@effortlessmotion/pdf-to-md 0.1.1 → 0.1.2

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.
Files changed (53) hide show
  1. package/dist/cjs/analyze.js.map +1 -1
  2. package/dist/cjs/audit.js.map +1 -1
  3. package/dist/cjs/cli/audit.js.map +1 -1
  4. package/dist/cjs/cli/extract.js.map +1 -1
  5. package/dist/cjs/cli/index.js.map +1 -1
  6. package/dist/cjs/cli/render.js.map +1 -1
  7. package/dist/cjs/cli/types.js.map +1 -1
  8. package/dist/cjs/cli.js.map +1 -1
  9. package/dist/cjs/collect.js.map +1 -1
  10. package/dist/cjs/dict.js.map +1 -1
  11. package/dist/cjs/emit.js.map +1 -1
  12. package/dist/cjs/engines/tesseract.js.map +1 -1
  13. package/dist/cjs/errors.js.map +1 -1
  14. package/dist/cjs/extract.js.map +1 -1
  15. package/dist/cjs/geometry.js.map +1 -1
  16. package/dist/cjs/index.js.map +1 -1
  17. package/dist/cjs/lines.js.map +1 -1
  18. package/dist/cjs/markdown.js.map +1 -1
  19. package/dist/cjs/page-marker.js.map +1 -1
  20. package/dist/cjs/pdf-open.js.map +1 -1
  21. package/dist/cjs/raster.js.map +1 -1
  22. package/dist/cjs/render-analysis.js.map +1 -1
  23. package/dist/cjs/report.js.map +1 -1
  24. package/dist/cjs/resolve.js.map +1 -1
  25. package/dist/cjs/types.js.map +1 -1
  26. package/dist/cjs/validate.js.map +1 -1
  27. package/dist/esm/analyze.js.map +1 -1
  28. package/dist/esm/audit.js.map +1 -1
  29. package/dist/esm/cli/audit.js.map +1 -1
  30. package/dist/esm/cli/extract.js.map +1 -1
  31. package/dist/esm/cli/index.js.map +1 -1
  32. package/dist/esm/cli/render.js.map +1 -1
  33. package/dist/esm/cli/types.js.map +1 -1
  34. package/dist/esm/cli.js.map +1 -1
  35. package/dist/esm/collect.js.map +1 -1
  36. package/dist/esm/dict.js.map +1 -1
  37. package/dist/esm/emit.js.map +1 -1
  38. package/dist/esm/engines/tesseract.js.map +1 -1
  39. package/dist/esm/errors.js.map +1 -1
  40. package/dist/esm/extract.js.map +1 -1
  41. package/dist/esm/geometry.js.map +1 -1
  42. package/dist/esm/index.js.map +1 -1
  43. package/dist/esm/lines.js.map +1 -1
  44. package/dist/esm/markdown.js.map +1 -1
  45. package/dist/esm/page-marker.js.map +1 -1
  46. package/dist/esm/pdf-open.js.map +1 -1
  47. package/dist/esm/raster.js.map +1 -1
  48. package/dist/esm/render-analysis.js.map +1 -1
  49. package/dist/esm/report.js.map +1 -1
  50. package/dist/esm/resolve.js.map +1 -1
  51. package/dist/esm/types.js.map +1 -1
  52. package/dist/esm/validate.js.map +1 -1
  53. package/package.json +10 -1
@@ -1 +1 @@
1
- {"version":3,"sources":["/Users/kevin/Projects/pdf-to-md/src/cli.ts"],"sourcesContent":["// Parsing and dispatch only. Commands live in src/cli/, one file each,\n// lazy-loaded — nothing tree- or dependency-heavy may be imported at the top of this file (`render` must never load the OCR engine, `--version` must load almost nothing).\nimport { readFileSync } from 'node:fs';\nimport path from 'node:path';\nimport { fileURLToPath } from 'node:url';\nimport { COMMANDS } from './cli/index.ts';\nimport type { Ctx } from './cli/types.ts';\n\n// __dirname (CJS) or its ESM equivalent, then two fixed hops up to the\n// package root. NOT `import.meta.resolve('pdf-to-md/package.json')` — Node's self-reference lookup throws ERR_MODULE_NOT_FOUND from a file this deep; a location-relative path is stable across both build targets.\nconst __dirname = path.dirname(typeof __filename !== 'undefined' ? __filename : fileURLToPath(import.meta.url));\n\nfunction packageVersion(): string {\n try {\n const pkgPath = path.resolve(__dirname, '..', '..', 'package.json');\n const pkg = JSON.parse(readFileSync(pkgPath, 'utf8')) as { version?: string };\n return pkg.version ?? 'unknown';\n } catch {\n return 'unknown';\n }\n}\n\n// Exported so the comparison is unit-testable without spawning a subprocess\n// with a faked Node version. null means unparseable input, not \"not below\".\nexport function isBelowNodeFloor(version: string, floor: string): boolean | null {\n const parts = (v: string) => v.replace(/^v/, '').replace(/^>=/, '').split('.').map(Number);\n const v = parts(version);\n const f = parts(floor);\n if (v.length !== 3 || f.length !== 3 || v.some(Number.isNaN) || f.some(Number.isNaN)) return null;\n const [rMaj, rMin, rPat] = v;\n const [fMaj, fMin, fPat] = f;\n return rMaj !== fMaj ? rMaj < fMaj : rMin !== fMin ? rMin < fMin : rPat < fPat;\n}\n\n// Reads engines.node from package.json (kept as the single source of truth,\n// not hardcoded here) and returns it iff process.version is below it.\nfunction belowNodeFloor(version = process.version): string | null {\n try {\n const pkgPath = path.resolve(__dirname, '..', '..', 'package.json');\n const pkg = JSON.parse(readFileSync(pkgPath, 'utf8')) as { engines?: { node?: string } };\n const floor = pkg.engines?.node;\n if (!floor) return null;\n return isBelowNodeFloor(version, floor) ? floor.replace(/^>=/, '') : null;\n } catch {\n return null;\n }\n}\n\nfunction usage(name: string): string {\n return (\n `usage: ${name} <file.pdf> [out.md] [flags...] convert a PDF to markdown\\n` +\n ` ${name} extract <file.pdf> [out.md] [flags...] (same, explicit)\\n` +\n ` ${name} render <file.pdf> <page>[-<page>] [--dpi|-d N (default 288)] [--out|-o DIR (default .)]\\n` +\n ` ${name} audit <words.jsonl> <file.md> severity-1 word-loss check\\n` +\n ` ${name} --version | --help\\n` +\n '\\n' +\n 'extract flags:\\n' +\n ' --ocr force the OCR path (image-only/scanned pages)\\n' +\n ' --no-ocr never fall back to OCR; keep sparse text output\\n' +\n ' --pages N[-M] convert only these 1-based source pages\\n' +\n ' --stdout write markdown to stdout instead of a file\\n' +\n ' --json print the full QA report as JSON\\n' +\n ` --page-markers emit a \"### pN\" heading per page\\n` +\n ' --format md|txt|raw output format (default md; raw = Analysis JSON incl. the\\n' +\n ' report; txt = reading-order plain text)\\n' +\n ' --words-json=FILE OCR path from a recognized-word dump (offline replay)\\n' +\n ' --debug-words=FILE dump every recognized word as JSON lines to FILE\\n' +\n ' --dpi=N OCR render resolution (default 288; raise for tiny print or\\n' +\n ' chart/label pages, >~400 measured worse on ordinary prose)\\n' +\n '\\n' +\n `extract tuning (measured defaults; see README's tuning table):\\n` +\n ' --float-max-words=N --float-max-width=F --float-margin=F --para-gap=F\\n' +\n ' --heading-scale=F --col-max-width=F --col-split=F --line-y-tol=F\\n' +\n ' --line-height-ratio=F\\n' +\n '\\n' +\n 'examples:\\n' +\n ` ${name} book.pdf # text layer -> book.md (auto-OCR if scanned)\\n` +\n ` ${name} scan.pdf out.md --ocr --json # force OCR, print the QA report\\n` +\n ` ${name} book.pdf --pages 27-31 --stdout # re-extract a page range to stdout\\n` +\n ` ${name} render book.pdf 27 # render p27 to PNG for eyeballing (288 dpi)\\n` +\n ` ${name} audit words.jsonl book.md # check no recognized word was lost`\n );\n}\n\n// Thrown errors -> exit 1 with the message verbatim; usage errors exit 2.\nexport default async function cli(argv: string[], name: string): Promise<void> {\n const floor = belowNodeFloor();\n if (floor) console.error(`${name}: Node ${process.version} is below the required >=${floor} — the OCR path can silently find no words instead of erroring; upgrade Node.`);\n\n if (argv[0] === '--version' || argv[0] === '-v') {\n console.log(packageVersion());\n return;\n }\n if (argv.length === 0 || argv[0] === '--help' || argv[0] === '-h') {\n // No args is a usage ERROR (bare `pdf-to-md` can't do anything); asking\n // for --help is not.\n if (argv.length === 0) {\n console.error(usage(name));\n process.exit(2);\n }\n console.log(`${name} v${packageVersion()}`);\n console.log('');\n console.log(usage(name));\n return;\n }\n\n const ctx: Ctx = {\n name,\n rest: argv.slice(1),\n usageError(message) {\n console.error(message);\n process.exit(2);\n },\n };\n\n try {\n const load = COMMANDS[argv[0]];\n if (load) {\n await (await load()).default(ctx);\n } else {\n // Compat shim: `pdf-to-md file.pdf [...]` — any first word that isn't\n // a reserved command is extract's own first argument.\n ctx.rest = argv;\n await (await COMMANDS.extract()).default(ctx);\n }\n } catch (err) {\n console.error((err as Error).message);\n process.exit(1);\n }\n}\n"],"names":["readFileSync","path","fileURLToPath","COMMANDS","__dirname","dirname","__filename","url","packageVersion","pkg","pkgPath","resolve","JSON","parse","version","isBelowNodeFloor","floor","parts","v","replace","split","map","Number","f","length","some","isNaN","rMaj","rMin","rPat","fMaj","fMin","fPat","belowNodeFloor","process","engines","node","usage","name","cli","argv","console","error","log","exit","ctx","rest","slice","usageError","message","load","default","extract","err"],"mappings":"AAAA,uEAAuE;AACvE,2KAA2K;AAC3K,SAASA,YAAY,QAAQ,UAAU;AACvC,OAAOC,UAAU,YAAY;AAC7B,SAASC,aAAa,QAAQ,WAAW;AACzC,SAASC,QAAQ,QAAQ,iBAAiB;AAG1C,uEAAuE;AACvE,oNAAoN;AACpN,MAAMC,YAAYH,KAAKI,OAAO,CAAC,OAAOC,eAAe,cAAcA,aAAaJ,cAAc,YAAYK,GAAG;AAE7G,SAASC;IACP,IAAI;YAGKC;QAFP,MAAMC,UAAUT,KAAKU,OAAO,CAACP,WAAW,MAAM,MAAM;QACpD,MAAMK,MAAMG,KAAKC,KAAK,CAACb,aAAaU,SAAS;QAC7C,QAAOD,eAAAA,IAAIK,OAAO,cAAXL,0BAAAA,eAAe;IACxB,EAAE,OAAM;QACN,OAAO;IACT;AACF;AAEA,4EAA4E;AAC5E,4EAA4E;AAC5E,OAAO,SAASM,iBAAiBD,OAAe,EAAEE,KAAa;IAC7D,MAAMC,QAAQ,CAACC,IAAcA,EAAEC,OAAO,CAAC,MAAM,IAAIA,OAAO,CAAC,OAAO,IAAIC,KAAK,CAAC,KAAKC,GAAG,CAACC;IACnF,MAAMJ,IAAID,MAAMH;IAChB,MAAMS,IAAIN,MAAMD;IAChB,IAAIE,EAAEM,MAAM,KAAK,KAAKD,EAAEC,MAAM,KAAK,KAAKN,EAAEO,IAAI,CAACH,OAAOI,KAAK,KAAKH,EAAEE,IAAI,CAACH,OAAOI,KAAK,GAAG,OAAO;IAC7F,MAAM,CAACC,MAAMC,MAAMC,KAAK,GAAGX;IAC3B,MAAM,CAACY,MAAMC,MAAMC,KAAK,GAAGT;IAC3B,OAAOI,SAASG,OAAOH,OAAOG,OAAOF,SAASG,OAAOH,OAAOG,OAAOF,OAAOG;AAC5E;AAEA,4EAA4E;AAC5E,sEAAsE;AACtE,SAASC,eAAenB,UAAUoB,QAAQpB,OAAO;IAC/C,IAAI;YAGYL;QAFd,MAAMC,UAAUT,KAAKU,OAAO,CAACP,WAAW,MAAM,MAAM;QACpD,MAAMK,MAAMG,KAAKC,KAAK,CAACb,aAAaU,SAAS;QAC7C,MAAMM,SAAQP,eAAAA,IAAI0B,OAAO,cAAX1B,mCAAAA,aAAa2B,IAAI;QAC/B,IAAI,CAACpB,OAAO,OAAO;QACnB,OAAOD,iBAAiBD,SAASE,SAASA,MAAMG,OAAO,CAAC,OAAO,MAAM;IACvE,EAAE,OAAM;QACN,OAAO;IACT;AACF;AAEA,SAASkB,MAAMC,IAAY;IACzB,OACE,CAAC,OAAO,EAAEA,KAAK,kEAAkE,CAAC,GAClF,CAAC,OAAO,EAAEA,KAAK,0DAA0D,CAAC,GAC1E,CAAC,OAAO,EAAEA,KAAK,0FAA0F,CAAC,GAC1G,CAAC,OAAO,EAAEA,KAAK,oEAAoE,CAAC,GACpF,CAAC,OAAO,EAAEA,KAAK,qBAAqB,CAAC,GACrC,OACA,qBACA,2EACA,6EACA,qEACA,wEACA,8DACA,CAAC,yDAAyD,CAAC,GAC3D,sFACA,qEACA,mFACA,8EACA,yFACA,wFACA,OACA,CAAC,gEAAgE,CAAC,GAClE,8EACA,yEACA,8BACA,OACA,gBACA,CAAC,EAAE,EAAEA,KAAK,gFAAgF,CAAC,GAC3F,CAAC,EAAE,EAAEA,KAAK,mEAAmE,CAAC,GAC9E,CAAC,EAAE,EAAEA,KAAK,sEAAsE,CAAC,GACjF,CAAC,EAAE,EAAEA,KAAK,+EAA+E,CAAC,GAC1F,CAAC,EAAE,EAAEA,KAAK,oEAAoE,CAAC;AAEnF;AAEA,0EAA0E;AAC1E,eAAe,eAAeC,IAAIC,IAAc,EAAEF,IAAY;IAC5D,MAAMtB,QAAQiB;IACd,IAAIjB,OAAOyB,QAAQC,KAAK,CAAC,GAAGJ,KAAK,OAAO,EAAEJ,QAAQpB,OAAO,CAAC,yBAAyB,EAAEE,MAAM,6EAA6E,CAAC;IAEzK,IAAIwB,IAAI,CAAC,EAAE,KAAK,eAAeA,IAAI,CAAC,EAAE,KAAK,MAAM;QAC/CC,QAAQE,GAAG,CAACnC;QACZ;IACF;IACA,IAAIgC,KAAKhB,MAAM,KAAK,KAAKgB,IAAI,CAAC,EAAE,KAAK,YAAYA,IAAI,CAAC,EAAE,KAAK,MAAM;QACjE,wEAAwE;QACxE,qBAAqB;QACrB,IAAIA,KAAKhB,MAAM,KAAK,GAAG;YACrBiB,QAAQC,KAAK,CAACL,MAAMC;YACpBJ,QAAQU,IAAI,CAAC;QACf;QACAH,QAAQE,GAAG,CAAC,GAAGL,KAAK,EAAE,EAAE9B,kBAAkB;QAC1CiC,QAAQE,GAAG,CAAC;QACZF,QAAQE,GAAG,CAACN,MAAMC;QAClB;IACF;IAEA,MAAMO,MAAW;QACfP;QACAQ,MAAMN,KAAKO,KAAK,CAAC;QACjBC,YAAWC,OAAO;YAChBR,QAAQC,KAAK,CAACO;YACdf,QAAQU,IAAI,CAAC;QACf;IACF;IAEA,IAAI;QACF,MAAMM,OAAO/C,QAAQ,CAACqC,IAAI,CAAC,EAAE,CAAC;QAC9B,IAAIU,MAAM;YACR,MAAM,AAAC,CAAA,MAAMA,MAAK,EAAGC,OAAO,CAACN;QAC/B,OAAO;YACL,sEAAsE;YACtE,sDAAsD;YACtDA,IAAIC,IAAI,GAAGN;YACX,MAAM,AAAC,CAAA,MAAMrC,SAASiD,OAAO,EAAC,EAAGD,OAAO,CAACN;QAC3C;IACF,EAAE,OAAOQ,KAAK;QACZZ,QAAQC,KAAK,CAAC,AAACW,IAAcJ,OAAO;QACpCf,QAAQU,IAAI,CAAC;IACf;AACF"}
1
+ {"version":3,"sources":["/Users/kevin/Dev/OpenSource/formats/pdf-to-md/src/cli.ts"],"sourcesContent":["// Parsing and dispatch only. Commands live in src/cli/, one file each,\n// lazy-loaded — nothing tree- or dependency-heavy may be imported at the top of this file (`render` must never load the OCR engine, `--version` must load almost nothing).\nimport { readFileSync } from 'node:fs';\nimport path from 'node:path';\nimport { fileURLToPath } from 'node:url';\nimport { COMMANDS } from './cli/index.ts';\nimport type { Ctx } from './cli/types.ts';\n\n// __dirname (CJS) or its ESM equivalent, then two fixed hops up to the\n// package root. NOT `import.meta.resolve('pdf-to-md/package.json')` — Node's self-reference lookup throws ERR_MODULE_NOT_FOUND from a file this deep; a location-relative path is stable across both build targets.\nconst __dirname = path.dirname(typeof __filename !== 'undefined' ? __filename : fileURLToPath(import.meta.url));\n\nfunction packageVersion(): string {\n try {\n const pkgPath = path.resolve(__dirname, '..', '..', 'package.json');\n const pkg = JSON.parse(readFileSync(pkgPath, 'utf8')) as { version?: string };\n return pkg.version ?? 'unknown';\n } catch {\n return 'unknown';\n }\n}\n\n// Exported so the comparison is unit-testable without spawning a subprocess\n// with a faked Node version. null means unparseable input, not \"not below\".\nexport function isBelowNodeFloor(version: string, floor: string): boolean | null {\n const parts = (v: string) => v.replace(/^v/, '').replace(/^>=/, '').split('.').map(Number);\n const v = parts(version);\n const f = parts(floor);\n if (v.length !== 3 || f.length !== 3 || v.some(Number.isNaN) || f.some(Number.isNaN)) return null;\n const [rMaj, rMin, rPat] = v;\n const [fMaj, fMin, fPat] = f;\n return rMaj !== fMaj ? rMaj < fMaj : rMin !== fMin ? rMin < fMin : rPat < fPat;\n}\n\n// Reads engines.node from package.json (kept as the single source of truth,\n// not hardcoded here) and returns it iff process.version is below it.\nfunction belowNodeFloor(version = process.version): string | null {\n try {\n const pkgPath = path.resolve(__dirname, '..', '..', 'package.json');\n const pkg = JSON.parse(readFileSync(pkgPath, 'utf8')) as { engines?: { node?: string } };\n const floor = pkg.engines?.node;\n if (!floor) return null;\n return isBelowNodeFloor(version, floor) ? floor.replace(/^>=/, '') : null;\n } catch {\n return null;\n }\n}\n\nfunction usage(name: string): string {\n return (\n `usage: ${name} <file.pdf> [out.md] [flags...] convert a PDF to markdown\\n` +\n ` ${name} extract <file.pdf> [out.md] [flags...] (same, explicit)\\n` +\n ` ${name} render <file.pdf> <page>[-<page>] [--dpi|-d N (default 288)] [--out|-o DIR (default .)]\\n` +\n ` ${name} audit <words.jsonl> <file.md> severity-1 word-loss check\\n` +\n ` ${name} --version | --help\\n` +\n '\\n' +\n 'extract flags:\\n' +\n ' --ocr force the OCR path (image-only/scanned pages)\\n' +\n ' --no-ocr never fall back to OCR; keep sparse text output\\n' +\n ' --pages N[-M] convert only these 1-based source pages\\n' +\n ' --stdout write markdown to stdout instead of a file\\n' +\n ' --json print the full QA report as JSON\\n' +\n ` --page-markers emit a \"### pN\" heading per page\\n` +\n ' --format md|txt|raw output format (default md; raw = Analysis JSON incl. the\\n' +\n ' report; txt = reading-order plain text)\\n' +\n ' --words-json=FILE OCR path from a recognized-word dump (offline replay)\\n' +\n ' --debug-words=FILE dump every recognized word as JSON lines to FILE\\n' +\n ' --dpi=N OCR render resolution (default 288; raise for tiny print or\\n' +\n ' chart/label pages, >~400 measured worse on ordinary prose)\\n' +\n '\\n' +\n `extract tuning (measured defaults; see README's tuning table):\\n` +\n ' --float-max-words=N --float-max-width=F --float-margin=F --para-gap=F\\n' +\n ' --heading-scale=F --col-max-width=F --col-split=F --line-y-tol=F\\n' +\n ' --line-height-ratio=F\\n' +\n '\\n' +\n 'examples:\\n' +\n ` ${name} book.pdf # text layer -> book.md (auto-OCR if scanned)\\n` +\n ` ${name} scan.pdf out.md --ocr --json # force OCR, print the QA report\\n` +\n ` ${name} book.pdf --pages 27-31 --stdout # re-extract a page range to stdout\\n` +\n ` ${name} render book.pdf 27 # render p27 to PNG for eyeballing (288 dpi)\\n` +\n ` ${name} audit words.jsonl book.md # check no recognized word was lost`\n );\n}\n\n// Thrown errors -> exit 1 with the message verbatim; usage errors exit 2.\nexport default async function cli(argv: string[], name: string): Promise<void> {\n const floor = belowNodeFloor();\n if (floor) console.error(`${name}: Node ${process.version} is below the required >=${floor} — the OCR path can silently find no words instead of erroring; upgrade Node.`);\n\n if (argv[0] === '--version' || argv[0] === '-v') {\n console.log(packageVersion());\n return;\n }\n if (argv.length === 0 || argv[0] === '--help' || argv[0] === '-h') {\n // No args is a usage ERROR (bare `pdf-to-md` can't do anything); asking\n // for --help is not.\n if (argv.length === 0) {\n console.error(usage(name));\n process.exit(2);\n }\n console.log(`${name} v${packageVersion()}`);\n console.log('');\n console.log(usage(name));\n return;\n }\n\n const ctx: Ctx = {\n name,\n rest: argv.slice(1),\n usageError(message) {\n console.error(message);\n process.exit(2);\n },\n };\n\n try {\n const load = COMMANDS[argv[0]];\n if (load) {\n await (await load()).default(ctx);\n } else {\n // Compat shim: `pdf-to-md file.pdf [...]` — any first word that isn't\n // a reserved command is extract's own first argument.\n ctx.rest = argv;\n await (await COMMANDS.extract()).default(ctx);\n }\n } catch (err) {\n console.error((err as Error).message);\n process.exit(1);\n }\n}\n"],"names":["readFileSync","path","fileURLToPath","COMMANDS","__dirname","dirname","__filename","url","packageVersion","pkg","pkgPath","resolve","JSON","parse","version","isBelowNodeFloor","floor","parts","v","replace","split","map","Number","f","length","some","isNaN","rMaj","rMin","rPat","fMaj","fMin","fPat","belowNodeFloor","process","engines","node","usage","name","cli","argv","console","error","log","exit","ctx","rest","slice","usageError","message","load","default","extract","err"],"mappings":"AAAA,uEAAuE;AACvE,2KAA2K;AAC3K,SAASA,YAAY,QAAQ,UAAU;AACvC,OAAOC,UAAU,YAAY;AAC7B,SAASC,aAAa,QAAQ,WAAW;AACzC,SAASC,QAAQ,QAAQ,iBAAiB;AAG1C,uEAAuE;AACvE,oNAAoN;AACpN,MAAMC,YAAYH,KAAKI,OAAO,CAAC,OAAOC,eAAe,cAAcA,aAAaJ,cAAc,YAAYK,GAAG;AAE7G,SAASC;IACP,IAAI;YAGKC;QAFP,MAAMC,UAAUT,KAAKU,OAAO,CAACP,WAAW,MAAM,MAAM;QACpD,MAAMK,MAAMG,KAAKC,KAAK,CAACb,aAAaU,SAAS;QAC7C,QAAOD,eAAAA,IAAIK,OAAO,cAAXL,0BAAAA,eAAe;IACxB,EAAE,OAAM;QACN,OAAO;IACT;AACF;AAEA,4EAA4E;AAC5E,4EAA4E;AAC5E,OAAO,SAASM,iBAAiBD,OAAe,EAAEE,KAAa;IAC7D,MAAMC,QAAQ,CAACC,IAAcA,EAAEC,OAAO,CAAC,MAAM,IAAIA,OAAO,CAAC,OAAO,IAAIC,KAAK,CAAC,KAAKC,GAAG,CAACC;IACnF,MAAMJ,IAAID,MAAMH;IAChB,MAAMS,IAAIN,MAAMD;IAChB,IAAIE,EAAEM,MAAM,KAAK,KAAKD,EAAEC,MAAM,KAAK,KAAKN,EAAEO,IAAI,CAACH,OAAOI,KAAK,KAAKH,EAAEE,IAAI,CAACH,OAAOI,KAAK,GAAG,OAAO;IAC7F,MAAM,CAACC,MAAMC,MAAMC,KAAK,GAAGX;IAC3B,MAAM,CAACY,MAAMC,MAAMC,KAAK,GAAGT;IAC3B,OAAOI,SAASG,OAAOH,OAAOG,OAAOF,SAASG,OAAOH,OAAOG,OAAOF,OAAOG;AAC5E;AAEA,4EAA4E;AAC5E,sEAAsE;AACtE,SAASC,eAAenB,UAAUoB,QAAQpB,OAAO;IAC/C,IAAI;YAGYL;QAFd,MAAMC,UAAUT,KAAKU,OAAO,CAACP,WAAW,MAAM,MAAM;QACpD,MAAMK,MAAMG,KAAKC,KAAK,CAACb,aAAaU,SAAS;QAC7C,MAAMM,SAAQP,eAAAA,IAAI0B,OAAO,cAAX1B,mCAAAA,aAAa2B,IAAI;QAC/B,IAAI,CAACpB,OAAO,OAAO;QACnB,OAAOD,iBAAiBD,SAASE,SAASA,MAAMG,OAAO,CAAC,OAAO,MAAM;IACvE,EAAE,OAAM;QACN,OAAO;IACT;AACF;AAEA,SAASkB,MAAMC,IAAY;IACzB,OACE,CAAC,OAAO,EAAEA,KAAK,kEAAkE,CAAC,GAClF,CAAC,OAAO,EAAEA,KAAK,0DAA0D,CAAC,GAC1E,CAAC,OAAO,EAAEA,KAAK,0FAA0F,CAAC,GAC1G,CAAC,OAAO,EAAEA,KAAK,oEAAoE,CAAC,GACpF,CAAC,OAAO,EAAEA,KAAK,qBAAqB,CAAC,GACrC,OACA,qBACA,2EACA,6EACA,qEACA,wEACA,8DACA,CAAC,yDAAyD,CAAC,GAC3D,sFACA,qEACA,mFACA,8EACA,yFACA,wFACA,OACA,CAAC,gEAAgE,CAAC,GAClE,8EACA,yEACA,8BACA,OACA,gBACA,CAAC,EAAE,EAAEA,KAAK,gFAAgF,CAAC,GAC3F,CAAC,EAAE,EAAEA,KAAK,mEAAmE,CAAC,GAC9E,CAAC,EAAE,EAAEA,KAAK,sEAAsE,CAAC,GACjF,CAAC,EAAE,EAAEA,KAAK,+EAA+E,CAAC,GAC1F,CAAC,EAAE,EAAEA,KAAK,oEAAoE,CAAC;AAEnF;AAEA,0EAA0E;AAC1E,eAAe,eAAeC,IAAIC,IAAc,EAAEF,IAAY;IAC5D,MAAMtB,QAAQiB;IACd,IAAIjB,OAAOyB,QAAQC,KAAK,CAAC,GAAGJ,KAAK,OAAO,EAAEJ,QAAQpB,OAAO,CAAC,yBAAyB,EAAEE,MAAM,6EAA6E,CAAC;IAEzK,IAAIwB,IAAI,CAAC,EAAE,KAAK,eAAeA,IAAI,CAAC,EAAE,KAAK,MAAM;QAC/CC,QAAQE,GAAG,CAACnC;QACZ;IACF;IACA,IAAIgC,KAAKhB,MAAM,KAAK,KAAKgB,IAAI,CAAC,EAAE,KAAK,YAAYA,IAAI,CAAC,EAAE,KAAK,MAAM;QACjE,wEAAwE;QACxE,qBAAqB;QACrB,IAAIA,KAAKhB,MAAM,KAAK,GAAG;YACrBiB,QAAQC,KAAK,CAACL,MAAMC;YACpBJ,QAAQU,IAAI,CAAC;QACf;QACAH,QAAQE,GAAG,CAAC,GAAGL,KAAK,EAAE,EAAE9B,kBAAkB;QAC1CiC,QAAQE,GAAG,CAAC;QACZF,QAAQE,GAAG,CAACN,MAAMC;QAClB;IACF;IAEA,MAAMO,MAAW;QACfP;QACAQ,MAAMN,KAAKO,KAAK,CAAC;QACjBC,YAAWC,OAAO;YAChBR,QAAQC,KAAK,CAACO;YACdf,QAAQU,IAAI,CAAC;QACf;IACF;IAEA,IAAI;QACF,MAAMM,OAAO/C,QAAQ,CAACqC,IAAI,CAAC,EAAE,CAAC;QAC9B,IAAIU,MAAM;YACR,MAAM,AAAC,CAAA,MAAMA,MAAK,EAAGC,OAAO,CAACN;QAC/B,OAAO;YACL,sEAAsE;YACtE,sDAAsD;YACtDA,IAAIC,IAAI,GAAGN;YACX,MAAM,AAAC,CAAA,MAAMrC,SAASiD,OAAO,EAAC,EAAGD,OAAO,CAACN;QAC3C;IACF,EAAE,OAAOQ,KAAK;QACZZ,QAAQC,KAAK,CAAC,AAACW,IAAcJ,OAAO;QACpCf,QAAQU,IAAI,CAAC;IACf;AACF"}
@@ -1 +1 @@
1
- {"version":3,"sources":["/Users/kevin/Projects/pdf-to-md/src/collect.ts"],"sourcesContent":["// Pass 1: load the PDF and collect positioned glyphs per page, plus the\n// document-wide char-weighted font-height histogram. Heights round to 0.1pt for stable histogram keys; y is flipped so sorting ascending reads top-down.\nimport type { PDFPageProxy } from 'pdfjs-dist/legacy/build/pdf.mjs';\nimport { OPS } from 'pdfjs-dist/legacy/build/pdf.mjs';\nimport { openDocument } from './pdf-open.ts';\nimport type { CollectedPage, CollectResult, Glyph } from './types.ts';\n\n// HYBRID-PAGE DETECTION (README's \"Design contract\"): a page can carry a\n// real text layer AND a separately-painted large image with its OWN baked-in text that `page.getTextContent()` never sees.\n//\n// Signal: scan the operator list for `OPS.paintImageXObject`, whose args are\n// `[objId, width, height]` — available synchronously, `large` is >=200x200px. Masks/nested Form XObjects aren't walked (accepted gap); missing dims fall back to \"op present\" as the signal.\nconst LARGE_IMAGE_MIN_PX = 200;\nasync function hasLargeImage(page: PDFPageProxy): Promise<boolean> {\n const { fnArray, argsArray } = await page.getOperatorList();\n for (let i = 0; i < fnArray.length; i++) {\n if (fnArray[i] !== OPS.paintImageXObject) continue;\n const [, w, h] = argsArray[i];\n if (typeof w === 'number' && typeof h === 'number') {\n if (w >= LARGE_IMAGE_MIN_PX && h >= LARGE_IMAGE_MIN_PX) return true;\n } else return true; // dims unavailable: coarser \"op present\" signal\n }\n return false;\n}\n\n// { first, last }: 1-based, inclusive page range (the CLI's --pages).\n// Out-of-range values CLAMP to the document's bounds; `page` on each returned entry is always the SOURCE page number, so markers/splicing stay correct.\n//\n// CAVEAT: `heightChars` is built ONLY from the pages actually collected, so\n// a `--pages` subset's heading levels can differ from a full run's — accepted since re-extracted pages get spliced back in for their prose, not diffed.\n\n// Cheap page-count-only load (no per-page text extraction) — the OCR path\n// needs it to resolve its own [first, last] clamping and print a \"### pN\" marker for every page in range, including ones with no recognized words.\nexport async function pageCount(src: string): Promise<number> {\n const doc = await openDocument(src);\n return doc.numPages;\n}\n\nexport async function collect(src: string, { first, last }: { first?: number; last?: number } = {}): Promise<CollectResult> {\n const doc = await openDocument(src);\n const from = Math.min(Math.max(1, first ?? 1), Math.max(1, doc.numPages));\n const to = Math.min(Math.max(from, last ?? doc.numPages), doc.numPages);\n const pages: CollectedPage[] = [];\n const heightChars = new Map<number, number>();\n for (let p = from; p <= to; p++) {\n const page = await doc.getPage(p);\n const { items } = await page.getTextContent();\n const vp = page.getViewport({ scale: 1 });\n const glyphs: Glyph[] = [];\n let textChars = 0;\n for (const it of items) {\n if (!('str' in it) || typeof it.str !== 'string') continue;\n textChars += it.str.length;\n if (!it.str.trim()) continue;\n const h = Math.round(Math.abs(it.transform[3]) * 10) / 10;\n glyphs.push({\n s: it.str,\n x: it.transform[4],\n y: vp.height - it.transform[5],\n w: it.width,\n h,\n });\n heightChars.set(h, (heightChars.get(h) || 0) + it.str.length);\n }\n // \"text layer used\" for hybrid-page purposes (page text > 80 chars) — a\n // page with only a couple of stray scraps isn't \"using\" the text layer even if it also carries a large image.\n const largeImage = textChars > 80 && (await hasLargeImage(page));\n pages.push({\n page: p,\n glyphs,\n width: vp.width,\n height: vp.height,\n textChars,\n largeImage,\n });\n }\n // `numPages` is the SELECTED count, not doc.numPages: the auto-OCR\n // chars/page floor must judge a `--pages` subset against its own page count, or a genuinely image-only page reads as merely \"sparse\".\n return { numPages: pages.length, pages, heightChars };\n}\n"],"names":["OPS","openDocument","LARGE_IMAGE_MIN_PX","hasLargeImage","page","fnArray","argsArray","getOperatorList","i","length","paintImageXObject","w","h","pageCount","src","doc","numPages","collect","first","last","from","Math","min","max","to","pages","heightChars","Map","p","getPage","items","getTextContent","vp","getViewport","scale","glyphs","textChars","it","str","trim","round","abs","transform","push","s","x","y","height","width","set","get","largeImage"],"mappings":"AAAA,wEAAwE;AACxE,yJAAyJ;AAEzJ,SAASA,GAAG,QAAQ,kCAAkC;AACtD,SAASC,YAAY,QAAQ,gBAAgB;AAG7C,yEAAyE;AACzE,2HAA2H;AAC3H,EAAE;AACF,6EAA6E;AAC7E,6LAA6L;AAC7L,MAAMC,qBAAqB;AAC3B,eAAeC,cAAcC,IAAkB;IAC7C,MAAM,EAAEC,OAAO,EAAEC,SAAS,EAAE,GAAG,MAAMF,KAAKG,eAAe;IACzD,IAAK,IAAIC,IAAI,GAAGA,IAAIH,QAAQI,MAAM,EAAED,IAAK;QACvC,IAAIH,OAAO,CAACG,EAAE,KAAKR,IAAIU,iBAAiB,EAAE;QAC1C,MAAM,GAAGC,GAAGC,EAAE,GAAGN,SAAS,CAACE,EAAE;QAC7B,IAAI,OAAOG,MAAM,YAAY,OAAOC,MAAM,UAAU;YAClD,IAAID,KAAKT,sBAAsBU,KAAKV,oBAAoB,OAAO;QACjE,OAAO,OAAO,MAAM,gDAAgD;IACtE;IACA,OAAO;AACT;AAEA,sEAAsE;AACtE,wJAAwJ;AACxJ,EAAE;AACF,4EAA4E;AAC5E,wJAAwJ;AAExJ,0EAA0E;AAC1E,mJAAmJ;AACnJ,OAAO,eAAeW,UAAUC,GAAW;IACzC,MAAMC,MAAM,MAAMd,aAAaa;IAC/B,OAAOC,IAAIC,QAAQ;AACrB;AAEA,OAAO,eAAeC,QAAQH,GAAW,EAAE,EAAEI,KAAK,EAAEC,IAAI,EAAqC,GAAG,CAAC,CAAC;IAChG,MAAMJ,MAAM,MAAMd,aAAaa;IAC/B,MAAMM,OAAOC,KAAKC,GAAG,CAACD,KAAKE,GAAG,CAAC,GAAGL,kBAAAA,mBAAAA,QAAS,IAAIG,KAAKE,GAAG,CAAC,GAAGR,IAAIC,QAAQ;IACvE,MAAMQ,KAAKH,KAAKC,GAAG,CAACD,KAAKE,GAAG,CAACH,MAAMD,iBAAAA,kBAAAA,OAAQJ,IAAIC,QAAQ,GAAGD,IAAIC,QAAQ;IACtE,MAAMS,QAAyB,EAAE;IACjC,MAAMC,cAAc,IAAIC;IACxB,IAAK,IAAIC,IAAIR,MAAMQ,KAAKJ,IAAII,IAAK;QAC/B,MAAMxB,OAAO,MAAMW,IAAIc,OAAO,CAACD;QAC/B,MAAM,EAAEE,KAAK,EAAE,GAAG,MAAM1B,KAAK2B,cAAc;QAC3C,MAAMC,KAAK5B,KAAK6B,WAAW,CAAC;YAAEC,OAAO;QAAE;QACvC,MAAMC,SAAkB,EAAE;QAC1B,IAAIC,YAAY;QAChB,KAAK,MAAMC,MAAMP,MAAO;YACtB,IAAI,CAAE,CAAA,SAASO,EAAC,KAAM,OAAOA,GAAGC,GAAG,KAAK,UAAU;YAClDF,aAAaC,GAAGC,GAAG,CAAC7B,MAAM;YAC1B,IAAI,CAAC4B,GAAGC,GAAG,CAACC,IAAI,IAAI;YACpB,MAAM3B,IAAIS,KAAKmB,KAAK,CAACnB,KAAKoB,GAAG,CAACJ,GAAGK,SAAS,CAAC,EAAE,IAAI,MAAM;YACvDP,OAAOQ,IAAI,CAAC;gBACVC,GAAGP,GAAGC,GAAG;gBACTO,GAAGR,GAAGK,SAAS,CAAC,EAAE;gBAClBI,GAAGd,GAAGe,MAAM,GAAGV,GAAGK,SAAS,CAAC,EAAE;gBAC9B/B,GAAG0B,GAAGW,KAAK;gBACXpC;YACF;YACAc,YAAYuB,GAAG,CAACrC,GAAG,AAACc,CAAAA,YAAYwB,GAAG,CAACtC,MAAM,CAAA,IAAKyB,GAAGC,GAAG,CAAC7B,MAAM;QAC9D;QACA,wEAAwE;QACxE,8GAA8G;QAC9G,MAAM0C,aAAaf,YAAY,MAAO,MAAMjC,cAAcC;QAC1DqB,MAAMkB,IAAI,CAAC;YACTvC,MAAMwB;YACNO;YACAa,OAAOhB,GAAGgB,KAAK;YACfD,QAAQf,GAAGe,MAAM;YACjBX;YACAe;QACF;IACF;IACA,mEAAmE;IACnE,sIAAsI;IACtI,OAAO;QAAEnC,UAAUS,MAAMhB,MAAM;QAAEgB;QAAOC;IAAY;AACtD"}
1
+ {"version":3,"sources":["/Users/kevin/Dev/OpenSource/formats/pdf-to-md/src/collect.ts"],"sourcesContent":["// Pass 1: load the PDF and collect positioned glyphs per page, plus the\n// document-wide char-weighted font-height histogram. Heights round to 0.1pt for stable histogram keys; y is flipped so sorting ascending reads top-down.\nimport type { PDFPageProxy } from 'pdfjs-dist/legacy/build/pdf.mjs';\nimport { OPS } from 'pdfjs-dist/legacy/build/pdf.mjs';\nimport { openDocument } from './pdf-open.ts';\nimport type { CollectedPage, CollectResult, Glyph } from './types.ts';\n\n// HYBRID-PAGE DETECTION (README's \"Design contract\"): a page can carry a\n// real text layer AND a separately-painted large image with its OWN baked-in text that `page.getTextContent()` never sees.\n//\n// Signal: scan the operator list for `OPS.paintImageXObject`, whose args are\n// `[objId, width, height]` — available synchronously, `large` is >=200x200px. Masks/nested Form XObjects aren't walked (accepted gap); missing dims fall back to \"op present\" as the signal.\nconst LARGE_IMAGE_MIN_PX = 200;\nasync function hasLargeImage(page: PDFPageProxy): Promise<boolean> {\n const { fnArray, argsArray } = await page.getOperatorList();\n for (let i = 0; i < fnArray.length; i++) {\n if (fnArray[i] !== OPS.paintImageXObject) continue;\n const [, w, h] = argsArray[i];\n if (typeof w === 'number' && typeof h === 'number') {\n if (w >= LARGE_IMAGE_MIN_PX && h >= LARGE_IMAGE_MIN_PX) return true;\n } else return true; // dims unavailable: coarser \"op present\" signal\n }\n return false;\n}\n\n// { first, last }: 1-based, inclusive page range (the CLI's --pages).\n// Out-of-range values CLAMP to the document's bounds; `page` on each returned entry is always the SOURCE page number, so markers/splicing stay correct.\n//\n// CAVEAT: `heightChars` is built ONLY from the pages actually collected, so\n// a `--pages` subset's heading levels can differ from a full run's — accepted since re-extracted pages get spliced back in for their prose, not diffed.\n\n// Cheap page-count-only load (no per-page text extraction) — the OCR path\n// needs it to resolve its own [first, last] clamping and print a \"### pN\" marker for every page in range, including ones with no recognized words.\nexport async function pageCount(src: string): Promise<number> {\n const doc = await openDocument(src);\n return doc.numPages;\n}\n\nexport async function collect(src: string, { first, last }: { first?: number; last?: number } = {}): Promise<CollectResult> {\n const doc = await openDocument(src);\n const from = Math.min(Math.max(1, first ?? 1), Math.max(1, doc.numPages));\n const to = Math.min(Math.max(from, last ?? doc.numPages), doc.numPages);\n const pages: CollectedPage[] = [];\n const heightChars = new Map<number, number>();\n for (let p = from; p <= to; p++) {\n const page = await doc.getPage(p);\n const { items } = await page.getTextContent();\n const vp = page.getViewport({ scale: 1 });\n const glyphs: Glyph[] = [];\n let textChars = 0;\n for (const it of items) {\n if (!('str' in it) || typeof it.str !== 'string') continue;\n textChars += it.str.length;\n if (!it.str.trim()) continue;\n const h = Math.round(Math.abs(it.transform[3]) * 10) / 10;\n glyphs.push({\n s: it.str,\n x: it.transform[4],\n y: vp.height - it.transform[5],\n w: it.width,\n h,\n });\n heightChars.set(h, (heightChars.get(h) || 0) + it.str.length);\n }\n // \"text layer used\" for hybrid-page purposes (page text > 80 chars) — a\n // page with only a couple of stray scraps isn't \"using\" the text layer even if it also carries a large image.\n const largeImage = textChars > 80 && (await hasLargeImage(page));\n pages.push({\n page: p,\n glyphs,\n width: vp.width,\n height: vp.height,\n textChars,\n largeImage,\n });\n }\n // `numPages` is the SELECTED count, not doc.numPages: the auto-OCR\n // chars/page floor must judge a `--pages` subset against its own page count, or a genuinely image-only page reads as merely \"sparse\".\n return { numPages: pages.length, pages, heightChars };\n}\n"],"names":["OPS","openDocument","LARGE_IMAGE_MIN_PX","hasLargeImage","page","fnArray","argsArray","getOperatorList","i","length","paintImageXObject","w","h","pageCount","src","doc","numPages","collect","first","last","from","Math","min","max","to","pages","heightChars","Map","p","getPage","items","getTextContent","vp","getViewport","scale","glyphs","textChars","it","str","trim","round","abs","transform","push","s","x","y","height","width","set","get","largeImage"],"mappings":"AAAA,wEAAwE;AACxE,yJAAyJ;AAEzJ,SAASA,GAAG,QAAQ,kCAAkC;AACtD,SAASC,YAAY,QAAQ,gBAAgB;AAG7C,yEAAyE;AACzE,2HAA2H;AAC3H,EAAE;AACF,6EAA6E;AAC7E,6LAA6L;AAC7L,MAAMC,qBAAqB;AAC3B,eAAeC,cAAcC,IAAkB;IAC7C,MAAM,EAAEC,OAAO,EAAEC,SAAS,EAAE,GAAG,MAAMF,KAAKG,eAAe;IACzD,IAAK,IAAIC,IAAI,GAAGA,IAAIH,QAAQI,MAAM,EAAED,IAAK;QACvC,IAAIH,OAAO,CAACG,EAAE,KAAKR,IAAIU,iBAAiB,EAAE;QAC1C,MAAM,GAAGC,GAAGC,EAAE,GAAGN,SAAS,CAACE,EAAE;QAC7B,IAAI,OAAOG,MAAM,YAAY,OAAOC,MAAM,UAAU;YAClD,IAAID,KAAKT,sBAAsBU,KAAKV,oBAAoB,OAAO;QACjE,OAAO,OAAO,MAAM,gDAAgD;IACtE;IACA,OAAO;AACT;AAEA,sEAAsE;AACtE,wJAAwJ;AACxJ,EAAE;AACF,4EAA4E;AAC5E,wJAAwJ;AAExJ,0EAA0E;AAC1E,mJAAmJ;AACnJ,OAAO,eAAeW,UAAUC,GAAW;IACzC,MAAMC,MAAM,MAAMd,aAAaa;IAC/B,OAAOC,IAAIC,QAAQ;AACrB;AAEA,OAAO,eAAeC,QAAQH,GAAW,EAAE,EAAEI,KAAK,EAAEC,IAAI,EAAqC,GAAG,CAAC,CAAC;IAChG,MAAMJ,MAAM,MAAMd,aAAaa;IAC/B,MAAMM,OAAOC,KAAKC,GAAG,CAACD,KAAKE,GAAG,CAAC,GAAGL,kBAAAA,mBAAAA,QAAS,IAAIG,KAAKE,GAAG,CAAC,GAAGR,IAAIC,QAAQ;IACvE,MAAMQ,KAAKH,KAAKC,GAAG,CAACD,KAAKE,GAAG,CAACH,MAAMD,iBAAAA,kBAAAA,OAAQJ,IAAIC,QAAQ,GAAGD,IAAIC,QAAQ;IACtE,MAAMS,QAAyB,EAAE;IACjC,MAAMC,cAAc,IAAIC;IACxB,IAAK,IAAIC,IAAIR,MAAMQ,KAAKJ,IAAII,IAAK;QAC/B,MAAMxB,OAAO,MAAMW,IAAIc,OAAO,CAACD;QAC/B,MAAM,EAAEE,KAAK,EAAE,GAAG,MAAM1B,KAAK2B,cAAc;QAC3C,MAAMC,KAAK5B,KAAK6B,WAAW,CAAC;YAAEC,OAAO;QAAE;QACvC,MAAMC,SAAkB,EAAE;QAC1B,IAAIC,YAAY;QAChB,KAAK,MAAMC,MAAMP,MAAO;YACtB,IAAI,CAAE,CAAA,SAASO,EAAC,KAAM,OAAOA,GAAGC,GAAG,KAAK,UAAU;YAClDF,aAAaC,GAAGC,GAAG,CAAC7B,MAAM;YAC1B,IAAI,CAAC4B,GAAGC,GAAG,CAACC,IAAI,IAAI;YACpB,MAAM3B,IAAIS,KAAKmB,KAAK,CAACnB,KAAKoB,GAAG,CAACJ,GAAGK,SAAS,CAAC,EAAE,IAAI,MAAM;YACvDP,OAAOQ,IAAI,CAAC;gBACVC,GAAGP,GAAGC,GAAG;gBACTO,GAAGR,GAAGK,SAAS,CAAC,EAAE;gBAClBI,GAAGd,GAAGe,MAAM,GAAGV,GAAGK,SAAS,CAAC,EAAE;gBAC9B/B,GAAG0B,GAAGW,KAAK;gBACXpC;YACF;YACAc,YAAYuB,GAAG,CAACrC,GAAG,AAACc,CAAAA,YAAYwB,GAAG,CAACtC,MAAM,CAAA,IAAKyB,GAAGC,GAAG,CAAC7B,MAAM;QAC9D;QACA,wEAAwE;QACxE,8GAA8G;QAC9G,MAAM0C,aAAaf,YAAY,MAAO,MAAMjC,cAAcC;QAC1DqB,MAAMkB,IAAI,CAAC;YACTvC,MAAMwB;YACNO;YACAa,OAAOhB,GAAGgB,KAAK;YACfD,QAAQf,GAAGe,MAAM;YACjBX;YACAe;QACF;IACF;IACA,mEAAmE;IACnE,sIAAsI;IACtI,OAAO;QAAEnC,UAAUS,MAAMhB,MAAM;QAAEgB;QAAOC;IAAY;AACtD"}
@@ -1 +1 @@
1
- {"version":3,"sources":["/Users/kevin/Projects/pdf-to-md/src/dict.ts"],"sourcesContent":["// The hyphen-rejoin dictionary, shared by src/emit.ts and src/markdown.ts.\n// Loaded once, lazily, on first use — never at import — so importing the library does no file IO on its own.\nimport { existsSync, readFileSync } from 'node:fs';\nimport wordListPath from 'word-list';\n\n// Bundled list (not the OS's /usr/share/dict/words) → identical rejoin decisions on every\n// OS/machine. PDF_TO_MD_DICT_PATH overrides it — the test seam for simulating no dictionary.\nexport const DICT_PATH = process.env.PDF_TO_MD_DICT_PATH || wordListPath;\n\nlet cache: { words: Set<string>; available: boolean } | null = null;\n\nfunction load(): { words: Set<string>; available: boolean } {\n if (!cache) {\n const available = existsSync(DICT_PATH);\n cache = { words: available ? new Set(readFileSync(DICT_PATH, 'utf8').toLowerCase().split('\\n')) : new Set(), available };\n }\n return cache;\n}\n\n/** The hyphen-rejoin dictionary; an empty set if none was found. */\nexport function getDict(): Set<string> {\n return load().words;\n}\n\n/** False when no dictionary file exists at DICT_PATH — every line-final\n * hyphen then stays unresolved. */\nexport function dictAvailable(): boolean {\n return load().available;\n}\n\n/** The notice callers surface via warn() when dictAvailable() is false. */\nexport function dictUnavailableWarning(): string {\n return `no dictionary at ${DICT_PATH} — line-final hyphens kept unresolved`;\n}\n"],"names":["existsSync","readFileSync","wordListPath","DICT_PATH","process","env","PDF_TO_MD_DICT_PATH","cache","load","available","words","Set","toLowerCase","split","getDict","dictAvailable","dictUnavailableWarning"],"mappings":"AAAA,2EAA2E;AAC3E,6GAA6G;AAC7G,SAASA,UAAU,EAAEC,YAAY,QAAQ,UAAU;AACnD,OAAOC,kBAAkB,YAAY;AAErC,0FAA0F;AAC1F,6FAA6F;AAC7F,OAAO,MAAMC,YAAYC,QAAQC,GAAG,CAACC,mBAAmB,IAAIJ,aAAa;AAEzE,IAAIK,QAA2D;AAE/D,SAASC;IACP,IAAI,CAACD,OAAO;QACV,MAAME,YAAYT,WAAWG;QAC7BI,QAAQ;YAAEG,OAAOD,YAAY,IAAIE,IAAIV,aAAaE,WAAW,QAAQS,WAAW,GAAGC,KAAK,CAAC,SAAS,IAAIF;YAAOF;QAAU;IACzH;IACA,OAAOF;AACT;AAEA,kEAAkE,GAClE,OAAO,SAASO;IACd,OAAON,OAAOE,KAAK;AACrB;AAEA;iCACiC,GACjC,OAAO,SAASK;IACd,OAAOP,OAAOC,SAAS;AACzB;AAEA,yEAAyE,GACzE,OAAO,SAASO;IACd,OAAO,CAAC,iBAAiB,EAAEb,UAAU,qCAAqC,CAAC;AAC7E"}
1
+ {"version":3,"sources":["/Users/kevin/Dev/OpenSource/formats/pdf-to-md/src/dict.ts"],"sourcesContent":["// The hyphen-rejoin dictionary, shared by src/emit.ts and src/markdown.ts.\n// Loaded once, lazily, on first use — never at import — so importing the library does no file IO on its own.\nimport { existsSync, readFileSync } from 'node:fs';\nimport wordListPath from 'word-list';\n\n// Bundled list (not the OS's /usr/share/dict/words) → identical rejoin decisions on every\n// OS/machine. PDF_TO_MD_DICT_PATH overrides it — the test seam for simulating no dictionary.\nexport const DICT_PATH = process.env.PDF_TO_MD_DICT_PATH || wordListPath;\n\nlet cache: { words: Set<string>; available: boolean } | null = null;\n\nfunction load(): { words: Set<string>; available: boolean } {\n if (!cache) {\n const available = existsSync(DICT_PATH);\n cache = { words: available ? new Set(readFileSync(DICT_PATH, 'utf8').toLowerCase().split('\\n')) : new Set(), available };\n }\n return cache;\n}\n\n/** The hyphen-rejoin dictionary; an empty set if none was found. */\nexport function getDict(): Set<string> {\n return load().words;\n}\n\n/** False when no dictionary file exists at DICT_PATH — every line-final\n * hyphen then stays unresolved. */\nexport function dictAvailable(): boolean {\n return load().available;\n}\n\n/** The notice callers surface via warn() when dictAvailable() is false. */\nexport function dictUnavailableWarning(): string {\n return `no dictionary at ${DICT_PATH} — line-final hyphens kept unresolved`;\n}\n"],"names":["existsSync","readFileSync","wordListPath","DICT_PATH","process","env","PDF_TO_MD_DICT_PATH","cache","load","available","words","Set","toLowerCase","split","getDict","dictAvailable","dictUnavailableWarning"],"mappings":"AAAA,2EAA2E;AAC3E,6GAA6G;AAC7G,SAASA,UAAU,EAAEC,YAAY,QAAQ,UAAU;AACnD,OAAOC,kBAAkB,YAAY;AAErC,0FAA0F;AAC1F,6FAA6F;AAC7F,OAAO,MAAMC,YAAYC,QAAQC,GAAG,CAACC,mBAAmB,IAAIJ,aAAa;AAEzE,IAAIK,QAA2D;AAE/D,SAASC;IACP,IAAI,CAACD,OAAO;QACV,MAAME,YAAYT,WAAWG;QAC7BI,QAAQ;YAAEG,OAAOD,YAAY,IAAIE,IAAIV,aAAaE,WAAW,QAAQS,WAAW,GAAGC,KAAK,CAAC,SAAS,IAAIF;YAAOF;QAAU;IACzH;IACA,OAAOF;AACT;AAEA,kEAAkE,GAClE,OAAO,SAASO;IACd,OAAON,OAAOE,KAAK;AACrB;AAEA;iCACiC,GACjC,OAAO,SAASK;IACd,OAAOP,OAAOC,SAAS;AACzB;AAEA,yEAAyE,GACzE,OAAO,SAASO;IACd,OAAO,CAAC,iBAAiB,EAAEb,UAAU,qCAAqC,CAAC;AAC7E"}
@@ -1 +1 @@
1
- {"version":3,"sources":["/Users/kevin/Projects/pdf-to-md/src/emit.ts"],"sourcesContent":["// emit.ts — OCR-path pass 3: lines -> markdown blocks.\nimport { getDict } from './dict.ts';\nimport type { OcrLine, PageOCR, Tuning, Word } from './geometry.ts';\nimport { bucket } from './geometry.ts';\n\n// Join a wrapped line onto the open paragraph, closing a line-final hyphen\n// only when the dictionary confirms the joined form is a word.\nfunction hyphenJoin(acc: string, next: string): string {\n if (acc.endsWith('-')) {\n const withoutDash = acc.slice(0, -1);\n const pen = withoutDash.at(-1);\n if (pen && /[A-Za-z0-9]/.test(pen)) {\n const head = withoutDash.split(' ').pop() ?? '';\n const firstTok = next.split(' ')[0] ?? '';\n const filtered = [...firstTok].filter((c) => /[A-Za-z-]/.test(c)).join('');\n const tail = filtered.split('-')[0] ?? '';\n const close = getDict().has((head + tail).toLowerCase());\n return (close ? withoutDash : acc) + next;\n }\n }\n return `${acc} ${next}`;\n}\n\n// Blocks with too few word-like tokens to trust as prose (see\n// UNSTRUCTURED_WORDLIKE_FRACTION below) are labeled `> [review: reason] text` instead of guessed at.\nconst REVIEW_REASON = {\n unstructuredLabels: 'unstructured labels',\n} as const;\n\n/** Share of word-like tokens (3+ letter run) below which a block counts as\n * unstructured. Measured (`review-blocks.test.ts`): junk 0.000-0.100 vs prose 0.743-0.857, cut 0.4. */\nexport const UNSTRUCTURED_WORDLIKE_FRACTION = 0.4;\nfunction isUnstructured(text: string): boolean {\n const tokens = text.split(/\\s+/).filter(Boolean);\n if (tokens.length === 0) return false; // never mark an empty block\n const wordLike = tokens.filter((t) => /[A-Za-z]{3}/.test(t)).length;\n return wordLike / tokens.length < UNSTRUCTURED_WORDLIKE_FRACTION;\n}\n\nfunction reviewLine(reason: string, text: string): string {\n return `> [review: ${reason}] ${text}`;\n}\n\n// One classified block, carrying constituent words for analyze()'s\n// block->word linkage; discriminated on `kind` so `reason`/`level` are compile-time guaranteed present on their variants.\nexport type EmitBlock =\n | { kind: 'heading'; text: string; level: number; words: Word[] }\n | { kind: 'paragraph'; text: string; words: Word[] }\n | { kind: 'quote'; text: string; words: Word[] }\n | {\n kind: 'review';\n text: string;\n /** The parseable reason inside `> [review: ...]`, carried on the block\n * so more reasons can be added without changing the render format. */\n reason: string;\n words: Word[];\n };\n\n// Renders one EmitBlock to the exact markdown emitPage used to push\n// directly — shared so extract.ts and analyze.ts produce byte-identical text.\nexport function renderEmitBlock(b: EmitBlock): string {\n switch (b.kind) {\n case 'heading':\n return `${'#'.repeat(b.level)} ${b.text}`;\n case 'quote':\n return `> ${b.text}`;\n case 'review':\n return reviewLine(b.reason, b.text);\n default:\n return b.text;\n }\n}\n\nexport function emitPage(p: PageOCR, levelOf: Map<number, number>, tuning: Tuning): EmitBlock[] {\n const blocks: EmitBlock[] = [];\n let para: OcrLine[] = [];\n const flush = () => {\n if (!para.length) return;\n const texts = para.map((l) => l.text);\n const merged = texts.slice(1).reduce(hyphenJoin, texts[0]);\n const words = para.flatMap((l) => l.words);\n // Same review-marker rule as the heading path below: no word may be\n // dropped or reordered here, only how the block renders changes.\n blocks.push(isUnstructured(merged) ? { kind: 'review', text: merged, reason: REVIEW_REASON.unstructuredLabels, words } : { kind: 'paragraph', text: merged, words });\n para = [];\n };\n\n let i = 0;\n while (i < p.lines.length) {\n const line: OcrLine = p.lines[i];\n const prev = i > 0 ? p.lines[i - 1] : null;\n if (prev && (line.col !== prev.col || prev.midY - line.midY > p.pitch * tuning.paraGap)) {\n // Join across a column break (instead of flushing) only when the column\n // changed, the paragraph dangles (no terminal punct), and the new text starts lowercase; a miss stays visible via pageStats.danglingLong.\n const danglingEnd = para.length > 0 && !/[.!?:;\"'”’)\\]]$/.test(para[para.length - 1].text);\n const continuationStart = /^[a-z]/.test(line.text);\n const joinAcrossColumns = line.col !== prev.col && danglingEnd && continuationStart && !levelOf.has(bucket(line.h));\n if (!joinAcrossColumns) flush();\n }\n\n if (levelOf.has(bucket(line.h))) {\n // Gather the run of consecutive heading-size lines: a wrapped title is\n // one heading, a long run is a pull quote in a bigger face.\n flush();\n const run: OcrLine[] = [line];\n let maxB = bucket(line.h);\n while (i + 1 < p.lines.length && levelOf.has(bucket(p.lines[i + 1].h)) && p.lines[i + 1].col === p.lines[i].col && p.lines[i].midY - p.lines[i + 1].midY < Math.max(p.lines[i].h, p.lines[i + 1].h) * 2.2) {\n i++;\n run.push(p.lines[i]);\n maxB = Math.max(maxB, bucket(p.lines[i].h));\n }\n const runTexts = run.map((l) => l.text);\n const merged = runTexts.slice(1).reduce(hyphenJoin, runTexts[0]);\n const words = run.flatMap((l) => l.words);\n // Below UNSTRUCTURED_WORDLIKE_FRACTION = chart labels/stray numerals,\n // not a heading; checked FIRST so a long numeral run isn't asserted as a pull quote.\n const junk = isUnstructured(merged);\n if (junk) blocks.push({ kind: 'review', text: merged, reason: REVIEW_REASON.unstructuredLabels, words });\n else if (merged.length > 150) blocks.push({ kind: 'quote', text: merged, words });\n else blocks.push({ kind: 'heading', text: merged, level: levelOf.get(maxB) as number, words });\n } else {\n para.push(line);\n }\n i++;\n }\n flush();\n return blocks;\n}\n"],"names":["getDict","bucket","hyphenJoin","acc","next","endsWith","withoutDash","slice","pen","at","test","filtered","head","split","pop","firstTok","filter","c","join","tail","close","has","toLowerCase","REVIEW_REASON","unstructuredLabels","UNSTRUCTURED_WORDLIKE_FRACTION","isUnstructured","text","tokens","Boolean","length","wordLike","t","reviewLine","reason","renderEmitBlock","b","kind","repeat","level","emitPage","p","levelOf","tuning","blocks","para","flush","texts","map","l","merged","reduce","words","flatMap","push","i","lines","line","prev","col","midY","pitch","paraGap","danglingEnd","continuationStart","joinAcrossColumns","h","run","maxB","Math","max","runTexts","junk","get"],"mappings":"AAAA,uDAAuD;AACvD,SAASA,OAAO,QAAQ,YAAY;AAEpC,SAASC,MAAM,QAAQ,gBAAgB;AAEvC,2EAA2E;AAC3E,+DAA+D;AAC/D,SAASC,WAAWC,GAAW,EAAEC,IAAY;IAC3C,IAAID,IAAIE,QAAQ,CAAC,MAAM;QACrB,MAAMC,cAAcH,IAAII,KAAK,CAAC,GAAG,CAAC;QAClC,MAAMC,MAAMF,YAAYG,EAAE,CAAC,CAAC;QAC5B,IAAID,OAAO,cAAcE,IAAI,CAACF,MAAM;gBACrBF,wBACIF,cAEJO;YAHb,MAAMC,QAAON,yBAAAA,YAAYO,KAAK,CAAC,KAAKC,GAAG,gBAA1BR,oCAAAA,yBAAgC;YAC7C,MAAMS,YAAWX,eAAAA,KAAKS,KAAK,CAAC,IAAI,CAAC,EAAE,cAAlBT,0BAAAA,eAAsB;YACvC,MAAMO,WAAW;mBAAII;aAAS,CAACC,MAAM,CAAC,CAACC,IAAM,YAAYP,IAAI,CAACO,IAAIC,IAAI,CAAC;YACvE,MAAMC,QAAOR,mBAAAA,SAASE,KAAK,CAAC,IAAI,CAAC,EAAE,cAAtBF,8BAAAA,mBAA0B;YACvC,MAAMS,QAAQpB,UAAUqB,GAAG,CAAC,AAACT,CAAAA,OAAOO,IAAG,EAAGG,WAAW;YACrD,OAAO,AAACF,CAAAA,QAAQd,cAAcH,GAAE,IAAKC;QACvC;IACF;IACA,OAAO,GAAGD,IAAI,CAAC,EAAEC,MAAM;AACzB;AAEA,8DAA8D;AAC9D,qGAAqG;AACrG,MAAMmB,gBAAgB;IACpBC,oBAAoB;AACtB;AAEA;qGACqG,GACrG,OAAO,MAAMC,iCAAiC,IAAI;AAClD,SAASC,eAAeC,IAAY;IAClC,MAAMC,SAASD,KAAKd,KAAK,CAAC,OAAOG,MAAM,CAACa;IACxC,IAAID,OAAOE,MAAM,KAAK,GAAG,OAAO,OAAO,4BAA4B;IACnE,MAAMC,WAAWH,OAAOZ,MAAM,CAAC,CAACgB,IAAM,cAActB,IAAI,CAACsB,IAAIF,MAAM;IACnE,OAAOC,WAAWH,OAAOE,MAAM,GAAGL;AACpC;AAEA,SAASQ,WAAWC,MAAc,EAAEP,IAAY;IAC9C,OAAO,CAAC,WAAW,EAAEO,OAAO,EAAE,EAAEP,MAAM;AACxC;AAiBA,oEAAoE;AACpE,8EAA8E;AAC9E,OAAO,SAASQ,gBAAgBC,CAAY;IAC1C,OAAQA,EAAEC,IAAI;QACZ,KAAK;YACH,OAAO,GAAG,IAAIC,MAAM,CAACF,EAAEG,KAAK,EAAE,CAAC,EAAEH,EAAET,IAAI,EAAE;QAC3C,KAAK;YACH,OAAO,CAAC,EAAE,EAAES,EAAET,IAAI,EAAE;QACtB,KAAK;YACH,OAAOM,WAAWG,EAAEF,MAAM,EAAEE,EAAET,IAAI;QACpC;YACE,OAAOS,EAAET,IAAI;IACjB;AACF;AAEA,OAAO,SAASa,SAASC,CAAU,EAAEC,OAA4B,EAAEC,MAAc;IAC/E,MAAMC,SAAsB,EAAE;IAC9B,IAAIC,OAAkB,EAAE;IACxB,MAAMC,QAAQ;QACZ,IAAI,CAACD,KAAKf,MAAM,EAAE;QAClB,MAAMiB,QAAQF,KAAKG,GAAG,CAAC,CAACC,IAAMA,EAAEtB,IAAI;QACpC,MAAMuB,SAASH,MAAMxC,KAAK,CAAC,GAAG4C,MAAM,CAACjD,YAAY6C,KAAK,CAAC,EAAE;QACzD,MAAMK,QAAQP,KAAKQ,OAAO,CAAC,CAACJ,IAAMA,EAAEG,KAAK;QACzC,oEAAoE;QACpE,iEAAiE;QACjER,OAAOU,IAAI,CAAC5B,eAAewB,UAAU;YAAEb,MAAM;YAAUV,MAAMuB;YAAQhB,QAAQX,cAAcC,kBAAkB;YAAE4B;QAAM,IAAI;YAAEf,MAAM;YAAaV,MAAMuB;YAAQE;QAAM;QAClKP,OAAO,EAAE;IACX;IAEA,IAAIU,IAAI;IACR,MAAOA,IAAId,EAAEe,KAAK,CAAC1B,MAAM,CAAE;QACzB,MAAM2B,OAAgBhB,EAAEe,KAAK,CAACD,EAAE;QAChC,MAAMG,OAAOH,IAAI,IAAId,EAAEe,KAAK,CAACD,IAAI,EAAE,GAAG;QACtC,IAAIG,QAASD,CAAAA,KAAKE,GAAG,KAAKD,KAAKC,GAAG,IAAID,KAAKE,IAAI,GAAGH,KAAKG,IAAI,GAAGnB,EAAEoB,KAAK,GAAGlB,OAAOmB,OAAO,AAAD,GAAI;YACvF,wEAAwE;YACxE,0IAA0I;YAC1I,MAAMC,cAAclB,KAAKf,MAAM,GAAG,KAAK,CAAC,kBAAkBpB,IAAI,CAACmC,IAAI,CAACA,KAAKf,MAAM,GAAG,EAAE,CAACH,IAAI;YACzF,MAAMqC,oBAAoB,SAAStD,IAAI,CAAC+C,KAAK9B,IAAI;YACjD,MAAMsC,oBAAoBR,KAAKE,GAAG,KAAKD,KAAKC,GAAG,IAAII,eAAeC,qBAAqB,CAACtB,QAAQrB,GAAG,CAACpB,OAAOwD,KAAKS,CAAC;YACjH,IAAI,CAACD,mBAAmBnB;QAC1B;QAEA,IAAIJ,QAAQrB,GAAG,CAACpB,OAAOwD,KAAKS,CAAC,IAAI;YAC/B,uEAAuE;YACvE,4DAA4D;YAC5DpB;YACA,MAAMqB,MAAiB;gBAACV;aAAK;YAC7B,IAAIW,OAAOnE,OAAOwD,KAAKS,CAAC;YACxB,MAAOX,IAAI,IAAId,EAAEe,KAAK,CAAC1B,MAAM,IAAIY,QAAQrB,GAAG,CAACpB,OAAOwC,EAAEe,KAAK,CAACD,IAAI,EAAE,CAACW,CAAC,MAAMzB,EAAEe,KAAK,CAACD,IAAI,EAAE,CAACI,GAAG,KAAKlB,EAAEe,KAAK,CAACD,EAAE,CAACI,GAAG,IAAIlB,EAAEe,KAAK,CAACD,EAAE,CAACK,IAAI,GAAGnB,EAAEe,KAAK,CAACD,IAAI,EAAE,CAACK,IAAI,GAAGS,KAAKC,GAAG,CAAC7B,EAAEe,KAAK,CAACD,EAAE,CAACW,CAAC,EAAEzB,EAAEe,KAAK,CAACD,IAAI,EAAE,CAACW,CAAC,IAAI,IAAK;gBACzMX;gBACAY,IAAIb,IAAI,CAACb,EAAEe,KAAK,CAACD,EAAE;gBACnBa,OAAOC,KAAKC,GAAG,CAACF,MAAMnE,OAAOwC,EAAEe,KAAK,CAACD,EAAE,CAACW,CAAC;YAC3C;YACA,MAAMK,WAAWJ,IAAInB,GAAG,CAAC,CAACC,IAAMA,EAAEtB,IAAI;YACtC,MAAMuB,SAASqB,SAAShE,KAAK,CAAC,GAAG4C,MAAM,CAACjD,YAAYqE,QAAQ,CAAC,EAAE;YAC/D,MAAMnB,QAAQe,IAAId,OAAO,CAAC,CAACJ,IAAMA,EAAEG,KAAK;YACxC,sEAAsE;YACtE,qFAAqF;YACrF,MAAMoB,OAAO9C,eAAewB;YAC5B,IAAIsB,MAAM5B,OAAOU,IAAI,CAAC;gBAAEjB,MAAM;gBAAUV,MAAMuB;gBAAQhB,QAAQX,cAAcC,kBAAkB;gBAAE4B;YAAM;iBACjG,IAAIF,OAAOpB,MAAM,GAAG,KAAKc,OAAOU,IAAI,CAAC;gBAAEjB,MAAM;gBAASV,MAAMuB;gBAAQE;YAAM;iBAC1ER,OAAOU,IAAI,CAAC;gBAAEjB,MAAM;gBAAWV,MAAMuB;gBAAQX,OAAOG,QAAQ+B,GAAG,CAACL;gBAAiBhB;YAAM;QAC9F,OAAO;YACLP,KAAKS,IAAI,CAACG;QACZ;QACAF;IACF;IACAT;IACA,OAAOF;AACT"}
1
+ {"version":3,"sources":["/Users/kevin/Dev/OpenSource/formats/pdf-to-md/src/emit.ts"],"sourcesContent":["// emit.ts — OCR-path pass 3: lines -> markdown blocks.\nimport { getDict } from './dict.ts';\nimport type { OcrLine, PageOCR, Tuning, Word } from './geometry.ts';\nimport { bucket } from './geometry.ts';\n\n// Join a wrapped line onto the open paragraph, closing a line-final hyphen\n// only when the dictionary confirms the joined form is a word.\nfunction hyphenJoin(acc: string, next: string): string {\n if (acc.endsWith('-')) {\n const withoutDash = acc.slice(0, -1);\n const pen = withoutDash.at(-1);\n if (pen && /[A-Za-z0-9]/.test(pen)) {\n const head = withoutDash.split(' ').pop() ?? '';\n const firstTok = next.split(' ')[0] ?? '';\n const filtered = [...firstTok].filter((c) => /[A-Za-z-]/.test(c)).join('');\n const tail = filtered.split('-')[0] ?? '';\n const close = getDict().has((head + tail).toLowerCase());\n return (close ? withoutDash : acc) + next;\n }\n }\n return `${acc} ${next}`;\n}\n\n// Blocks with too few word-like tokens to trust as prose (see\n// UNSTRUCTURED_WORDLIKE_FRACTION below) are labeled `> [review: reason] text` instead of guessed at.\nconst REVIEW_REASON = {\n unstructuredLabels: 'unstructured labels',\n} as const;\n\n/** Share of word-like tokens (3+ letter run) below which a block counts as\n * unstructured. Measured (`review-blocks.test.ts`): junk 0.000-0.100 vs prose 0.743-0.857, cut 0.4. */\nexport const UNSTRUCTURED_WORDLIKE_FRACTION = 0.4;\nfunction isUnstructured(text: string): boolean {\n const tokens = text.split(/\\s+/).filter(Boolean);\n if (tokens.length === 0) return false; // never mark an empty block\n const wordLike = tokens.filter((t) => /[A-Za-z]{3}/.test(t)).length;\n return wordLike / tokens.length < UNSTRUCTURED_WORDLIKE_FRACTION;\n}\n\nfunction reviewLine(reason: string, text: string): string {\n return `> [review: ${reason}] ${text}`;\n}\n\n// One classified block, carrying constituent words for analyze()'s\n// block->word linkage; discriminated on `kind` so `reason`/`level` are compile-time guaranteed present on their variants.\nexport type EmitBlock =\n | { kind: 'heading'; text: string; level: number; words: Word[] }\n | { kind: 'paragraph'; text: string; words: Word[] }\n | { kind: 'quote'; text: string; words: Word[] }\n | {\n kind: 'review';\n text: string;\n /** The parseable reason inside `> [review: ...]`, carried on the block\n * so more reasons can be added without changing the render format. */\n reason: string;\n words: Word[];\n };\n\n// Renders one EmitBlock to the exact markdown emitPage used to push\n// directly — shared so extract.ts and analyze.ts produce byte-identical text.\nexport function renderEmitBlock(b: EmitBlock): string {\n switch (b.kind) {\n case 'heading':\n return `${'#'.repeat(b.level)} ${b.text}`;\n case 'quote':\n return `> ${b.text}`;\n case 'review':\n return reviewLine(b.reason, b.text);\n default:\n return b.text;\n }\n}\n\nexport function emitPage(p: PageOCR, levelOf: Map<number, number>, tuning: Tuning): EmitBlock[] {\n const blocks: EmitBlock[] = [];\n let para: OcrLine[] = [];\n const flush = () => {\n if (!para.length) return;\n const texts = para.map((l) => l.text);\n const merged = texts.slice(1).reduce(hyphenJoin, texts[0]);\n const words = para.flatMap((l) => l.words);\n // Same review-marker rule as the heading path below: no word may be\n // dropped or reordered here, only how the block renders changes.\n blocks.push(isUnstructured(merged) ? { kind: 'review', text: merged, reason: REVIEW_REASON.unstructuredLabels, words } : { kind: 'paragraph', text: merged, words });\n para = [];\n };\n\n let i = 0;\n while (i < p.lines.length) {\n const line: OcrLine = p.lines[i];\n const prev = i > 0 ? p.lines[i - 1] : null;\n if (prev && (line.col !== prev.col || prev.midY - line.midY > p.pitch * tuning.paraGap)) {\n // Join across a column break (instead of flushing) only when the column\n // changed, the paragraph dangles (no terminal punct), and the new text starts lowercase; a miss stays visible via pageStats.danglingLong.\n const danglingEnd = para.length > 0 && !/[.!?:;\"'”’)\\]]$/.test(para[para.length - 1].text);\n const continuationStart = /^[a-z]/.test(line.text);\n const joinAcrossColumns = line.col !== prev.col && danglingEnd && continuationStart && !levelOf.has(bucket(line.h));\n if (!joinAcrossColumns) flush();\n }\n\n if (levelOf.has(bucket(line.h))) {\n // Gather the run of consecutive heading-size lines: a wrapped title is\n // one heading, a long run is a pull quote in a bigger face.\n flush();\n const run: OcrLine[] = [line];\n let maxB = bucket(line.h);\n while (i + 1 < p.lines.length && levelOf.has(bucket(p.lines[i + 1].h)) && p.lines[i + 1].col === p.lines[i].col && p.lines[i].midY - p.lines[i + 1].midY < Math.max(p.lines[i].h, p.lines[i + 1].h) * 2.2) {\n i++;\n run.push(p.lines[i]);\n maxB = Math.max(maxB, bucket(p.lines[i].h));\n }\n const runTexts = run.map((l) => l.text);\n const merged = runTexts.slice(1).reduce(hyphenJoin, runTexts[0]);\n const words = run.flatMap((l) => l.words);\n // Below UNSTRUCTURED_WORDLIKE_FRACTION = chart labels/stray numerals,\n // not a heading; checked FIRST so a long numeral run isn't asserted as a pull quote.\n const junk = isUnstructured(merged);\n if (junk) blocks.push({ kind: 'review', text: merged, reason: REVIEW_REASON.unstructuredLabels, words });\n else if (merged.length > 150) blocks.push({ kind: 'quote', text: merged, words });\n else blocks.push({ kind: 'heading', text: merged, level: levelOf.get(maxB) as number, words });\n } else {\n para.push(line);\n }\n i++;\n }\n flush();\n return blocks;\n}\n"],"names":["getDict","bucket","hyphenJoin","acc","next","endsWith","withoutDash","slice","pen","at","test","filtered","head","split","pop","firstTok","filter","c","join","tail","close","has","toLowerCase","REVIEW_REASON","unstructuredLabels","UNSTRUCTURED_WORDLIKE_FRACTION","isUnstructured","text","tokens","Boolean","length","wordLike","t","reviewLine","reason","renderEmitBlock","b","kind","repeat","level","emitPage","p","levelOf","tuning","blocks","para","flush","texts","map","l","merged","reduce","words","flatMap","push","i","lines","line","prev","col","midY","pitch","paraGap","danglingEnd","continuationStart","joinAcrossColumns","h","run","maxB","Math","max","runTexts","junk","get"],"mappings":"AAAA,uDAAuD;AACvD,SAASA,OAAO,QAAQ,YAAY;AAEpC,SAASC,MAAM,QAAQ,gBAAgB;AAEvC,2EAA2E;AAC3E,+DAA+D;AAC/D,SAASC,WAAWC,GAAW,EAAEC,IAAY;IAC3C,IAAID,IAAIE,QAAQ,CAAC,MAAM;QACrB,MAAMC,cAAcH,IAAII,KAAK,CAAC,GAAG,CAAC;QAClC,MAAMC,MAAMF,YAAYG,EAAE,CAAC,CAAC;QAC5B,IAAID,OAAO,cAAcE,IAAI,CAACF,MAAM;gBACrBF,wBACIF,cAEJO;YAHb,MAAMC,QAAON,yBAAAA,YAAYO,KAAK,CAAC,KAAKC,GAAG,gBAA1BR,oCAAAA,yBAAgC;YAC7C,MAAMS,YAAWX,eAAAA,KAAKS,KAAK,CAAC,IAAI,CAAC,EAAE,cAAlBT,0BAAAA,eAAsB;YACvC,MAAMO,WAAW;mBAAII;aAAS,CAACC,MAAM,CAAC,CAACC,IAAM,YAAYP,IAAI,CAACO,IAAIC,IAAI,CAAC;YACvE,MAAMC,QAAOR,mBAAAA,SAASE,KAAK,CAAC,IAAI,CAAC,EAAE,cAAtBF,8BAAAA,mBAA0B;YACvC,MAAMS,QAAQpB,UAAUqB,GAAG,CAAC,AAACT,CAAAA,OAAOO,IAAG,EAAGG,WAAW;YACrD,OAAO,AAACF,CAAAA,QAAQd,cAAcH,GAAE,IAAKC;QACvC;IACF;IACA,OAAO,GAAGD,IAAI,CAAC,EAAEC,MAAM;AACzB;AAEA,8DAA8D;AAC9D,qGAAqG;AACrG,MAAMmB,gBAAgB;IACpBC,oBAAoB;AACtB;AAEA;qGACqG,GACrG,OAAO,MAAMC,iCAAiC,IAAI;AAClD,SAASC,eAAeC,IAAY;IAClC,MAAMC,SAASD,KAAKd,KAAK,CAAC,OAAOG,MAAM,CAACa;IACxC,IAAID,OAAOE,MAAM,KAAK,GAAG,OAAO,OAAO,4BAA4B;IACnE,MAAMC,WAAWH,OAAOZ,MAAM,CAAC,CAACgB,IAAM,cAActB,IAAI,CAACsB,IAAIF,MAAM;IACnE,OAAOC,WAAWH,OAAOE,MAAM,GAAGL;AACpC;AAEA,SAASQ,WAAWC,MAAc,EAAEP,IAAY;IAC9C,OAAO,CAAC,WAAW,EAAEO,OAAO,EAAE,EAAEP,MAAM;AACxC;AAiBA,oEAAoE;AACpE,8EAA8E;AAC9E,OAAO,SAASQ,gBAAgBC,CAAY;IAC1C,OAAQA,EAAEC,IAAI;QACZ,KAAK;YACH,OAAO,GAAG,IAAIC,MAAM,CAACF,EAAEG,KAAK,EAAE,CAAC,EAAEH,EAAET,IAAI,EAAE;QAC3C,KAAK;YACH,OAAO,CAAC,EAAE,EAAES,EAAET,IAAI,EAAE;QACtB,KAAK;YACH,OAAOM,WAAWG,EAAEF,MAAM,EAAEE,EAAET,IAAI;QACpC;YACE,OAAOS,EAAET,IAAI;IACjB;AACF;AAEA,OAAO,SAASa,SAASC,CAAU,EAAEC,OAA4B,EAAEC,MAAc;IAC/E,MAAMC,SAAsB,EAAE;IAC9B,IAAIC,OAAkB,EAAE;IACxB,MAAMC,QAAQ;QACZ,IAAI,CAACD,KAAKf,MAAM,EAAE;QAClB,MAAMiB,QAAQF,KAAKG,GAAG,CAAC,CAACC,IAAMA,EAAEtB,IAAI;QACpC,MAAMuB,SAASH,MAAMxC,KAAK,CAAC,GAAG4C,MAAM,CAACjD,YAAY6C,KAAK,CAAC,EAAE;QACzD,MAAMK,QAAQP,KAAKQ,OAAO,CAAC,CAACJ,IAAMA,EAAEG,KAAK;QACzC,oEAAoE;QACpE,iEAAiE;QACjER,OAAOU,IAAI,CAAC5B,eAAewB,UAAU;YAAEb,MAAM;YAAUV,MAAMuB;YAAQhB,QAAQX,cAAcC,kBAAkB;YAAE4B;QAAM,IAAI;YAAEf,MAAM;YAAaV,MAAMuB;YAAQE;QAAM;QAClKP,OAAO,EAAE;IACX;IAEA,IAAIU,IAAI;IACR,MAAOA,IAAId,EAAEe,KAAK,CAAC1B,MAAM,CAAE;QACzB,MAAM2B,OAAgBhB,EAAEe,KAAK,CAACD,EAAE;QAChC,MAAMG,OAAOH,IAAI,IAAId,EAAEe,KAAK,CAACD,IAAI,EAAE,GAAG;QACtC,IAAIG,QAASD,CAAAA,KAAKE,GAAG,KAAKD,KAAKC,GAAG,IAAID,KAAKE,IAAI,GAAGH,KAAKG,IAAI,GAAGnB,EAAEoB,KAAK,GAAGlB,OAAOmB,OAAO,AAAD,GAAI;YACvF,wEAAwE;YACxE,0IAA0I;YAC1I,MAAMC,cAAclB,KAAKf,MAAM,GAAG,KAAK,CAAC,kBAAkBpB,IAAI,CAACmC,IAAI,CAACA,KAAKf,MAAM,GAAG,EAAE,CAACH,IAAI;YACzF,MAAMqC,oBAAoB,SAAStD,IAAI,CAAC+C,KAAK9B,IAAI;YACjD,MAAMsC,oBAAoBR,KAAKE,GAAG,KAAKD,KAAKC,GAAG,IAAII,eAAeC,qBAAqB,CAACtB,QAAQrB,GAAG,CAACpB,OAAOwD,KAAKS,CAAC;YACjH,IAAI,CAACD,mBAAmBnB;QAC1B;QAEA,IAAIJ,QAAQrB,GAAG,CAACpB,OAAOwD,KAAKS,CAAC,IAAI;YAC/B,uEAAuE;YACvE,4DAA4D;YAC5DpB;YACA,MAAMqB,MAAiB;gBAACV;aAAK;YAC7B,IAAIW,OAAOnE,OAAOwD,KAAKS,CAAC;YACxB,MAAOX,IAAI,IAAId,EAAEe,KAAK,CAAC1B,MAAM,IAAIY,QAAQrB,GAAG,CAACpB,OAAOwC,EAAEe,KAAK,CAACD,IAAI,EAAE,CAACW,CAAC,MAAMzB,EAAEe,KAAK,CAACD,IAAI,EAAE,CAACI,GAAG,KAAKlB,EAAEe,KAAK,CAACD,EAAE,CAACI,GAAG,IAAIlB,EAAEe,KAAK,CAACD,EAAE,CAACK,IAAI,GAAGnB,EAAEe,KAAK,CAACD,IAAI,EAAE,CAACK,IAAI,GAAGS,KAAKC,GAAG,CAAC7B,EAAEe,KAAK,CAACD,EAAE,CAACW,CAAC,EAAEzB,EAAEe,KAAK,CAACD,IAAI,EAAE,CAACW,CAAC,IAAI,IAAK;gBACzMX;gBACAY,IAAIb,IAAI,CAACb,EAAEe,KAAK,CAACD,EAAE;gBACnBa,OAAOC,KAAKC,GAAG,CAACF,MAAMnE,OAAOwC,EAAEe,KAAK,CAACD,EAAE,CAACW,CAAC;YAC3C;YACA,MAAMK,WAAWJ,IAAInB,GAAG,CAAC,CAACC,IAAMA,EAAEtB,IAAI;YACtC,MAAMuB,SAASqB,SAAShE,KAAK,CAAC,GAAG4C,MAAM,CAACjD,YAAYqE,QAAQ,CAAC,EAAE;YAC/D,MAAMnB,QAAQe,IAAId,OAAO,CAAC,CAACJ,IAAMA,EAAEG,KAAK;YACxC,sEAAsE;YACtE,qFAAqF;YACrF,MAAMoB,OAAO9C,eAAewB;YAC5B,IAAIsB,MAAM5B,OAAOU,IAAI,CAAC;gBAAEjB,MAAM;gBAAUV,MAAMuB;gBAAQhB,QAAQX,cAAcC,kBAAkB;gBAAE4B;YAAM;iBACjG,IAAIF,OAAOpB,MAAM,GAAG,KAAKc,OAAOU,IAAI,CAAC;gBAAEjB,MAAM;gBAASV,MAAMuB;gBAAQE;YAAM;iBAC1ER,OAAOU,IAAI,CAAC;gBAAEjB,MAAM;gBAAWV,MAAMuB;gBAAQX,OAAOG,QAAQ+B,GAAG,CAACL;gBAAiBhB;YAAM;QAC9F,OAAO;YACLP,KAAKS,IAAI,CAACG;QACZ;QACAF;IACF;IACAT;IACA,OAAOF;AACT"}
@@ -1 +1 @@
1
- {"version":3,"sources":["/Users/kevin/Projects/pdf-to-md/src/engines/tesseract.ts"],"sourcesContent":["// engines/tesseract.ts — THE OCR engine: tesseract.js on every platform.\n//\n// Seam contract: words come back in normalized 0-1 page coordinates, origin\n// bottom-left, y increasing upward — the y-flip happens here, and nowhere else.\n//\n// Pure WASM, in-process, no native dependency: no crash containment needed.\nimport { mkdirSync } from 'node:fs';\nimport os from 'node:os';\nimport path from 'node:path';\nimport type { Page as TesseractPage } from 'tesseract.js';\nimport { createWorker, PSM } from 'tesseract.js';\nimport type { PageWord, Word } from '../geometry.ts';\nimport { foldTypographic } from '../lines.ts';\nimport { awaitPdfOpen, openPdfForRender } from '../pdf-open.ts';\nimport { renderLoadedPage } from '../raster.ts';\nimport { resolvePackagePath } from '../resolve.ts';\n\n// 288 dpi: the render resolution the engine's word-overlap numbers were\n// measured at (test/fixtures/ocr-engine-poc/) — the known-good operating point.\nexport const OCR_RENDER_DPI = 288;\n\n// Chart pages starve PSM.AUTO (unrecognized text can't be marked for review).\n// Prose-fraction retry trigger, measured (sparse-retry.test.ts): chart 0.442 vs prose 0.767-0.938, cut 0.6.\nexport const SPARSE_RETRY_PROSE_FRACTION = 0.6;\n\n/** Share of `words` whose text contains a run of 3+ letters (`src/emit.ts`'s isUnstructured canary, applied per-page instead of per-block). */\nfunction proseFraction(words: { text: string }[]): number {\n if (words.length === 0) return 0;\n return words.filter((w) => /[A-Za-z]{3}/.test(w.text)).length / words.length;\n}\n\n/** Does this page's PSM.AUTO pass look chart-like enough to re-recognize\n * with PSM.SPARSE_TEXT? A page with 0 words never retries (nothing to take a prose fraction of). */\nexport function shouldRetrySparse(words: { text: string }[]): boolean {\n return words.length > 0 && proseFraction(words) < SPARSE_RETRY_PROSE_FRACTION;\n}\n\n// The locally installed traineddata (npm package @tesseract.js-data/eng),\n// resolved dynamically so depcheck won't flag it \"unused\" — it's a genuine runtime dependency, listed in .depcheckrc's `ignores`.\nconst TESSDATA_DIR = path.join(path.dirname(resolvePackagePath('@tesseract.js-data/eng/package.json')), '4.0.0');\n\n/** Per-page recognition progress, delivered once per page as it finishes.\n * Engine chatter goes through `onWarning`/`warnings` instead — this is timing/count data only. */\nexport type OcrProgressEvent = { type: 'page'; page: number; words: number; ms: number };\n\nexport interface RecognizeResult {\n words: PageWord[];\n /** pages (1-based) where the engine returned zero words */\n emptyPages: number[];\n /** wall-clock ms per page (render + recognize), keyed by page number */\n msPerPage: Map<number, number>;\n}\n\n/** tesseract.js's nested blocks/paragraphs/lines/words -> this repo's flat Word contract (normalize + y-flip). Same shape regardless of which PSM produced `page`, so both the AUTO pass and the SPARSE_TEXT retry share it. */\nfunction wordsFromPage(page: TesseractPage, width: number, height: number): Word[] {\n const pageWords: Word[] = [];\n for (const block of page.blocks ?? []) {\n for (const para of block.paragraphs ?? []) {\n for (const line of para.lines ?? []) {\n for (const w of line.words ?? []) {\n const { x0, y0, x1, y1 } = w.bbox;\n pageWords.push({\n // Same typographic folding as the text path (src/lines.ts) — both\n // paths must agree on how a printed apostrophe is spelled.\n text: foldTypographic(w.text),\n x: x0 / width,\n y: (height - y1) / height, // y-flip: top-left px -> bottom-left normalized\n w: (x1 - x0) / width,\n h: (y1 - y0) / height,\n // Plumbed through unmodified — tesseract.js's own 0-100\n // per-word confidence. See src/report.ts's LOW_CONFIDENCE_THRESHOLD.\n confidence: w.confidence,\n });\n }\n }\n }\n }\n return pageWords;\n}\n\n// tesseract.js-core's WASM build writes native diagnostics straight to this\n// process's stdout/stderr, with no override to silence them at the source.\n//\n// Suppression must span the whole run: worker-thread stdio arrives\n// asynchronously, on later event-loop turns; the synchronous carve-out below (`withRealStdio`) is the only safe exit.\nlet suppressDepth = 0;\nlet savedStdoutWrite: typeof process.stdout.write | null = null;\nlet savedStderrWrite: typeof process.stderr.write | null = null;\nasync function withSuppressedStdio<T>(fn: () => Promise<T>): Promise<T> {\n if (suppressDepth === 0) {\n savedStdoutWrite = process.stdout.write.bind(process.stdout);\n savedStderrWrite = process.stderr.write.bind(process.stderr);\n process.stdout.write = (() => true) as typeof process.stdout.write;\n process.stderr.write = (() => true) as typeof process.stderr.write;\n }\n suppressDepth++;\n try {\n return await fn();\n } finally {\n suppressDepth--;\n if (suppressDepth === 0) {\n if (savedStdoutWrite) process.stdout.write = savedStdoutWrite;\n if (savedStderrWrite) process.stderr.write = savedStderrWrite;\n savedStdoutWrite = null;\n savedStderrWrite = null;\n }\n }\n}\n\n/** Pierces an active `withSuppressedStdio` region SYNCHRONOUSLY to let this\n * module's own onProgress/onWarning calls through. No-op outside any suppressed region. */\nfunction withRealStdio<T>(fn: () => T): T {\n if (suppressDepth === 0) return fn();\n const realOut = savedStdoutWrite;\n const realErr = savedStderrWrite;\n if (realOut) process.stdout.write = realOut;\n if (realErr) process.stderr.write = realErr;\n try {\n return fn();\n } finally {\n process.stdout.write = (() => true) as typeof process.stdout.write;\n process.stderr.write = (() => true) as typeof process.stderr.write;\n }\n}\n\n/** Recognize words on pages [first..last] of a PDF (renders at `dpi`,\n * default OCR_RENDER_DPI). Progress via `onProgress`; diagnostics via `onWarning` — this module never writes to stdout/stderr itself. */\nexport async function recognizePdf(pdfPath: string, first: number, last: number, { dpi = OCR_RENDER_DPI, onProgress, onWarning }: { dpi?: number; onProgress?: (e: OcrProgressEvent) => void; onWarning?: (message: string) => void } = {}): Promise<RecognizeResult> {\n // tesseract.js writes its decompressed traineddata cache here by default,\n // dropping a model file into the user's cwd — redirect it under tmpdir() instead.\n const cachePath = path.join(os.tmpdir(), 'pdf-to-md-engine');\n mkdirSync(cachePath, { recursive: true });\n\n const words: PageWord[] = [];\n const emptyPages: number[] = [];\n const msPerPage = new Map<number, number>();\n\n // ONE continuous suppressed region for the whole run; onProgress/onWarning\n // below go through `withRealStdio`, the only sanctioned way out of it.\n await withSuppressedStdio(async () => {\n // OEM 1 (LSTM only, tesseract.js's own default): measured 99.0% recall\n // on the hard benchmark pages (2026-08-16).\n const worker = await createWorker('eng', 1, {\n langPath: TESSDATA_DIR,\n cachePath,\n });\n try {\n // NOT optional: tesseract.js's own default PSM (SINGLE_BLOCK) measured\n // 3-8 percentage points worse than AUTO on two-column pages.\n await worker.setParameters({ tessedit_pageseg_mode: PSM.AUTO });\n\n const loadingTask = openPdfForRender(pdfPath);\n try {\n const pdf = await awaitPdfOpen(loadingTask, pdfPath);\n for (let pn = first; pn <= Math.min(last, pdf.numPages); pn++) {\n const t0 = Date.now();\n const { png, width, height } = await renderLoadedPage(pdf, pn, dpi);\n const { data } = await worker.recognize(png, {}, { blocks: true });\n let pageWords: Word[] = wordsFromPage(data, width, height);\n\n // Adaptive second pass: re-recognizes the SAME rendered PNG, so\n // this costs one extra recognize() call only on chart-like pages.\n if (shouldRetrySparse(pageWords)) {\n const autoFraction = proseFraction(pageWords);\n const autoWords = pageWords;\n // setParameters on the shared worker is reversible, so one\n // in-process worker can alternate PSM per page instead of paying for a second worker.\n await worker.setParameters({ tessedit_pageseg_mode: PSM.SPARSE_TEXT });\n const sparse = await worker.recognize(png, {}, { blocks: true });\n const sparseWords = wordsFromPage(sparse.data, width, height);\n await worker.setParameters({ tessedit_pageseg_mode: PSM.AUTO });\n\n if (sparseWords.length > autoWords.length) {\n pageWords = sparseWords;\n const autoCount = autoWords.length;\n const sparseCount = sparseWords.length;\n withRealStdio(() => onWarning?.(`p${pn}: chart-like page (prose fraction ${autoFraction.toFixed(2)}) — retried with sparse-text segmentation, ${autoCount} -> ${sparseCount} words`));\n }\n }\n\n const ms = Date.now() - t0;\n for (const w of pageWords) words.push({ page: pn, ...w });\n if (pageWords.length === 0) emptyPages.push(pn);\n msPerPage.set(pn, ms);\n const wordCount = pageWords.length;\n withRealStdio(() => onProgress?.({ type: 'page', page: pn, words: wordCount, ms }));\n }\n } finally {\n await loadingTask.destroy();\n }\n } finally {\n await worker.terminate();\n }\n });\n\n return { words, emptyPages, msPerPage };\n}\n"],"names":["mkdirSync","os","path","createWorker","PSM","foldTypographic","awaitPdfOpen","openPdfForRender","renderLoadedPage","resolvePackagePath","OCR_RENDER_DPI","SPARSE_RETRY_PROSE_FRACTION","proseFraction","words","length","filter","w","test","text","shouldRetrySparse","TESSDATA_DIR","join","dirname","wordsFromPage","page","width","height","pageWords","block","blocks","para","paragraphs","line","lines","x0","y0","x1","y1","bbox","push","x","y","h","confidence","suppressDepth","savedStdoutWrite","savedStderrWrite","withSuppressedStdio","fn","process","stdout","write","bind","stderr","withRealStdio","realOut","realErr","recognizePdf","pdfPath","first","last","dpi","onProgress","onWarning","cachePath","tmpdir","recursive","emptyPages","msPerPage","Map","worker","langPath","setParameters","tessedit_pageseg_mode","AUTO","loadingTask","pdf","pn","Math","min","numPages","t0","Date","now","png","data","recognize","autoFraction","autoWords","SPARSE_TEXT","sparse","sparseWords","autoCount","sparseCount","toFixed","ms","set","wordCount","type","destroy","terminate"],"mappings":"AAAA,yEAAyE;AACzE,EAAE;AACF,4EAA4E;AAC5E,gFAAgF;AAChF,EAAE;AACF,4EAA4E;AAC5E,SAASA,SAAS,QAAQ,UAAU;AACpC,OAAOC,QAAQ,UAAU;AACzB,OAAOC,UAAU,YAAY;AAE7B,SAASC,YAAY,EAAEC,GAAG,QAAQ,eAAe;AAEjD,SAASC,eAAe,QAAQ,cAAc;AAC9C,SAASC,YAAY,EAAEC,gBAAgB,QAAQ,iBAAiB;AAChE,SAASC,gBAAgB,QAAQ,eAAe;AAChD,SAASC,kBAAkB,QAAQ,gBAAgB;AAEnD,wEAAwE;AACxE,gFAAgF;AAChF,OAAO,MAAMC,iBAAiB,IAAI;AAElC,8EAA8E;AAC9E,4GAA4G;AAC5G,OAAO,MAAMC,8BAA8B,IAAI;AAE/C,6IAA6I,GAC7I,SAASC,cAAcC,KAAyB;IAC9C,IAAIA,MAAMC,MAAM,KAAK,GAAG,OAAO;IAC/B,OAAOD,MAAME,MAAM,CAAC,CAACC,IAAM,cAAcC,IAAI,CAACD,EAAEE,IAAI,GAAGJ,MAAM,GAAGD,MAAMC,MAAM;AAC9E;AAEA;kGACkG,GAClG,OAAO,SAASK,kBAAkBN,KAAyB;IACzD,OAAOA,MAAMC,MAAM,GAAG,KAAKF,cAAcC,SAASF;AACpD;AAEA,0EAA0E;AAC1E,kIAAkI;AAClI,MAAMS,eAAelB,KAAKmB,IAAI,CAACnB,KAAKoB,OAAO,CAACb,mBAAmB,yCAAyC;AAcxG,8NAA8N,GAC9N,SAASc,cAAcC,IAAmB,EAAEC,KAAa,EAAEC,MAAc;QAEnDF;IADpB,MAAMG,YAAoB,EAAE;IAC5B,KAAK,MAAMC,UAASJ,eAAAA,KAAKK,MAAM,cAAXL,0BAAAA,eAAe,EAAE,CAAE;YAClBI;QAAnB,KAAK,MAAME,SAAQF,oBAAAA,MAAMG,UAAU,cAAhBH,+BAAAA,oBAAoB,EAAE,CAAE;gBACtBE;YAAnB,KAAK,MAAME,SAAQF,cAAAA,KAAKG,KAAK,cAAVH,yBAAAA,cAAc,EAAE,CAAE;oBACnBE;gBAAhB,KAAK,MAAMhB,MAAKgB,cAAAA,KAAKnB,KAAK,cAAVmB,yBAAAA,cAAc,EAAE,CAAE;oBAChC,MAAM,EAAEE,EAAE,EAAEC,EAAE,EAAEC,EAAE,EAAEC,EAAE,EAAE,GAAGrB,EAAEsB,IAAI;oBACjCX,UAAUY,IAAI,CAAC;wBACb,kEAAkE;wBAClE,2DAA2D;wBAC3DrB,MAAMb,gBAAgBW,EAAEE,IAAI;wBAC5BsB,GAAGN,KAAKT;wBACRgB,GAAG,AAACf,CAAAA,SAASW,EAAC,IAAKX;wBACnBV,GAAG,AAACoB,CAAAA,KAAKF,EAAC,IAAKT;wBACfiB,GAAG,AAACL,CAAAA,KAAKF,EAAC,IAAKT;wBACf,wDAAwD;wBACxD,qEAAqE;wBACrEiB,YAAY3B,EAAE2B,UAAU;oBAC1B;gBACF;YACF;QACF;IACF;IACA,OAAOhB;AACT;AAEA,4EAA4E;AAC5E,2EAA2E;AAC3E,EAAE;AACF,mEAAmE;AACnE,sHAAsH;AACtH,IAAIiB,gBAAgB;AACpB,IAAIC,mBAAuD;AAC3D,IAAIC,mBAAuD;AAC3D,eAAeC,oBAAuBC,EAAoB;IACxD,IAAIJ,kBAAkB,GAAG;QACvBC,mBAAmBI,QAAQC,MAAM,CAACC,KAAK,CAACC,IAAI,CAACH,QAAQC,MAAM;QAC3DJ,mBAAmBG,QAAQI,MAAM,CAACF,KAAK,CAACC,IAAI,CAACH,QAAQI,MAAM;QAC3DJ,QAAQC,MAAM,CAACC,KAAK,GAAI,IAAM;QAC9BF,QAAQI,MAAM,CAACF,KAAK,GAAI,IAAM;IAChC;IACAP;IACA,IAAI;QACF,OAAO,MAAMI;IACf,SAAU;QACRJ;QACA,IAAIA,kBAAkB,GAAG;YACvB,IAAIC,kBAAkBI,QAAQC,MAAM,CAACC,KAAK,GAAGN;YAC7C,IAAIC,kBAAkBG,QAAQI,MAAM,CAACF,KAAK,GAAGL;YAC7CD,mBAAmB;YACnBC,mBAAmB;QACrB;IACF;AACF;AAEA;yFACyF,GACzF,SAASQ,cAAiBN,EAAW;IACnC,IAAIJ,kBAAkB,GAAG,OAAOI;IAChC,MAAMO,UAAUV;IAChB,MAAMW,UAAUV;IAChB,IAAIS,SAASN,QAAQC,MAAM,CAACC,KAAK,GAAGI;IACpC,IAAIC,SAASP,QAAQI,MAAM,CAACF,KAAK,GAAGK;IACpC,IAAI;QACF,OAAOR;IACT,SAAU;QACRC,QAAQC,MAAM,CAACC,KAAK,GAAI,IAAM;QAC9BF,QAAQI,MAAM,CAACF,KAAK,GAAI,IAAM;IAChC;AACF;AAEA;uIACuI,GACvI,OAAO,eAAeM,aAAaC,OAAe,EAAEC,KAAa,EAAEC,IAAY,EAAE,EAAEC,MAAMnD,cAAc,EAAEoD,UAAU,EAAEC,SAAS,EAAuG,GAAG,CAAC,CAAC;IACxO,0EAA0E;IAC1E,kFAAkF;IAClF,MAAMC,YAAY9D,KAAKmB,IAAI,CAACpB,GAAGgE,MAAM,IAAI;IACzCjE,UAAUgE,WAAW;QAAEE,WAAW;IAAK;IAEvC,MAAMrD,QAAoB,EAAE;IAC5B,MAAMsD,aAAuB,EAAE;IAC/B,MAAMC,YAAY,IAAIC;IAEtB,2EAA2E;IAC3E,uEAAuE;IACvE,MAAMtB,oBAAoB;QACxB,uEAAuE;QACvE,4CAA4C;QAC5C,MAAMuB,SAAS,MAAMnE,aAAa,OAAO,GAAG;YAC1CoE,UAAUnD;YACV4C;QACF;QACA,IAAI;YACF,uEAAuE;YACvE,6DAA6D;YAC7D,MAAMM,OAAOE,aAAa,CAAC;gBAAEC,uBAAuBrE,IAAIsE,IAAI;YAAC;YAE7D,MAAMC,cAAcpE,iBAAiBmD;YACrC,IAAI;gBACF,MAAMkB,MAAM,MAAMtE,aAAaqE,aAAajB;gBAC5C,IAAK,IAAImB,KAAKlB,OAAOkB,MAAMC,KAAKC,GAAG,CAACnB,MAAMgB,IAAII,QAAQ,GAAGH,KAAM;oBAC7D,MAAMI,KAAKC,KAAKC,GAAG;oBACnB,MAAM,EAAEC,GAAG,EAAE3D,KAAK,EAAEC,MAAM,EAAE,GAAG,MAAMlB,iBAAiBoE,KAAKC,IAAIhB;oBAC/D,MAAM,EAAEwB,IAAI,EAAE,GAAG,MAAMf,OAAOgB,SAAS,CAACF,KAAK,CAAC,GAAG;wBAAEvD,QAAQ;oBAAK;oBAChE,IAAIF,YAAoBJ,cAAc8D,MAAM5D,OAAOC;oBAEnD,gEAAgE;oBAChE,kEAAkE;oBAClE,IAAIP,kBAAkBQ,YAAY;wBAChC,MAAM4D,eAAe3E,cAAce;wBACnC,MAAM6D,YAAY7D;wBAClB,2DAA2D;wBAC3D,sFAAsF;wBACtF,MAAM2C,OAAOE,aAAa,CAAC;4BAAEC,uBAAuBrE,IAAIqF,WAAW;wBAAC;wBACpE,MAAMC,SAAS,MAAMpB,OAAOgB,SAAS,CAACF,KAAK,CAAC,GAAG;4BAAEvD,QAAQ;wBAAK;wBAC9D,MAAM8D,cAAcpE,cAAcmE,OAAOL,IAAI,EAAE5D,OAAOC;wBACtD,MAAM4C,OAAOE,aAAa,CAAC;4BAAEC,uBAAuBrE,IAAIsE,IAAI;wBAAC;wBAE7D,IAAIiB,YAAY7E,MAAM,GAAG0E,UAAU1E,MAAM,EAAE;4BACzCa,YAAYgE;4BACZ,MAAMC,YAAYJ,UAAU1E,MAAM;4BAClC,MAAM+E,cAAcF,YAAY7E,MAAM;4BACtCwC,cAAc,IAAMS,sBAAAA,gCAAAA,UAAY,CAAC,CAAC,EAAEc,GAAG,kCAAkC,EAAEU,aAAaO,OAAO,CAAC,GAAG,2CAA2C,EAAEF,UAAU,IAAI,EAAEC,YAAY,MAAM,CAAC;wBACrL;oBACF;oBAEA,MAAME,KAAKb,KAAKC,GAAG,KAAKF;oBACxB,KAAK,MAAMjE,KAAKW,UAAWd,MAAM0B,IAAI,CAAC;wBAAEf,MAAMqD;wBAAI,GAAG7D,CAAC;oBAAC;oBACvD,IAAIW,UAAUb,MAAM,KAAK,GAAGqD,WAAW5B,IAAI,CAACsC;oBAC5CT,UAAU4B,GAAG,CAACnB,IAAIkB;oBAClB,MAAME,YAAYtE,UAAUb,MAAM;oBAClCwC,cAAc,IAAMQ,uBAAAA,iCAAAA,WAAa;4BAAEoC,MAAM;4BAAQ1E,MAAMqD;4BAAIhE,OAAOoF;4BAAWF;wBAAG;gBAClF;YACF,SAAU;gBACR,MAAMpB,YAAYwB,OAAO;YAC3B;QACF,SAAU;YACR,MAAM7B,OAAO8B,SAAS;QACxB;IACF;IAEA,OAAO;QAAEvF;QAAOsD;QAAYC;IAAU;AACxC"}
1
+ {"version":3,"sources":["/Users/kevin/Dev/OpenSource/formats/pdf-to-md/src/engines/tesseract.ts"],"sourcesContent":["// engines/tesseract.ts — THE OCR engine: tesseract.js on every platform.\n//\n// Seam contract: words come back in normalized 0-1 page coordinates, origin\n// bottom-left, y increasing upward — the y-flip happens here, and nowhere else.\n//\n// Pure WASM, in-process, no native dependency: no crash containment needed.\nimport { mkdirSync } from 'node:fs';\nimport os from 'node:os';\nimport path from 'node:path';\nimport type { Page as TesseractPage } from 'tesseract.js';\nimport { createWorker, PSM } from 'tesseract.js';\nimport type { PageWord, Word } from '../geometry.ts';\nimport { foldTypographic } from '../lines.ts';\nimport { awaitPdfOpen, openPdfForRender } from '../pdf-open.ts';\nimport { renderLoadedPage } from '../raster.ts';\nimport { resolvePackagePath } from '../resolve.ts';\n\n// 288 dpi: the render resolution the engine's word-overlap numbers were\n// measured at (test/fixtures/ocr-engine-poc/) — the known-good operating point.\nexport const OCR_RENDER_DPI = 288;\n\n// Chart pages starve PSM.AUTO (unrecognized text can't be marked for review).\n// Prose-fraction retry trigger, measured (sparse-retry.test.ts): chart 0.442 vs prose 0.767-0.938, cut 0.6.\nexport const SPARSE_RETRY_PROSE_FRACTION = 0.6;\n\n/** Share of `words` whose text contains a run of 3+ letters (`src/emit.ts`'s isUnstructured canary, applied per-page instead of per-block). */\nfunction proseFraction(words: { text: string }[]): number {\n if (words.length === 0) return 0;\n return words.filter((w) => /[A-Za-z]{3}/.test(w.text)).length / words.length;\n}\n\n/** Does this page's PSM.AUTO pass look chart-like enough to re-recognize\n * with PSM.SPARSE_TEXT? A page with 0 words never retries (nothing to take a prose fraction of). */\nexport function shouldRetrySparse(words: { text: string }[]): boolean {\n return words.length > 0 && proseFraction(words) < SPARSE_RETRY_PROSE_FRACTION;\n}\n\n// The locally installed traineddata (npm package @tesseract.js-data/eng),\n// resolved dynamically so depcheck won't flag it \"unused\" — it's a genuine runtime dependency, listed in .depcheckrc's `ignores`.\nconst TESSDATA_DIR = path.join(path.dirname(resolvePackagePath('@tesseract.js-data/eng/package.json')), '4.0.0');\n\n/** Per-page recognition progress, delivered once per page as it finishes.\n * Engine chatter goes through `onWarning`/`warnings` instead — this is timing/count data only. */\nexport type OcrProgressEvent = { type: 'page'; page: number; words: number; ms: number };\n\nexport interface RecognizeResult {\n words: PageWord[];\n /** pages (1-based) where the engine returned zero words */\n emptyPages: number[];\n /** wall-clock ms per page (render + recognize), keyed by page number */\n msPerPage: Map<number, number>;\n}\n\n/** tesseract.js's nested blocks/paragraphs/lines/words -> this repo's flat Word contract (normalize + y-flip). Same shape regardless of which PSM produced `page`, so both the AUTO pass and the SPARSE_TEXT retry share it. */\nfunction wordsFromPage(page: TesseractPage, width: number, height: number): Word[] {\n const pageWords: Word[] = [];\n for (const block of page.blocks ?? []) {\n for (const para of block.paragraphs ?? []) {\n for (const line of para.lines ?? []) {\n for (const w of line.words ?? []) {\n const { x0, y0, x1, y1 } = w.bbox;\n pageWords.push({\n // Same typographic folding as the text path (src/lines.ts) — both\n // paths must agree on how a printed apostrophe is spelled.\n text: foldTypographic(w.text),\n x: x0 / width,\n y: (height - y1) / height, // y-flip: top-left px -> bottom-left normalized\n w: (x1 - x0) / width,\n h: (y1 - y0) / height,\n // Plumbed through unmodified — tesseract.js's own 0-100\n // per-word confidence. See src/report.ts's LOW_CONFIDENCE_THRESHOLD.\n confidence: w.confidence,\n });\n }\n }\n }\n }\n return pageWords;\n}\n\n// tesseract.js-core's WASM build writes native diagnostics straight to this\n// process's stdout/stderr, with no override to silence them at the source.\n//\n// Suppression must span the whole run: worker-thread stdio arrives\n// asynchronously, on later event-loop turns; the synchronous carve-out below (`withRealStdio`) is the only safe exit.\nlet suppressDepth = 0;\nlet savedStdoutWrite: typeof process.stdout.write | null = null;\nlet savedStderrWrite: typeof process.stderr.write | null = null;\nasync function withSuppressedStdio<T>(fn: () => Promise<T>): Promise<T> {\n if (suppressDepth === 0) {\n savedStdoutWrite = process.stdout.write.bind(process.stdout);\n savedStderrWrite = process.stderr.write.bind(process.stderr);\n process.stdout.write = (() => true) as typeof process.stdout.write;\n process.stderr.write = (() => true) as typeof process.stderr.write;\n }\n suppressDepth++;\n try {\n return await fn();\n } finally {\n suppressDepth--;\n if (suppressDepth === 0) {\n if (savedStdoutWrite) process.stdout.write = savedStdoutWrite;\n if (savedStderrWrite) process.stderr.write = savedStderrWrite;\n savedStdoutWrite = null;\n savedStderrWrite = null;\n }\n }\n}\n\n/** Pierces an active `withSuppressedStdio` region SYNCHRONOUSLY to let this\n * module's own onProgress/onWarning calls through. No-op outside any suppressed region. */\nfunction withRealStdio<T>(fn: () => T): T {\n if (suppressDepth === 0) return fn();\n const realOut = savedStdoutWrite;\n const realErr = savedStderrWrite;\n if (realOut) process.stdout.write = realOut;\n if (realErr) process.stderr.write = realErr;\n try {\n return fn();\n } finally {\n process.stdout.write = (() => true) as typeof process.stdout.write;\n process.stderr.write = (() => true) as typeof process.stderr.write;\n }\n}\n\n/** Recognize words on pages [first..last] of a PDF (renders at `dpi`,\n * default OCR_RENDER_DPI). Progress via `onProgress`; diagnostics via `onWarning` — this module never writes to stdout/stderr itself. */\nexport async function recognizePdf(pdfPath: string, first: number, last: number, { dpi = OCR_RENDER_DPI, onProgress, onWarning }: { dpi?: number; onProgress?: (e: OcrProgressEvent) => void; onWarning?: (message: string) => void } = {}): Promise<RecognizeResult> {\n // tesseract.js writes its decompressed traineddata cache here by default,\n // dropping a model file into the user's cwd — redirect it under tmpdir() instead.\n const cachePath = path.join(os.tmpdir(), 'pdf-to-md-engine');\n mkdirSync(cachePath, { recursive: true });\n\n const words: PageWord[] = [];\n const emptyPages: number[] = [];\n const msPerPage = new Map<number, number>();\n\n // ONE continuous suppressed region for the whole run; onProgress/onWarning\n // below go through `withRealStdio`, the only sanctioned way out of it.\n await withSuppressedStdio(async () => {\n // OEM 1 (LSTM only, tesseract.js's own default): measured 99.0% recall\n // on the hard benchmark pages (2026-08-16).\n const worker = await createWorker('eng', 1, {\n langPath: TESSDATA_DIR,\n cachePath,\n });\n try {\n // NOT optional: tesseract.js's own default PSM (SINGLE_BLOCK) measured\n // 3-8 percentage points worse than AUTO on two-column pages.\n await worker.setParameters({ tessedit_pageseg_mode: PSM.AUTO });\n\n const loadingTask = openPdfForRender(pdfPath);\n try {\n const pdf = await awaitPdfOpen(loadingTask, pdfPath);\n for (let pn = first; pn <= Math.min(last, pdf.numPages); pn++) {\n const t0 = Date.now();\n const { png, width, height } = await renderLoadedPage(pdf, pn, dpi);\n const { data } = await worker.recognize(png, {}, { blocks: true });\n let pageWords: Word[] = wordsFromPage(data, width, height);\n\n // Adaptive second pass: re-recognizes the SAME rendered PNG, so\n // this costs one extra recognize() call only on chart-like pages.\n if (shouldRetrySparse(pageWords)) {\n const autoFraction = proseFraction(pageWords);\n const autoWords = pageWords;\n // setParameters on the shared worker is reversible, so one\n // in-process worker can alternate PSM per page instead of paying for a second worker.\n await worker.setParameters({ tessedit_pageseg_mode: PSM.SPARSE_TEXT });\n const sparse = await worker.recognize(png, {}, { blocks: true });\n const sparseWords = wordsFromPage(sparse.data, width, height);\n await worker.setParameters({ tessedit_pageseg_mode: PSM.AUTO });\n\n if (sparseWords.length > autoWords.length) {\n pageWords = sparseWords;\n const autoCount = autoWords.length;\n const sparseCount = sparseWords.length;\n withRealStdio(() => onWarning?.(`p${pn}: chart-like page (prose fraction ${autoFraction.toFixed(2)}) — retried with sparse-text segmentation, ${autoCount} -> ${sparseCount} words`));\n }\n }\n\n const ms = Date.now() - t0;\n for (const w of pageWords) words.push({ page: pn, ...w });\n if (pageWords.length === 0) emptyPages.push(pn);\n msPerPage.set(pn, ms);\n const wordCount = pageWords.length;\n withRealStdio(() => onProgress?.({ type: 'page', page: pn, words: wordCount, ms }));\n }\n } finally {\n await loadingTask.destroy();\n }\n } finally {\n await worker.terminate();\n }\n });\n\n return { words, emptyPages, msPerPage };\n}\n"],"names":["mkdirSync","os","path","createWorker","PSM","foldTypographic","awaitPdfOpen","openPdfForRender","renderLoadedPage","resolvePackagePath","OCR_RENDER_DPI","SPARSE_RETRY_PROSE_FRACTION","proseFraction","words","length","filter","w","test","text","shouldRetrySparse","TESSDATA_DIR","join","dirname","wordsFromPage","page","width","height","pageWords","block","blocks","para","paragraphs","line","lines","x0","y0","x1","y1","bbox","push","x","y","h","confidence","suppressDepth","savedStdoutWrite","savedStderrWrite","withSuppressedStdio","fn","process","stdout","write","bind","stderr","withRealStdio","realOut","realErr","recognizePdf","pdfPath","first","last","dpi","onProgress","onWarning","cachePath","tmpdir","recursive","emptyPages","msPerPage","Map","worker","langPath","setParameters","tessedit_pageseg_mode","AUTO","loadingTask","pdf","pn","Math","min","numPages","t0","Date","now","png","data","recognize","autoFraction","autoWords","SPARSE_TEXT","sparse","sparseWords","autoCount","sparseCount","toFixed","ms","set","wordCount","type","destroy","terminate"],"mappings":"AAAA,yEAAyE;AACzE,EAAE;AACF,4EAA4E;AAC5E,gFAAgF;AAChF,EAAE;AACF,4EAA4E;AAC5E,SAASA,SAAS,QAAQ,UAAU;AACpC,OAAOC,QAAQ,UAAU;AACzB,OAAOC,UAAU,YAAY;AAE7B,SAASC,YAAY,EAAEC,GAAG,QAAQ,eAAe;AAEjD,SAASC,eAAe,QAAQ,cAAc;AAC9C,SAASC,YAAY,EAAEC,gBAAgB,QAAQ,iBAAiB;AAChE,SAASC,gBAAgB,QAAQ,eAAe;AAChD,SAASC,kBAAkB,QAAQ,gBAAgB;AAEnD,wEAAwE;AACxE,gFAAgF;AAChF,OAAO,MAAMC,iBAAiB,IAAI;AAElC,8EAA8E;AAC9E,4GAA4G;AAC5G,OAAO,MAAMC,8BAA8B,IAAI;AAE/C,6IAA6I,GAC7I,SAASC,cAAcC,KAAyB;IAC9C,IAAIA,MAAMC,MAAM,KAAK,GAAG,OAAO;IAC/B,OAAOD,MAAME,MAAM,CAAC,CAACC,IAAM,cAAcC,IAAI,CAACD,EAAEE,IAAI,GAAGJ,MAAM,GAAGD,MAAMC,MAAM;AAC9E;AAEA;kGACkG,GAClG,OAAO,SAASK,kBAAkBN,KAAyB;IACzD,OAAOA,MAAMC,MAAM,GAAG,KAAKF,cAAcC,SAASF;AACpD;AAEA,0EAA0E;AAC1E,kIAAkI;AAClI,MAAMS,eAAelB,KAAKmB,IAAI,CAACnB,KAAKoB,OAAO,CAACb,mBAAmB,yCAAyC;AAcxG,8NAA8N,GAC9N,SAASc,cAAcC,IAAmB,EAAEC,KAAa,EAAEC,MAAc;QAEnDF;IADpB,MAAMG,YAAoB,EAAE;IAC5B,KAAK,MAAMC,UAASJ,eAAAA,KAAKK,MAAM,cAAXL,0BAAAA,eAAe,EAAE,CAAE;YAClBI;QAAnB,KAAK,MAAME,SAAQF,oBAAAA,MAAMG,UAAU,cAAhBH,+BAAAA,oBAAoB,EAAE,CAAE;gBACtBE;YAAnB,KAAK,MAAME,SAAQF,cAAAA,KAAKG,KAAK,cAAVH,yBAAAA,cAAc,EAAE,CAAE;oBACnBE;gBAAhB,KAAK,MAAMhB,MAAKgB,cAAAA,KAAKnB,KAAK,cAAVmB,yBAAAA,cAAc,EAAE,CAAE;oBAChC,MAAM,EAAEE,EAAE,EAAEC,EAAE,EAAEC,EAAE,EAAEC,EAAE,EAAE,GAAGrB,EAAEsB,IAAI;oBACjCX,UAAUY,IAAI,CAAC;wBACb,kEAAkE;wBAClE,2DAA2D;wBAC3DrB,MAAMb,gBAAgBW,EAAEE,IAAI;wBAC5BsB,GAAGN,KAAKT;wBACRgB,GAAG,AAACf,CAAAA,SAASW,EAAC,IAAKX;wBACnBV,GAAG,AAACoB,CAAAA,KAAKF,EAAC,IAAKT;wBACfiB,GAAG,AAACL,CAAAA,KAAKF,EAAC,IAAKT;wBACf,wDAAwD;wBACxD,qEAAqE;wBACrEiB,YAAY3B,EAAE2B,UAAU;oBAC1B;gBACF;YACF;QACF;IACF;IACA,OAAOhB;AACT;AAEA,4EAA4E;AAC5E,2EAA2E;AAC3E,EAAE;AACF,mEAAmE;AACnE,sHAAsH;AACtH,IAAIiB,gBAAgB;AACpB,IAAIC,mBAAuD;AAC3D,IAAIC,mBAAuD;AAC3D,eAAeC,oBAAuBC,EAAoB;IACxD,IAAIJ,kBAAkB,GAAG;QACvBC,mBAAmBI,QAAQC,MAAM,CAACC,KAAK,CAACC,IAAI,CAACH,QAAQC,MAAM;QAC3DJ,mBAAmBG,QAAQI,MAAM,CAACF,KAAK,CAACC,IAAI,CAACH,QAAQI,MAAM;QAC3DJ,QAAQC,MAAM,CAACC,KAAK,GAAI,IAAM;QAC9BF,QAAQI,MAAM,CAACF,KAAK,GAAI,IAAM;IAChC;IACAP;IACA,IAAI;QACF,OAAO,MAAMI;IACf,SAAU;QACRJ;QACA,IAAIA,kBAAkB,GAAG;YACvB,IAAIC,kBAAkBI,QAAQC,MAAM,CAACC,KAAK,GAAGN;YAC7C,IAAIC,kBAAkBG,QAAQI,MAAM,CAACF,KAAK,GAAGL;YAC7CD,mBAAmB;YACnBC,mBAAmB;QACrB;IACF;AACF;AAEA;yFACyF,GACzF,SAASQ,cAAiBN,EAAW;IACnC,IAAIJ,kBAAkB,GAAG,OAAOI;IAChC,MAAMO,UAAUV;IAChB,MAAMW,UAAUV;IAChB,IAAIS,SAASN,QAAQC,MAAM,CAACC,KAAK,GAAGI;IACpC,IAAIC,SAASP,QAAQI,MAAM,CAACF,KAAK,GAAGK;IACpC,IAAI;QACF,OAAOR;IACT,SAAU;QACRC,QAAQC,MAAM,CAACC,KAAK,GAAI,IAAM;QAC9BF,QAAQI,MAAM,CAACF,KAAK,GAAI,IAAM;IAChC;AACF;AAEA;uIACuI,GACvI,OAAO,eAAeM,aAAaC,OAAe,EAAEC,KAAa,EAAEC,IAAY,EAAE,EAAEC,MAAMnD,cAAc,EAAEoD,UAAU,EAAEC,SAAS,EAAuG,GAAG,CAAC,CAAC;IACxO,0EAA0E;IAC1E,kFAAkF;IAClF,MAAMC,YAAY9D,KAAKmB,IAAI,CAACpB,GAAGgE,MAAM,IAAI;IACzCjE,UAAUgE,WAAW;QAAEE,WAAW;IAAK;IAEvC,MAAMrD,QAAoB,EAAE;IAC5B,MAAMsD,aAAuB,EAAE;IAC/B,MAAMC,YAAY,IAAIC;IAEtB,2EAA2E;IAC3E,uEAAuE;IACvE,MAAMtB,oBAAoB;QACxB,uEAAuE;QACvE,4CAA4C;QAC5C,MAAMuB,SAAS,MAAMnE,aAAa,OAAO,GAAG;YAC1CoE,UAAUnD;YACV4C;QACF;QACA,IAAI;YACF,uEAAuE;YACvE,6DAA6D;YAC7D,MAAMM,OAAOE,aAAa,CAAC;gBAAEC,uBAAuBrE,IAAIsE,IAAI;YAAC;YAE7D,MAAMC,cAAcpE,iBAAiBmD;YACrC,IAAI;gBACF,MAAMkB,MAAM,MAAMtE,aAAaqE,aAAajB;gBAC5C,IAAK,IAAImB,KAAKlB,OAAOkB,MAAMC,KAAKC,GAAG,CAACnB,MAAMgB,IAAII,QAAQ,GAAGH,KAAM;oBAC7D,MAAMI,KAAKC,KAAKC,GAAG;oBACnB,MAAM,EAAEC,GAAG,EAAE3D,KAAK,EAAEC,MAAM,EAAE,GAAG,MAAMlB,iBAAiBoE,KAAKC,IAAIhB;oBAC/D,MAAM,EAAEwB,IAAI,EAAE,GAAG,MAAMf,OAAOgB,SAAS,CAACF,KAAK,CAAC,GAAG;wBAAEvD,QAAQ;oBAAK;oBAChE,IAAIF,YAAoBJ,cAAc8D,MAAM5D,OAAOC;oBAEnD,gEAAgE;oBAChE,kEAAkE;oBAClE,IAAIP,kBAAkBQ,YAAY;wBAChC,MAAM4D,eAAe3E,cAAce;wBACnC,MAAM6D,YAAY7D;wBAClB,2DAA2D;wBAC3D,sFAAsF;wBACtF,MAAM2C,OAAOE,aAAa,CAAC;4BAAEC,uBAAuBrE,IAAIqF,WAAW;wBAAC;wBACpE,MAAMC,SAAS,MAAMpB,OAAOgB,SAAS,CAACF,KAAK,CAAC,GAAG;4BAAEvD,QAAQ;wBAAK;wBAC9D,MAAM8D,cAAcpE,cAAcmE,OAAOL,IAAI,EAAE5D,OAAOC;wBACtD,MAAM4C,OAAOE,aAAa,CAAC;4BAAEC,uBAAuBrE,IAAIsE,IAAI;wBAAC;wBAE7D,IAAIiB,YAAY7E,MAAM,GAAG0E,UAAU1E,MAAM,EAAE;4BACzCa,YAAYgE;4BACZ,MAAMC,YAAYJ,UAAU1E,MAAM;4BAClC,MAAM+E,cAAcF,YAAY7E,MAAM;4BACtCwC,cAAc,IAAMS,sBAAAA,gCAAAA,UAAY,CAAC,CAAC,EAAEc,GAAG,kCAAkC,EAAEU,aAAaO,OAAO,CAAC,GAAG,2CAA2C,EAAEF,UAAU,IAAI,EAAEC,YAAY,MAAM,CAAC;wBACrL;oBACF;oBAEA,MAAME,KAAKb,KAAKC,GAAG,KAAKF;oBACxB,KAAK,MAAMjE,KAAKW,UAAWd,MAAM0B,IAAI,CAAC;wBAAEf,MAAMqD;wBAAI,GAAG7D,CAAC;oBAAC;oBACvD,IAAIW,UAAUb,MAAM,KAAK,GAAGqD,WAAW5B,IAAI,CAACsC;oBAC5CT,UAAU4B,GAAG,CAACnB,IAAIkB;oBAClB,MAAME,YAAYtE,UAAUb,MAAM;oBAClCwC,cAAc,IAAMQ,uBAAAA,iCAAAA,WAAa;4BAAEoC,MAAM;4BAAQ1E,MAAMqD;4BAAIhE,OAAOoF;4BAAWF;wBAAG;gBAClF;YACF,SAAU;gBACR,MAAMpB,YAAYwB,OAAO;YAC3B;QACF,SAAU;YACR,MAAM7B,OAAO8B,SAAS;QACxB;IACF;IAEA,OAAO;QAAEvF;QAAOsD;QAAYC;IAAU;AACxC"}
@@ -1 +1 @@
1
- {"version":3,"sources":["/Users/kevin/Projects/pdf-to-md/src/errors.ts"],"sourcesContent":["// Library code throws; only src/cli.ts prints and exits.\n\n/** The closed set of error codes every `PdfToMdError` carries — switch on\n * `code` instead of parsing `message`. */\nexport type PdfToMdErrorCode =\n /** input missing/unreadable/not a PDF */\n | 'PDF_OPEN'\n /** malformed page selection (validated before any work) */\n | 'PAGE_RANGE'\n /** rendering an opened, in-range page to PNG failed */\n | 'PAGE_RENDER'\n /** the OCR engine crashed twice on one page (message names it) */\n | 'OCR_PAGE_FAILED'\n /** malformed word-dump input (--words-json / extractOcr words) */\n | 'WORDS_INPUT'\n /** malformed Analysis handed to toMarkdown/toText (e.g. an out-of-range\n * wordIndexes entry) — validated fail-fast, whole object checked before any output. */\n | 'ANALYSIS_INPUT';\n\n/** The error every entry point throws — never a bare `Error`. */\nexport class PdfToMdError extends Error {\n code: PdfToMdErrorCode;\n\n constructor(code: PdfToMdErrorCode, message: string) {\n super(message);\n this.name = 'PdfToMdError';\n this.code = code;\n }\n}\n"],"names":["PdfToMdError","Error","code","message","name"],"mappings":"AAAA,yDAAyD;AAEzD;wCACwC,GAgBxC,+DAA+D,GAC/D,OAAO,MAAMA,qBAAqBC;IAGhC,YAAYC,IAAsB,EAAEC,OAAe,CAAE;QACnD,KAAK,CAACA;QACN,IAAI,CAACC,IAAI,GAAG;QACZ,IAAI,CAACF,IAAI,GAAGA;IACd;AACF"}
1
+ {"version":3,"sources":["/Users/kevin/Dev/OpenSource/formats/pdf-to-md/src/errors.ts"],"sourcesContent":["// Library code throws; only src/cli.ts prints and exits.\n\n/** The closed set of error codes every `PdfToMdError` carries — switch on\n * `code` instead of parsing `message`. */\nexport type PdfToMdErrorCode =\n /** input missing/unreadable/not a PDF */\n | 'PDF_OPEN'\n /** malformed page selection (validated before any work) */\n | 'PAGE_RANGE'\n /** rendering an opened, in-range page to PNG failed */\n | 'PAGE_RENDER'\n /** the OCR engine crashed twice on one page (message names it) */\n | 'OCR_PAGE_FAILED'\n /** malformed word-dump input (--words-json / extractOcr words) */\n | 'WORDS_INPUT'\n /** malformed Analysis handed to toMarkdown/toText (e.g. an out-of-range\n * wordIndexes entry) — validated fail-fast, whole object checked before any output. */\n | 'ANALYSIS_INPUT';\n\n/** The error every entry point throws — never a bare `Error`. */\nexport class PdfToMdError extends Error {\n code: PdfToMdErrorCode;\n\n constructor(code: PdfToMdErrorCode, message: string) {\n super(message);\n this.name = 'PdfToMdError';\n this.code = code;\n }\n}\n"],"names":["PdfToMdError","Error","code","message","name"],"mappings":"AAAA,yDAAyD;AAEzD;wCACwC,GAgBxC,+DAA+D,GAC/D,OAAO,MAAMA,qBAAqBC;IAGhC,YAAYC,IAAsB,EAAEC,OAAe,CAAE;QACnD,KAAK,CAACA;QACN,IAAI,CAACC,IAAI,GAAG;QACZ,IAAI,CAACF,IAAI,GAAGA;IACd;AACF"}
@@ -1 +1 @@
1
- {"version":3,"sources":["/Users/kevin/Projects/pdf-to-md/src/extract.ts"],"sourcesContent":["// The library layer for PDF -> markdown conversion: extractText/extractOcr/\n// pdfToMarkdown each take plain args, return plain data, and throw PdfToMdError — no process.exit, no console.*, no file writing.\n//\n// Warnings are DATA: every notice is returned in `warnings` verbatim and also\n// delivered incrementally via `onWarning` — the exact strings are load-bearing (the test suite asserts them).\nimport { buildOcrAnalysisPages } from './analyze.ts';\nimport { collect } from './collect.ts';\nimport { dictAvailable, dictUnavailableWarning } from './dict.ts';\nimport type { OcrProgressEvent } from './engines/tesseract.ts';\nimport type { PageWord, Tuning } from './geometry.ts';\nimport { toLines } from './lines.ts';\nimport { buildMarkdown } from './markdown.ts';\nimport { renderPagesToMarkdown } from './render-analysis.ts';\nimport { AUTO_OCR_CHARS_PER_PAGE, buildOcrReport, buildReport } from './report.ts';\nimport type { PageLines, Report } from './types.ts';\nimport { validatePages } from './validate.ts';\n\nexport type { OcrProgressEvent } from './engines/tesseract.ts';\n\n// A plain-JS caller (or `.ts` with `as any`) gets no compile-time check on\n// `pages`, so it's validated once at runtime — see src/validate.ts's `validatePages`.\n\n/** One recognized word with its page — the OCR replay input contract (same\n * shape as `geometry.ts`'s `PageWord`, re-exported under this module's public name). */\nexport type OcrWordInput = PageWord;\n\n/** A 1-based, inclusive page range — the shape every entry point's `pages`\n * option takes (the CLI's own `--pages N[-M]` parses into this). */\nexport interface PageSelection {\n first: number;\n /** inclusive; defaults to `first` */\n last?: number;\n}\n\n/** extractText/extractOcr/pdfToMarkdown's common return shape. */\nexport interface ExtractResult {\n markdown: string;\n /** per-document and per-page QA metrics — see `Report` (`src/types.ts`) */\n report: Report;\n /** every notice also delivered incrementally via `onWarning`, collected\n * here in full (hybrid pages, auto-OCR-fallback, empty OCR pages, sparse-retry) */\n warnings: string[];\n}\n\ninterface CommonOptions {\n /** convert only this page range (1-based, inclusive); defaults to the\n * whole document */\n pages?: PageSelection;\n /** delivered as each warning happens (also returned, in full, as\n * `warnings` on the resolved result) */\n onWarning?: (message: string) => void;\n}\n\n/** Options for {@link extractText}. */\nexport interface ExtractTextOptions extends CommonOptions {\n /** emit a `### pN` marker before each page's content (output formatting\n * only — `report.pageStats` is always populated regardless) */\n pageMarkers?: boolean;\n}\n\n/** Options for {@link extractOcr}. */\nexport interface ExtractOcrOptions extends CommonOptions {\n /** override any subset of the OCR-path geometry defaults — see\n * DEFAULT_TUNING (`src/geometry.ts`) for each knob's measured default */\n tuning?: Partial<Tuning>;\n /** page-render resolution for engine recognition (default 288, measured;\n * >~400 measured WORSE on ordinary pages — not a \"bigger is safer\" knob) */\n dpi?: number;\n /** marks report.ocrFallback — set by pdfToMarkdown's auto-fallback */\n fallback?: boolean;\n /** one event per page as recognition finishes (timing/count only — engine\n * chatter and the sparse-retry notice arrive via `onWarning` instead) */\n onProgress?: (e: OcrProgressEvent) => void;\n /** every recognized word entering the geometry, delivered once after\n * recognition — the severity-1 audit instrument (--debug-words) */\n onWords?: (words: OcrWordInput[]) => void;\n}\n\n/** Options for {@link pdfToMarkdown}. */\nexport interface PdfToMarkdownOptions extends ExtractTextOptions, Omit<ExtractOcrOptions, 'fallback'> {\n /** force the OCR path even when a text layer exists */\n ocr?: boolean;\n /** never fall back to OCR, even for image-only pages */\n noOcr?: boolean;\n}\n\n/**\n * Text-layer extraction. Returns markdown + QA report + warnings (hybrid\n * pages carrying large images whose baked-in text the text layer can never see). Makes no OCR decision — see pdfToMarkdown for the fallback.\n *\n * @param pdfPath path to the PDF file\n * @param opts see {@link ExtractTextOptions}\n */\n// One collector per entry point: collects into `warnings`, streams via\n// onWarning, and emits the dictionary-unavailable notice exactly once up front.\nfunction warnCollector(onWarning?: (message: string) => void): { warnings: string[]; warn: (m: string) => void } {\n const warnings: string[] = [];\n const warn = (m: string) => {\n warnings.push(m);\n onWarning?.(m);\n };\n if (!dictAvailable()) warn(dictUnavailableWarning());\n return { warnings, warn };\n}\n\nexport async function extractText(pdfPath: string, opts: ExtractTextOptions = {}): Promise<ExtractResult> {\n validatePages(opts.pages);\n const { warnings, warn } = warnCollector(opts.onWarning);\n\n const range = opts.pages ? { first: opts.pages.first, last: opts.pages.last ?? opts.pages.first } : {};\n const { numPages, pages, heightChars } = await collect(pdfPath, range);\n const pageLines: PageLines[] = pages.map((p) => ({\n page: p.page,\n lines: toLines(p),\n }));\n const pageMarkers = opts.pageMarkers ?? false;\n const { md, stats, bodyH, headingSizes } = buildMarkdown(pageLines, heightChars, { pageMarkers });\n // pageStats needs \"### pN\" markers to segment content; when the caller\n // didn't request them, build a throwaway marked variant (same heightChars) purely to measure stats from, so pageStats is always populated.\n const statsMd = pageMarkers ? md : buildMarkdown(pageLines, heightChars, { pageMarkers: true }).md;\n\n // Advisory, not prescriptive: `hasLargeImage` fires on any large image,\n // even one with no baked-in text, where `--ocr` recovers nothing and can\n // actively degrade a correct result (measured: \"I\" -> \"|\", stray em-dashes) — so this tells the caller to COMPARE the OCR rerun, never blindly replace.\n const hybridPages = pages.filter((p) => p.largeImage).map((p) => p.page);\n for (const n of hybridPages) {\n warn(`p${n}: text layer used but page carries large image(s); if it has baked-in text this output won't have it — compare (don't blindly replace) against --pages ${n} --ocr, which can itself introduce errors`);\n }\n\n const report = buildReport({\n numPages,\n md,\n statsMd,\n stats,\n bodyH,\n headingSizes,\n hybridPages,\n });\n return { markdown: md, report, warnings };\n}\n\n/**\n * OCR extraction. Input is either a PDF (pages render at `dpi` and the\n * engine recognizes them, imported lazily) or an already-recognized word dump (the offline replay input — see `--words-json`).\n *\n * @param input `{ pdfPath }` to recognize, or `{ words }` to replay a dump\n * @param opts see {@link ExtractOcrOptions}\n */\nexport async function extractOcr(input: { pdfPath: string } | { words: OcrWordInput[] }, opts: ExtractOcrOptions = {}): Promise<ExtractResult> {\n validatePages(opts.pages);\n const { warnings, warn } = warnCollector(opts.onWarning);\n\n // Shared with analyze(pdf, {path:'ocr'}) — see that module's header for\n // why this makes the two byte-identical by construction, not by hand.\n const { pages, words } = await buildOcrAnalysisPages(input, opts, warn);\n const md = renderPagesToMarkdown(pages, { forcePageMarkers: true, pageMarkers: true });\n\n const report = buildOcrReport({\n md,\n fallback: opts.fallback,\n // Confidence is on `words` — see report.ts's low-confidence detector for\n // how it handles a confidence-less replay honestly.\n words,\n });\n return { markdown: md, report, warnings };\n}\n\n/** pdfToMarkdown's auto-OCR fallback predicate, shared with the CLI's\n * --format=raw|txt path so the two call sites can never drift apart. Module-internal: not re-exported by `src/index.ts`. */\nexport function needsOcrFallback(charsPerPage: number): boolean {\n return charsPerPage < AUTO_OCR_CHARS_PER_PAGE;\n}\n\n/** The exact stderr/warnings notice printed when needsOcrFallback fires —\n * one copy of the string for both call sites (asserted verbatim by the suite). */\nexport function ocrFallbackNotice(charsPerPage: number): string {\n return `text layer yields ~${charsPerPage} chars/page (< ${AUTO_OCR_CHARS_PER_PAGE}) — pages are images of text; falling back to OCR (suppress with --no-ocr)`;\n}\n\n/**\n * The headline API: extract a PDF to markdown, text layer first, falling\n * back to OCR when the text layer is empty enough the pages are certainly images of text (< `AUTO_OCR_CHARS_PER_PAGE` chars/page) — the same decision the CLI makes.\n *\n * @param pdfPath path to the PDF file\n * @param opts see {@link PdfToMarkdownOptions}\n */\nexport async function pdfToMarkdown(pdfPath: string, opts: PdfToMarkdownOptions = {}): Promise<ExtractResult> {\n if (opts.ocr) return extractOcr({ pdfPath }, opts);\n\n // Warning DELIVERY is deferred on the text pass: if the fallback fires,\n // those hybrid warnings describe output being discarded, so only a kept text result replays them.\n const text = await extractText(pdfPath, { ...opts, onWarning: undefined });\n if (needsOcrFallback(text.report.charsPerPage) && !opts.noOcr) {\n const notice = ocrFallbackNotice(text.report.charsPerPage);\n opts.onWarning?.(notice);\n const ocr = await extractOcr({ pdfPath }, { ...opts, fallback: true });\n ocr.warnings.unshift(notice);\n return ocr;\n }\n for (const w of text.warnings) opts.onWarning?.(w);\n return text;\n}\n"],"names":["buildOcrAnalysisPages","collect","dictAvailable","dictUnavailableWarning","toLines","buildMarkdown","renderPagesToMarkdown","AUTO_OCR_CHARS_PER_PAGE","buildOcrReport","buildReport","validatePages","warnCollector","onWarning","warnings","warn","m","push","extractText","pdfPath","opts","pages","range","first","last","numPages","heightChars","pageLines","map","p","page","lines","pageMarkers","md","stats","bodyH","headingSizes","statsMd","hybridPages","filter","largeImage","n","report","markdown","extractOcr","input","words","forcePageMarkers","fallback","needsOcrFallback","charsPerPage","ocrFallbackNotice","pdfToMarkdown","ocr","text","undefined","noOcr","notice","unshift","w"],"mappings":"AAAA,4EAA4E;AAC5E,kIAAkI;AAClI,EAAE;AACF,8EAA8E;AAC9E,8GAA8G;AAC9G,SAASA,qBAAqB,QAAQ,eAAe;AACrD,SAASC,OAAO,QAAQ,eAAe;AACvC,SAASC,aAAa,EAAEC,sBAAsB,QAAQ,YAAY;AAGlE,SAASC,OAAO,QAAQ,aAAa;AACrC,SAASC,aAAa,QAAQ,gBAAgB;AAC9C,SAASC,qBAAqB,QAAQ,uBAAuB;AAC7D,SAASC,uBAAuB,EAAEC,cAAc,EAAEC,WAAW,QAAQ,cAAc;AAEnF,SAASC,aAAa,QAAQ,gBAAgB;AAuE9C;;;;;;CAMC,GACD,uEAAuE;AACvE,gFAAgF;AAChF,SAASC,cAAcC,SAAqC;IAC1D,MAAMC,WAAqB,EAAE;IAC7B,MAAMC,OAAO,CAACC;QACZF,SAASG,IAAI,CAACD;QACdH,sBAAAA,gCAAAA,UAAYG;IACd;IACA,IAAI,CAACb,iBAAiBY,KAAKX;IAC3B,OAAO;QAAEU;QAAUC;IAAK;AAC1B;AAEA,OAAO,eAAeG,YAAYC,OAAe,EAAEC,OAA2B,CAAC,CAAC;QAIlBA,kBAMxCA;IATpBT,cAAcS,KAAKC,KAAK;IACxB,MAAM,EAAEP,QAAQ,EAAEC,IAAI,EAAE,GAAGH,cAAcQ,KAAKP,SAAS;IAEvD,MAAMS,QAAQF,KAAKC,KAAK,GAAG;QAAEE,OAAOH,KAAKC,KAAK,CAACE,KAAK;QAAEC,IAAI,GAAEJ,mBAAAA,KAAKC,KAAK,CAACG,IAAI,cAAfJ,8BAAAA,mBAAmBA,KAAKC,KAAK,CAACE,KAAK;IAAC,IAAI,CAAC;IACrG,MAAM,EAAEE,QAAQ,EAAEJ,KAAK,EAAEK,WAAW,EAAE,GAAG,MAAMxB,QAAQiB,SAASG;IAChE,MAAMK,YAAyBN,MAAMO,GAAG,CAAC,CAACC,IAAO,CAAA;YAC/CC,MAAMD,EAAEC,IAAI;YACZC,OAAO1B,QAAQwB;QACjB,CAAA;IACA,MAAMG,eAAcZ,oBAAAA,KAAKY,WAAW,cAAhBZ,+BAAAA,oBAAoB;IACxC,MAAM,EAAEa,EAAE,EAAEC,KAAK,EAAEC,KAAK,EAAEC,YAAY,EAAE,GAAG9B,cAAcqB,WAAWD,aAAa;QAAEM;IAAY;IAC/F,uEAAuE;IACvE,2IAA2I;IAC3I,MAAMK,UAAUL,cAAcC,KAAK3B,cAAcqB,WAAWD,aAAa;QAAEM,aAAa;IAAK,GAAGC,EAAE;IAElG,wEAAwE;IACxE,yEAAyE;IACzE,wJAAwJ;IACxJ,MAAMK,cAAcjB,MAAMkB,MAAM,CAAC,CAACV,IAAMA,EAAEW,UAAU,EAAEZ,GAAG,CAAC,CAACC,IAAMA,EAAEC,IAAI;IACvE,KAAK,MAAMW,KAAKH,YAAa;QAC3BvB,KAAK,CAAC,CAAC,EAAE0B,EAAE,uJAAuJ,EAAEA,EAAE,yCAAyC,CAAC;IAClN;IAEA,MAAMC,SAAShC,YAAY;QACzBe;QACAQ;QACAI;QACAH;QACAC;QACAC;QACAE;IACF;IACA,OAAO;QAAEK,UAAUV;QAAIS;QAAQ5B;IAAS;AAC1C;AAEA;;;;;;CAMC,GACD,OAAO,eAAe8B,WAAWC,KAAsD,EAAEzB,OAA0B,CAAC,CAAC;IACnHT,cAAcS,KAAKC,KAAK;IACxB,MAAM,EAAEP,QAAQ,EAAEC,IAAI,EAAE,GAAGH,cAAcQ,KAAKP,SAAS;IAEvD,wEAAwE;IACxE,sEAAsE;IACtE,MAAM,EAAEQ,KAAK,EAAEyB,KAAK,EAAE,GAAG,MAAM7C,sBAAsB4C,OAAOzB,MAAML;IAClE,MAAMkB,KAAK1B,sBAAsBc,OAAO;QAAE0B,kBAAkB;QAAMf,aAAa;IAAK;IAEpF,MAAMU,SAASjC,eAAe;QAC5BwB;QACAe,UAAU5B,KAAK4B,QAAQ;QACvB,yEAAyE;QACzE,oDAAoD;QACpDF;IACF;IACA,OAAO;QAAEH,UAAUV;QAAIS;QAAQ5B;IAAS;AAC1C;AAEA;0HAC0H,GAC1H,OAAO,SAASmC,iBAAiBC,YAAoB;IACnD,OAAOA,eAAe1C;AACxB;AAEA;gFACgF,GAChF,OAAO,SAAS2C,kBAAkBD,YAAoB;IACpD,OAAO,CAAC,mBAAmB,EAAEA,aAAa,eAAe,EAAE1C,wBAAwB,0EAA0E,CAAC;AAChK;AAEA;;;;;;CAMC,GACD,OAAO,eAAe4C,cAAcjC,OAAe,EAAEC,OAA6B,CAAC,CAAC;QAanDA;IAZ/B,IAAIA,KAAKiC,GAAG,EAAE,OAAOT,WAAW;QAAEzB;IAAQ,GAAGC;IAE7C,wEAAwE;IACxE,kGAAkG;IAClG,MAAMkC,OAAO,MAAMpC,YAAYC,SAAS;QAAE,GAAGC,IAAI;QAAEP,WAAW0C;IAAU;IACxE,IAAIN,iBAAiBK,KAAKZ,MAAM,CAACQ,YAAY,KAAK,CAAC9B,KAAKoC,KAAK,EAAE;YAE7DpC;QADA,MAAMqC,SAASN,kBAAkBG,KAAKZ,MAAM,CAACQ,YAAY;SACzD9B,mBAAAA,KAAKP,SAAS,cAAdO,uCAAAA,sBAAAA,MAAiBqC;QACjB,MAAMJ,MAAM,MAAMT,WAAW;YAAEzB;QAAQ,GAAG;YAAE,GAAGC,IAAI;YAAE4B,UAAU;QAAK;QACpEK,IAAIvC,QAAQ,CAAC4C,OAAO,CAACD;QACrB,OAAOJ;IACT;IACA,KAAK,MAAMM,KAAKL,KAAKxC,QAAQ,EAAEM,kBAAAA,KAAKP,SAAS,cAAdO,sCAAAA,qBAAAA,MAAiBuC;IAChD,OAAOL;AACT"}
1
+ {"version":3,"sources":["/Users/kevin/Dev/OpenSource/formats/pdf-to-md/src/extract.ts"],"sourcesContent":["// The library layer for PDF -> markdown conversion: extractText/extractOcr/\n// pdfToMarkdown each take plain args, return plain data, and throw PdfToMdError — no process.exit, no console.*, no file writing.\n//\n// Warnings are DATA: every notice is returned in `warnings` verbatim and also\n// delivered incrementally via `onWarning` — the exact strings are load-bearing (the test suite asserts them).\nimport { buildOcrAnalysisPages } from './analyze.ts';\nimport { collect } from './collect.ts';\nimport { dictAvailable, dictUnavailableWarning } from './dict.ts';\nimport type { OcrProgressEvent } from './engines/tesseract.ts';\nimport type { PageWord, Tuning } from './geometry.ts';\nimport { toLines } from './lines.ts';\nimport { buildMarkdown } from './markdown.ts';\nimport { renderPagesToMarkdown } from './render-analysis.ts';\nimport { AUTO_OCR_CHARS_PER_PAGE, buildOcrReport, buildReport } from './report.ts';\nimport type { PageLines, Report } from './types.ts';\nimport { validatePages } from './validate.ts';\n\nexport type { OcrProgressEvent } from './engines/tesseract.ts';\n\n// A plain-JS caller (or `.ts` with `as any`) gets no compile-time check on\n// `pages`, so it's validated once at runtime — see src/validate.ts's `validatePages`.\n\n/** One recognized word with its page — the OCR replay input contract (same\n * shape as `geometry.ts`'s `PageWord`, re-exported under this module's public name). */\nexport type OcrWordInput = PageWord;\n\n/** A 1-based, inclusive page range — the shape every entry point's `pages`\n * option takes (the CLI's own `--pages N[-M]` parses into this). */\nexport interface PageSelection {\n first: number;\n /** inclusive; defaults to `first` */\n last?: number;\n}\n\n/** extractText/extractOcr/pdfToMarkdown's common return shape. */\nexport interface ExtractResult {\n markdown: string;\n /** per-document and per-page QA metrics — see `Report` (`src/types.ts`) */\n report: Report;\n /** every notice also delivered incrementally via `onWarning`, collected\n * here in full (hybrid pages, auto-OCR-fallback, empty OCR pages, sparse-retry) */\n warnings: string[];\n}\n\ninterface CommonOptions {\n /** convert only this page range (1-based, inclusive); defaults to the\n * whole document */\n pages?: PageSelection;\n /** delivered as each warning happens (also returned, in full, as\n * `warnings` on the resolved result) */\n onWarning?: (message: string) => void;\n}\n\n/** Options for {@link extractText}. */\nexport interface ExtractTextOptions extends CommonOptions {\n /** emit a `### pN` marker before each page's content (output formatting\n * only — `report.pageStats` is always populated regardless) */\n pageMarkers?: boolean;\n}\n\n/** Options for {@link extractOcr}. */\nexport interface ExtractOcrOptions extends CommonOptions {\n /** override any subset of the OCR-path geometry defaults — see\n * DEFAULT_TUNING (`src/geometry.ts`) for each knob's measured default */\n tuning?: Partial<Tuning>;\n /** page-render resolution for engine recognition (default 288, measured;\n * >~400 measured WORSE on ordinary pages — not a \"bigger is safer\" knob) */\n dpi?: number;\n /** marks report.ocrFallback — set by pdfToMarkdown's auto-fallback */\n fallback?: boolean;\n /** one event per page as recognition finishes (timing/count only — engine\n * chatter and the sparse-retry notice arrive via `onWarning` instead) */\n onProgress?: (e: OcrProgressEvent) => void;\n /** every recognized word entering the geometry, delivered once after\n * recognition — the severity-1 audit instrument (--debug-words) */\n onWords?: (words: OcrWordInput[]) => void;\n}\n\n/** Options for {@link pdfToMarkdown}. */\nexport interface PdfToMarkdownOptions extends ExtractTextOptions, Omit<ExtractOcrOptions, 'fallback'> {\n /** force the OCR path even when a text layer exists */\n ocr?: boolean;\n /** never fall back to OCR, even for image-only pages */\n noOcr?: boolean;\n}\n\n/**\n * Text-layer extraction. Returns markdown + QA report + warnings (hybrid\n * pages carrying large images whose baked-in text the text layer can never see). Makes no OCR decision — see pdfToMarkdown for the fallback.\n *\n * @param pdfPath path to the PDF file\n * @param opts see {@link ExtractTextOptions}\n */\n// One collector per entry point: collects into `warnings`, streams via\n// onWarning, and emits the dictionary-unavailable notice exactly once up front.\nfunction warnCollector(onWarning?: (message: string) => void): { warnings: string[]; warn: (m: string) => void } {\n const warnings: string[] = [];\n const warn = (m: string) => {\n warnings.push(m);\n onWarning?.(m);\n };\n if (!dictAvailable()) warn(dictUnavailableWarning());\n return { warnings, warn };\n}\n\nexport async function extractText(pdfPath: string, opts: ExtractTextOptions = {}): Promise<ExtractResult> {\n validatePages(opts.pages);\n const { warnings, warn } = warnCollector(opts.onWarning);\n\n const range = opts.pages ? { first: opts.pages.first, last: opts.pages.last ?? opts.pages.first } : {};\n const { numPages, pages, heightChars } = await collect(pdfPath, range);\n const pageLines: PageLines[] = pages.map((p) => ({\n page: p.page,\n lines: toLines(p),\n }));\n const pageMarkers = opts.pageMarkers ?? false;\n const { md, stats, bodyH, headingSizes } = buildMarkdown(pageLines, heightChars, { pageMarkers });\n // pageStats needs \"### pN\" markers to segment content; when the caller\n // didn't request them, build a throwaway marked variant (same heightChars) purely to measure stats from, so pageStats is always populated.\n const statsMd = pageMarkers ? md : buildMarkdown(pageLines, heightChars, { pageMarkers: true }).md;\n\n // Advisory, not prescriptive: `hasLargeImage` fires on any large image,\n // even one with no baked-in text, where `--ocr` recovers nothing and can\n // actively degrade a correct result (measured: \"I\" -> \"|\", stray em-dashes) — so this tells the caller to COMPARE the OCR rerun, never blindly replace.\n const hybridPages = pages.filter((p) => p.largeImage).map((p) => p.page);\n for (const n of hybridPages) {\n warn(`p${n}: text layer used but page carries large image(s); if it has baked-in text this output won't have it — compare (don't blindly replace) against --pages ${n} --ocr, which can itself introduce errors`);\n }\n\n const report = buildReport({\n numPages,\n md,\n statsMd,\n stats,\n bodyH,\n headingSizes,\n hybridPages,\n });\n return { markdown: md, report, warnings };\n}\n\n/**\n * OCR extraction. Input is either a PDF (pages render at `dpi` and the\n * engine recognizes them, imported lazily) or an already-recognized word dump (the offline replay input — see `--words-json`).\n *\n * @param input `{ pdfPath }` to recognize, or `{ words }` to replay a dump\n * @param opts see {@link ExtractOcrOptions}\n */\nexport async function extractOcr(input: { pdfPath: string } | { words: OcrWordInput[] }, opts: ExtractOcrOptions = {}): Promise<ExtractResult> {\n validatePages(opts.pages);\n const { warnings, warn } = warnCollector(opts.onWarning);\n\n // Shared with analyze(pdf, {path:'ocr'}) — see that module's header for\n // why this makes the two byte-identical by construction, not by hand.\n const { pages, words } = await buildOcrAnalysisPages(input, opts, warn);\n const md = renderPagesToMarkdown(pages, { forcePageMarkers: true, pageMarkers: true });\n\n const report = buildOcrReport({\n md,\n fallback: opts.fallback,\n // Confidence is on `words` — see report.ts's low-confidence detector for\n // how it handles a confidence-less replay honestly.\n words,\n });\n return { markdown: md, report, warnings };\n}\n\n/** pdfToMarkdown's auto-OCR fallback predicate, shared with the CLI's\n * --format=raw|txt path so the two call sites can never drift apart. Module-internal: not re-exported by `src/index.ts`. */\nexport function needsOcrFallback(charsPerPage: number): boolean {\n return charsPerPage < AUTO_OCR_CHARS_PER_PAGE;\n}\n\n/** The exact stderr/warnings notice printed when needsOcrFallback fires —\n * one copy of the string for both call sites (asserted verbatim by the suite). */\nexport function ocrFallbackNotice(charsPerPage: number): string {\n return `text layer yields ~${charsPerPage} chars/page (< ${AUTO_OCR_CHARS_PER_PAGE}) — pages are images of text; falling back to OCR (suppress with --no-ocr)`;\n}\n\n/**\n * The headline API: extract a PDF to markdown, text layer first, falling\n * back to OCR when the text layer is empty enough the pages are certainly images of text (< `AUTO_OCR_CHARS_PER_PAGE` chars/page) — the same decision the CLI makes.\n *\n * @param pdfPath path to the PDF file\n * @param opts see {@link PdfToMarkdownOptions}\n */\nexport async function pdfToMarkdown(pdfPath: string, opts: PdfToMarkdownOptions = {}): Promise<ExtractResult> {\n if (opts.ocr) return extractOcr({ pdfPath }, opts);\n\n // Warning DELIVERY is deferred on the text pass: if the fallback fires,\n // those hybrid warnings describe output being discarded, so only a kept text result replays them.\n const text = await extractText(pdfPath, { ...opts, onWarning: undefined });\n if (needsOcrFallback(text.report.charsPerPage) && !opts.noOcr) {\n const notice = ocrFallbackNotice(text.report.charsPerPage);\n opts.onWarning?.(notice);\n const ocr = await extractOcr({ pdfPath }, { ...opts, fallback: true });\n ocr.warnings.unshift(notice);\n return ocr;\n }\n for (const w of text.warnings) opts.onWarning?.(w);\n return text;\n}\n"],"names":["buildOcrAnalysisPages","collect","dictAvailable","dictUnavailableWarning","toLines","buildMarkdown","renderPagesToMarkdown","AUTO_OCR_CHARS_PER_PAGE","buildOcrReport","buildReport","validatePages","warnCollector","onWarning","warnings","warn","m","push","extractText","pdfPath","opts","pages","range","first","last","numPages","heightChars","pageLines","map","p","page","lines","pageMarkers","md","stats","bodyH","headingSizes","statsMd","hybridPages","filter","largeImage","n","report","markdown","extractOcr","input","words","forcePageMarkers","fallback","needsOcrFallback","charsPerPage","ocrFallbackNotice","pdfToMarkdown","ocr","text","undefined","noOcr","notice","unshift","w"],"mappings":"AAAA,4EAA4E;AAC5E,kIAAkI;AAClI,EAAE;AACF,8EAA8E;AAC9E,8GAA8G;AAC9G,SAASA,qBAAqB,QAAQ,eAAe;AACrD,SAASC,OAAO,QAAQ,eAAe;AACvC,SAASC,aAAa,EAAEC,sBAAsB,QAAQ,YAAY;AAGlE,SAASC,OAAO,QAAQ,aAAa;AACrC,SAASC,aAAa,QAAQ,gBAAgB;AAC9C,SAASC,qBAAqB,QAAQ,uBAAuB;AAC7D,SAASC,uBAAuB,EAAEC,cAAc,EAAEC,WAAW,QAAQ,cAAc;AAEnF,SAASC,aAAa,QAAQ,gBAAgB;AAuE9C;;;;;;CAMC,GACD,uEAAuE;AACvE,gFAAgF;AAChF,SAASC,cAAcC,SAAqC;IAC1D,MAAMC,WAAqB,EAAE;IAC7B,MAAMC,OAAO,CAACC;QACZF,SAASG,IAAI,CAACD;QACdH,sBAAAA,gCAAAA,UAAYG;IACd;IACA,IAAI,CAACb,iBAAiBY,KAAKX;IAC3B,OAAO;QAAEU;QAAUC;IAAK;AAC1B;AAEA,OAAO,eAAeG,YAAYC,OAAe,EAAEC,OAA2B,CAAC,CAAC;QAIlBA,kBAMxCA;IATpBT,cAAcS,KAAKC,KAAK;IACxB,MAAM,EAAEP,QAAQ,EAAEC,IAAI,EAAE,GAAGH,cAAcQ,KAAKP,SAAS;IAEvD,MAAMS,QAAQF,KAAKC,KAAK,GAAG;QAAEE,OAAOH,KAAKC,KAAK,CAACE,KAAK;QAAEC,IAAI,GAAEJ,mBAAAA,KAAKC,KAAK,CAACG,IAAI,cAAfJ,8BAAAA,mBAAmBA,KAAKC,KAAK,CAACE,KAAK;IAAC,IAAI,CAAC;IACrG,MAAM,EAAEE,QAAQ,EAAEJ,KAAK,EAAEK,WAAW,EAAE,GAAG,MAAMxB,QAAQiB,SAASG;IAChE,MAAMK,YAAyBN,MAAMO,GAAG,CAAC,CAACC,IAAO,CAAA;YAC/CC,MAAMD,EAAEC,IAAI;YACZC,OAAO1B,QAAQwB;QACjB,CAAA;IACA,MAAMG,eAAcZ,oBAAAA,KAAKY,WAAW,cAAhBZ,+BAAAA,oBAAoB;IACxC,MAAM,EAAEa,EAAE,EAAEC,KAAK,EAAEC,KAAK,EAAEC,YAAY,EAAE,GAAG9B,cAAcqB,WAAWD,aAAa;QAAEM;IAAY;IAC/F,uEAAuE;IACvE,2IAA2I;IAC3I,MAAMK,UAAUL,cAAcC,KAAK3B,cAAcqB,WAAWD,aAAa;QAAEM,aAAa;IAAK,GAAGC,EAAE;IAElG,wEAAwE;IACxE,yEAAyE;IACzE,wJAAwJ;IACxJ,MAAMK,cAAcjB,MAAMkB,MAAM,CAAC,CAACV,IAAMA,EAAEW,UAAU,EAAEZ,GAAG,CAAC,CAACC,IAAMA,EAAEC,IAAI;IACvE,KAAK,MAAMW,KAAKH,YAAa;QAC3BvB,KAAK,CAAC,CAAC,EAAE0B,EAAE,uJAAuJ,EAAEA,EAAE,yCAAyC,CAAC;IAClN;IAEA,MAAMC,SAAShC,YAAY;QACzBe;QACAQ;QACAI;QACAH;QACAC;QACAC;QACAE;IACF;IACA,OAAO;QAAEK,UAAUV;QAAIS;QAAQ5B;IAAS;AAC1C;AAEA;;;;;;CAMC,GACD,OAAO,eAAe8B,WAAWC,KAAsD,EAAEzB,OAA0B,CAAC,CAAC;IACnHT,cAAcS,KAAKC,KAAK;IACxB,MAAM,EAAEP,QAAQ,EAAEC,IAAI,EAAE,GAAGH,cAAcQ,KAAKP,SAAS;IAEvD,wEAAwE;IACxE,sEAAsE;IACtE,MAAM,EAAEQ,KAAK,EAAEyB,KAAK,EAAE,GAAG,MAAM7C,sBAAsB4C,OAAOzB,MAAML;IAClE,MAAMkB,KAAK1B,sBAAsBc,OAAO;QAAE0B,kBAAkB;QAAMf,aAAa;IAAK;IAEpF,MAAMU,SAASjC,eAAe;QAC5BwB;QACAe,UAAU5B,KAAK4B,QAAQ;QACvB,yEAAyE;QACzE,oDAAoD;QACpDF;IACF;IACA,OAAO;QAAEH,UAAUV;QAAIS;QAAQ5B;IAAS;AAC1C;AAEA;0HAC0H,GAC1H,OAAO,SAASmC,iBAAiBC,YAAoB;IACnD,OAAOA,eAAe1C;AACxB;AAEA;gFACgF,GAChF,OAAO,SAAS2C,kBAAkBD,YAAoB;IACpD,OAAO,CAAC,mBAAmB,EAAEA,aAAa,eAAe,EAAE1C,wBAAwB,0EAA0E,CAAC;AAChK;AAEA;;;;;;CAMC,GACD,OAAO,eAAe4C,cAAcjC,OAAe,EAAEC,OAA6B,CAAC,CAAC;QAanDA;IAZ/B,IAAIA,KAAKiC,GAAG,EAAE,OAAOT,WAAW;QAAEzB;IAAQ,GAAGC;IAE7C,wEAAwE;IACxE,kGAAkG;IAClG,MAAMkC,OAAO,MAAMpC,YAAYC,SAAS;QAAE,GAAGC,IAAI;QAAEP,WAAW0C;IAAU;IACxE,IAAIN,iBAAiBK,KAAKZ,MAAM,CAACQ,YAAY,KAAK,CAAC9B,KAAKoC,KAAK,EAAE;YAE7DpC;QADA,MAAMqC,SAASN,kBAAkBG,KAAKZ,MAAM,CAACQ,YAAY;SACzD9B,mBAAAA,KAAKP,SAAS,cAAdO,uCAAAA,sBAAAA,MAAiBqC;QACjB,MAAMJ,MAAM,MAAMT,WAAW;YAAEzB;QAAQ,GAAG;YAAE,GAAGC,IAAI;YAAE4B,UAAU;QAAK;QACpEK,IAAIvC,QAAQ,CAAC4C,OAAO,CAACD;QACrB,OAAOJ;IACT;IACA,KAAK,MAAMM,KAAKL,KAAKxC,QAAQ,EAAEM,kBAAAA,KAAKP,SAAS,cAAdO,sCAAAA,qBAAAA,MAAiBuC;IAChD,OAAOL;AACT"}
@@ -1 +1 @@
1
- {"version":3,"sources":["/Users/kevin/Projects/pdf-to-md/src/geometry.ts"],"sourcesContent":["// geometry.ts — OCR-path pass 2: words -> lines/columns/floats. Input is\n// flat words (normalized page coords, origin bottom-left, y up), not pre-clustered line observations.\nexport interface Word {\n text: string;\n x: number; // rect.minX\n y: number; // rect.minY\n w: number; // rect.width\n h: number; // rect.height\n // 0-100 (tesseract.js's own `confidence` field, not `conf`). Optional —\n // passthrough only; see report.ts's LOW_CONFIDENCE_THRESHOLD for the consumer.\n confidence?: number;\n}\n\n/** A `Word` with its page number — the shared shape behind both the OCR\n * engine's own output (tesseract.ts) and the replay input contract (extract.ts's OcrWordInput). */\nexport interface PageWord extends Word {\n page: number;\n}\n\nexport interface OcrLine {\n text: string;\n x: number; // rect.minX, used for float-alignment and indentation checks\n width: number; // rect.width, used for the float width check\n midY: number;\n h: number; // char-weighted dominant constituent height\n col: number;\n // Constituent words, left-to-right (same order as `text`'s tokens) — carries\n // word identity through row-building for analyze()'s block->word linkage.\n words: Word[];\n}\n\nexport interface PageOCR {\n lines: OcrLine[];\n // Decorative fragment candidates (src/emit.ts's floats aside), kept as full\n // OcrLine so callers recover both the printed text and constituent words.\n floats: OcrLine[];\n pitch: number;\n}\n\n/** Tuning knobs for the OCR-path geometry pass (CLI: `--float-*`/`--para-gap`/\n * `--heading-scale`/`--col-*`/`--line-*`). See DEFAULT_TUNING for defaults. */\nexport interface Tuning {\n /** fragments with more words than this are body text, not decoration */\n floatMaxWords: number;\n /** fraction of page width above which a fragment is too wide to be\n * decoration */\n floatMaxWidth: number;\n /** how far a fragment's left edge must sit from its column edge to be\n * decoration */\n floatMargin: number;\n /** a vertical gap this many times the line pitch is a paragraph break */\n paraGap: number;\n /** lines taller than this times body height are headings */\n headingScale: number;\n /** boxes wider than this are excluded from gutter detection */\n colMaxWidth: number;\n /** fractional x of the two-column midpoint */\n colSplit: number;\n /** vertical distance within which two boxes are the same line */\n lineYTol: number;\n /** same-line boxes with a larger height ratio don't merge */\n lineHeightRatio: number;\n}\n\n/** Measured defaults; derivation in `test/unit/tuning.test.ts`. lineHeightRatio\n * 2.4 = pair-F1 plateau edge (0.924@1.6 -> 0.995@2.4, 9,471 pairs, 2026-08-16). */\nexport const DEFAULT_TUNING: Tuning = {\n floatMaxWords: 3,\n floatMaxWidth: 0.15,\n floatMargin: 0.008,\n paraGap: 1.6,\n headingScale: 2.0,\n colMaxWidth: 0.55,\n colSplit: 0.5,\n lineYTol: 0.006,\n lineHeightRatio: 2.4,\n};\n\n// Two-column decision (test/unit/column-break.test.ts): gutter ink <=0.017\n// (2-col) vs >=0.047 (1-col) measured over 100 baseline pages, cut 0.03.\nexport const MAX_GUTTER_INK_FRAC = 0.03;\n\n// Near-empty pages (e.g. a 2-word title page) have no gutter ink either, so\n// a word-count floor is needed too; genuine columns measure >=76 words.\nexport const MIN_COL_WORDS = 10;\n\n// Measured: a thin data rail is 11% of body words; the thinnest genuine\n// column is 31% — 0.2 sits between.\nexport const MIN_COL_SHARE = 0.2;\n\ninterface Rect {\n x: number;\n y: number;\n w: number;\n h: number;\n}\ntype Range = readonly [number, number];\n\nconst maxX = (r: Rect): number => r.x + r.w;\nconst maxY = (r: Rect): number => r.y + r.h;\nconst midX = (r: Rect): number => r.x + r.w / 2;\nconst midY = (r: Rect): number => r.y + r.h / 2;\n\nfunction union(a: Rect, b: Rect): Rect {\n const x0 = Math.min(a.x, b.x);\n const y0 = Math.min(a.y, b.y);\n const x1 = Math.max(maxX(a), maxX(b));\n const y1 = Math.max(maxY(a), maxY(b));\n return { x: x0, y: y0, w: x1 - x0, h: y1 - y0 };\n}\n\n// Overlap of a rect's x-span with a column x-range.\nfunction xOverlap(r: Rect, range: Range): number {\n return Math.max(0, Math.min(maxX(r), range[1]) - Math.max(r.x, range[0]));\n}\n\nfunction median(xs: number[]): number | undefined {\n if (!xs.length) return undefined;\n return [...xs].sort((a, b) => a - b)[Math.floor(xs.length / 2)];\n}\n\n// Char-weighted dominant height: most characters wins (ties keep first) —\n// same rule as src/lines.ts, so a small merged fragment can't promote a heading.\nfunction dominantHeight(parts: Word[]): number {\n let best = parts[0];\n for (let i = 1; i < parts.length; i++) {\n if (parts[i].text.length > best.text.length) best = parts[i];\n }\n return best.h;\n}\n\nfunction assignColumn(rect: Rect, colRanges: Range[]): number {\n if (colRanges.length !== 2) return 0;\n const l = xOverlap(rect, colRanges[0]);\n const r = xOverlap(rect, colRanges[1]);\n if (l !== r) return l > r ? 0 : 1;\n const centers = colRanges.map((rg) => (rg[0] + rg[1]) / 2);\n return Math.abs(midX(rect) - centers[0]) <= Math.abs(midX(rect) - centers[1]) ? 0 : 1;\n}\n\n// A column x-range from full-width body rows: short paragraph-final rows and\n// decorative fragments must not stretch it.\nfunction range(rects: Rect[]): Range | undefined {\n if (rects.length < 2) return undefined;\n return [Math.min(...rects.map((r) => r.x)), Math.max(...rects.map(maxX))];\n}\n\ninterface Row {\n rect: Rect;\n words: Word[];\n band: number; // -1 (gutter band / n/a), 0 (left or single column), 1 (right)\n}\n\n// Groups same-visual-line words via lineYTol/lineHeightRatio, requiring band\n// agreement; searches ALL open rows since per-word y drifts ~0.001 across a\n// printed line, enough to interleave two lines' words in y-sorted order.\nfunction buildRows(words: Word[], tuning: Tuning, bandOf: (w: Word) => number): Row[] {\n const rows: Row[] = [];\n // Sort order only affects which compatible row a tie lands in (every row\n // is searched) — top-to-bottom, left-to-right keeps results deterministic.\n const sorted = [...words].sort((a, b) => midY(b) - midY(a) || a.x - b.x);\n for (const w of sorted) {\n const band = bandOf(w);\n const target = rows.find((r) => r.band === band && Math.abs(midY(r.rect) - midY(w)) <= tuning.lineYTol && Math.max(r.rect.h, w.h) < Math.min(r.rect.h, w.h) * tuning.lineHeightRatio);\n if (target) {\n target.words.push(w);\n target.rect = union(target.rect, w);\n } else {\n rows.push({ rect: { x: w.x, y: w.y, w: w.w, h: w.h }, words: [w], band });\n }\n }\n return rows;\n}\n\nfunction makeLine(parts: Word[], rect: Rect, col: number): OcrLine {\n const sorted = [...parts].sort((a, b) => a.x - b.x);\n const text = sorted.map((p) => p.text).join(' ');\n return {\n text,\n x: rect.x,\n width: rect.w,\n midY: midY(rect),\n h: dominantHeight(parts),\n col,\n words: sorted,\n };\n}\n\n// Word-geometry pass (render/recognize happens upstream, in the OCR engine).\n// Body-height words band left/right before row-grouping; heading-height words group across the full width, so a title stays one line.\nexport function buildPageOCR(words: Word[], tuning: Tuning): PageOCR {\n if (!words.length) return { lines: [], floats: [], pitch: 0.012 };\n\n const colBand = 0.02;\n const colLeftCut = tuning.colSplit - colBand;\n const colRightCut = tuning.colSplit + colBand;\n\n const medianH = median(words.map((w) => w.h)) ?? 0.017;\n // Same literal `medianH * 2` used elsewhere for this gate — not\n // `tuning.headingScale`, a separate constant that happens to share the 2.0 default.\n const tallGate = medianH * 2;\n\n // Two-column decision (test/unit/column-break.test.ts): gutter ink <=0.017\n // (2-col) vs >=0.047 (1-col), cut 0.03; floors >=0.2 share/>=10 words per side.\n const bodyWords = words.filter((w) => w.h < tallGate);\n const gutterInk = bodyWords.filter((w) => w.x < tuning.colSplit && maxX(w) > tuning.colSplit).length;\n const leftPop = bodyWords.filter((w) => midX(w) < colLeftCut).length;\n const rightPop = bodyWords.filter((w) => midX(w) > colRightCut).length;\n const twoCol = leftPop >= MIN_COL_WORDS && rightPop >= MIN_COL_WORDS && leftPop >= bodyWords.length * MIN_COL_SHARE && rightPop >= bodyWords.length * MIN_COL_SHARE && gutterInk <= bodyWords.length * MAX_GUTTER_INK_FRAC;\n\n // Band by word CENTER, no dead band: a word box can touch the split exactly\n // (a box can touch the split exactly) and still belongs to its center's column — the twoCol\n // gate above already guarantees near-zero gutter ink.\n const bandOf = (w: Word): number => {\n if (!twoCol) return 0;\n return midX(w) < tuning.colSplit ? 0 : 1;\n };\n const tallRows = buildRows(\n words.filter((w) => w.h >= tallGate),\n tuning,\n () => 0\n );\n const normalRows = buildRows(\n words.filter((w) => w.h < tallGate),\n tuning,\n bandOf\n );\n const rows = [...tallRows, ...normalRows];\n\n // Column ranges from body rows (height <= medianH*1.6): two-col ranges use\n // rows 0.2-0.55 wide; single-col has no upper bound (full-width rows ARE the column).\n let colRanges: Range[];\n if (twoCol) {\n const wideRows = rows.filter((r) => r.rect.w >= 0.2 && r.rect.w <= 0.55 && r.rect.h <= medianH * 1.6);\n const left = range(wideRows.filter((r) => r.band === 0).map((r) => r.rect));\n const right = range(wideRows.filter((r) => r.band === 1).map((r) => r.rect));\n colRanges = [left, right].filter((r): r is Range => r !== undefined);\n } else {\n const wideRows = rows.filter((r) => r.rect.w >= 0.2 && r.rect.h <= medianH * 1.6);\n const single = range(wideRows.map((r) => r.rect));\n colRanges = single !== undefined ? [single] : [];\n }\n\n const lines: OcrLine[] = rows.map((row) => makeLine(row.words, row.rect, assignColumn(row.rect, colRanges)));\n lines.sort((a, b) => a.col - b.col || b.midY - a.midY);\n\n // Float detection: short+narrow+misaligned, rescued into body when it\n // hangs under a body line like a paragraph continuation (same col, gap<=paraGap x pitch, x-overlap) — protects a centered paragraph's short last line.\n const preGaps: number[] = [];\n for (let k = 1; k < lines.length; k++) {\n if (lines[k].col === lines[k - 1].col) {\n const g = lines[k - 1].midY - lines[k].midY;\n if (g > 0.0005) preGaps.push(g);\n }\n }\n const provisionalPitch = median(preGaps) ?? 0.012;\n const colStarts = new Map<number, number>();\n colRanges.forEach((r, i) => {\n colStarts.set(i, r[0]);\n });\n const body: OcrLine[] = [];\n const floatCandidates: OcrLine[] = [];\n // Display-size junk (e.g. \"1/3\" at 68pt, 0.179 wide vs the 0.15 cap)\n // escapes the narrow float test; caught here by junk text + isolation (no adjacent tall line within 2.2x height).\n const tallLines = lines.filter((l) => l.h >= tallGate);\n for (const line of lines) {\n const wordCount = line.text.split(' ').filter(Boolean).length;\n const start = colStarts.get(line.col);\n const short = wordCount <= tuning.floatMaxWords && start !== undefined && Math.abs(line.x - start) > tuning.floatMargin;\n const displayJunkFragment =\n short &&\n line.h >= tallGate &&\n !/[A-Za-z]{3}/.test(line.text) &&\n // Same-column only, matching emitPage's heading-run gather — a heading\n // in the OTHER column can be y-adjacent without composing with this one.\n !tallLines.some((other) => other !== line && other.col === line.col && Math.abs(other.midY - line.midY) < Math.max(other.h, line.h) * 2.2);\n const looksDecorative = displayJunkFragment || (short && line.width < tuning.floatMaxWidth);\n const paragraphAdjacent =\n looksDecorative &&\n body.some((b) => {\n if (b.col !== line.col) return false;\n const gap = b.midY - line.midY;\n if (gap <= 0 || gap > tuning.paraGap * provisionalPitch) return false;\n const overlap = Math.min(b.x + b.width, line.x + line.width) - Math.max(b.x, line.x);\n return overlap >= line.width * 0.5;\n });\n if (looksDecorative && !paragraphAdjacent) {\n floatCandidates.push(line);\n } else {\n body.push(line);\n }\n }\n // One float per candidate row, in line order — no clustering of adjacent\n // candidates: replay against the frozen baseline regressed ~30 pages when tried.\n const floats = floatCandidates;\n\n // Line pitch = median vertical gap between consecutive same-column lines.\n const gaps: number[] = [];\n for (let k = 1; k < body.length; k++) {\n if (body[k].col === body[k - 1].col) {\n const g = body[k - 1].midY - body[k].midY;\n if (g > 0.0005) gaps.push(g);\n }\n }\n return { lines: body, floats, pitch: median(gaps) ?? 0.012 };\n}\n\n// Document-wide height histogram -> heading levels\n\n// Bucket heights to 0.002 of page height so histogram keys are stable.\nexport function bucket(h: number): number {\n return Math.round(h * 500) / 500;\n}\n\n// `pages` are the already-body-only (post-float-filtering) PageOCR results.\nexport function headingLevels(pages: PageOCR[], tuning: Tuning): { bodyH: number; levelOf: Map<number, number> } {\n const chars = new Map<number, number>();\n for (const p of pages) {\n for (const l of p.lines) {\n const b = bucket(l.h);\n chars.set(b, (chars.get(b) ?? 0) + l.text.length);\n }\n }\n // Most-common height wins; ties broken by first-encountered (deterministic;\n // char-count ties across real documents are not observed).\n let bodyH = 0;\n let bestCount = -1;\n for (const [k, v] of chars) {\n if (v > bestCount) {\n bestCount = v;\n bodyH = k;\n }\n }\n\n const levelOf = new Map<number, number>();\n let level = 3; // \"### pN\" page markers occupy level 3\n let rep: number | null = null;\n const sizes = [...chars.keys()].filter((s) => s > bodyH * tuning.headingScale).sort((a, b) => b - a);\n for (const s of sizes) {\n if (rep === null || s < rep * 0.8) {\n level = Math.min(6, level + 1);\n rep = s;\n }\n levelOf.set(s, level);\n }\n return { bodyH, levelOf };\n}\n"],"names":["DEFAULT_TUNING","floatMaxWords","floatMaxWidth","floatMargin","paraGap","headingScale","colMaxWidth","colSplit","lineYTol","lineHeightRatio","MAX_GUTTER_INK_FRAC","MIN_COL_WORDS","MIN_COL_SHARE","maxX","r","x","w","maxY","y","h","midX","midY","union","a","b","x0","Math","min","y0","x1","max","y1","xOverlap","range","median","xs","length","undefined","sort","floor","dominantHeight","parts","best","i","text","assignColumn","rect","colRanges","l","centers","map","rg","abs","rects","buildRows","words","tuning","bandOf","rows","sorted","band","target","find","push","makeLine","col","p","join","width","buildPageOCR","lines","floats","pitch","colBand","colLeftCut","colRightCut","medianH","tallGate","bodyWords","filter","gutterInk","leftPop","rightPop","twoCol","tallRows","normalRows","wideRows","left","right","single","row","preGaps","k","g","provisionalPitch","colStarts","Map","forEach","set","body","floatCandidates","tallLines","line","wordCount","split","Boolean","start","get","short","displayJunkFragment","test","some","other","looksDecorative","paragraphAdjacent","gap","overlap","gaps","bucket","round","headingLevels","pages","chars","bodyH","bestCount","v","levelOf","level","rep","sizes","keys","s"],"mappings":"AAAA,yEAAyE;AACzE,sGAAsG;AA+DtG;iFACiF,GACjF,OAAO,MAAMA,iBAAyB;IACpCC,eAAe;IACfC,eAAe;IACfC,aAAa;IACbC,SAAS;IACTC,cAAc;IACdC,aAAa;IACbC,UAAU;IACVC,UAAU;IACVC,iBAAiB;AACnB,EAAE;AAEF,2EAA2E;AAC3E,yEAAyE;AACzE,OAAO,MAAMC,sBAAsB,KAAK;AAExC,4EAA4E;AAC5E,wEAAwE;AACxE,OAAO,MAAMC,gBAAgB,GAAG;AAEhC,wEAAwE;AACxE,oCAAoC;AACpC,OAAO,MAAMC,gBAAgB,IAAI;AAUjC,MAAMC,OAAO,CAACC,IAAoBA,EAAEC,CAAC,GAAGD,EAAEE,CAAC;AAC3C,MAAMC,OAAO,CAACH,IAAoBA,EAAEI,CAAC,GAAGJ,EAAEK,CAAC;AAC3C,MAAMC,OAAO,CAACN,IAAoBA,EAAEC,CAAC,GAAGD,EAAEE,CAAC,GAAG;AAC9C,MAAMK,OAAO,CAACP,IAAoBA,EAAEI,CAAC,GAAGJ,EAAEK,CAAC,GAAG;AAE9C,SAASG,MAAMC,CAAO,EAAEC,CAAO;IAC7B,MAAMC,KAAKC,KAAKC,GAAG,CAACJ,EAAER,CAAC,EAAES,EAAET,CAAC;IAC5B,MAAMa,KAAKF,KAAKC,GAAG,CAACJ,EAAEL,CAAC,EAAEM,EAAEN,CAAC;IAC5B,MAAMW,KAAKH,KAAKI,GAAG,CAACjB,KAAKU,IAAIV,KAAKW;IAClC,MAAMO,KAAKL,KAAKI,GAAG,CAACb,KAAKM,IAAIN,KAAKO;IAClC,OAAO;QAAET,GAAGU;QAAIP,GAAGU;QAAIZ,GAAGa,KAAKJ;QAAIN,GAAGY,KAAKH;IAAG;AAChD;AAEA,oDAAoD;AACpD,SAASI,SAASlB,CAAO,EAAEmB,KAAY;IACrC,OAAOP,KAAKI,GAAG,CAAC,GAAGJ,KAAKC,GAAG,CAACd,KAAKC,IAAImB,KAAK,CAAC,EAAE,IAAIP,KAAKI,GAAG,CAAChB,EAAEC,CAAC,EAAEkB,KAAK,CAAC,EAAE;AACzE;AAEA,SAASC,OAAOC,EAAY;IAC1B,IAAI,CAACA,GAAGC,MAAM,EAAE,OAAOC;IACvB,OAAO;WAAIF;KAAG,CAACG,IAAI,CAAC,CAACf,GAAGC,IAAMD,IAAIC,EAAE,CAACE,KAAKa,KAAK,CAACJ,GAAGC,MAAM,GAAG,GAAG;AACjE;AAEA,0EAA0E;AAC1E,iFAAiF;AACjF,SAASI,eAAeC,KAAa;IACnC,IAAIC,OAAOD,KAAK,CAAC,EAAE;IACnB,IAAK,IAAIE,IAAI,GAAGA,IAAIF,MAAML,MAAM,EAAEO,IAAK;QACrC,IAAIF,KAAK,CAACE,EAAE,CAACC,IAAI,CAACR,MAAM,GAAGM,KAAKE,IAAI,CAACR,MAAM,EAAEM,OAAOD,KAAK,CAACE,EAAE;IAC9D;IACA,OAAOD,KAAKvB,CAAC;AACf;AAEA,SAAS0B,aAAaC,IAAU,EAAEC,SAAkB;IAClD,IAAIA,UAAUX,MAAM,KAAK,GAAG,OAAO;IACnC,MAAMY,IAAIhB,SAASc,MAAMC,SAAS,CAAC,EAAE;IACrC,MAAMjC,IAAIkB,SAASc,MAAMC,SAAS,CAAC,EAAE;IACrC,IAAIC,MAAMlC,GAAG,OAAOkC,IAAIlC,IAAI,IAAI;IAChC,MAAMmC,UAAUF,UAAUG,GAAG,CAAC,CAACC,KAAO,AAACA,CAAAA,EAAE,CAAC,EAAE,GAAGA,EAAE,CAAC,EAAE,AAAD,IAAK;IACxD,OAAOzB,KAAK0B,GAAG,CAAChC,KAAK0B,QAAQG,OAAO,CAAC,EAAE,KAAKvB,KAAK0B,GAAG,CAAChC,KAAK0B,QAAQG,OAAO,CAAC,EAAE,IAAI,IAAI;AACtF;AAEA,6EAA6E;AAC7E,4CAA4C;AAC5C,SAAShB,MAAMoB,KAAa;IAC1B,IAAIA,MAAMjB,MAAM,GAAG,GAAG,OAAOC;IAC7B,OAAO;QAACX,KAAKC,GAAG,IAAI0B,MAAMH,GAAG,CAAC,CAACpC,IAAMA,EAAEC,CAAC;QAAIW,KAAKI,GAAG,IAAIuB,MAAMH,GAAG,CAACrC;KAAO;AAC3E;AAQA,6EAA6E;AAC7E,4EAA4E;AAC5E,yEAAyE;AACzE,SAASyC,UAAUC,KAAa,EAAEC,MAAc,EAAEC,MAA2B;IAC3E,MAAMC,OAAc,EAAE;IACtB,yEAAyE;IACzE,2EAA2E;IAC3E,MAAMC,SAAS;WAAIJ;KAAM,CAACjB,IAAI,CAAC,CAACf,GAAGC,IAAMH,KAAKG,KAAKH,KAAKE,MAAMA,EAAER,CAAC,GAAGS,EAAET,CAAC;IACvE,KAAK,MAAMC,KAAK2C,OAAQ;QACtB,MAAMC,OAAOH,OAAOzC;QACpB,MAAM6C,SAASH,KAAKI,IAAI,CAAC,CAAChD,IAAMA,EAAE8C,IAAI,KAAKA,QAAQlC,KAAK0B,GAAG,CAAC/B,KAAKP,EAAEgC,IAAI,IAAIzB,KAAKL,OAAOwC,OAAOhD,QAAQ,IAAIkB,KAAKI,GAAG,CAAChB,EAAEgC,IAAI,CAAC3B,CAAC,EAAEH,EAAEG,CAAC,IAAIO,KAAKC,GAAG,CAACb,EAAEgC,IAAI,CAAC3B,CAAC,EAAEH,EAAEG,CAAC,IAAIqC,OAAO/C,eAAe;QACpL,IAAIoD,QAAQ;YACVA,OAAON,KAAK,CAACQ,IAAI,CAAC/C;YAClB6C,OAAOf,IAAI,GAAGxB,MAAMuC,OAAOf,IAAI,EAAE9B;QACnC,OAAO;YACL0C,KAAKK,IAAI,CAAC;gBAAEjB,MAAM;oBAAE/B,GAAGC,EAAED,CAAC;oBAAEG,GAAGF,EAAEE,CAAC;oBAAEF,GAAGA,EAAEA,CAAC;oBAAEG,GAAGH,EAAEG,CAAC;gBAAC;gBAAGoC,OAAO;oBAACvC;iBAAE;gBAAE4C;YAAK;QACzE;IACF;IACA,OAAOF;AACT;AAEA,SAASM,SAASvB,KAAa,EAAEK,IAAU,EAAEmB,GAAW;IACtD,MAAMN,SAAS;WAAIlB;KAAM,CAACH,IAAI,CAAC,CAACf,GAAGC,IAAMD,EAAER,CAAC,GAAGS,EAAET,CAAC;IAClD,MAAM6B,OAAOe,OAAOT,GAAG,CAAC,CAACgB,IAAMA,EAAEtB,IAAI,EAAEuB,IAAI,CAAC;IAC5C,OAAO;QACLvB;QACA7B,GAAG+B,KAAK/B,CAAC;QACTqD,OAAOtB,KAAK9B,CAAC;QACbK,MAAMA,KAAKyB;QACX3B,GAAGqB,eAAeC;QAClBwB;QACAV,OAAOI;IACT;AACF;AAEA,6EAA6E;AAC7E,sIAAsI;AACtI,OAAO,SAASU,aAAad,KAAa,EAAEC,MAAc;QAOxCtB,SA0DSA,UAiDYA;IAjHrC,IAAI,CAACqB,MAAMnB,MAAM,EAAE,OAAO;QAAEkC,OAAO,EAAE;QAAEC,QAAQ,EAAE;QAAEC,OAAO;IAAM;IAEhE,MAAMC,UAAU;IAChB,MAAMC,aAAalB,OAAOjD,QAAQ,GAAGkE;IACrC,MAAME,cAAcnB,OAAOjD,QAAQ,GAAGkE;IAEtC,MAAMG,WAAU1C,UAAAA,OAAOqB,MAAML,GAAG,CAAC,CAAClC,IAAMA,EAAEG,CAAC,gBAA3Be,qBAAAA,UAAiC;IACjD,gEAAgE;IAChE,oFAAoF;IACpF,MAAM2C,WAAWD,UAAU;IAE3B,2EAA2E;IAC3E,gFAAgF;IAChF,MAAME,YAAYvB,MAAMwB,MAAM,CAAC,CAAC/D,IAAMA,EAAEG,CAAC,GAAG0D;IAC5C,MAAMG,YAAYF,UAAUC,MAAM,CAAC,CAAC/D,IAAMA,EAAED,CAAC,GAAGyC,OAAOjD,QAAQ,IAAIM,KAAKG,KAAKwC,OAAOjD,QAAQ,EAAE6B,MAAM;IACpG,MAAM6C,UAAUH,UAAUC,MAAM,CAAC,CAAC/D,IAAMI,KAAKJ,KAAK0D,YAAYtC,MAAM;IACpE,MAAM8C,WAAWJ,UAAUC,MAAM,CAAC,CAAC/D,IAAMI,KAAKJ,KAAK2D,aAAavC,MAAM;IACtE,MAAM+C,SAASF,WAAWtE,iBAAiBuE,YAAYvE,iBAAiBsE,WAAWH,UAAU1C,MAAM,GAAGxB,iBAAiBsE,YAAYJ,UAAU1C,MAAM,GAAGxB,iBAAiBoE,aAAaF,UAAU1C,MAAM,GAAG1B;IAEvM,4EAA4E;IAC5E,4FAA4F;IAC5F,sDAAsD;IACtD,MAAM+C,SAAS,CAACzC;QACd,IAAI,CAACmE,QAAQ,OAAO;QACpB,OAAO/D,KAAKJ,KAAKwC,OAAOjD,QAAQ,GAAG,IAAI;IACzC;IACA,MAAM6E,WAAW9B,UACfC,MAAMwB,MAAM,CAAC,CAAC/D,IAAMA,EAAEG,CAAC,IAAI0D,WAC3BrB,QACA,IAAM;IAER,MAAM6B,aAAa/B,UACjBC,MAAMwB,MAAM,CAAC,CAAC/D,IAAMA,EAAEG,CAAC,GAAG0D,WAC1BrB,QACAC;IAEF,MAAMC,OAAO;WAAI0B;WAAaC;KAAW;IAEzC,2EAA2E;IAC3E,sFAAsF;IACtF,IAAItC;IACJ,IAAIoC,QAAQ;QACV,MAAMG,WAAW5B,KAAKqB,MAAM,CAAC,CAACjE,IAAMA,EAAEgC,IAAI,CAAC9B,CAAC,IAAI,OAAOF,EAAEgC,IAAI,CAAC9B,CAAC,IAAI,QAAQF,EAAEgC,IAAI,CAAC3B,CAAC,IAAIyD,UAAU;QACjG,MAAMW,OAAOtD,MAAMqD,SAASP,MAAM,CAAC,CAACjE,IAAMA,EAAE8C,IAAI,KAAK,GAAGV,GAAG,CAAC,CAACpC,IAAMA,EAAEgC,IAAI;QACzE,MAAM0C,QAAQvD,MAAMqD,SAASP,MAAM,CAAC,CAACjE,IAAMA,EAAE8C,IAAI,KAAK,GAAGV,GAAG,CAAC,CAACpC,IAAMA,EAAEgC,IAAI;QAC1EC,YAAY;YAACwC;YAAMC;SAAM,CAACT,MAAM,CAAC,CAACjE,IAAkBA,MAAMuB;IAC5D,OAAO;QACL,MAAMiD,WAAW5B,KAAKqB,MAAM,CAAC,CAACjE,IAAMA,EAAEgC,IAAI,CAAC9B,CAAC,IAAI,OAAOF,EAAEgC,IAAI,CAAC3B,CAAC,IAAIyD,UAAU;QAC7E,MAAMa,SAASxD,MAAMqD,SAASpC,GAAG,CAAC,CAACpC,IAAMA,EAAEgC,IAAI;QAC/CC,YAAY0C,WAAWpD,YAAY;YAACoD;SAAO,GAAG,EAAE;IAClD;IAEA,MAAMnB,QAAmBZ,KAAKR,GAAG,CAAC,CAACwC,MAAQ1B,SAAS0B,IAAInC,KAAK,EAAEmC,IAAI5C,IAAI,EAAED,aAAa6C,IAAI5C,IAAI,EAAEC;IAChGuB,MAAMhC,IAAI,CAAC,CAACf,GAAGC,IAAMD,EAAE0C,GAAG,GAAGzC,EAAEyC,GAAG,IAAIzC,EAAEH,IAAI,GAAGE,EAAEF,IAAI;IAErD,sEAAsE;IACtE,uJAAuJ;IACvJ,MAAMsE,UAAoB,EAAE;IAC5B,IAAK,IAAIC,IAAI,GAAGA,IAAItB,MAAMlC,MAAM,EAAEwD,IAAK;QACrC,IAAItB,KAAK,CAACsB,EAAE,CAAC3B,GAAG,KAAKK,KAAK,CAACsB,IAAI,EAAE,CAAC3B,GAAG,EAAE;YACrC,MAAM4B,IAAIvB,KAAK,CAACsB,IAAI,EAAE,CAACvE,IAAI,GAAGiD,KAAK,CAACsB,EAAE,CAACvE,IAAI;YAC3C,IAAIwE,IAAI,QAAQF,QAAQ5B,IAAI,CAAC8B;QAC/B;IACF;IACA,MAAMC,oBAAmB5D,WAAAA,OAAOyD,sBAAPzD,sBAAAA,WAAmB;IAC5C,MAAM6D,YAAY,IAAIC;IACtBjD,UAAUkD,OAAO,CAAC,CAACnF,GAAG6B;QACpBoD,UAAUG,GAAG,CAACvD,GAAG7B,CAAC,CAAC,EAAE;IACvB;IACA,MAAMqF,OAAkB,EAAE;IAC1B,MAAMC,kBAA6B,EAAE;IACrC,qEAAqE;IACrE,kHAAkH;IAClH,MAAMC,YAAY/B,MAAMS,MAAM,CAAC,CAAC/B,IAAMA,EAAE7B,CAAC,IAAI0D;IAC7C,KAAK,MAAMyB,QAAQhC,MAAO;QACxB,MAAMiC,YAAYD,KAAK1D,IAAI,CAAC4D,KAAK,CAAC,KAAKzB,MAAM,CAAC0B,SAASrE,MAAM;QAC7D,MAAMsE,QAAQX,UAAUY,GAAG,CAACL,KAAKrC,GAAG;QACpC,MAAM2C,QAAQL,aAAa/C,OAAOvD,aAAa,IAAIyG,UAAUrE,aAAaX,KAAK0B,GAAG,CAACkD,KAAKvF,CAAC,GAAG2F,SAASlD,OAAOrD,WAAW;QACvH,MAAM0G,sBACJD,SACAN,KAAKnF,CAAC,IAAI0D,YACV,CAAC,cAAciC,IAAI,CAACR,KAAK1D,IAAI,KAC7B,uEAAuE;QACvE,yEAAyE;QACzE,CAACyD,UAAUU,IAAI,CAAC,CAACC,QAAUA,UAAUV,QAAQU,MAAM/C,GAAG,KAAKqC,KAAKrC,GAAG,IAAIvC,KAAK0B,GAAG,CAAC4D,MAAM3F,IAAI,GAAGiF,KAAKjF,IAAI,IAAIK,KAAKI,GAAG,CAACkF,MAAM7F,CAAC,EAAEmF,KAAKnF,CAAC,IAAI;QACxI,MAAM8F,kBAAkBJ,uBAAwBD,SAASN,KAAKlC,KAAK,GAAGZ,OAAOtD,aAAa;QAC1F,MAAMgH,oBACJD,mBACAd,KAAKY,IAAI,CAAC,CAACvF;YACT,IAAIA,EAAEyC,GAAG,KAAKqC,KAAKrC,GAAG,EAAE,OAAO;YAC/B,MAAMkD,MAAM3F,EAAEH,IAAI,GAAGiF,KAAKjF,IAAI;YAC9B,IAAI8F,OAAO,KAAKA,MAAM3D,OAAOpD,OAAO,GAAG0F,kBAAkB,OAAO;YAChE,MAAMsB,UAAU1F,KAAKC,GAAG,CAACH,EAAET,CAAC,GAAGS,EAAE4C,KAAK,EAAEkC,KAAKvF,CAAC,GAAGuF,KAAKlC,KAAK,IAAI1C,KAAKI,GAAG,CAACN,EAAET,CAAC,EAAEuF,KAAKvF,CAAC;YACnF,OAAOqG,WAAWd,KAAKlC,KAAK,GAAG;QACjC;QACF,IAAI6C,mBAAmB,CAACC,mBAAmB;YACzCd,gBAAgBrC,IAAI,CAACuC;QACvB,OAAO;YACLH,KAAKpC,IAAI,CAACuC;QACZ;IACF;IACA,yEAAyE;IACzE,iFAAiF;IACjF,MAAM/B,SAAS6B;IAEf,0EAA0E;IAC1E,MAAMiB,OAAiB,EAAE;IACzB,IAAK,IAAIzB,IAAI,GAAGA,IAAIO,KAAK/D,MAAM,EAAEwD,IAAK;QACpC,IAAIO,IAAI,CAACP,EAAE,CAAC3B,GAAG,KAAKkC,IAAI,CAACP,IAAI,EAAE,CAAC3B,GAAG,EAAE;YACnC,MAAM4B,IAAIM,IAAI,CAACP,IAAI,EAAE,CAACvE,IAAI,GAAG8E,IAAI,CAACP,EAAE,CAACvE,IAAI;YACzC,IAAIwE,IAAI,QAAQwB,KAAKtD,IAAI,CAAC8B;QAC5B;IACF;IACA,OAAO;QAAEvB,OAAO6B;QAAM5B;QAAQC,KAAK,GAAEtC,WAAAA,OAAOmF,mBAAPnF,sBAAAA,WAAgB;IAAM;AAC7D;AAEA,mDAAmD;AAEnD,uEAAuE;AACvE,OAAO,SAASoF,OAAOnG,CAAS;IAC9B,OAAOO,KAAK6F,KAAK,CAACpG,IAAI,OAAO;AAC/B;AAEA,4EAA4E;AAC5E,OAAO,SAASqG,cAAcC,KAAgB,EAAEjE,MAAc;IAC5D,MAAMkE,QAAQ,IAAI1B;IAClB,KAAK,MAAM9B,KAAKuD,MAAO;QACrB,KAAK,MAAMzE,KAAKkB,EAAEI,KAAK,CAAE;gBAEToD;YADd,MAAMlG,IAAI8F,OAAOtE,EAAE7B,CAAC;YACpBuG,MAAMxB,GAAG,CAAC1E,GAAG,EAACkG,aAAAA,MAAMf,GAAG,CAACnF,gBAAVkG,wBAAAA,aAAgB,KAAK1E,EAAEJ,IAAI,CAACR,MAAM;QAClD;IACF;IACA,4EAA4E;IAC5E,2DAA2D;IAC3D,IAAIuF,QAAQ;IACZ,IAAIC,YAAY,CAAC;IACjB,KAAK,MAAM,CAAChC,GAAGiC,EAAE,IAAIH,MAAO;QAC1B,IAAIG,IAAID,WAAW;YACjBA,YAAYC;YACZF,QAAQ/B;QACV;IACF;IAEA,MAAMkC,UAAU,IAAI9B;IACpB,IAAI+B,QAAQ,GAAG,uCAAuC;IACtD,IAAIC,MAAqB;IACzB,MAAMC,QAAQ;WAAIP,MAAMQ,IAAI;KAAG,CAACnD,MAAM,CAAC,CAACoD,IAAMA,IAAIR,QAAQnE,OAAOnD,YAAY,EAAEiC,IAAI,CAAC,CAACf,GAAGC,IAAMA,IAAID;IAClG,KAAK,MAAM4G,KAAKF,MAAO;QACrB,IAAID,QAAQ,QAAQG,IAAIH,MAAM,KAAK;YACjCD,QAAQrG,KAAKC,GAAG,CAAC,GAAGoG,QAAQ;YAC5BC,MAAMG;QACR;QACAL,QAAQ5B,GAAG,CAACiC,GAAGJ;IACjB;IACA,OAAO;QAAEJ;QAAOG;IAAQ;AAC1B"}
1
+ {"version":3,"sources":["/Users/kevin/Dev/OpenSource/formats/pdf-to-md/src/geometry.ts"],"sourcesContent":["// geometry.ts — OCR-path pass 2: words -> lines/columns/floats. Input is\n// flat words (normalized page coords, origin bottom-left, y up), not pre-clustered line observations.\nexport interface Word {\n text: string;\n x: number; // rect.minX\n y: number; // rect.minY\n w: number; // rect.width\n h: number; // rect.height\n // 0-100 (tesseract.js's own `confidence` field, not `conf`). Optional —\n // passthrough only; see report.ts's LOW_CONFIDENCE_THRESHOLD for the consumer.\n confidence?: number;\n}\n\n/** A `Word` with its page number — the shared shape behind both the OCR\n * engine's own output (tesseract.ts) and the replay input contract (extract.ts's OcrWordInput). */\nexport interface PageWord extends Word {\n page: number;\n}\n\nexport interface OcrLine {\n text: string;\n x: number; // rect.minX, used for float-alignment and indentation checks\n width: number; // rect.width, used for the float width check\n midY: number;\n h: number; // char-weighted dominant constituent height\n col: number;\n // Constituent words, left-to-right (same order as `text`'s tokens) — carries\n // word identity through row-building for analyze()'s block->word linkage.\n words: Word[];\n}\n\nexport interface PageOCR {\n lines: OcrLine[];\n // Decorative fragment candidates (src/emit.ts's floats aside), kept as full\n // OcrLine so callers recover both the printed text and constituent words.\n floats: OcrLine[];\n pitch: number;\n}\n\n/** Tuning knobs for the OCR-path geometry pass (CLI: `--float-*`/`--para-gap`/\n * `--heading-scale`/`--col-*`/`--line-*`). See DEFAULT_TUNING for defaults. */\nexport interface Tuning {\n /** fragments with more words than this are body text, not decoration */\n floatMaxWords: number;\n /** fraction of page width above which a fragment is too wide to be\n * decoration */\n floatMaxWidth: number;\n /** how far a fragment's left edge must sit from its column edge to be\n * decoration */\n floatMargin: number;\n /** a vertical gap this many times the line pitch is a paragraph break */\n paraGap: number;\n /** lines taller than this times body height are headings */\n headingScale: number;\n /** boxes wider than this are excluded from gutter detection */\n colMaxWidth: number;\n /** fractional x of the two-column midpoint */\n colSplit: number;\n /** vertical distance within which two boxes are the same line */\n lineYTol: number;\n /** same-line boxes with a larger height ratio don't merge */\n lineHeightRatio: number;\n}\n\n/** Measured defaults; derivation in `test/unit/tuning.test.ts`. lineHeightRatio\n * 2.4 = pair-F1 plateau edge (0.924@1.6 -> 0.995@2.4, 9,471 pairs, 2026-08-16). */\nexport const DEFAULT_TUNING: Tuning = {\n floatMaxWords: 3,\n floatMaxWidth: 0.15,\n floatMargin: 0.008,\n paraGap: 1.6,\n headingScale: 2.0,\n colMaxWidth: 0.55,\n colSplit: 0.5,\n lineYTol: 0.006,\n lineHeightRatio: 2.4,\n};\n\n// Two-column decision (test/unit/column-break.test.ts): gutter ink <=0.017\n// (2-col) vs >=0.047 (1-col) measured over 100 baseline pages, cut 0.03.\nexport const MAX_GUTTER_INK_FRAC = 0.03;\n\n// Near-empty pages (e.g. a 2-word title page) have no gutter ink either, so\n// a word-count floor is needed too; genuine columns measure >=76 words.\nexport const MIN_COL_WORDS = 10;\n\n// Measured: a thin data rail is 11% of body words; the thinnest genuine\n// column is 31% — 0.2 sits between.\nexport const MIN_COL_SHARE = 0.2;\n\ninterface Rect {\n x: number;\n y: number;\n w: number;\n h: number;\n}\ntype Range = readonly [number, number];\n\nconst maxX = (r: Rect): number => r.x + r.w;\nconst maxY = (r: Rect): number => r.y + r.h;\nconst midX = (r: Rect): number => r.x + r.w / 2;\nconst midY = (r: Rect): number => r.y + r.h / 2;\n\nfunction union(a: Rect, b: Rect): Rect {\n const x0 = Math.min(a.x, b.x);\n const y0 = Math.min(a.y, b.y);\n const x1 = Math.max(maxX(a), maxX(b));\n const y1 = Math.max(maxY(a), maxY(b));\n return { x: x0, y: y0, w: x1 - x0, h: y1 - y0 };\n}\n\n// Overlap of a rect's x-span with a column x-range.\nfunction xOverlap(r: Rect, range: Range): number {\n return Math.max(0, Math.min(maxX(r), range[1]) - Math.max(r.x, range[0]));\n}\n\nfunction median(xs: number[]): number | undefined {\n if (!xs.length) return undefined;\n return [...xs].sort((a, b) => a - b)[Math.floor(xs.length / 2)];\n}\n\n// Char-weighted dominant height: most characters wins (ties keep first) —\n// same rule as src/lines.ts, so a small merged fragment can't promote a heading.\nfunction dominantHeight(parts: Word[]): number {\n let best = parts[0];\n for (let i = 1; i < parts.length; i++) {\n if (parts[i].text.length > best.text.length) best = parts[i];\n }\n return best.h;\n}\n\nfunction assignColumn(rect: Rect, colRanges: Range[]): number {\n if (colRanges.length !== 2) return 0;\n const l = xOverlap(rect, colRanges[0]);\n const r = xOverlap(rect, colRanges[1]);\n if (l !== r) return l > r ? 0 : 1;\n const centers = colRanges.map((rg) => (rg[0] + rg[1]) / 2);\n return Math.abs(midX(rect) - centers[0]) <= Math.abs(midX(rect) - centers[1]) ? 0 : 1;\n}\n\n// A column x-range from full-width body rows: short paragraph-final rows and\n// decorative fragments must not stretch it.\nfunction range(rects: Rect[]): Range | undefined {\n if (rects.length < 2) return undefined;\n return [Math.min(...rects.map((r) => r.x)), Math.max(...rects.map(maxX))];\n}\n\ninterface Row {\n rect: Rect;\n words: Word[];\n band: number; // -1 (gutter band / n/a), 0 (left or single column), 1 (right)\n}\n\n// Groups same-visual-line words via lineYTol/lineHeightRatio, requiring band\n// agreement; searches ALL open rows since per-word y drifts ~0.001 across a\n// printed line, enough to interleave two lines' words in y-sorted order.\nfunction buildRows(words: Word[], tuning: Tuning, bandOf: (w: Word) => number): Row[] {\n const rows: Row[] = [];\n // Sort order only affects which compatible row a tie lands in (every row\n // is searched) — top-to-bottom, left-to-right keeps results deterministic.\n const sorted = [...words].sort((a, b) => midY(b) - midY(a) || a.x - b.x);\n for (const w of sorted) {\n const band = bandOf(w);\n const target = rows.find((r) => r.band === band && Math.abs(midY(r.rect) - midY(w)) <= tuning.lineYTol && Math.max(r.rect.h, w.h) < Math.min(r.rect.h, w.h) * tuning.lineHeightRatio);\n if (target) {\n target.words.push(w);\n target.rect = union(target.rect, w);\n } else {\n rows.push({ rect: { x: w.x, y: w.y, w: w.w, h: w.h }, words: [w], band });\n }\n }\n return rows;\n}\n\nfunction makeLine(parts: Word[], rect: Rect, col: number): OcrLine {\n const sorted = [...parts].sort((a, b) => a.x - b.x);\n const text = sorted.map((p) => p.text).join(' ');\n return {\n text,\n x: rect.x,\n width: rect.w,\n midY: midY(rect),\n h: dominantHeight(parts),\n col,\n words: sorted,\n };\n}\n\n// Word-geometry pass (render/recognize happens upstream, in the OCR engine).\n// Body-height words band left/right before row-grouping; heading-height words group across the full width, so a title stays one line.\nexport function buildPageOCR(words: Word[], tuning: Tuning): PageOCR {\n if (!words.length) return { lines: [], floats: [], pitch: 0.012 };\n\n const colBand = 0.02;\n const colLeftCut = tuning.colSplit - colBand;\n const colRightCut = tuning.colSplit + colBand;\n\n const medianH = median(words.map((w) => w.h)) ?? 0.017;\n // Same literal `medianH * 2` used elsewhere for this gate — not\n // `tuning.headingScale`, a separate constant that happens to share the 2.0 default.\n const tallGate = medianH * 2;\n\n // Two-column decision (test/unit/column-break.test.ts): gutter ink <=0.017\n // (2-col) vs >=0.047 (1-col), cut 0.03; floors >=0.2 share/>=10 words per side.\n const bodyWords = words.filter((w) => w.h < tallGate);\n const gutterInk = bodyWords.filter((w) => w.x < tuning.colSplit && maxX(w) > tuning.colSplit).length;\n const leftPop = bodyWords.filter((w) => midX(w) < colLeftCut).length;\n const rightPop = bodyWords.filter((w) => midX(w) > colRightCut).length;\n const twoCol = leftPop >= MIN_COL_WORDS && rightPop >= MIN_COL_WORDS && leftPop >= bodyWords.length * MIN_COL_SHARE && rightPop >= bodyWords.length * MIN_COL_SHARE && gutterInk <= bodyWords.length * MAX_GUTTER_INK_FRAC;\n\n // Band by word CENTER, no dead band: a word box can touch the split exactly\n // (a box can touch the split exactly) and still belongs to its center's column — the twoCol\n // gate above already guarantees near-zero gutter ink.\n const bandOf = (w: Word): number => {\n if (!twoCol) return 0;\n return midX(w) < tuning.colSplit ? 0 : 1;\n };\n const tallRows = buildRows(\n words.filter((w) => w.h >= tallGate),\n tuning,\n () => 0\n );\n const normalRows = buildRows(\n words.filter((w) => w.h < tallGate),\n tuning,\n bandOf\n );\n const rows = [...tallRows, ...normalRows];\n\n // Column ranges from body rows (height <= medianH*1.6): two-col ranges use\n // rows 0.2-0.55 wide; single-col has no upper bound (full-width rows ARE the column).\n let colRanges: Range[];\n if (twoCol) {\n const wideRows = rows.filter((r) => r.rect.w >= 0.2 && r.rect.w <= 0.55 && r.rect.h <= medianH * 1.6);\n const left = range(wideRows.filter((r) => r.band === 0).map((r) => r.rect));\n const right = range(wideRows.filter((r) => r.band === 1).map((r) => r.rect));\n colRanges = [left, right].filter((r): r is Range => r !== undefined);\n } else {\n const wideRows = rows.filter((r) => r.rect.w >= 0.2 && r.rect.h <= medianH * 1.6);\n const single = range(wideRows.map((r) => r.rect));\n colRanges = single !== undefined ? [single] : [];\n }\n\n const lines: OcrLine[] = rows.map((row) => makeLine(row.words, row.rect, assignColumn(row.rect, colRanges)));\n lines.sort((a, b) => a.col - b.col || b.midY - a.midY);\n\n // Float detection: short+narrow+misaligned, rescued into body when it\n // hangs under a body line like a paragraph continuation (same col, gap<=paraGap x pitch, x-overlap) — protects a centered paragraph's short last line.\n const preGaps: number[] = [];\n for (let k = 1; k < lines.length; k++) {\n if (lines[k].col === lines[k - 1].col) {\n const g = lines[k - 1].midY - lines[k].midY;\n if (g > 0.0005) preGaps.push(g);\n }\n }\n const provisionalPitch = median(preGaps) ?? 0.012;\n const colStarts = new Map<number, number>();\n colRanges.forEach((r, i) => {\n colStarts.set(i, r[0]);\n });\n const body: OcrLine[] = [];\n const floatCandidates: OcrLine[] = [];\n // Display-size junk (e.g. \"1/3\" at 68pt, 0.179 wide vs the 0.15 cap)\n // escapes the narrow float test; caught here by junk text + isolation (no adjacent tall line within 2.2x height).\n const tallLines = lines.filter((l) => l.h >= tallGate);\n for (const line of lines) {\n const wordCount = line.text.split(' ').filter(Boolean).length;\n const start = colStarts.get(line.col);\n const short = wordCount <= tuning.floatMaxWords && start !== undefined && Math.abs(line.x - start) > tuning.floatMargin;\n const displayJunkFragment =\n short &&\n line.h >= tallGate &&\n !/[A-Za-z]{3}/.test(line.text) &&\n // Same-column only, matching emitPage's heading-run gather — a heading\n // in the OTHER column can be y-adjacent without composing with this one.\n !tallLines.some((other) => other !== line && other.col === line.col && Math.abs(other.midY - line.midY) < Math.max(other.h, line.h) * 2.2);\n const looksDecorative = displayJunkFragment || (short && line.width < tuning.floatMaxWidth);\n const paragraphAdjacent =\n looksDecorative &&\n body.some((b) => {\n if (b.col !== line.col) return false;\n const gap = b.midY - line.midY;\n if (gap <= 0 || gap > tuning.paraGap * provisionalPitch) return false;\n const overlap = Math.min(b.x + b.width, line.x + line.width) - Math.max(b.x, line.x);\n return overlap >= line.width * 0.5;\n });\n if (looksDecorative && !paragraphAdjacent) {\n floatCandidates.push(line);\n } else {\n body.push(line);\n }\n }\n // One float per candidate row, in line order — no clustering of adjacent\n // candidates: replay against the frozen baseline regressed ~30 pages when tried.\n const floats = floatCandidates;\n\n // Line pitch = median vertical gap between consecutive same-column lines.\n const gaps: number[] = [];\n for (let k = 1; k < body.length; k++) {\n if (body[k].col === body[k - 1].col) {\n const g = body[k - 1].midY - body[k].midY;\n if (g > 0.0005) gaps.push(g);\n }\n }\n return { lines: body, floats, pitch: median(gaps) ?? 0.012 };\n}\n\n// Document-wide height histogram -> heading levels\n\n// Bucket heights to 0.002 of page height so histogram keys are stable.\nexport function bucket(h: number): number {\n return Math.round(h * 500) / 500;\n}\n\n// `pages` are the already-body-only (post-float-filtering) PageOCR results.\nexport function headingLevels(pages: PageOCR[], tuning: Tuning): { bodyH: number; levelOf: Map<number, number> } {\n const chars = new Map<number, number>();\n for (const p of pages) {\n for (const l of p.lines) {\n const b = bucket(l.h);\n chars.set(b, (chars.get(b) ?? 0) + l.text.length);\n }\n }\n // Most-common height wins; ties broken by first-encountered (deterministic;\n // char-count ties across real documents are not observed).\n let bodyH = 0;\n let bestCount = -1;\n for (const [k, v] of chars) {\n if (v > bestCount) {\n bestCount = v;\n bodyH = k;\n }\n }\n\n const levelOf = new Map<number, number>();\n let level = 3; // \"### pN\" page markers occupy level 3\n let rep: number | null = null;\n const sizes = [...chars.keys()].filter((s) => s > bodyH * tuning.headingScale).sort((a, b) => b - a);\n for (const s of sizes) {\n if (rep === null || s < rep * 0.8) {\n level = Math.min(6, level + 1);\n rep = s;\n }\n levelOf.set(s, level);\n }\n return { bodyH, levelOf };\n}\n"],"names":["DEFAULT_TUNING","floatMaxWords","floatMaxWidth","floatMargin","paraGap","headingScale","colMaxWidth","colSplit","lineYTol","lineHeightRatio","MAX_GUTTER_INK_FRAC","MIN_COL_WORDS","MIN_COL_SHARE","maxX","r","x","w","maxY","y","h","midX","midY","union","a","b","x0","Math","min","y0","x1","max","y1","xOverlap","range","median","xs","length","undefined","sort","floor","dominantHeight","parts","best","i","text","assignColumn","rect","colRanges","l","centers","map","rg","abs","rects","buildRows","words","tuning","bandOf","rows","sorted","band","target","find","push","makeLine","col","p","join","width","buildPageOCR","lines","floats","pitch","colBand","colLeftCut","colRightCut","medianH","tallGate","bodyWords","filter","gutterInk","leftPop","rightPop","twoCol","tallRows","normalRows","wideRows","left","right","single","row","preGaps","k","g","provisionalPitch","colStarts","Map","forEach","set","body","floatCandidates","tallLines","line","wordCount","split","Boolean","start","get","short","displayJunkFragment","test","some","other","looksDecorative","paragraphAdjacent","gap","overlap","gaps","bucket","round","headingLevels","pages","chars","bodyH","bestCount","v","levelOf","level","rep","sizes","keys","s"],"mappings":"AAAA,yEAAyE;AACzE,sGAAsG;AA+DtG;iFACiF,GACjF,OAAO,MAAMA,iBAAyB;IACpCC,eAAe;IACfC,eAAe;IACfC,aAAa;IACbC,SAAS;IACTC,cAAc;IACdC,aAAa;IACbC,UAAU;IACVC,UAAU;IACVC,iBAAiB;AACnB,EAAE;AAEF,2EAA2E;AAC3E,yEAAyE;AACzE,OAAO,MAAMC,sBAAsB,KAAK;AAExC,4EAA4E;AAC5E,wEAAwE;AACxE,OAAO,MAAMC,gBAAgB,GAAG;AAEhC,wEAAwE;AACxE,oCAAoC;AACpC,OAAO,MAAMC,gBAAgB,IAAI;AAUjC,MAAMC,OAAO,CAACC,IAAoBA,EAAEC,CAAC,GAAGD,EAAEE,CAAC;AAC3C,MAAMC,OAAO,CAACH,IAAoBA,EAAEI,CAAC,GAAGJ,EAAEK,CAAC;AAC3C,MAAMC,OAAO,CAACN,IAAoBA,EAAEC,CAAC,GAAGD,EAAEE,CAAC,GAAG;AAC9C,MAAMK,OAAO,CAACP,IAAoBA,EAAEI,CAAC,GAAGJ,EAAEK,CAAC,GAAG;AAE9C,SAASG,MAAMC,CAAO,EAAEC,CAAO;IAC7B,MAAMC,KAAKC,KAAKC,GAAG,CAACJ,EAAER,CAAC,EAAES,EAAET,CAAC;IAC5B,MAAMa,KAAKF,KAAKC,GAAG,CAACJ,EAAEL,CAAC,EAAEM,EAAEN,CAAC;IAC5B,MAAMW,KAAKH,KAAKI,GAAG,CAACjB,KAAKU,IAAIV,KAAKW;IAClC,MAAMO,KAAKL,KAAKI,GAAG,CAACb,KAAKM,IAAIN,KAAKO;IAClC,OAAO;QAAET,GAAGU;QAAIP,GAAGU;QAAIZ,GAAGa,KAAKJ;QAAIN,GAAGY,KAAKH;IAAG;AAChD;AAEA,oDAAoD;AACpD,SAASI,SAASlB,CAAO,EAAEmB,KAAY;IACrC,OAAOP,KAAKI,GAAG,CAAC,GAAGJ,KAAKC,GAAG,CAACd,KAAKC,IAAImB,KAAK,CAAC,EAAE,IAAIP,KAAKI,GAAG,CAAChB,EAAEC,CAAC,EAAEkB,KAAK,CAAC,EAAE;AACzE;AAEA,SAASC,OAAOC,EAAY;IAC1B,IAAI,CAACA,GAAGC,MAAM,EAAE,OAAOC;IACvB,OAAO;WAAIF;KAAG,CAACG,IAAI,CAAC,CAACf,GAAGC,IAAMD,IAAIC,EAAE,CAACE,KAAKa,KAAK,CAACJ,GAAGC,MAAM,GAAG,GAAG;AACjE;AAEA,0EAA0E;AAC1E,iFAAiF;AACjF,SAASI,eAAeC,KAAa;IACnC,IAAIC,OAAOD,KAAK,CAAC,EAAE;IACnB,IAAK,IAAIE,IAAI,GAAGA,IAAIF,MAAML,MAAM,EAAEO,IAAK;QACrC,IAAIF,KAAK,CAACE,EAAE,CAACC,IAAI,CAACR,MAAM,GAAGM,KAAKE,IAAI,CAACR,MAAM,EAAEM,OAAOD,KAAK,CAACE,EAAE;IAC9D;IACA,OAAOD,KAAKvB,CAAC;AACf;AAEA,SAAS0B,aAAaC,IAAU,EAAEC,SAAkB;IAClD,IAAIA,UAAUX,MAAM,KAAK,GAAG,OAAO;IACnC,MAAMY,IAAIhB,SAASc,MAAMC,SAAS,CAAC,EAAE;IACrC,MAAMjC,IAAIkB,SAASc,MAAMC,SAAS,CAAC,EAAE;IACrC,IAAIC,MAAMlC,GAAG,OAAOkC,IAAIlC,IAAI,IAAI;IAChC,MAAMmC,UAAUF,UAAUG,GAAG,CAAC,CAACC,KAAO,AAACA,CAAAA,EAAE,CAAC,EAAE,GAAGA,EAAE,CAAC,EAAE,AAAD,IAAK;IACxD,OAAOzB,KAAK0B,GAAG,CAAChC,KAAK0B,QAAQG,OAAO,CAAC,EAAE,KAAKvB,KAAK0B,GAAG,CAAChC,KAAK0B,QAAQG,OAAO,CAAC,EAAE,IAAI,IAAI;AACtF;AAEA,6EAA6E;AAC7E,4CAA4C;AAC5C,SAAShB,MAAMoB,KAAa;IAC1B,IAAIA,MAAMjB,MAAM,GAAG,GAAG,OAAOC;IAC7B,OAAO;QAACX,KAAKC,GAAG,IAAI0B,MAAMH,GAAG,CAAC,CAACpC,IAAMA,EAAEC,CAAC;QAAIW,KAAKI,GAAG,IAAIuB,MAAMH,GAAG,CAACrC;KAAO;AAC3E;AAQA,6EAA6E;AAC7E,4EAA4E;AAC5E,yEAAyE;AACzE,SAASyC,UAAUC,KAAa,EAAEC,MAAc,EAAEC,MAA2B;IAC3E,MAAMC,OAAc,EAAE;IACtB,yEAAyE;IACzE,2EAA2E;IAC3E,MAAMC,SAAS;WAAIJ;KAAM,CAACjB,IAAI,CAAC,CAACf,GAAGC,IAAMH,KAAKG,KAAKH,KAAKE,MAAMA,EAAER,CAAC,GAAGS,EAAET,CAAC;IACvE,KAAK,MAAMC,KAAK2C,OAAQ;QACtB,MAAMC,OAAOH,OAAOzC;QACpB,MAAM6C,SAASH,KAAKI,IAAI,CAAC,CAAChD,IAAMA,EAAE8C,IAAI,KAAKA,QAAQlC,KAAK0B,GAAG,CAAC/B,KAAKP,EAAEgC,IAAI,IAAIzB,KAAKL,OAAOwC,OAAOhD,QAAQ,IAAIkB,KAAKI,GAAG,CAAChB,EAAEgC,IAAI,CAAC3B,CAAC,EAAEH,EAAEG,CAAC,IAAIO,KAAKC,GAAG,CAACb,EAAEgC,IAAI,CAAC3B,CAAC,EAAEH,EAAEG,CAAC,IAAIqC,OAAO/C,eAAe;QACpL,IAAIoD,QAAQ;YACVA,OAAON,KAAK,CAACQ,IAAI,CAAC/C;YAClB6C,OAAOf,IAAI,GAAGxB,MAAMuC,OAAOf,IAAI,EAAE9B;QACnC,OAAO;YACL0C,KAAKK,IAAI,CAAC;gBAAEjB,MAAM;oBAAE/B,GAAGC,EAAED,CAAC;oBAAEG,GAAGF,EAAEE,CAAC;oBAAEF,GAAGA,EAAEA,CAAC;oBAAEG,GAAGH,EAAEG,CAAC;gBAAC;gBAAGoC,OAAO;oBAACvC;iBAAE;gBAAE4C;YAAK;QACzE;IACF;IACA,OAAOF;AACT;AAEA,SAASM,SAASvB,KAAa,EAAEK,IAAU,EAAEmB,GAAW;IACtD,MAAMN,SAAS;WAAIlB;KAAM,CAACH,IAAI,CAAC,CAACf,GAAGC,IAAMD,EAAER,CAAC,GAAGS,EAAET,CAAC;IAClD,MAAM6B,OAAOe,OAAOT,GAAG,CAAC,CAACgB,IAAMA,EAAEtB,IAAI,EAAEuB,IAAI,CAAC;IAC5C,OAAO;QACLvB;QACA7B,GAAG+B,KAAK/B,CAAC;QACTqD,OAAOtB,KAAK9B,CAAC;QACbK,MAAMA,KAAKyB;QACX3B,GAAGqB,eAAeC;QAClBwB;QACAV,OAAOI;IACT;AACF;AAEA,6EAA6E;AAC7E,sIAAsI;AACtI,OAAO,SAASU,aAAad,KAAa,EAAEC,MAAc;QAOxCtB,SA0DSA,UAiDYA;IAjHrC,IAAI,CAACqB,MAAMnB,MAAM,EAAE,OAAO;QAAEkC,OAAO,EAAE;QAAEC,QAAQ,EAAE;QAAEC,OAAO;IAAM;IAEhE,MAAMC,UAAU;IAChB,MAAMC,aAAalB,OAAOjD,QAAQ,GAAGkE;IACrC,MAAME,cAAcnB,OAAOjD,QAAQ,GAAGkE;IAEtC,MAAMG,WAAU1C,UAAAA,OAAOqB,MAAML,GAAG,CAAC,CAAClC,IAAMA,EAAEG,CAAC,gBAA3Be,qBAAAA,UAAiC;IACjD,gEAAgE;IAChE,oFAAoF;IACpF,MAAM2C,WAAWD,UAAU;IAE3B,2EAA2E;IAC3E,gFAAgF;IAChF,MAAME,YAAYvB,MAAMwB,MAAM,CAAC,CAAC/D,IAAMA,EAAEG,CAAC,GAAG0D;IAC5C,MAAMG,YAAYF,UAAUC,MAAM,CAAC,CAAC/D,IAAMA,EAAED,CAAC,GAAGyC,OAAOjD,QAAQ,IAAIM,KAAKG,KAAKwC,OAAOjD,QAAQ,EAAE6B,MAAM;IACpG,MAAM6C,UAAUH,UAAUC,MAAM,CAAC,CAAC/D,IAAMI,KAAKJ,KAAK0D,YAAYtC,MAAM;IACpE,MAAM8C,WAAWJ,UAAUC,MAAM,CAAC,CAAC/D,IAAMI,KAAKJ,KAAK2D,aAAavC,MAAM;IACtE,MAAM+C,SAASF,WAAWtE,iBAAiBuE,YAAYvE,iBAAiBsE,WAAWH,UAAU1C,MAAM,GAAGxB,iBAAiBsE,YAAYJ,UAAU1C,MAAM,GAAGxB,iBAAiBoE,aAAaF,UAAU1C,MAAM,GAAG1B;IAEvM,4EAA4E;IAC5E,4FAA4F;IAC5F,sDAAsD;IACtD,MAAM+C,SAAS,CAACzC;QACd,IAAI,CAACmE,QAAQ,OAAO;QACpB,OAAO/D,KAAKJ,KAAKwC,OAAOjD,QAAQ,GAAG,IAAI;IACzC;IACA,MAAM6E,WAAW9B,UACfC,MAAMwB,MAAM,CAAC,CAAC/D,IAAMA,EAAEG,CAAC,IAAI0D,WAC3BrB,QACA,IAAM;IAER,MAAM6B,aAAa/B,UACjBC,MAAMwB,MAAM,CAAC,CAAC/D,IAAMA,EAAEG,CAAC,GAAG0D,WAC1BrB,QACAC;IAEF,MAAMC,OAAO;WAAI0B;WAAaC;KAAW;IAEzC,2EAA2E;IAC3E,sFAAsF;IACtF,IAAItC;IACJ,IAAIoC,QAAQ;QACV,MAAMG,WAAW5B,KAAKqB,MAAM,CAAC,CAACjE,IAAMA,EAAEgC,IAAI,CAAC9B,CAAC,IAAI,OAAOF,EAAEgC,IAAI,CAAC9B,CAAC,IAAI,QAAQF,EAAEgC,IAAI,CAAC3B,CAAC,IAAIyD,UAAU;QACjG,MAAMW,OAAOtD,MAAMqD,SAASP,MAAM,CAAC,CAACjE,IAAMA,EAAE8C,IAAI,KAAK,GAAGV,GAAG,CAAC,CAACpC,IAAMA,EAAEgC,IAAI;QACzE,MAAM0C,QAAQvD,MAAMqD,SAASP,MAAM,CAAC,CAACjE,IAAMA,EAAE8C,IAAI,KAAK,GAAGV,GAAG,CAAC,CAACpC,IAAMA,EAAEgC,IAAI;QAC1EC,YAAY;YAACwC;YAAMC;SAAM,CAACT,MAAM,CAAC,CAACjE,IAAkBA,MAAMuB;IAC5D,OAAO;QACL,MAAMiD,WAAW5B,KAAKqB,MAAM,CAAC,CAACjE,IAAMA,EAAEgC,IAAI,CAAC9B,CAAC,IAAI,OAAOF,EAAEgC,IAAI,CAAC3B,CAAC,IAAIyD,UAAU;QAC7E,MAAMa,SAASxD,MAAMqD,SAASpC,GAAG,CAAC,CAACpC,IAAMA,EAAEgC,IAAI;QAC/CC,YAAY0C,WAAWpD,YAAY;YAACoD;SAAO,GAAG,EAAE;IAClD;IAEA,MAAMnB,QAAmBZ,KAAKR,GAAG,CAAC,CAACwC,MAAQ1B,SAAS0B,IAAInC,KAAK,EAAEmC,IAAI5C,IAAI,EAAED,aAAa6C,IAAI5C,IAAI,EAAEC;IAChGuB,MAAMhC,IAAI,CAAC,CAACf,GAAGC,IAAMD,EAAE0C,GAAG,GAAGzC,EAAEyC,GAAG,IAAIzC,EAAEH,IAAI,GAAGE,EAAEF,IAAI;IAErD,sEAAsE;IACtE,uJAAuJ;IACvJ,MAAMsE,UAAoB,EAAE;IAC5B,IAAK,IAAIC,IAAI,GAAGA,IAAItB,MAAMlC,MAAM,EAAEwD,IAAK;QACrC,IAAItB,KAAK,CAACsB,EAAE,CAAC3B,GAAG,KAAKK,KAAK,CAACsB,IAAI,EAAE,CAAC3B,GAAG,EAAE;YACrC,MAAM4B,IAAIvB,KAAK,CAACsB,IAAI,EAAE,CAACvE,IAAI,GAAGiD,KAAK,CAACsB,EAAE,CAACvE,IAAI;YAC3C,IAAIwE,IAAI,QAAQF,QAAQ5B,IAAI,CAAC8B;QAC/B;IACF;IACA,MAAMC,oBAAmB5D,WAAAA,OAAOyD,sBAAPzD,sBAAAA,WAAmB;IAC5C,MAAM6D,YAAY,IAAIC;IACtBjD,UAAUkD,OAAO,CAAC,CAACnF,GAAG6B;QACpBoD,UAAUG,GAAG,CAACvD,GAAG7B,CAAC,CAAC,EAAE;IACvB;IACA,MAAMqF,OAAkB,EAAE;IAC1B,MAAMC,kBAA6B,EAAE;IACrC,qEAAqE;IACrE,kHAAkH;IAClH,MAAMC,YAAY/B,MAAMS,MAAM,CAAC,CAAC/B,IAAMA,EAAE7B,CAAC,IAAI0D;IAC7C,KAAK,MAAMyB,QAAQhC,MAAO;QACxB,MAAMiC,YAAYD,KAAK1D,IAAI,CAAC4D,KAAK,CAAC,KAAKzB,MAAM,CAAC0B,SAASrE,MAAM;QAC7D,MAAMsE,QAAQX,UAAUY,GAAG,CAACL,KAAKrC,GAAG;QACpC,MAAM2C,QAAQL,aAAa/C,OAAOvD,aAAa,IAAIyG,UAAUrE,aAAaX,KAAK0B,GAAG,CAACkD,KAAKvF,CAAC,GAAG2F,SAASlD,OAAOrD,WAAW;QACvH,MAAM0G,sBACJD,SACAN,KAAKnF,CAAC,IAAI0D,YACV,CAAC,cAAciC,IAAI,CAACR,KAAK1D,IAAI,KAC7B,uEAAuE;QACvE,yEAAyE;QACzE,CAACyD,UAAUU,IAAI,CAAC,CAACC,QAAUA,UAAUV,QAAQU,MAAM/C,GAAG,KAAKqC,KAAKrC,GAAG,IAAIvC,KAAK0B,GAAG,CAAC4D,MAAM3F,IAAI,GAAGiF,KAAKjF,IAAI,IAAIK,KAAKI,GAAG,CAACkF,MAAM7F,CAAC,EAAEmF,KAAKnF,CAAC,IAAI;QACxI,MAAM8F,kBAAkBJ,uBAAwBD,SAASN,KAAKlC,KAAK,GAAGZ,OAAOtD,aAAa;QAC1F,MAAMgH,oBACJD,mBACAd,KAAKY,IAAI,CAAC,CAACvF;YACT,IAAIA,EAAEyC,GAAG,KAAKqC,KAAKrC,GAAG,EAAE,OAAO;YAC/B,MAAMkD,MAAM3F,EAAEH,IAAI,GAAGiF,KAAKjF,IAAI;YAC9B,IAAI8F,OAAO,KAAKA,MAAM3D,OAAOpD,OAAO,GAAG0F,kBAAkB,OAAO;YAChE,MAAMsB,UAAU1F,KAAKC,GAAG,CAACH,EAAET,CAAC,GAAGS,EAAE4C,KAAK,EAAEkC,KAAKvF,CAAC,GAAGuF,KAAKlC,KAAK,IAAI1C,KAAKI,GAAG,CAACN,EAAET,CAAC,EAAEuF,KAAKvF,CAAC;YACnF,OAAOqG,WAAWd,KAAKlC,KAAK,GAAG;QACjC;QACF,IAAI6C,mBAAmB,CAACC,mBAAmB;YACzCd,gBAAgBrC,IAAI,CAACuC;QACvB,OAAO;YACLH,KAAKpC,IAAI,CAACuC;QACZ;IACF;IACA,yEAAyE;IACzE,iFAAiF;IACjF,MAAM/B,SAAS6B;IAEf,0EAA0E;IAC1E,MAAMiB,OAAiB,EAAE;IACzB,IAAK,IAAIzB,IAAI,GAAGA,IAAIO,KAAK/D,MAAM,EAAEwD,IAAK;QACpC,IAAIO,IAAI,CAACP,EAAE,CAAC3B,GAAG,KAAKkC,IAAI,CAACP,IAAI,EAAE,CAAC3B,GAAG,EAAE;YACnC,MAAM4B,IAAIM,IAAI,CAACP,IAAI,EAAE,CAACvE,IAAI,GAAG8E,IAAI,CAACP,EAAE,CAACvE,IAAI;YACzC,IAAIwE,IAAI,QAAQwB,KAAKtD,IAAI,CAAC8B;QAC5B;IACF;IACA,OAAO;QAAEvB,OAAO6B;QAAM5B;QAAQC,KAAK,GAAEtC,WAAAA,OAAOmF,mBAAPnF,sBAAAA,WAAgB;IAAM;AAC7D;AAEA,mDAAmD;AAEnD,uEAAuE;AACvE,OAAO,SAASoF,OAAOnG,CAAS;IAC9B,OAAOO,KAAK6F,KAAK,CAACpG,IAAI,OAAO;AAC/B;AAEA,4EAA4E;AAC5E,OAAO,SAASqG,cAAcC,KAAgB,EAAEjE,MAAc;IAC5D,MAAMkE,QAAQ,IAAI1B;IAClB,KAAK,MAAM9B,KAAKuD,MAAO;QACrB,KAAK,MAAMzE,KAAKkB,EAAEI,KAAK,CAAE;gBAEToD;YADd,MAAMlG,IAAI8F,OAAOtE,EAAE7B,CAAC;YACpBuG,MAAMxB,GAAG,CAAC1E,GAAG,EAACkG,aAAAA,MAAMf,GAAG,CAACnF,gBAAVkG,wBAAAA,aAAgB,KAAK1E,EAAEJ,IAAI,CAACR,MAAM;QAClD;IACF;IACA,4EAA4E;IAC5E,2DAA2D;IAC3D,IAAIuF,QAAQ;IACZ,IAAIC,YAAY,CAAC;IACjB,KAAK,MAAM,CAAChC,GAAGiC,EAAE,IAAIH,MAAO;QAC1B,IAAIG,IAAID,WAAW;YACjBA,YAAYC;YACZF,QAAQ/B;QACV;IACF;IAEA,MAAMkC,UAAU,IAAI9B;IACpB,IAAI+B,QAAQ,GAAG,uCAAuC;IACtD,IAAIC,MAAqB;IACzB,MAAMC,QAAQ;WAAIP,MAAMQ,IAAI;KAAG,CAACnD,MAAM,CAAC,CAACoD,IAAMA,IAAIR,QAAQnE,OAAOnD,YAAY,EAAEiC,IAAI,CAAC,CAACf,GAAGC,IAAMA,IAAID;IAClG,KAAK,MAAM4G,KAAKF,MAAO;QACrB,IAAID,QAAQ,QAAQG,IAAIH,MAAM,KAAK;YACjCD,QAAQrG,KAAKC,GAAG,CAAC,GAAGoG,QAAQ;YAC5BC,MAAMG;QACR;QACAL,QAAQ5B,GAAG,CAACiC,GAAGJ;IACjB;IACA,OAAO;QAAEJ;QAAOG;IAAQ;AAC1B"}
@@ -1 +1 @@
1
- {"version":3,"sources":["/Users/kevin/Projects/pdf-to-md/src/index.ts"],"sourcesContent":["// Public library API. Deliberately small: every export is a stability\n// promise; internals stay module-private behind these.\n\nexport type { AnalyzeOcrOptions, AnalyzeOptions, AnalyzeTextOptions } from './analyze.ts';\nexport { analyze } from './analyze.ts';\nexport type { AuditResult } from './audit.ts';\nexport { auditWords } from './audit.ts';\nexport type { PdfToMdErrorCode } from './errors.ts';\nexport { PdfToMdError } from './errors.ts';\nexport type {\n ExtractOcrOptions,\n ExtractResult,\n ExtractTextOptions,\n OcrProgressEvent,\n OcrWordInput,\n PageSelection,\n PdfToMarkdownOptions,\n} from './extract.ts';\nexport { extractOcr, extractText, pdfToMarkdown } from './extract.ts';\nexport type { PageWord, Tuning } from './geometry.ts';\nexport { DEFAULT_TUNING } from './geometry.ts';\nexport type { RenderPageResult } from './raster.ts';\nexport { renderPageToPNG } from './raster.ts';\nexport { toMarkdown, toText } from './render-analysis.ts';\nexport { OUTPUT_VERSION } from './report.ts';\nexport type { Analysis, AnalysisBlock, AnalysisPage, AnalysisWord, PageStatsRow, Report } from './types.ts';\n"],"names":["analyze","auditWords","PdfToMdError","extractOcr","extractText","pdfToMarkdown","DEFAULT_TUNING","renderPageToPNG","toMarkdown","toText","OUTPUT_VERSION"],"mappings":"AAAA,sEAAsE;AACtE,uDAAuD;AAGvD,SAASA,OAAO,QAAQ,eAAe;AAEvC,SAASC,UAAU,QAAQ,aAAa;AAExC,SAASC,YAAY,QAAQ,cAAc;AAU3C,SAASC,UAAU,EAAEC,WAAW,EAAEC,aAAa,QAAQ,eAAe;AAEtE,SAASC,cAAc,QAAQ,gBAAgB;AAE/C,SAASC,eAAe,QAAQ,cAAc;AAC9C,SAASC,UAAU,EAAEC,MAAM,QAAQ,uBAAuB;AAC1D,SAASC,cAAc,QAAQ,cAAc"}
1
+ {"version":3,"sources":["/Users/kevin/Dev/OpenSource/formats/pdf-to-md/src/index.ts"],"sourcesContent":["// Public library API. Deliberately small: every export is a stability\n// promise; internals stay module-private behind these.\n\nexport type { AnalyzeOcrOptions, AnalyzeOptions, AnalyzeTextOptions } from './analyze.ts';\nexport { analyze } from './analyze.ts';\nexport type { AuditResult } from './audit.ts';\nexport { auditWords } from './audit.ts';\nexport type { PdfToMdErrorCode } from './errors.ts';\nexport { PdfToMdError } from './errors.ts';\nexport type {\n ExtractOcrOptions,\n ExtractResult,\n ExtractTextOptions,\n OcrProgressEvent,\n OcrWordInput,\n PageSelection,\n PdfToMarkdownOptions,\n} from './extract.ts';\nexport { extractOcr, extractText, pdfToMarkdown } from './extract.ts';\nexport type { PageWord, Tuning } from './geometry.ts';\nexport { DEFAULT_TUNING } from './geometry.ts';\nexport type { RenderPageResult } from './raster.ts';\nexport { renderPageToPNG } from './raster.ts';\nexport { toMarkdown, toText } from './render-analysis.ts';\nexport { OUTPUT_VERSION } from './report.ts';\nexport type { Analysis, AnalysisBlock, AnalysisPage, AnalysisWord, PageStatsRow, Report } from './types.ts';\n"],"names":["analyze","auditWords","PdfToMdError","extractOcr","extractText","pdfToMarkdown","DEFAULT_TUNING","renderPageToPNG","toMarkdown","toText","OUTPUT_VERSION"],"mappings":"AAAA,sEAAsE;AACtE,uDAAuD;AAGvD,SAASA,OAAO,QAAQ,eAAe;AAEvC,SAASC,UAAU,QAAQ,aAAa;AAExC,SAASC,YAAY,QAAQ,cAAc;AAU3C,SAASC,UAAU,EAAEC,WAAW,EAAEC,aAAa,QAAQ,eAAe;AAEtE,SAASC,cAAc,QAAQ,gBAAgB;AAE/C,SAASC,eAAe,QAAQ,cAAc;AAC9C,SAASC,UAAU,EAAEC,MAAM,QAAQ,uBAAuB;AAC1D,SAASC,cAAc,QAAQ,cAAc"}
@@ -1 +1 @@
1
- {"version":3,"sources":["/Users/kevin/Projects/pdf-to-md/src/lines.ts"],"sourcesContent":["// Pass 2: glyphs -> visual lines — de-braids two-column pages (spanning\n// lines like headings sort with the left column), joins words by measured horizontal gap, and folds typographic codepoints to ASCII so search doesn't miss half the hits.\nimport type { CollectedPage, Glyph, Line } from './types.ts';\n\nconst FOLD: Record<string, string> = {\n ʼ: \"'\",\n '’': \"'\",\n '‘': \"'\",\n '“': '\"',\n '”': '\"',\n '–': '-',\n};\n// Exported for the OCR engine seam (src/engines/tesseract.ts): both paths\n// must fold identically or the same printed apostrophe surfaces as two spellings depending on which path read the page.\nexport const foldTypographic = (s: string): string => s.replace(/[ʼ‘’“”–]/g, (c) => FOLD[c]);\nconst normalize = foldTypographic;\n\n// Intermediate row while grouping glyphs by visual line, before the final\n// text/height reduction below.\ninterface LineDraft {\n col: number;\n y: number;\n parts: Glyph[];\n}\n\nexport function toLines({ glyphs, width }: Pick<CollectedPage, 'glyphs' | 'width'>): Line[] {\n if (!glyphs.length) return [];\n\n const mid = width / 2;\n // A real two-column page has a gutter: clusters each side, almost nothing crossing mid.\n // Label/value layouts also cluster but their values cross mid — de-braiding would reorder them.\n const twoCol = glyphs.filter((g) => g.x + g.w < mid).length > glyphs.length * 0.25 && glyphs.filter((g) => g.x > mid).length > glyphs.length * 0.25 && glyphs.filter((g) => g.x < mid && g.x + g.w > mid).length < glyphs.length * 0.15;\n\n // Group on y with a tolerance that scales with font size.\n const lines: LineDraft[] = [];\n for (const g of [...glyphs].sort((a, b) => a.y - b.y || a.x - b.x)) {\n // NOTE: start-x assignment can misfile a gutter-straddling glyph; an\n // overlap-based alternative was tried and reshuffled stray numerals with\n // no measured win, so start-x stays (src/geometry.ts's OCR path assigns by x-span overlap instead, where the mis-assignment is measured).\n const col = twoCol && g.x > mid ? 1 : 0;\n const last = lines.at(-1);\n if (last && last.col === col && Math.abs(last.y - g.y) < Math.max(2, g.h * 0.5)) {\n last.parts.push(g);\n last.y = (last.y + g.y) / 2;\n } else lines.push({ col, y: g.y, parts: [g] });\n }\n lines.sort((a, b) => a.col - b.col || a.y - b.y); // column, then down the page\n\n return lines\n .map((l): Line => {\n const parts = l.parts.sort((a, b) => a.x - b.x);\n let t = '';\n let prev: Glyph | null = null;\n for (const cur of parts) {\n // insert a space when the measured gap is wider than intra-word kerning\n if (prev && cur.x - (prev.x + prev.w) > prev.h * 0.12 && !/\\s$/.test(t) && !/^\\s/.test(cur.s)) t += ' ';\n t += cur.s;\n prev = cur;\n }\n // Line height = the char-weighted dominant height, not the max: a small\n // decorative large-face numeral must not promote a body line to a heading.\n const byH = new Map<number, number>();\n for (const p of parts) byH.set(p.h, (byH.get(p.h) || 0) + p.s.length);\n return {\n text: normalize(t).replace(/\\s+/g, ' ').trim(),\n x: parts[0].x,\n y: l.y,\n h: [...byH.entries()].sort((a, b) => b[1] - a[1])[0][0],\n col: l.col,\n };\n })\n .filter((l) => l.text);\n}\n"],"names":["FOLD","ʼ","foldTypographic","s","replace","c","normalize","toLines","glyphs","width","length","mid","twoCol","filter","g","x","w","lines","sort","a","b","y","col","last","at","Math","abs","max","h","parts","push","map","l","t","prev","cur","test","byH","Map","p","set","get","text","trim","entries"],"mappings":"AAAA,wEAAwE;AACxE,0KAA0K;AAG1K,MAAMA,OAA+B;IACnCC,GAAG;IACH,KAAK;IACL,KAAK;IACL,KAAK;IACL,KAAK;IACL,KAAK;AACP;AACA,0EAA0E;AAC1E,wHAAwH;AACxH,OAAO,MAAMC,kBAAkB,CAACC,IAAsBA,EAAEC,OAAO,CAAC,aAAa,CAACC,IAAML,IAAI,CAACK,EAAE,EAAE;AAC7F,MAAMC,YAAYJ;AAUlB,OAAO,SAASK,QAAQ,EAAEC,MAAM,EAAEC,KAAK,EAA2C;IAChF,IAAI,CAACD,OAAOE,MAAM,EAAE,OAAO,EAAE;IAE7B,MAAMC,MAAMF,QAAQ;IACpB,wFAAwF;IACxF,gGAAgG;IAChG,MAAMG,SAASJ,OAAOK,MAAM,CAAC,CAACC,IAAMA,EAAEC,CAAC,GAAGD,EAAEE,CAAC,GAAGL,KAAKD,MAAM,GAAGF,OAAOE,MAAM,GAAG,QAAQF,OAAOK,MAAM,CAAC,CAACC,IAAMA,EAAEC,CAAC,GAAGJ,KAAKD,MAAM,GAAGF,OAAOE,MAAM,GAAG,QAAQF,OAAOK,MAAM,CAAC,CAACC,IAAMA,EAAEC,CAAC,GAAGJ,OAAOG,EAAEC,CAAC,GAAGD,EAAEE,CAAC,GAAGL,KAAKD,MAAM,GAAGF,OAAOE,MAAM,GAAG;IAEnO,0DAA0D;IAC1D,MAAMO,QAAqB,EAAE;IAC7B,KAAK,MAAMH,KAAK;WAAIN;KAAO,CAACU,IAAI,CAAC,CAACC,GAAGC,IAAMD,EAAEE,CAAC,GAAGD,EAAEC,CAAC,IAAIF,EAAEJ,CAAC,GAAGK,EAAEL,CAAC,EAAG;QAClE,qEAAqE;QACrE,yEAAyE;QACzE,0IAA0I;QAC1I,MAAMO,MAAMV,UAAUE,EAAEC,CAAC,GAAGJ,MAAM,IAAI;QACtC,MAAMY,OAAON,MAAMO,EAAE,CAAC,CAAC;QACvB,IAAID,QAAQA,KAAKD,GAAG,KAAKA,OAAOG,KAAKC,GAAG,CAACH,KAAKF,CAAC,GAAGP,EAAEO,CAAC,IAAII,KAAKE,GAAG,CAAC,GAAGb,EAAEc,CAAC,GAAG,MAAM;YAC/EL,KAAKM,KAAK,CAACC,IAAI,CAAChB;YAChBS,KAAKF,CAAC,GAAG,AAACE,CAAAA,KAAKF,CAAC,GAAGP,EAAEO,CAAC,AAADA,IAAK;QAC5B,OAAOJ,MAAMa,IAAI,CAAC;YAAER;YAAKD,GAAGP,EAAEO,CAAC;YAAEQ,OAAO;gBAACf;aAAE;QAAC;IAC9C;IACAG,MAAMC,IAAI,CAAC,CAACC,GAAGC,IAAMD,EAAEG,GAAG,GAAGF,EAAEE,GAAG,IAAIH,EAAEE,CAAC,GAAGD,EAAEC,CAAC,GAAG,6BAA6B;IAE/E,OAAOJ,MACJc,GAAG,CAAC,CAACC;QACJ,MAAMH,QAAQG,EAAEH,KAAK,CAACX,IAAI,CAAC,CAACC,GAAGC,IAAMD,EAAEJ,CAAC,GAAGK,EAAEL,CAAC;QAC9C,IAAIkB,IAAI;QACR,IAAIC,OAAqB;QACzB,KAAK,MAAMC,OAAON,MAAO;YACvB,wEAAwE;YACxE,IAAIK,QAAQC,IAAIpB,CAAC,GAAImB,CAAAA,KAAKnB,CAAC,GAAGmB,KAAKlB,CAAC,AAADA,IAAKkB,KAAKN,CAAC,GAAG,QAAQ,CAAC,MAAMQ,IAAI,CAACH,MAAM,CAAC,MAAMG,IAAI,CAACD,IAAIhC,CAAC,GAAG8B,KAAK;YACpGA,KAAKE,IAAIhC,CAAC;YACV+B,OAAOC;QACT;QACA,wEAAwE;QACxE,2EAA2E;QAC3E,MAAME,MAAM,IAAIC;QAChB,KAAK,MAAMC,KAAKV,MAAOQ,IAAIG,GAAG,CAACD,EAAEX,CAAC,EAAE,AAACS,CAAAA,IAAII,GAAG,CAACF,EAAEX,CAAC,KAAK,CAAA,IAAKW,EAAEpC,CAAC,CAACO,MAAM;QACpE,OAAO;YACLgC,MAAMpC,UAAU2B,GAAG7B,OAAO,CAAC,QAAQ,KAAKuC,IAAI;YAC5C5B,GAAGc,KAAK,CAAC,EAAE,CAACd,CAAC;YACbM,GAAGW,EAAEX,CAAC;YACNO,GAAG;mBAAIS,IAAIO,OAAO;aAAG,CAAC1B,IAAI,CAAC,CAACC,GAAGC,IAAMA,CAAC,CAAC,EAAE,GAAGD,CAAC,CAAC,EAAE,CAAC,CAAC,EAAE,CAAC,EAAE;YACvDG,KAAKU,EAAEV,GAAG;QACZ;IACF,GACCT,MAAM,CAAC,CAACmB,IAAMA,EAAEU,IAAI;AACzB"}
1
+ {"version":3,"sources":["/Users/kevin/Dev/OpenSource/formats/pdf-to-md/src/lines.ts"],"sourcesContent":["// Pass 2: glyphs -> visual lines — de-braids two-column pages (spanning\n// lines like headings sort with the left column), joins words by measured horizontal gap, and folds typographic codepoints to ASCII so search doesn't miss half the hits.\nimport type { CollectedPage, Glyph, Line } from './types.ts';\n\nconst FOLD: Record<string, string> = {\n ʼ: \"'\",\n '’': \"'\",\n '‘': \"'\",\n '“': '\"',\n '”': '\"',\n '–': '-',\n};\n// Exported for the OCR engine seam (src/engines/tesseract.ts): both paths\n// must fold identically or the same printed apostrophe surfaces as two spellings depending on which path read the page.\nexport const foldTypographic = (s: string): string => s.replace(/[ʼ‘’“”–]/g, (c) => FOLD[c]);\nconst normalize = foldTypographic;\n\n// Intermediate row while grouping glyphs by visual line, before the final\n// text/height reduction below.\ninterface LineDraft {\n col: number;\n y: number;\n parts: Glyph[];\n}\n\nexport function toLines({ glyphs, width }: Pick<CollectedPage, 'glyphs' | 'width'>): Line[] {\n if (!glyphs.length) return [];\n\n const mid = width / 2;\n // A real two-column page has a gutter: clusters each side, almost nothing crossing mid.\n // Label/value layouts also cluster but their values cross mid — de-braiding would reorder them.\n const twoCol = glyphs.filter((g) => g.x + g.w < mid).length > glyphs.length * 0.25 && glyphs.filter((g) => g.x > mid).length > glyphs.length * 0.25 && glyphs.filter((g) => g.x < mid && g.x + g.w > mid).length < glyphs.length * 0.15;\n\n // Group on y with a tolerance that scales with font size.\n const lines: LineDraft[] = [];\n for (const g of [...glyphs].sort((a, b) => a.y - b.y || a.x - b.x)) {\n // NOTE: start-x assignment can misfile a gutter-straddling glyph; an\n // overlap-based alternative was tried and reshuffled stray numerals with\n // no measured win, so start-x stays (src/geometry.ts's OCR path assigns by x-span overlap instead, where the mis-assignment is measured).\n const col = twoCol && g.x > mid ? 1 : 0;\n const last = lines.at(-1);\n if (last && last.col === col && Math.abs(last.y - g.y) < Math.max(2, g.h * 0.5)) {\n last.parts.push(g);\n last.y = (last.y + g.y) / 2;\n } else lines.push({ col, y: g.y, parts: [g] });\n }\n lines.sort((a, b) => a.col - b.col || a.y - b.y); // column, then down the page\n\n return lines\n .map((l): Line => {\n const parts = l.parts.sort((a, b) => a.x - b.x);\n let t = '';\n let prev: Glyph | null = null;\n for (const cur of parts) {\n // insert a space when the measured gap is wider than intra-word kerning\n if (prev && cur.x - (prev.x + prev.w) > prev.h * 0.12 && !/\\s$/.test(t) && !/^\\s/.test(cur.s)) t += ' ';\n t += cur.s;\n prev = cur;\n }\n // Line height = the char-weighted dominant height, not the max: a small\n // decorative large-face numeral must not promote a body line to a heading.\n const byH = new Map<number, number>();\n for (const p of parts) byH.set(p.h, (byH.get(p.h) || 0) + p.s.length);\n return {\n text: normalize(t).replace(/\\s+/g, ' ').trim(),\n x: parts[0].x,\n y: l.y,\n h: [...byH.entries()].sort((a, b) => b[1] - a[1])[0][0],\n col: l.col,\n };\n })\n .filter((l) => l.text);\n}\n"],"names":["FOLD","ʼ","foldTypographic","s","replace","c","normalize","toLines","glyphs","width","length","mid","twoCol","filter","g","x","w","lines","sort","a","b","y","col","last","at","Math","abs","max","h","parts","push","map","l","t","prev","cur","test","byH","Map","p","set","get","text","trim","entries"],"mappings":"AAAA,wEAAwE;AACxE,0KAA0K;AAG1K,MAAMA,OAA+B;IACnCC,GAAG;IACH,KAAK;IACL,KAAK;IACL,KAAK;IACL,KAAK;IACL,KAAK;AACP;AACA,0EAA0E;AAC1E,wHAAwH;AACxH,OAAO,MAAMC,kBAAkB,CAACC,IAAsBA,EAAEC,OAAO,CAAC,aAAa,CAACC,IAAML,IAAI,CAACK,EAAE,EAAE;AAC7F,MAAMC,YAAYJ;AAUlB,OAAO,SAASK,QAAQ,EAAEC,MAAM,EAAEC,KAAK,EAA2C;IAChF,IAAI,CAACD,OAAOE,MAAM,EAAE,OAAO,EAAE;IAE7B,MAAMC,MAAMF,QAAQ;IACpB,wFAAwF;IACxF,gGAAgG;IAChG,MAAMG,SAASJ,OAAOK,MAAM,CAAC,CAACC,IAAMA,EAAEC,CAAC,GAAGD,EAAEE,CAAC,GAAGL,KAAKD,MAAM,GAAGF,OAAOE,MAAM,GAAG,QAAQF,OAAOK,MAAM,CAAC,CAACC,IAAMA,EAAEC,CAAC,GAAGJ,KAAKD,MAAM,GAAGF,OAAOE,MAAM,GAAG,QAAQF,OAAOK,MAAM,CAAC,CAACC,IAAMA,EAAEC,CAAC,GAAGJ,OAAOG,EAAEC,CAAC,GAAGD,EAAEE,CAAC,GAAGL,KAAKD,MAAM,GAAGF,OAAOE,MAAM,GAAG;IAEnO,0DAA0D;IAC1D,MAAMO,QAAqB,EAAE;IAC7B,KAAK,MAAMH,KAAK;WAAIN;KAAO,CAACU,IAAI,CAAC,CAACC,GAAGC,IAAMD,EAAEE,CAAC,GAAGD,EAAEC,CAAC,IAAIF,EAAEJ,CAAC,GAAGK,EAAEL,CAAC,EAAG;QAClE,qEAAqE;QACrE,yEAAyE;QACzE,0IAA0I;QAC1I,MAAMO,MAAMV,UAAUE,EAAEC,CAAC,GAAGJ,MAAM,IAAI;QACtC,MAAMY,OAAON,MAAMO,EAAE,CAAC,CAAC;QACvB,IAAID,QAAQA,KAAKD,GAAG,KAAKA,OAAOG,KAAKC,GAAG,CAACH,KAAKF,CAAC,GAAGP,EAAEO,CAAC,IAAII,KAAKE,GAAG,CAAC,GAAGb,EAAEc,CAAC,GAAG,MAAM;YAC/EL,KAAKM,KAAK,CAACC,IAAI,CAAChB;YAChBS,KAAKF,CAAC,GAAG,AAACE,CAAAA,KAAKF,CAAC,GAAGP,EAAEO,CAAC,AAADA,IAAK;QAC5B,OAAOJ,MAAMa,IAAI,CAAC;YAAER;YAAKD,GAAGP,EAAEO,CAAC;YAAEQ,OAAO;gBAACf;aAAE;QAAC;IAC9C;IACAG,MAAMC,IAAI,CAAC,CAACC,GAAGC,IAAMD,EAAEG,GAAG,GAAGF,EAAEE,GAAG,IAAIH,EAAEE,CAAC,GAAGD,EAAEC,CAAC,GAAG,6BAA6B;IAE/E,OAAOJ,MACJc,GAAG,CAAC,CAACC;QACJ,MAAMH,QAAQG,EAAEH,KAAK,CAACX,IAAI,CAAC,CAACC,GAAGC,IAAMD,EAAEJ,CAAC,GAAGK,EAAEL,CAAC;QAC9C,IAAIkB,IAAI;QACR,IAAIC,OAAqB;QACzB,KAAK,MAAMC,OAAON,MAAO;YACvB,wEAAwE;YACxE,IAAIK,QAAQC,IAAIpB,CAAC,GAAImB,CAAAA,KAAKnB,CAAC,GAAGmB,KAAKlB,CAAC,AAADA,IAAKkB,KAAKN,CAAC,GAAG,QAAQ,CAAC,MAAMQ,IAAI,CAACH,MAAM,CAAC,MAAMG,IAAI,CAACD,IAAIhC,CAAC,GAAG8B,KAAK;YACpGA,KAAKE,IAAIhC,CAAC;YACV+B,OAAOC;QACT;QACA,wEAAwE;QACxE,2EAA2E;QAC3E,MAAME,MAAM,IAAIC;QAChB,KAAK,MAAMC,KAAKV,MAAOQ,IAAIG,GAAG,CAACD,EAAEX,CAAC,EAAE,AAACS,CAAAA,IAAII,GAAG,CAACF,EAAEX,CAAC,KAAK,CAAA,IAAKW,EAAEpC,CAAC,CAACO,MAAM;QACpE,OAAO;YACLgC,MAAMpC,UAAU2B,GAAG7B,OAAO,CAAC,QAAQ,KAAKuC,IAAI;YAC5C5B,GAAGc,KAAK,CAAC,EAAE,CAACd,CAAC;YACbM,GAAGW,EAAEX,CAAC;YACNO,GAAG;mBAAIS,IAAIO,OAAO;aAAG,CAAC1B,IAAI,CAAC,CAACC,GAAGC,IAAMA,CAAC,CAAC,EAAE,GAAGD,CAAC,CAAC,EAAE,CAAC,CAAC,EAAE,CAAC,EAAE;YACvDG,KAAKU,EAAEV,GAAG;QACZ;IACF,GACCT,MAAM,CAAC,CAACmB,IAAMA,EAAEU,IAAI;AACzB"}
@@ -1 +1 @@
1
- {"version":3,"sources":["/Users/kevin/Projects/pdf-to-md/src/markdown.ts"],"sourcesContent":["// Pass 3: lines -> markdown. Structure is measured, not guessed — heading\n// levels come from the font-height histogram, hyphen-closes only against a dictionary (see the inline comments below for each rule).\nimport { getDict } from './dict.ts';\nimport { pageMarkerLine } from './page-marker.ts';\nimport type { BuildMarkdownOptions, Line, MarkdownResult, MarkdownStats, PageLines } from './types.ts';\n\nconst GLYPH = /^[•◦▪▫●○·‣⁃]$/;\nconst BULLETS = /^(?:[•◦▪▫●○·‣⁃]\\s*)+/;\n\nexport function buildMarkdown(pageLines: PageLines[], heightChars: Map<number, number>, opts: BuildMarkdownOptions = {}): MarkdownResult {\n // body height = char-weighted mode; headings are the rarer, larger sizes.\n // The 1.35 multiplier keeps emphasized body text out of the outline.\n const bodyH = [...heightChars.entries()].sort((a, b) => b[1] - a[1])[0]?.[0] ?? 0;\n const headingSizes = [...heightChars.keys()].filter((h) => h > bodyH * 1.35).sort((a, b) => b - a);\n // with page markers on, content headings start below the page ### level;\n // sizes beyond the ###### budget share level 6 rather than being dropped\n const offset = opts.pageMarkers ? 3 : 0;\n const levelOf = new Map(headingSizes.slice(0, 6).map((h, i) => [h, Math.min(6, i + 1 + offset)]));\n\n const out: string[] = [];\n const stats: MarkdownStats = {\n headings: 0,\n listItems: 0,\n paragraphs: 0,\n joinedPairs: 0,\n reflowed: 0,\n callouts: 0,\n closedHyphens: 0,\n };\n let para: string[] = [];\n\n // a label line ending in ':' whose next line is a bare number belongs on one line\n function attachValue(label: string, lines: Line[], get: () => number, set: (n: number) => void): string {\n const j = get() + 1;\n if (/[::]\\s*$/.test(label) && j < lines.length && /^[0-9][\\d.,]*%?$/.test(lines[j].text)) {\n set(j);\n stats.joinedPairs++;\n return `${label} ${lines[j].text}`;\n }\n return label;\n }\n const flush = () => {\n if (!para.length) return;\n out.push(para.join(' ').replace(/\\s+([,.;:%])/g, '$1'));\n out.push('');\n stats.paragraphs++;\n para = [];\n };\n // append a wrapped line to the open paragraph, resolving a line-final hyphen\n const appendToPara = (text: string) => {\n const prev = para.at(-1);\n if (prev && /\\w-$/.test(prev)) {\n // a bare \" -\" is punctuation, not a wrapped word\n const head = prev.replace(/-$/, '').split(/\\s/).pop() as string;\n const tail = (text.split(/\\s/)[0] ?? '').replace(/[^A-Za-z-]/g, '').split('-')[0];\n const close = getDict().has((head + tail).toLowerCase());\n if (close) stats.closedHyphens++;\n para[para.length - 1] = prev.replace(/-$/, close ? '' : '-') + text;\n } else para.push(text);\n };\n\n for (const { page, lines } of pageLines) {\n if (opts.pageMarkers) {\n flush();\n out.push(pageMarkerLine(page));\n out.push('');\n }\n if (!lines.length) continue;\n const indentBase = Math.min(...lines.map((l) => l.x));\n\n for (let i = 0; i < lines.length; i++) {\n let { text, h, x } = lines[i];\n const prevLine = i > 0 ? lines[i - 1] : null;\n // a column change or a vertical jump wider than a blank line ends the paragraph\n if (para.length && prevLine && (lines[i].col !== prevLine.col || lines[i].y - prevLine.y > h * 2.2)) flush();\n\n // A large-font line opens a run: gather the consecutive same-size lines\n // below it (a title wraps across lines). A short merged run is one heading; a long one is a callout/pull-quote in a bigger face — prose, not structure — and becomes a blockquote.\n if (levelOf.has(h) && !BULLETS.test(text)) {\n const run = [text];\n while (i + 1 < lines.length && lines[i + 1].h === h && lines[i + 1].col === lines[i].col && lines[i + 1].y - lines[i].y < h * 2.2 && !BULLETS.test(lines[i + 1].text)) run.push(lines[++i].text);\n // join wrapped title lines, resolving line-final hyphens like body text\n const merged = run.reduce((acc, part) => (/\\w-$/.test(acc) ? acc.replace(/-$/, getDict().has(((acc.replace(/-$/, '').split(/\\s/).pop() as string) + (part.split(/\\s/)[0] ?? '').replace(/[^A-Za-z-]/g, '').split('-')[0]).toLowerCase()) ? '' : '-') + part : `${acc} ${part}`)).replace(/\\s+([,.;:%])/g, '$1');\n flush();\n if (merged.length > 150) {\n out.push(`> ${merged}`);\n out.push('');\n stats.callouts++;\n } else {\n out.push(`${'#'.repeat(levelOf.get(h) as number)} ${merged}`);\n out.push('');\n stats.headings++;\n }\n continue;\n }\n\n // Lines can carry several stacked glyphs (\"• ▪\") from nested source lists.\n const stripped = text.replace(BULLETS, '').trim();\n const wasBullet = stripped !== text;\n if (wasBullet) text = stripped;\n\n // bullet with nothing left on the line: attach the following line's text\n if ((!text || GLYPH.test(text)) && i + 1 < lines.length) {\n flush();\n const depth = Math.max(0, Math.round((x - indentBase) / 18));\n let label = lines[++i].text.replace(BULLETS, '').trim();\n while (!label && i + 1 < lines.length)\n // stacked glyphs\n label = lines[++i].text.replace(BULLETS, '').trim();\n label = attachValue(\n label,\n lines,\n () => i,\n (n) => {\n i = n;\n }\n );\n out.push(`${' '.repeat(depth)}- ${label}`);\n stats.listItems++;\n continue;\n }\n // glyph(s) followed by text on the same line\n if (wasBullet) {\n flush();\n const depth = Math.max(0, Math.round((x - indentBase) / 18));\n out.push(\n `${' '.repeat(depth)}- ${attachValue(\n text,\n lines,\n () => i,\n (n) => {\n i = n;\n }\n )}`\n );\n stats.listItems++;\n continue;\n }\n // `label:` on its own line followed by a bare value -> keep together\n if (/[::]\\s*$/.test(text) && i + 1 < lines.length && /^[0-9][\\d.,]*%?$/.test(lines[i + 1].text)) {\n flush();\n out.push(`- ${text} ${lines[++i].text}`);\n stats.joinedPairs++;\n continue;\n }\n // otherwise body text: re-flow wrapped lines into a paragraph\n appendToPara(text);\n if (/[.!?:]$/.test(text)) flush();\n else stats.reflowed++;\n }\n flush();\n }\n flush();\n\n const md = `${out\n .join('\\n')\n .replace(/\\n{3,}/g, '\\n\\n')\n .trim()}\\n`;\n return { md, stats, bodyH, headingSizes };\n}\n"],"names":["getDict","pageMarkerLine","GLYPH","BULLETS","buildMarkdown","pageLines","heightChars","opts","bodyH","entries","sort","a","b","headingSizes","keys","filter","h","offset","pageMarkers","levelOf","Map","slice","map","i","Math","min","out","stats","headings","listItems","paragraphs","joinedPairs","reflowed","callouts","closedHyphens","para","attachValue","label","lines","get","set","j","test","length","text","flush","push","join","replace","appendToPara","prev","at","head","split","pop","tail","close","has","toLowerCase","page","indentBase","l","x","prevLine","col","y","run","merged","reduce","acc","part","repeat","stripped","trim","wasBullet","depth","max","round","n","md"],"mappings":"AAAA,0EAA0E;AAC1E,qIAAqI;AACrI,SAASA,OAAO,QAAQ,YAAY;AACpC,SAASC,cAAc,QAAQ,mBAAmB;AAGlD,MAAMC,QAAQ;AACd,MAAMC,UAAU;AAEhB,OAAO,SAASC,cAAcC,SAAsB,EAAEC,WAAgC,EAAEC,OAA6B,CAAC,CAAC;;QAGvG;IAFd,0EAA0E;IAC1E,qEAAqE;IACrE,MAAMC,iBAAQ,SAAA;WAAIF,YAAYG,OAAO;KAAG,CAACC,IAAI,CAAC,CAACC,GAAGC,IAAMA,CAAC,CAAC,EAAE,GAAGD,CAAC,CAAC,EAAE,CAAC,CAAC,EAAE,cAAzD,6BAAA,MAA2D,CAAC,EAAE,uCAAI;IAChF,MAAME,eAAe;WAAIP,YAAYQ,IAAI;KAAG,CAACC,MAAM,CAAC,CAACC,IAAMA,IAAIR,QAAQ,MAAME,IAAI,CAAC,CAACC,GAAGC,IAAMA,IAAID;IAChG,yEAAyE;IACzE,yEAAyE;IACzE,MAAMM,SAASV,KAAKW,WAAW,GAAG,IAAI;IACtC,MAAMC,UAAU,IAAIC,IAAIP,aAAaQ,KAAK,CAAC,GAAG,GAAGC,GAAG,CAAC,CAACN,GAAGO,IAAM;YAACP;YAAGQ,KAAKC,GAAG,CAAC,GAAGF,IAAI,IAAIN;SAAQ;IAE/F,MAAMS,MAAgB,EAAE;IACxB,MAAMC,QAAuB;QAC3BC,UAAU;QACVC,WAAW;QACXC,YAAY;QACZC,aAAa;QACbC,UAAU;QACVC,UAAU;QACVC,eAAe;IACjB;IACA,IAAIC,OAAiB,EAAE;IAEvB,kFAAkF;IAClF,SAASC,YAAYC,KAAa,EAAEC,KAAa,EAAEC,GAAiB,EAAEC,GAAwB;QAC5F,MAAMC,IAAIF,QAAQ;QAClB,IAAI,WAAWG,IAAI,CAACL,UAAUI,IAAIH,MAAMK,MAAM,IAAI,mBAAmBD,IAAI,CAACJ,KAAK,CAACG,EAAE,CAACG,IAAI,GAAG;YACxFJ,IAAIC;YACJd,MAAMI,WAAW;YACjB,OAAO,GAAGM,MAAM,CAAC,EAAEC,KAAK,CAACG,EAAE,CAACG,IAAI,EAAE;QACpC;QACA,OAAOP;IACT;IACA,MAAMQ,QAAQ;QACZ,IAAI,CAACV,KAAKQ,MAAM,EAAE;QAClBjB,IAAIoB,IAAI,CAACX,KAAKY,IAAI,CAAC,KAAKC,OAAO,CAAC,iBAAiB;QACjDtB,IAAIoB,IAAI,CAAC;QACTnB,MAAMG,UAAU;QAChBK,OAAO,EAAE;IACX;IACA,6EAA6E;IAC7E,MAAMc,eAAe,CAACL;QACpB,MAAMM,OAAOf,KAAKgB,EAAE,CAAC,CAAC;QACtB,IAAID,QAAQ,OAAOR,IAAI,CAACQ,OAAO;gBAGfN;YAFd,iDAAiD;YACjD,MAAMQ,OAAOF,KAAKF,OAAO,CAAC,MAAM,IAAIK,KAAK,CAAC,MAAMC,GAAG;YACnD,MAAMC,OAAO,EAACX,eAAAA,KAAKS,KAAK,CAAC,KAAK,CAAC,EAAE,cAAnBT,0BAAAA,eAAuB,IAAII,OAAO,CAAC,eAAe,IAAIK,KAAK,CAAC,IAAI,CAAC,EAAE;YACjF,MAAMG,QAAQxD,UAAUyD,GAAG,CAAC,AAACL,CAAAA,OAAOG,IAAG,EAAGG,WAAW;YACrD,IAAIF,OAAO7B,MAAMO,aAAa;YAC9BC,IAAI,CAACA,KAAKQ,MAAM,GAAG,EAAE,GAAGO,KAAKF,OAAO,CAAC,MAAMQ,QAAQ,KAAK,OAAOZ;QACjE,OAAOT,KAAKW,IAAI,CAACF;IACnB;IAEA,KAAK,MAAM,EAAEe,IAAI,EAAErB,KAAK,EAAE,IAAIjC,UAAW;QACvC,IAAIE,KAAKW,WAAW,EAAE;YACpB2B;YACAnB,IAAIoB,IAAI,CAAC7C,eAAe0D;YACxBjC,IAAIoB,IAAI,CAAC;QACX;QACA,IAAI,CAACR,MAAMK,MAAM,EAAE;QACnB,MAAMiB,aAAapC,KAAKC,GAAG,IAAIa,MAAMhB,GAAG,CAAC,CAACuC,IAAMA,EAAEC,CAAC;QAEnD,IAAK,IAAIvC,IAAI,GAAGA,IAAIe,MAAMK,MAAM,EAAEpB,IAAK;YACrC,IAAI,EAAEqB,IAAI,EAAE5B,CAAC,EAAE8C,CAAC,EAAE,GAAGxB,KAAK,CAACf,EAAE;YAC7B,MAAMwC,WAAWxC,IAAI,IAAIe,KAAK,CAACf,IAAI,EAAE,GAAG;YACxC,gFAAgF;YAChF,IAAIY,KAAKQ,MAAM,IAAIoB,YAAazB,CAAAA,KAAK,CAACf,EAAE,CAACyC,GAAG,KAAKD,SAASC,GAAG,IAAI1B,KAAK,CAACf,EAAE,CAAC0C,CAAC,GAAGF,SAASE,CAAC,GAAGjD,IAAI,GAAE,GAAI6B;YAErG,wEAAwE;YACxE,mLAAmL;YACnL,IAAI1B,QAAQsC,GAAG,CAACzC,MAAM,CAACb,QAAQuC,IAAI,CAACE,OAAO;gBACzC,MAAMsB,MAAM;oBAACtB;iBAAK;gBAClB,MAAOrB,IAAI,IAAIe,MAAMK,MAAM,IAAIL,KAAK,CAACf,IAAI,EAAE,CAACP,CAAC,KAAKA,KAAKsB,KAAK,CAACf,IAAI,EAAE,CAACyC,GAAG,KAAK1B,KAAK,CAACf,EAAE,CAACyC,GAAG,IAAI1B,KAAK,CAACf,IAAI,EAAE,CAAC0C,CAAC,GAAG3B,KAAK,CAACf,EAAE,CAAC0C,CAAC,GAAGjD,IAAI,OAAO,CAACb,QAAQuC,IAAI,CAACJ,KAAK,CAACf,IAAI,EAAE,CAACqB,IAAI,EAAGsB,IAAIpB,IAAI,CAACR,KAAK,CAAC,EAAEf,EAAE,CAACqB,IAAI;gBAC/L,wEAAwE;gBACxE,MAAMuB,SAASD,IAAIE,MAAM,CAAC,CAACC,KAAKC;wBAAqHA;2BAA3G,OAAO5B,IAAI,CAAC2B,OAAOA,IAAIrB,OAAO,CAAC,MAAMhD,UAAUyD,GAAG,CAAC,AAAC,CAAA,AAACY,IAAIrB,OAAO,CAAC,MAAM,IAAIK,KAAK,CAAC,MAAMC,GAAG,KAAgB,EAACgB,eAAAA,KAAKjB,KAAK,CAAC,KAAK,CAAC,EAAE,cAAnBiB,0BAAAA,eAAuB,IAAItB,OAAO,CAAC,eAAe,IAAIK,KAAK,CAAC,IAAI,CAAC,EAAE,AAAD,EAAGK,WAAW,MAAM,KAAK,OAAOY,OAAO,GAAGD,IAAI,CAAC,EAAEC,MAAM;mBAAGtB,OAAO,CAAC,iBAAiB;gBAC1SH;gBACA,IAAIsB,OAAOxB,MAAM,GAAG,KAAK;oBACvBjB,IAAIoB,IAAI,CAAC,CAAC,EAAE,EAAEqB,QAAQ;oBACtBzC,IAAIoB,IAAI,CAAC;oBACTnB,MAAMM,QAAQ;gBAChB,OAAO;oBACLP,IAAIoB,IAAI,CAAC,GAAG,IAAIyB,MAAM,CAACpD,QAAQoB,GAAG,CAACvB,IAAc,CAAC,EAAEmD,QAAQ;oBAC5DzC,IAAIoB,IAAI,CAAC;oBACTnB,MAAMC,QAAQ;gBAChB;gBACA;YACF;YAEA,2EAA2E;YAC3E,MAAM4C,WAAW5B,KAAKI,OAAO,CAAC7C,SAAS,IAAIsE,IAAI;YAC/C,MAAMC,YAAYF,aAAa5B;YAC/B,IAAI8B,WAAW9B,OAAO4B;YAEtB,yEAAyE;YACzE,IAAI,AAAC,CAAA,CAAC5B,QAAQ1C,MAAMwC,IAAI,CAACE,KAAI,KAAMrB,IAAI,IAAIe,MAAMK,MAAM,EAAE;gBACvDE;gBACA,MAAM8B,QAAQnD,KAAKoD,GAAG,CAAC,GAAGpD,KAAKqD,KAAK,CAAC,AAACf,CAAAA,IAAIF,UAAS,IAAK;gBACxD,IAAIvB,QAAQC,KAAK,CAAC,EAAEf,EAAE,CAACqB,IAAI,CAACI,OAAO,CAAC7C,SAAS,IAAIsE,IAAI;gBACrD,MAAO,CAACpC,SAASd,IAAI,IAAIe,MAAMK,MAAM,CACnC,iBAAiB;gBACjBN,QAAQC,KAAK,CAAC,EAAEf,EAAE,CAACqB,IAAI,CAACI,OAAO,CAAC7C,SAAS,IAAIsE,IAAI;gBACnDpC,QAAQD,YACNC,OACAC,OACA,IAAMf,GACN,CAACuD;oBACCvD,IAAIuD;gBACN;gBAEFpD,IAAIoB,IAAI,CAAC,GAAG,KAAKyB,MAAM,CAACI,OAAO,EAAE,EAAEtC,OAAO;gBAC1CV,MAAME,SAAS;gBACf;YACF;YACA,6CAA6C;YAC7C,IAAI6C,WAAW;gBACb7B;gBACA,MAAM8B,QAAQnD,KAAKoD,GAAG,CAAC,GAAGpD,KAAKqD,KAAK,CAAC,AAACf,CAAAA,IAAIF,UAAS,IAAK;gBACxDlC,IAAIoB,IAAI,CACN,GAAG,KAAKyB,MAAM,CAACI,OAAO,EAAE,EAAEvC,YACxBQ,MACAN,OACA,IAAMf,GACN,CAACuD;oBACCvD,IAAIuD;gBACN,IACC;gBAELnD,MAAME,SAAS;gBACf;YACF;YACA,qEAAqE;YACrE,IAAI,WAAWa,IAAI,CAACE,SAASrB,IAAI,IAAIe,MAAMK,MAAM,IAAI,mBAAmBD,IAAI,CAACJ,KAAK,CAACf,IAAI,EAAE,CAACqB,IAAI,GAAG;gBAC/FC;gBACAnB,IAAIoB,IAAI,CAAC,CAAC,EAAE,EAAEF,KAAK,CAAC,EAAEN,KAAK,CAAC,EAAEf,EAAE,CAACqB,IAAI,EAAE;gBACvCjB,MAAMI,WAAW;gBACjB;YACF;YACA,8DAA8D;YAC9DkB,aAAaL;YACb,IAAI,UAAUF,IAAI,CAACE,OAAOC;iBACrBlB,MAAMK,QAAQ;QACrB;QACAa;IACF;IACAA;IAEA,MAAMkC,KAAK,GAAGrD,IACXqB,IAAI,CAAC,MACLC,OAAO,CAAC,WAAW,QACnByB,IAAI,GAAG,EAAE,CAAC;IACb,OAAO;QAAEM;QAAIpD;QAAOnB;QAAOK;IAAa;AAC1C"}
1
+ {"version":3,"sources":["/Users/kevin/Dev/OpenSource/formats/pdf-to-md/src/markdown.ts"],"sourcesContent":["// Pass 3: lines -> markdown. Structure is measured, not guessed — heading\n// levels come from the font-height histogram, hyphen-closes only against a dictionary (see the inline comments below for each rule).\nimport { getDict } from './dict.ts';\nimport { pageMarkerLine } from './page-marker.ts';\nimport type { BuildMarkdownOptions, Line, MarkdownResult, MarkdownStats, PageLines } from './types.ts';\n\nconst GLYPH = /^[•◦▪▫●○·‣⁃]$/;\nconst BULLETS = /^(?:[•◦▪▫●○·‣⁃]\\s*)+/;\n\nexport function buildMarkdown(pageLines: PageLines[], heightChars: Map<number, number>, opts: BuildMarkdownOptions = {}): MarkdownResult {\n // body height = char-weighted mode; headings are the rarer, larger sizes.\n // The 1.35 multiplier keeps emphasized body text out of the outline.\n const bodyH = [...heightChars.entries()].sort((a, b) => b[1] - a[1])[0]?.[0] ?? 0;\n const headingSizes = [...heightChars.keys()].filter((h) => h > bodyH * 1.35).sort((a, b) => b - a);\n // with page markers on, content headings start below the page ### level;\n // sizes beyond the ###### budget share level 6 rather than being dropped\n const offset = opts.pageMarkers ? 3 : 0;\n const levelOf = new Map(headingSizes.slice(0, 6).map((h, i) => [h, Math.min(6, i + 1 + offset)]));\n\n const out: string[] = [];\n const stats: MarkdownStats = {\n headings: 0,\n listItems: 0,\n paragraphs: 0,\n joinedPairs: 0,\n reflowed: 0,\n callouts: 0,\n closedHyphens: 0,\n };\n let para: string[] = [];\n\n // a label line ending in ':' whose next line is a bare number belongs on one line\n function attachValue(label: string, lines: Line[], get: () => number, set: (n: number) => void): string {\n const j = get() + 1;\n if (/[::]\\s*$/.test(label) && j < lines.length && /^[0-9][\\d.,]*%?$/.test(lines[j].text)) {\n set(j);\n stats.joinedPairs++;\n return `${label} ${lines[j].text}`;\n }\n return label;\n }\n const flush = () => {\n if (!para.length) return;\n out.push(para.join(' ').replace(/\\s+([,.;:%])/g, '$1'));\n out.push('');\n stats.paragraphs++;\n para = [];\n };\n // append a wrapped line to the open paragraph, resolving a line-final hyphen\n const appendToPara = (text: string) => {\n const prev = para.at(-1);\n if (prev && /\\w-$/.test(prev)) {\n // a bare \" -\" is punctuation, not a wrapped word\n const head = prev.replace(/-$/, '').split(/\\s/).pop() as string;\n const tail = (text.split(/\\s/)[0] ?? '').replace(/[^A-Za-z-]/g, '').split('-')[0];\n const close = getDict().has((head + tail).toLowerCase());\n if (close) stats.closedHyphens++;\n para[para.length - 1] = prev.replace(/-$/, close ? '' : '-') + text;\n } else para.push(text);\n };\n\n for (const { page, lines } of pageLines) {\n if (opts.pageMarkers) {\n flush();\n out.push(pageMarkerLine(page));\n out.push('');\n }\n if (!lines.length) continue;\n const indentBase = Math.min(...lines.map((l) => l.x));\n\n for (let i = 0; i < lines.length; i++) {\n let { text, h, x } = lines[i];\n const prevLine = i > 0 ? lines[i - 1] : null;\n // a column change or a vertical jump wider than a blank line ends the paragraph\n if (para.length && prevLine && (lines[i].col !== prevLine.col || lines[i].y - prevLine.y > h * 2.2)) flush();\n\n // A large-font line opens a run: gather the consecutive same-size lines\n // below it (a title wraps across lines). A short merged run is one heading; a long one is a callout/pull-quote in a bigger face — prose, not structure — and becomes a blockquote.\n if (levelOf.has(h) && !BULLETS.test(text)) {\n const run = [text];\n while (i + 1 < lines.length && lines[i + 1].h === h && lines[i + 1].col === lines[i].col && lines[i + 1].y - lines[i].y < h * 2.2 && !BULLETS.test(lines[i + 1].text)) run.push(lines[++i].text);\n // join wrapped title lines, resolving line-final hyphens like body text\n const merged = run.reduce((acc, part) => (/\\w-$/.test(acc) ? acc.replace(/-$/, getDict().has(((acc.replace(/-$/, '').split(/\\s/).pop() as string) + (part.split(/\\s/)[0] ?? '').replace(/[^A-Za-z-]/g, '').split('-')[0]).toLowerCase()) ? '' : '-') + part : `${acc} ${part}`)).replace(/\\s+([,.;:%])/g, '$1');\n flush();\n if (merged.length > 150) {\n out.push(`> ${merged}`);\n out.push('');\n stats.callouts++;\n } else {\n out.push(`${'#'.repeat(levelOf.get(h) as number)} ${merged}`);\n out.push('');\n stats.headings++;\n }\n continue;\n }\n\n // Lines can carry several stacked glyphs (\"• ▪\") from nested source lists.\n const stripped = text.replace(BULLETS, '').trim();\n const wasBullet = stripped !== text;\n if (wasBullet) text = stripped;\n\n // bullet with nothing left on the line: attach the following line's text\n if ((!text || GLYPH.test(text)) && i + 1 < lines.length) {\n flush();\n const depth = Math.max(0, Math.round((x - indentBase) / 18));\n let label = lines[++i].text.replace(BULLETS, '').trim();\n while (!label && i + 1 < lines.length)\n // stacked glyphs\n label = lines[++i].text.replace(BULLETS, '').trim();\n label = attachValue(\n label,\n lines,\n () => i,\n (n) => {\n i = n;\n }\n );\n out.push(`${' '.repeat(depth)}- ${label}`);\n stats.listItems++;\n continue;\n }\n // glyph(s) followed by text on the same line\n if (wasBullet) {\n flush();\n const depth = Math.max(0, Math.round((x - indentBase) / 18));\n out.push(\n `${' '.repeat(depth)}- ${attachValue(\n text,\n lines,\n () => i,\n (n) => {\n i = n;\n }\n )}`\n );\n stats.listItems++;\n continue;\n }\n // `label:` on its own line followed by a bare value -> keep together\n if (/[::]\\s*$/.test(text) && i + 1 < lines.length && /^[0-9][\\d.,]*%?$/.test(lines[i + 1].text)) {\n flush();\n out.push(`- ${text} ${lines[++i].text}`);\n stats.joinedPairs++;\n continue;\n }\n // otherwise body text: re-flow wrapped lines into a paragraph\n appendToPara(text);\n if (/[.!?:]$/.test(text)) flush();\n else stats.reflowed++;\n }\n flush();\n }\n flush();\n\n const md = `${out\n .join('\\n')\n .replace(/\\n{3,}/g, '\\n\\n')\n .trim()}\\n`;\n return { md, stats, bodyH, headingSizes };\n}\n"],"names":["getDict","pageMarkerLine","GLYPH","BULLETS","buildMarkdown","pageLines","heightChars","opts","bodyH","entries","sort","a","b","headingSizes","keys","filter","h","offset","pageMarkers","levelOf","Map","slice","map","i","Math","min","out","stats","headings","listItems","paragraphs","joinedPairs","reflowed","callouts","closedHyphens","para","attachValue","label","lines","get","set","j","test","length","text","flush","push","join","replace","appendToPara","prev","at","head","split","pop","tail","close","has","toLowerCase","page","indentBase","l","x","prevLine","col","y","run","merged","reduce","acc","part","repeat","stripped","trim","wasBullet","depth","max","round","n","md"],"mappings":"AAAA,0EAA0E;AAC1E,qIAAqI;AACrI,SAASA,OAAO,QAAQ,YAAY;AACpC,SAASC,cAAc,QAAQ,mBAAmB;AAGlD,MAAMC,QAAQ;AACd,MAAMC,UAAU;AAEhB,OAAO,SAASC,cAAcC,SAAsB,EAAEC,WAAgC,EAAEC,OAA6B,CAAC,CAAC;;QAGvG;IAFd,0EAA0E;IAC1E,qEAAqE;IACrE,MAAMC,iBAAQ,SAAA;WAAIF,YAAYG,OAAO;KAAG,CAACC,IAAI,CAAC,CAACC,GAAGC,IAAMA,CAAC,CAAC,EAAE,GAAGD,CAAC,CAAC,EAAE,CAAC,CAAC,EAAE,cAAzD,6BAAA,MAA2D,CAAC,EAAE,uCAAI;IAChF,MAAME,eAAe;WAAIP,YAAYQ,IAAI;KAAG,CAACC,MAAM,CAAC,CAACC,IAAMA,IAAIR,QAAQ,MAAME,IAAI,CAAC,CAACC,GAAGC,IAAMA,IAAID;IAChG,yEAAyE;IACzE,yEAAyE;IACzE,MAAMM,SAASV,KAAKW,WAAW,GAAG,IAAI;IACtC,MAAMC,UAAU,IAAIC,IAAIP,aAAaQ,KAAK,CAAC,GAAG,GAAGC,GAAG,CAAC,CAACN,GAAGO,IAAM;YAACP;YAAGQ,KAAKC,GAAG,CAAC,GAAGF,IAAI,IAAIN;SAAQ;IAE/F,MAAMS,MAAgB,EAAE;IACxB,MAAMC,QAAuB;QAC3BC,UAAU;QACVC,WAAW;QACXC,YAAY;QACZC,aAAa;QACbC,UAAU;QACVC,UAAU;QACVC,eAAe;IACjB;IACA,IAAIC,OAAiB,EAAE;IAEvB,kFAAkF;IAClF,SAASC,YAAYC,KAAa,EAAEC,KAAa,EAAEC,GAAiB,EAAEC,GAAwB;QAC5F,MAAMC,IAAIF,QAAQ;QAClB,IAAI,WAAWG,IAAI,CAACL,UAAUI,IAAIH,MAAMK,MAAM,IAAI,mBAAmBD,IAAI,CAACJ,KAAK,CAACG,EAAE,CAACG,IAAI,GAAG;YACxFJ,IAAIC;YACJd,MAAMI,WAAW;YACjB,OAAO,GAAGM,MAAM,CAAC,EAAEC,KAAK,CAACG,EAAE,CAACG,IAAI,EAAE;QACpC;QACA,OAAOP;IACT;IACA,MAAMQ,QAAQ;QACZ,IAAI,CAACV,KAAKQ,MAAM,EAAE;QAClBjB,IAAIoB,IAAI,CAACX,KAAKY,IAAI,CAAC,KAAKC,OAAO,CAAC,iBAAiB;QACjDtB,IAAIoB,IAAI,CAAC;QACTnB,MAAMG,UAAU;QAChBK,OAAO,EAAE;IACX;IACA,6EAA6E;IAC7E,MAAMc,eAAe,CAACL;QACpB,MAAMM,OAAOf,KAAKgB,EAAE,CAAC,CAAC;QACtB,IAAID,QAAQ,OAAOR,IAAI,CAACQ,OAAO;gBAGfN;YAFd,iDAAiD;YACjD,MAAMQ,OAAOF,KAAKF,OAAO,CAAC,MAAM,IAAIK,KAAK,CAAC,MAAMC,GAAG;YACnD,MAAMC,OAAO,EAACX,eAAAA,KAAKS,KAAK,CAAC,KAAK,CAAC,EAAE,cAAnBT,0BAAAA,eAAuB,IAAII,OAAO,CAAC,eAAe,IAAIK,KAAK,CAAC,IAAI,CAAC,EAAE;YACjF,MAAMG,QAAQxD,UAAUyD,GAAG,CAAC,AAACL,CAAAA,OAAOG,IAAG,EAAGG,WAAW;YACrD,IAAIF,OAAO7B,MAAMO,aAAa;YAC9BC,IAAI,CAACA,KAAKQ,MAAM,GAAG,EAAE,GAAGO,KAAKF,OAAO,CAAC,MAAMQ,QAAQ,KAAK,OAAOZ;QACjE,OAAOT,KAAKW,IAAI,CAACF;IACnB;IAEA,KAAK,MAAM,EAAEe,IAAI,EAAErB,KAAK,EAAE,IAAIjC,UAAW;QACvC,IAAIE,KAAKW,WAAW,EAAE;YACpB2B;YACAnB,IAAIoB,IAAI,CAAC7C,eAAe0D;YACxBjC,IAAIoB,IAAI,CAAC;QACX;QACA,IAAI,CAACR,MAAMK,MAAM,EAAE;QACnB,MAAMiB,aAAapC,KAAKC,GAAG,IAAIa,MAAMhB,GAAG,CAAC,CAACuC,IAAMA,EAAEC,CAAC;QAEnD,IAAK,IAAIvC,IAAI,GAAGA,IAAIe,MAAMK,MAAM,EAAEpB,IAAK;YACrC,IAAI,EAAEqB,IAAI,EAAE5B,CAAC,EAAE8C,CAAC,EAAE,GAAGxB,KAAK,CAACf,EAAE;YAC7B,MAAMwC,WAAWxC,IAAI,IAAIe,KAAK,CAACf,IAAI,EAAE,GAAG;YACxC,gFAAgF;YAChF,IAAIY,KAAKQ,MAAM,IAAIoB,YAAazB,CAAAA,KAAK,CAACf,EAAE,CAACyC,GAAG,KAAKD,SAASC,GAAG,IAAI1B,KAAK,CAACf,EAAE,CAAC0C,CAAC,GAAGF,SAASE,CAAC,GAAGjD,IAAI,GAAE,GAAI6B;YAErG,wEAAwE;YACxE,mLAAmL;YACnL,IAAI1B,QAAQsC,GAAG,CAACzC,MAAM,CAACb,QAAQuC,IAAI,CAACE,OAAO;gBACzC,MAAMsB,MAAM;oBAACtB;iBAAK;gBAClB,MAAOrB,IAAI,IAAIe,MAAMK,MAAM,IAAIL,KAAK,CAACf,IAAI,EAAE,CAACP,CAAC,KAAKA,KAAKsB,KAAK,CAACf,IAAI,EAAE,CAACyC,GAAG,KAAK1B,KAAK,CAACf,EAAE,CAACyC,GAAG,IAAI1B,KAAK,CAACf,IAAI,EAAE,CAAC0C,CAAC,GAAG3B,KAAK,CAACf,EAAE,CAAC0C,CAAC,GAAGjD,IAAI,OAAO,CAACb,QAAQuC,IAAI,CAACJ,KAAK,CAACf,IAAI,EAAE,CAACqB,IAAI,EAAGsB,IAAIpB,IAAI,CAACR,KAAK,CAAC,EAAEf,EAAE,CAACqB,IAAI;gBAC/L,wEAAwE;gBACxE,MAAMuB,SAASD,IAAIE,MAAM,CAAC,CAACC,KAAKC;wBAAqHA;2BAA3G,OAAO5B,IAAI,CAAC2B,OAAOA,IAAIrB,OAAO,CAAC,MAAMhD,UAAUyD,GAAG,CAAC,AAAC,CAAA,AAACY,IAAIrB,OAAO,CAAC,MAAM,IAAIK,KAAK,CAAC,MAAMC,GAAG,KAAgB,EAACgB,eAAAA,KAAKjB,KAAK,CAAC,KAAK,CAAC,EAAE,cAAnBiB,0BAAAA,eAAuB,IAAItB,OAAO,CAAC,eAAe,IAAIK,KAAK,CAAC,IAAI,CAAC,EAAE,AAAD,EAAGK,WAAW,MAAM,KAAK,OAAOY,OAAO,GAAGD,IAAI,CAAC,EAAEC,MAAM;mBAAGtB,OAAO,CAAC,iBAAiB;gBAC1SH;gBACA,IAAIsB,OAAOxB,MAAM,GAAG,KAAK;oBACvBjB,IAAIoB,IAAI,CAAC,CAAC,EAAE,EAAEqB,QAAQ;oBACtBzC,IAAIoB,IAAI,CAAC;oBACTnB,MAAMM,QAAQ;gBAChB,OAAO;oBACLP,IAAIoB,IAAI,CAAC,GAAG,IAAIyB,MAAM,CAACpD,QAAQoB,GAAG,CAACvB,IAAc,CAAC,EAAEmD,QAAQ;oBAC5DzC,IAAIoB,IAAI,CAAC;oBACTnB,MAAMC,QAAQ;gBAChB;gBACA;YACF;YAEA,2EAA2E;YAC3E,MAAM4C,WAAW5B,KAAKI,OAAO,CAAC7C,SAAS,IAAIsE,IAAI;YAC/C,MAAMC,YAAYF,aAAa5B;YAC/B,IAAI8B,WAAW9B,OAAO4B;YAEtB,yEAAyE;YACzE,IAAI,AAAC,CAAA,CAAC5B,QAAQ1C,MAAMwC,IAAI,CAACE,KAAI,KAAMrB,IAAI,IAAIe,MAAMK,MAAM,EAAE;gBACvDE;gBACA,MAAM8B,QAAQnD,KAAKoD,GAAG,CAAC,GAAGpD,KAAKqD,KAAK,CAAC,AAACf,CAAAA,IAAIF,UAAS,IAAK;gBACxD,IAAIvB,QAAQC,KAAK,CAAC,EAAEf,EAAE,CAACqB,IAAI,CAACI,OAAO,CAAC7C,SAAS,IAAIsE,IAAI;gBACrD,MAAO,CAACpC,SAASd,IAAI,IAAIe,MAAMK,MAAM,CACnC,iBAAiB;gBACjBN,QAAQC,KAAK,CAAC,EAAEf,EAAE,CAACqB,IAAI,CAACI,OAAO,CAAC7C,SAAS,IAAIsE,IAAI;gBACnDpC,QAAQD,YACNC,OACAC,OACA,IAAMf,GACN,CAACuD;oBACCvD,IAAIuD;gBACN;gBAEFpD,IAAIoB,IAAI,CAAC,GAAG,KAAKyB,MAAM,CAACI,OAAO,EAAE,EAAEtC,OAAO;gBAC1CV,MAAME,SAAS;gBACf;YACF;YACA,6CAA6C;YAC7C,IAAI6C,WAAW;gBACb7B;gBACA,MAAM8B,QAAQnD,KAAKoD,GAAG,CAAC,GAAGpD,KAAKqD,KAAK,CAAC,AAACf,CAAAA,IAAIF,UAAS,IAAK;gBACxDlC,IAAIoB,IAAI,CACN,GAAG,KAAKyB,MAAM,CAACI,OAAO,EAAE,EAAEvC,YACxBQ,MACAN,OACA,IAAMf,GACN,CAACuD;oBACCvD,IAAIuD;gBACN,IACC;gBAELnD,MAAME,SAAS;gBACf;YACF;YACA,qEAAqE;YACrE,IAAI,WAAWa,IAAI,CAACE,SAASrB,IAAI,IAAIe,MAAMK,MAAM,IAAI,mBAAmBD,IAAI,CAACJ,KAAK,CAACf,IAAI,EAAE,CAACqB,IAAI,GAAG;gBAC/FC;gBACAnB,IAAIoB,IAAI,CAAC,CAAC,EAAE,EAAEF,KAAK,CAAC,EAAEN,KAAK,CAAC,EAAEf,EAAE,CAACqB,IAAI,EAAE;gBACvCjB,MAAMI,WAAW;gBACjB;YACF;YACA,8DAA8D;YAC9DkB,aAAaL;YACb,IAAI,UAAUF,IAAI,CAACE,OAAOC;iBACrBlB,MAAMK,QAAQ;QACrB;QACAa;IACF;IACAA;IAEA,MAAMkC,KAAK,GAAGrD,IACXqB,IAAI,CAAC,MACLC,OAAO,CAAC,WAAW,QACnByB,IAAI,GAAG,EAAE,CAAC;IACb,OAAO;QAAEM;QAAIpD;QAAOnB;QAAOK;IAAa;AAC1C"}