[KnifeLemon]

What does this change?

Implements the Project 4 item "Layout engine: paragraph layout and line breaking": a paragraph goes in, positioned and shaped lines come out. Everything is in src/Filee.Engines/Typesetting. No converter uses it yet; the render quality harness now measures it against rhwp (DocumentRenderers.Candidate is set).

• Line breaking (LineBreaker.cs): UAX Add e-book conversion, a built-in HTML reader and a calibre package #14 at Unicode 16.0, rules LB1–LB31.
  • It passes all 16,672 cases of Unicode's LineBreakTest.txt. I checked that locally and didn't commit the file; LineBreakerTests covers each rule family instead.
  • Tailorings:
    • 한글 어절 (breakNonLatinWord KEEP_WORD) keeps Hangul words whole.
    • 글자 (BREAK_WORD) breaks between syllables.
    • Latin words can break anywhere (breakLatinWord BREAK_WORD).
    • A document's own characters forbidden at the start or end of a line (Word w:noLineBreaksBefore / After) are honoured.
• Paragraph text (ParagraphText.cs): the inlines become one string: tabs, line breaks as U+2028, links, bookmarks, and a list marker in front.
  • Each character gets its style from its 한글 language slot (font, 장평, 자간, relative size, 글자 위치) or from Word's run font, scale, spacing and raise.
• Layout (ParagraphLayout.cs, output in LineBox.cs):
  • Pipeline: items of one style, font, script and direction are shaped; lines are filled greedily and break at the last opportunity. Each line's parts are shaped again with their context.
  • Line fitting:
    • Spaces at a line end hang past it.
    • A word wider than the line breaks between clusters.
    • 최소 공백 (condense) lets the spaces shrink to fit.
    • lineWrap KEEP and SQUEEZE hold the paragraph on one line.
  • Alignment: left, right and centre; justify widens the spaces (or the clusters on a line with no spaces), and distribute (배분) widens the last line too. First-line and hanging indents.
  • Tabs: left, right, centre and decimal stops with leaders, measured from the left indent as 한글 does, with default stops every 40 pt. A hanging indent works as a stop, so a list marker hangs.
  • Line height:
    • 한글's percent, fixed, at-least and between-lines rules measure from the font size, with the baseline at 85% of the text height.
    • Word's multiple and 한글 "글꼴에 어울리는 줄 높이" use the fonts' OS/2 win metrics.
  • Characters: ratio, letter spacing, raise, super / subscript, colour, shade, underline, strike-through, links and bookmarks. Synthetic bold and italic are drawn through new GlyphRun.Embolden / SkewX.
  • Direction: a simplified bidi pass sets levels, and parts are reordered by rule L2. Runs stay in logical order in the page, so the PDF's text reads as written. Full UBA stays with the bidi item.
• Model: HCharLayout gains FontSpace and Kerning, read from HWPX useFontSpace / useKerning.
  • With "글꼴에 어울리는 빈칸" off (as in our template), a space is half the font size wide.
  • With useKerning="0", the font's kerning is off.
  • These two settings made the line breaks match rhwp (below).
• PDF writer fix: subset cmaps now keep only the characters the text has. CJK fonts map 文 and the Kangxi radical ⽂ to one glyph, and the ToUnicode table took the lower code point, so Japanese and Chinese text copied out as radicals.
• Interim page flow (FileeRenderer.cs): sections' pages, paragraphs stacked with a new page when full, and headings as the PDF outline. Table cells are stacked as paragraphs for now. The page composer item replaces this flow.
• Docs: a new section "Filee renderer: line breaking and paragraph layout" in ENGINES.md; RENDER-QUALITY.md updated.

Against 한글 itself

HangulLineBreakTests (a private-corpus test, run by build/render-compare.ps1) compares where 한글 broke lines with where Filee breaks them.

• Input: HWPX files that 한글 saved. They store the start and width of every line (hp:linesegarray).
• Method: each plain-text paragraph is laid out at the width 한글 used, with the file's shapes, and its line starts are compared with 한글's.

Result on the four HWPX files attached to Filee's issues (#49, #51, #65), kept out of the repository:

	Matching 한글
Paragraphs	150 of 152 (98.7%)
Line starts	161 of 163 (98.8%)

The two that differ:

• one uses a font that isn't installed (210 이솝우화);
• one is a 75-character run of - wider than the line, which 한글 keeps on one line.

자간 is a share of each character's own width (장평 applied), not of the font size. A space or a narrowed character closes up less. Before this change, two lines with negative 자간 took one or two characters more than 한글's; now they break where 한글 broke them.

Reader fix: 한글 writes breakNonLatinWord the other way round from its names. BREAK_WORD is 어절 and KEEP_WORD is 글자. Two pieces of evidence:

• In the binary HWP of the same documents, bit 7 of the paragraph shape (한글 단위; 1 = 글자) is set exactly where the HWPX says KEEP_WORD.
• In plain KEEP_WORD paragraphs, 한글's stored lines break inside words 6 times in 9 ("외국인등록|번호", "공동이|용을").

HwpxReader now maps it that way; breakLatinWord agrees with its names.

Against rhwp (feature matrix, Windows fonts)

Feature	Same lines	Note
script-zh, script-ja, lists, tabs	100%	same baselines too
line-spacing-* (5 alignments)	87.5%	rhwp splits a word in the distributed paragraph
script-ko	60%	rhwp splits some words even in 어절 paragraphs ("않습니|다.")
ar, he, hi, gu, th	allowed	rhwp draws these unshaped or at fixed widths; Filee's text now round-trips 100%
tables, headers-footers, notes, columns	low	left to the page composer

The thresholds in render-thresholds.json are today's values, each with its reason.

Also: the main sync brought Italian and Indonesian without error.document_too_complex (a builtin-engines key). Both now have it; that is what failed the first CI run.

Tests

• LineBreakerTests: 34 cases.
• ParagraphLayoutTests: 38 cases, with the pinned Noto fonts of build/fetch-fonts.ps1. Expected widths come from the shaper, so no magic numbers.
• A FontSubsetterTests case for the cmap fix.
• Engines suite locally: 2299 passed, 0 failed.

Not in this PR

• CJK typography (금칙 beyond UAX Add e-book conversion, a built-in HTML reader and a calibre package #14, autospacing, grid).
• Full UBA.
• Dictionary breaking for Thai.
• Paragraph borders.
• Honouring stored HWPX line segments.

These have their own board items.

🤖 Generated with Claude Code