI;AAC/B,UAAI,MAAM,MAAM;AACd,cAAM,UAAU,qBAAqB,IAAI,OAAO;AAChD,cAAM,YAAY,qBAAqB,IAAI,SAAS;AACpD,cAAM,cAAc,qBAAqB,IAAI,WAAW;AACxD,cAAM,YAAY,GAAG,aAAa,WAAW;AAC7C,cAAM,UAAU,qBAAqB,IAAI,SAAS;AAClD,cAAM,SAAS,SAAS,aAAa,QAAQ,KAAK,SAAS,aAAa,OAAO;AAC/E,cAAM,OAAO,UAAU,OAAO,YAAY,IAAI,MAAM,IAAI;AACxD,eAAO,OAAO,IAAI,IAAI;AAAA,UACpB,QAAQ,OAAO,GAAG,aAAa,QAAQ,CAAC,KAAK;AAAA,UAC7C,MAAM,qBAAqB,IAAI,MAAM,KAAK;AAAA,UAC1C,QAAQ,qBAAqB,IAAI,QAAQ,KAAK;AAAA,UAC9C,WAAW,eAAe,SAAS,YAAY,aAAa,MAAM,KAAK,YAAY;AAAA,UACnF,OAAO,aAAa,cAAc,aAAa,UAAU,YAAY,MAAM,SAAS,YAAY;AAAA,UAChG,OAAO,OAAO,SAAS,aAAa,QAAQ,CAAC,KAAK;AAAA,UAClD,SAAS,OAAO,WAAW,aAAa,QAAQ,CAAC,KAAK;AAAA,UACtD,YAAY,OAAO,kBAAkB,IAAI,IAAI;AAAA,UAC7C;AAAA,QACF,CAAC;AAAA,MACH;AAAA,IACF,WAAW,QAAQ,SAAS;AAC1B,YAAM,KAAK,GAAG,aAAa,IAAI;AAC/B,UAAI,MAAM,QAAQ,GAAG,aAAa,aAAa,MAAM,IAAK,WAAU,IAAI,EAAE;AAAA,IAC5E,WAAW,QAAQ,UAAU;AAC3B,YAAM,KAAK,GAAG,aAAa,IAAI;AAC/B,UAAI,MAAM,MAAM;AACd,cAAM,QAAQ,qBAAqB,IAAI,OAAO;AAC9C,eAAO,UAAU,IAAI,IAAK,OAAO,aAAa,YAAY,KAAmB,SAAS;AACtF,eAAO,SAAS,IAAI,IAAI,cAAc,EAAE,CAAC;AACzC,cAAM,MAAM,GAAG,aAAa,YAAY;AACxC,YAAI,OAAO,KAAM,YAAW,IAAI,IAAI,GAAG;AAAA,MACzC;AAAA,IACF,WAAW,QAAQ,cAAc;AAC/B,YAAM,KAAK,GAAG,aAAa,IAAI;AAC/B,UAAI,MAAM,MAAM;AACd,cAAM,KAAuB;AAAA,UAC3B,MAAM,UAAU,qBAAqB,IAAI,YAAY,CAAC;AAAA,UACtD,OAAO,UAAU,qBAAqB,IAAI,aAAa,CAAC;AAAA,UACxD,KAAK,UAAU,qBAAqB,IAAI,WAAW,CAAC;AAAA,UACpD,QAAQ,UAAU,qBAAqB,IAAI,cAAc,CAAC;AAAA,QAC5D;AACA,cAAM,YAAY,qBAAqB,IAAI,WAAW;AACtD,cAAM,WAAW,YAAY,qBAAqB,WAAW,UAAU,IAAI;AAC3E,cAAM,OAAO,UAAU,aAAa,WAAW;AAC/C,YAAI,QAAQ,KAAK,YAAY,MAAM,OAAQ,IAAG,OAAO;AACrD,eAAO,WAAW,IAAI,IAAI,EAAE;AAAA,MAC9B;AAAA,IACF;AACA,UAAM,WAAW,GAAG;AACpB,aAAS,IAAI,GAAG,IAAI,SAAS,QAAQ,KAAK;AACxC,YAAM,KAAK,SAAS,CAAC;AACrB,UAAI,GAAG,aAAa,EAAG,MAAK,EAAa;AAAA,IAC3C;AAAA,EACF;AACA,OAAK,IAAI;AACT,aAAW,CAAC,IAAI,GAAG,KAAK,YAAY;AAClC,UAAM,IAAI,OAAO,SAAS,IAAI,EAAE;AAChC,QAAI,KAAK,UAAU,IAAI,GAAG,EAAG,GAAE,cAAc;AAAA,EAC/C;AACA,SAAO;AACT;","names":[]}
|
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
"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; }
|
|
2
2
|
|
|
3
|
-
var
|
|
3
|
+
var _chunkEMGF6EEVcjs = require('./chunk-EMGF6EEV.cjs');
|
|
4
4
|
|
|
5
5
|
// src/pdf/formula/models.ts
|
|
6
6
|
var _crypto = require('crypto');
|
|
@@ -23,7 +23,7 @@ function isOfflineMode() {
|
|
|
23
23
|
}
|
|
24
24
|
function assertNetworkAllowed(what, hint) {
|
|
25
25
|
if (!isOfflineMode()) return;
|
|
26
|
-
throw new (0,
|
|
26
|
+
throw new (0, _chunkEMGF6EEVcjs.KordocError)(
|
|
27
27
|
`\uD3D0\uC1C4\uB9DD \uBAA8\uB4DC(KORDOC_OFFLINE)\uC5D0\uC11C \uCC28\uB2E8\uB428: ${what}` + (hint ? ` \u2014 ${hint}` : "")
|
|
28
28
|
);
|
|
29
29
|
}
|
|
@@ -266,4 +266,4 @@ async function sha256OfFile(p) {
|
|
|
266
266
|
|
|
267
267
|
|
|
268
268
|
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;
|
|
269
|
-
//# sourceMappingURL=chunk-
|
|
269
|
+
//# sourceMappingURL=chunk-5VVS2QIJ.cjs.map
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"sources":["/Users/chris_gomdori/workspace/kordoc/dist/chunk-EQZI77PP.cjs","../src/pdf/formula/models.ts","../src/shared/offline.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;ACDxD;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;AAEU,IAAA;AACpC,sBAAA;AACX,QAAA;AACY,QAAA;AACL,QAAA;AACA,QAAA;AACE,QAAA;AACV,MAAA;AACD,MAAA;AACF,IAAA;AAGI,IAAA;AACoB,MAAA;AAChB,IAAA;AAER,IAAA;AAEgD,IAAA;AAClD,EAAA;AACF;AAGuF;AACV,EAAA;AAC7E;AAGsG;AACpE,EAAA;AACI,EAAA;AACK,EAAA;AACU,EAAA;AACe,oBAAA;AAChE,IAAA;AACF,EAAA;AACI,EAAA;AACoB,IAAA;AAChB,EAAA;AAER,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;AAG6B,EAAA;AACsB,EAAA;AAEhB,EAAA;AACxB,IAAA;AAAA;AAEO,MAAA;AAChB,IAAA;AACD,EAAA;AAC2B,EAAA;AAChB,IAAA;AAC0C,MAAA;AACpD,IAAA;AACF,EAAA;AAEmD,EAAA;AACQ,EAAA;AAC1C,EAAA;AAEoB,EAAA;AACjC,EAAA;AACyF,IAAA;AACzC,IAAA;AAC5B,MAAA;AACP,sBAAA;AACX,QAAA;AACA,QAAA;AACA,QAAA;AACO,QAAA;AACR,MAAA;AACF,IAAA;AACwB,IAAA;AACf,EAAA;AACN,IAAA;AACmB,MAAA;AACf,IAAA;AAER,IAAA;AAC+D,IAAA;AACjE,EAAA;AAEa,kBAAA;AACX,IAAA;AACA,IAAA;AACA,IAAA;AACO,IAAA;AACR,EAAA;AAGG,EAAA;AACA,EAAA;AACkC,IAAA;AAC1B,EAAA;AACN,IAAA;AACmB,MAAA;AACf,IAAA;AAER,IAAA;AACiE,IAAA;AACnE,EAAA;AAE4B,EAAA;AACtB,IAAA;AACmB,MAAA;AACf,IAAA;AAER,IAAA;AACU,IAAA;AAC4D,MAAA;AACtE,IAAA;AACF,EAAA;AAEgC,EAAA;AACnB,kBAAA;AACX,IAAA;AACA,IAAA;AACA,IAAA;AACO,IAAA;AACR,EAAA;AACH;AAEwD;AACzB,EAAA;AACI,EAAA;AACY,EAAA;AACZ,IAAA;AACf,MAAA;AAEhB,IAAA;AACD,EAAA;AACoB,EAAA;AACvB;AD1E4F;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA","file":"/Users/chris_gomdori/workspace/kordoc/dist/chunk-EQZI77PP.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\"\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\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 // 기존 파일 있지만 SHA 불일치 → 삭제\n try {\n await unlink(localPath)\n } catch {\n // 없을 수 있음\n }\n\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 if (await isExistingValid(localPath, spec.sha256)) {\n onProgress?.({ spec, downloaded: 0, total: null, phase: \"skip\" })\n return\n }\n try {\n await unlink(localPath)\n } catch {\n // 재다운로드 전 손상/잔존 파일 정리 best-effort — 애초에 없으면 실패해도 무방\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 const partPath = `${localPath}.part`\n await mkdir(dirname(localPath), { recursive: true })\n\n const resp = await fetch(spec.url, {\n headers: {\n // HF CDN 은 UA 없으면 가끔 403 을 뱉는다\n \"User-Agent\": \"kordoc-formula-ocr/1.0 (+https://github.com/chrisryugj/kordoc)\",\n },\n })\n if (!resp.ok || !resp.body) {\n throw new Error(\n `${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 try {\n await unlink(partPath)\n } catch {\n // .part 임시파일 정리 best-effort — 본 에러(스트리밍 실패)를 가리지 않도록 무시\n }\n throw new Error(`${spec.name} 스트리밍 실패: ${(e as Error).message}`)\n }\n\n onProgress?.({\n spec,\n downloaded,\n total,\n phase: \"verify\",\n })\n\n // SHA 검증\n let actual: string\n try {\n actual = await sha256OfFile(partPath)\n } catch (e) {\n try {\n await unlink(partPath)\n } catch {\n // .part 임시파일 정리 best-effort — 본 에러(SHA 계산 실패)를 가리지 않도록 무시\n }\n throw new Error(`${spec.name} SHA 계산 실패: ${(e as Error).message}`)\n }\n\n if (actual !== spec.sha256) {\n try {\n await unlink(partPath)\n } catch {\n // 오염된 .part 정리 best-effort — 본 에러(SHA mismatch)를 가리지 않도록 무시\n }\n throw new Error(\n `${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}\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"]}
|
|
1
|
+
{"version":3,"sources":["/Users/chris_gomdori/workspace/kordoc/dist/chunk-5VVS2QIJ.cjs","../src/pdf/formula/models.ts","../src/shared/offline.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;ACDxD;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;AAEU,IAAA;AACpC,sBAAA;AACX,QAAA;AACY,QAAA;AACL,QAAA;AACA,QAAA;AACE,QAAA;AACV,MAAA;AACD,MAAA;AACF,IAAA;AAGI,IAAA;AACoB,MAAA;AAChB,IAAA;AAER,IAAA;AAEgD,IAAA;AAClD,EAAA;AACF;AAGuF;AACV,EAAA;AAC7E;AAGsG;AACpE,EAAA;AACI,EAAA;AACK,EAAA;AACU,EAAA;AACe,oBAAA;AAChE,IAAA;AACF,EAAA;AACI,EAAA;AACoB,IAAA;AAChB,EAAA;AAER,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;AAG6B,EAAA;AACsB,EAAA;AAEhB,EAAA;AACxB,IAAA;AAAA;AAEO,MAAA;AAChB,IAAA;AACD,EAAA;AAC2B,EAAA;AAChB,IAAA;AAC0C,MAAA;AACpD,IAAA;AACF,EAAA;AAEmD,EAAA;AACQ,EAAA;AAC1C,EAAA;AAEoB,EAAA;AACjC,EAAA;AACyF,IAAA;AACzC,IAAA;AAC5B,MAAA;AACP,sBAAA;AACX,QAAA;AACA,QAAA;AACA,QAAA;AACO,QAAA;AACR,MAAA;AACF,IAAA;AACwB,IAAA;AACf,EAAA;AACN,IAAA;AACmB,MAAA;AACf,IAAA;AAER,IAAA;AAC+D,IAAA;AACjE,EAAA;AAEa,kBAAA;AACX,IAAA;AACA,IAAA;AACA,IAAA;AACO,IAAA;AACR,EAAA;AAGG,EAAA;AACA,EAAA;AACkC,IAAA;AAC1B,EAAA;AACN,IAAA;AACmB,MAAA;AACf,IAAA;AAER,IAAA;AACiE,IAAA;AACnE,EAAA;AAE4B,EAAA;AACtB,IAAA;AACmB,MAAA;AACf,IAAA;AAER,IAAA;AACU,IAAA;AAC4D,MAAA;AACtE,IAAA;AACF,EAAA;AAEgC,EAAA;AACnB,kBAAA;AACX,IAAA;AACA,IAAA;AACA,IAAA;AACO,IAAA;AACR,EAAA;AACH;AAEwD;AACzB,EAAA;AACI,EAAA;AACY,EAAA;AACZ,IAAA;AACf,MAAA;AAEhB,IAAA;AACD,EAAA;AACoB,EAAA;AACvB;AD1E4F;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA","file":"/Users/chris_gomdori/workspace/kordoc/dist/chunk-5VVS2QIJ.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\"\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\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 // 기존 파일 있지만 SHA 불일치 → 삭제\n try {\n await unlink(localPath)\n } catch {\n // 없을 수 있음\n }\n\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 if (await isExistingValid(localPath, spec.sha256)) {\n onProgress?.({ spec, downloaded: 0, total: null, phase: \"skip\" })\n return\n }\n try {\n await unlink(localPath)\n } catch {\n // 재다운로드 전 손상/잔존 파일 정리 best-effort — 애초에 없으면 실패해도 무방\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 const partPath = `${localPath}.part`\n await mkdir(dirname(localPath), { recursive: true })\n\n const resp = await fetch(spec.url, {\n headers: {\n // HF CDN 은 UA 없으면 가끔 403 을 뱉는다\n \"User-Agent\": \"kordoc-formula-ocr/1.0 (+https://github.com/chrisryugj/kordoc)\",\n },\n })\n if (!resp.ok || !resp.body) {\n throw new Error(\n `${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 try {\n await unlink(partPath)\n } catch {\n // .part 임시파일 정리 best-effort — 본 에러(스트리밍 실패)를 가리지 않도록 무시\n }\n throw new Error(`${spec.name} 스트리밍 실패: ${(e as Error).message}`)\n }\n\n onProgress?.({\n spec,\n downloaded,\n total,\n phase: \"verify\",\n })\n\n // SHA 검증\n let actual: string\n try {\n actual = await sha256OfFile(partPath)\n } catch (e) {\n try {\n await unlink(partPath)\n } catch {\n // .part 임시파일 정리 best-effort — 본 에러(SHA 계산 실패)를 가리지 않도록 무시\n }\n throw new Error(`${spec.name} SHA 계산 실패: ${(e as Error).message}`)\n }\n\n if (actual !== spec.sha256) {\n try {\n await unlink(partPath)\n } catch {\n // 오염된 .part 정리 best-effort — 본 에러(SHA mismatch)를 가리지 않도록 무시\n }\n throw new Error(\n `${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}\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"]}
|