DOCX: Hangul lines break at syllable boundaries, splitting words Word keeps whole

[developer0hye]

Summary

Hangul lines are broken at syllable boundaries. Word breaks at eojeol
(space-delimited word) boundaries on ordinary lines and only falls back to
character-level breaking to keep a justified line from going too loose. The
result is that a Korean word is split across two lines where Word moves the
whole word down.

Neither w:wordWrap nor w:kinsoku appears in the source, and settings.xml
carries only compatibilityMode=15, so Word's defaults are in force.

This is Word-specific behaviour, not a general CJK rule: PowerPoint itself
breaks Korean mid-word, and office2pdf's PPTX output matches PowerPoint's break
points exactly on the same corpus.

Measured

tests/golden_mocks/business/sources/docx/02_contract_ko.docx, page 1:

paragraph	Word	office2pdf
제1조 line 1 ends / line 2 begins	…필요한 제반 사항을 / 정함을 목적으로 한다.	…사항을 정 / 함을
제4조 line 1 ends / line 2 begins	…VAT 별도)으로 / 하며, 을은…	…으로 하 / 며, 을은…

One syllable (10.5pt) migrates in each case.

Word was not out of room. On 제1조 the trailing space of Word's line 1 ends at
507.99pt and "정" would end at 518.49pt, well inside the 524.47pt right margin —
Word had 6pt of space and still moved the whole word down. Line width: Word
493.80 vs office2pdf 508.11.

The trigger shows up in the same file's two justified paragraphs
(<w:jc w:val="both"/>): there Word does break mid-eojeol, and identically to
office2pdf (both split 보관|한다.). So character-level breaking is Word's
fallback for justified lines, not its default.

Corroborated across every wrapped body paragraph of an internal test document,
where office2pdf breaks 제공하는 → 제 / 공하는, 지원하는 → 지원하 / 는,
발화량 → 발 / 화량, 발생할 → 발생 / 할 and 103,125 / 원 while Word breaks at the
space each time. That is not a metrics difference: matched wrapped lines agree
to ≤0.1pt, both sides use a 70.8pt left margin and the same 453.6pt text width,
and the constant 3.7pt difference on paragraph-final lines is Word's trailing
space glyph.

Reproduction

cargo run --release -p office2pdf-cli -- \
    tests/golden_mocks/business/sources/docx/02_contract_ko.docx -o out.pdf

Compare the line endings of 제1조 and 제4조 against Word.

Expected

On a non-justified line, a Hangul eojeol is treated as an unbreakable unit: 제1조
line 1 ends at "…제반 사항을" and 정함을 begins line 2 whole.

Actual

Every syllable boundary is a break opportunity, so the last word of the line is
split and one syllable is left behind on the line above.

Found in the 2026-07-29 full-corpus visual audit at c1adfbd (native Office GT, 150 DPI, mutool-trace geometry).

Visual comparison

Compare

GT left, office2pdf right, 150 DPI, same page.

[developer0hye]

Fixed by #731. Non-justified DOCX Hangul paragraphs now break at eojeol boundaries: each qualifying word is emitted inside an inline #box, which is the one thing UAX #14 cannot break through — text(lang: "ko"), par(linebreaks:) and text(costs:) were each measured to leave the breakpoints untouched, because typst-layout builds its ICU4X segmenters with default options and never consults Lang::KOREAN. A U+2060 marker also works and was rejected: it lands in the text layer (#664).

Fresh GT comparison on 02_contract_ko page 1: 제1조 line 1 now ends …필요한 제반 사항을 and 제4조 ends …VAT 별도)으로, both matching Word (main split them as …사항을 정 / …으로 하). Matched lines 17 → 21 of 21, re-wraps 1 → 0, page-1 text line-for-line identical to GT, AE 5% 50,388 → 47,825. The justified paragraphs still split 보관|한다 exactly as Word does.

Not one baseline moved on any Korean fixture (1151 shared baselines on the technical brief), the text layer is character-identical to main, and of 377 corpus files only the five Korean DOCX changed — every PPTX and XLSX byte-identical.

One residual: 10_research_report_ko page 2 keeps 표시되어야 whole where its GT splits it. That fixture defines no Normal style, so its unstyled paragraphs inherit the Word application's wordWrap=0 default — native Word one-factor probes identified w:wordWrap as the exact discriminator, now tracked in #730. Other deviations on the audited page: #724, #649, #646, #628, #728, #630.