0.26.0: HWP → HTML on our own reader (replaces pyhwp) + HWP reader fixes

[CocoRoF]

Why

XGEN's document source viewer rendered .hwp through pyhwp's hwp5html (AGPL-3.0). XGEN keeps no AGPL dependency, so this adds an HTML renderer on our own HWP reader. Comparing it against hwp5html on real documents also exposed reader bugs that already affected workflow's HWP preview (DOCX path).

New: documents/legacy/hwp_html.hwp_to_html(content) -> str

One self-contained HTML page built directly from the reader's model (no DOCX detour, no subprocess):

• fonts, sizes, colors, alignment, margins, hanging indents, line spacing
• merged cells, column widths, cell and table border/fill
• pictures, headers/footers, master pages, footnotes
• inline objects kept at their position in the line

Safe to inject unsanitized (the frontend does):

• text and attributes are escaped
• font names are stripped to [\w .-]
• every CSS rule is scoped under .hwp-doc
• only png/jpeg/gif/bmp are embedded (bmp/tif are converted to png), with a 64 MB image cap

Reader fixes (shared with hwp_to_docx)

• Pictures dropped: BinData compression is per item (spec table 18). png/jpg stored raw were inflated, failed, and were silently skipped (11 of 15 pictures in one real press release).
• Wrong strikethrough: per table 35, strikethrough is bits 18-20. Underline kind 2 was read as strike.
• Misplaced styles: char-shape positions are raw word positions (controls count as 8 words), so styles after section or field controls were misplaced.
• Missing text:
  • hyphen, 묶음 빈칸 and 고정폭 빈칸 are now text
  • auto numbers are rendered (표 1)
  • 덧말 and 글자 겹침 text is kept
  • footnotes and endnotes are kept
  • inline (treat-as-char) objects are placed in the line
  • nested paragraph lists and text-box attachments are kept
• Layout:
  • paragraph margins and indent
  • distribute alignment
  • Korean line breaking by 어절
  • landscape pages
  • table border/fill and caption side
  • per-section headers
• Distribution documents (배포용) open per Hancom's published spec, and copy/print protection carries over. DRM and certificate-encrypted files are refused by name. Decompression is capped at 256 MB.

Verification

• 31 HWP files: 8 format fixtures and 23 real documents (government reports, a 10 MB handbook, exams, the HWP spec itself).
  • hwp5html fails on 5 of them; ours fails on none.
  • Every file converts on both the HTML and DOCX paths.
  • Coverage of doc2chunk's chunk text (what citation highlighting searches) is ≥ hwp5html on every file.
  • Image and table counts are ≥ hwp5html.
• Tests:
  • new tests/unit/test_hwp_html.py (25 tests, fixtures built per spec)
  • the rich fixture's strike encoding is corrected to table 35
  • full suite: 957 passed, 1 skipped

본 제품은 한글과컴퓨터의 글 문서 파일(.hwp) 공개 문서를 참고하여 개발하였습니다.

🤖 Generated with Claude Code