penpot/render-wasm/docs/cjk-layout.md

Chinese and Korean core layout

Paragraph text-language selects a layout policy. Writing mode remains a whole-shape value. UI availability never changes either attribute.

Named Chinese presets

Tags	Preset
zh, zh-Hans, zh-Hant, unknown regions	Original basic punctuation policy
zh-Hans-CN, zh-Hant-CN, zh-CN, zh-SG	Mainland: adjacent punctuation compression; quarter-em Han/Latin gaps
zh-Hant-TW, zh-TW	Taiwan: solid punctuation; quarter-em Han/Latin gaps
zh-Hant-HK, zh-HK, zh-MO	Hong Kong: centered punctuation spacing; quarter-em Han/Latin gaps

Script does not imply region. Changing a profile does not convert characters. These presets select a limited set of styles; they do not claim that every publication in a region uses the same settings.

Compression applies to eligible adjacent full-width punctuation. Vertical colon, semicolon, question and exclamation marks keep their full frame. Proportional horizontal punctuation is not compressed again. Vertical cells retain at least a half-em frame. The Taiwan and Hong Kong presets allow a single comma or full stop to hang by at most half an em vertically; adjacent punctuation and horizontal text do not hang.

Chinese justification expands eligible character gaps by at most half an em. The final line/column remains unexpanded. These bounds are explicit POC choices, not numeric requirements imposed by a regional standard.

Korean modes

• Missing korean-wrap, or syllable: syllable wrapping with punctuation rules.
• keep-words: keep adjoining letters and digits together across scripts and style runs. Punctuation keeps its own opportunities and restrictions.
• spaces-only: preserve the earlier experiment. It remains readable in the UI for saved text but is not offered for new selections by default.

Fitting words move whole. Overlong words may use safe grapheme boundaries. Nonbreaking spaces/joiners, numeric prefixes/suffixes, paired dashes/ellipses, annotations and punctuation restrictions remain separate from letter ties. Korean justification first expands spaces, at most half an em per gap. Syllable mode may then expand eligible character gaps by at most one eighth of an em. Word mode never expands inside a word. Original spaces remain in source text; no Japanese word-space substitution applies.

Japanese line adjustment

Japanese justification uses the JLREQ §3.8.4 ordered expansion in both orientations (jlreq_ordered_expansion). Gaps widen by pair-table stage: Western word spaces, then Japanese/Western gaps up to half an em, then solid kana and kanji up to a quarter em. Slack left after the three stages spreads evenly over the same gaps. Punctuation aki, gaps beside a ruby base, gaps inside a Western word and gaps across a warichu note take none. The final line or column stays unexpanded; distribution keeps its own path.

Squeezing (oikomi) follows the line adjustment setting in both orientations. Push-in first squeezes whenever the aki on the line can hold the next unit. Push-out first squeezes only when pushing out cannot resolve a kinsoku break, which horizontally means the break would move back over more than four kinsoku pairs. Horizontal text keeps a unit wider than the box whole on its own line instead. Push-out only never squeezes.

Horizontal emergency wrapping and offsets

SkParagraph can ignore word joins during emergency wrapping. The horizontal plan measures candidate units and replaces joins with zero-width break opportunities only inside overlong units, at safe grapheme boundaries. Both marker characters occupy three UTF-8 bytes and one UTF-16 code unit. Their offset maps therefore stay identical for drawing, hit testing, carets, selection, text paths and export. Width participates in the plan cache key; resizing a box restores whole-word joins when the word fits again.

If an indivisible Chinese or Korean punctuation/annotation unit exceeds the box width, the plan lets it overflow together. It first measures surrounding text at the original width, then records those line starts with layout-only line separators before giving Skia enough physical width for the unit. Ordinary text keeps its original wrap budget. Source text stays unchanged; the plan's offset map accounts for every inserted separator.

Painting, strokes, caret/selection geometry, hit testing, annotations, paths and exports share the plan and its alignment offset. The width cache also rebuilds these breaks on resize. Legacy Korean spaces-only keeps its earlier behavior; the supported modes do not claim complete CSS keep-all conformance.

Pinyin readings

The Chinese layout panel offers typed Pinyin readings when the selected paragraph has an explicit Chinese profile. In the WASM text editor, select one character or a group, in any text styles, then enter its reading. A reading over spans of several styles, such as characters with and without emphasis marks, gives them one ruby-group; the renderer draws one base over the spans of a group (Paragraph::ruby_groups), whose first span holds the reading and its presentation. To edit an existing reading, select its whole base. Selections across paragraphs, or that cut a reading or warichu note, do not offer a reading field, which prevents repeated readings. For character-by-character readings, annotate each character separately: neighbours with the same reading but no shared group stay separate bases.

Readings use the saved ruby attributes and the shared annotation renderer. They support size, alignment, overhang, side, hiding and annotation clearance. Adjacent reading spans recover the style covering each base even when Skia merges their styles. Emphasis marks retain their per-base placement across wraps; text bounds include the marks so individual exports keep them whole. Without a text selection, presentation changes affect existing readings. Horizontal readings form one Latin group above or below the base; vertical readings form one sideways Latin group beside it. A reading longer than its base overhangs an adjacent comma, full stop or middle dot (or kana) without its own reading, by at most one ruby character and half the neighbour, in both orientations. Han neighbours take no overhang. What overhang cannot absorb, or all of it under ruby-overhang none, spreads the base. Pronunciation is entered by the author; the editor does not generate or distribute syllables.

Composed and combining tone marks retain their source text and shaping. Glyphs sharing a cluster retain the same complete source range. Tests use src/fonts/pinyin-test.ttf, an 8 KB subset of the repository's Work Sans Regular font, with letters, umlauts and combining tone marks. The subset retains its copyright and SIL Open Font License metadata; its family name is Pinyin Test. FontTools generated it from the Pinyin test strings.

Emphasis and underline positions

Chinese profiles expose emphasis shape, emphasis position, underline position and annotation clearance independently of Pinyin. Position controls do not turn on underlining; use the existing underline control for that.

Optional span text-emphasis-position stores auto, over-right, over-left, under-right or under-left. The first part selects the horizontal side; the second selects the vertical side. Auto uses below/right for Chinese and preserves above/right for Japanese, Korean and plain text.

Optional text-underline-position stores auto, under, left, right, under-left or under-right. Under places horizontal underlines below the base rect; left/right select the vertical side. Chinese auto uses below/left. Other auto values preserve the earlier baseline offset and vertical left side. Explicit values can select other styles without changing the language. These are supported POC choices, not a claim of complete decoration support.

Automatic clearance reserves each occupied side. Stacked annotations retain the largest reading and mark sizes across mixed-font spans, preserving the prior Japanese column widths. Underlines remain next to the base; readings move outside an underline on their side, and emphasis moves outside a reading on its side. The underline band reserves 0.18 em, a POC bound around the existing 1/9-em gap and stroke width. Disabling clearance retains fixed line-height behavior and can allow overlap/clipping.

Horizontal marks are shaped once per span and placed by their measured box (advance by ascent and descent at half the base size). Paint, text bounds and selection use that one box, so bounds cover exactly what paints. Position data carries the mark's em box centred on it, because SVG export draws the mark at a font size of that box's width. Vertical marks already centre the glyph and their position-data box on the same point.

Decoration edits that intersect a note, or a reading without a group, apply to its whole base, so changing a side cannot duplicate its reading on split spans. A grouped reading keeps one base when an edit splits its spans. Other text styles remain unchanged. Selection snapshots, Plugin API properties, CSS handoff and saved content retain the independent positions. CSS auto positions inherit the Chinese paragraph defaults; explicit positions emit the corresponding CSS keyword pairs. The binary writer uses former span header bytes 14/15; header size and prior offsets stay unchanged. Unknown bytes decode to auto. Both WASM consumers share that writer.

Defaults and positions were checked on 2026-10-20 against CSS Text Decoration, emphasis and underline positions. Japanese/Korean vertical auto underlines retain this branch's existing left side rather than adopting the spec's suggested right-side style.

Evidence and release checks

Renderer tests cover regional geometry, horizontal/vertical wrapping, style boundaries, decomposed Jamo, punctuation, justification, stable offsets and distinct wrap protocol bytes. Editor splits preserve the complete wrap mode. Both WASM consumers use the same writer and byte values: 0 syllable, 1 experiment, 2 keep-words; unknown values fall back to syllable wrapping. The renderer decodes the byte through the exported RawKoreanWrap enum.

Real Pinyin, Zhuyin and Korean IME checks and native-reader review remain release checks. Font glyph coverage and localized forms still depend on the chosen fonts. Dedicated Zhuyin layout is described below.

References checked on 2026-10-19:

• CLREQ draft, 2026-10-07, sections 6.1 and 6.3.2: punctuation boundaries, compression and hanging.
• KLREQ draft, 2026-03-21, sections 7.1 and 7.2: syllable/word wrapping and justification restrictions.
• CSS Text 3: letter-breaking policy is distinct from punctuation and overflow handling.

Mandarin Zhuyin annotations

An optional span flag, ruby-zhuyin, selects dedicated right-side Zhuyin. Absent or false retains normal ruby. Each span maps one Han character (optionally with one variation selector) to one typed Mandarin syllable in ruby. Enter 1–3 U+3105–U+3129 symbols and at most one tone: U+02C9, U+02CA, U+02C7, U+02CB or U+02D9. Neutral tone may precede or follow the symbols; other tones must follow them. The editor validates this syntax; it does not check pronunciation, infer mappings or generate readings. Unsupported API content keeps its text and falls back to ordinary ruby.

Symbols remain upright to the base's right in horizontal and vertical text. Their size is 0.3 em. A 0.6 em strip holds symbols and tones, including an empty tone lane for unmarked first tone. Non-neutral tones sit outside the upper right of the last symbol; neutral tone sits above the symbol stack. The stack centres vertically on the base. These rules follow CLREQ sections 5.5.3.2–5.5.3.3, retrieved 2026-10-20: https://www.w3.org/TR/clreq/. The strip is wider than the reference's half-em minimum to contain a full-width tone fallback glyph.

Horizontal layout reserves the strip in the inline advance. Vertical Auto clearance reserves it in the right gutter, outside a same-side underline and inside same-side emphasis. Fixed clearance may allow overlaps, as with ordinary ruby. Zhuyin ignores ruby size, alignment, overhang and side; hiding a reading removes its room without clearing saved text or mode.

The mode uses bit 1 of span header byte 13; bit 0 keeps over/under for normal ruby. Header size stays unchanged. Unknown flag bits fall back to ordinary over ruby. Position-data direction 5 carries a single upright annotation glyph with offsets into the original ruby string and its drawing baseline. Both layout paths use it for SVG; canvas, PNG, PDF and vertical outlines use the same shaped glyph positions. HTML handoff uses positioned symbol boxes; font metrics can make it differ from the renderer's ink-centred placement.

This slice supports Mandarin, one base per span. Group-to-syllable mapping, dialectal symbols/tones, automatic pronunciation, native IME acceptance and native-reader approval remain outside this slice. Changing a base into a group preserves the reading and uses normal ruby until a single base is restored. Existing Pinyin and Japanese readings keep their saved behavior.

Renderer modules

• shapes/text_policy.rs: TextLayoutPolicy (Plain, Japanese, Chinese(ChineseRegion), Korean(KoreanWrap)), parsed once per paragraph from LanguageTag. Layout code asks it named questions.
• shapes/jlreq.rs: JLREQ classes (JlreqClass), pair rules and kinsoku.
• shapes/line_breaks.rs: horizontal layout-text rewrite, OffsetMap and PolicyBoundaries.
• shapes/cjk_chars.rs: shared Han, kana, hangul and punctuation ranges.
• shapes/text_attrs.rs, text_ruby.rs, text_warichu.rs and text_allocation.rs: span attributes, ruby rules, warichu and the justify/squeeze allocators both axes share.
• shapes/text_shaping/: glyph shaping, orientation and font tables.
• text_vertical/flow/: JLREQ column planning, with the pairwise policies in flow/pairwise.rs. Horizontal and vertical helpers use the same jlreq_* and pairwise_* prefixes.
• wasm/text/layout_attrs.rs: the Raw* enums the frontend writes, including RawKoreanWrap.