- **Japanese (`ja`) and Korean (`ko`) draft catalogs** (`crates/ui-egui/locales/ja.toml`, `ko.toml`): every UI string (1,351 keys), command label (519) and menu wording (29), machine-drafted from our English (Japanese: concise, no desu/masu endings, full-width punctuation where Japanese UI text uses it; Korean: no hamnida endings, instructions in -seyo). Placeholders identical to English, `|context` keys translated per context, plural entries give only `other`. Marked `status = "draft - needs native review"` and registered as drafts (not in `COMPLETE`), like fr/de/pt-BR.
- **Plurals:** `plural_one` returns false for ja and ko (always "other").
- **Language picker and detection:** Preferences > Language lists the two languages by their own names; System picks `ja*` / `ko*` from the OS language list; `pref.language` doc lists `ja|ko` (engine spec and docs/mcp.md).
- **CJK UI fonts** (`crates/ui-egui/src/cjk.rs`): egui's fonts and Inter have no CJK glyphs. After the existing system font scan, the desktop app works out which CJK scripts (kana, Hangul, Han) the UI shows: the UI language's, the language picker's while Preferences is open, typed/IME text, and the open file's names and comments (re-read when the tab's revision changes, not mid-drag, and never again once all three scripts have a face). For each needed script it appends one face to egui's fallback list, the UI language's script first: Windows Yu Gothic UI / Meiryo / Malgun Gothic / Microsoft YaHei, macOS Hiragino Sans / Apple SD Gothic Neo / PingFang SC, Linux/BSD Noto Sans CJK, Source Han Sans, Nanum, WenQuanYi; the craft-fonts build's embedded BIZ UDPGothic is tried first (free). Han without kana outside a Japanese/Korean UI reads as Chinese (Chinese faces first). English users load no CJK face. No font files are added to the repo.
- **Read once, share the bytes:** new `FontDb::find_one` reads only the one face a fallback needs instead of every file of the family (Yu Gothic UI is 4 files of ~14 MB); `theme::font_data` hands egui the font database's shared bytes instead of copying them (also used for the existing Hebrew/Arabic fallbacks). The scan's existing 64 MB per-file cap applies. Any failure degrades to the previous behaviour (tofu at worst), never a panic.
- **Web:** no system fonts there. `crates/text/build.rs` now embeds craft-fonts' Noto Sans CJK into the wasm build when `FIGCRAFT_EMBED_CJK` is set (it covers kana, Hangul and Han; +16 MB) and reruns when that variable changes. Without it, `cjk::can_show` keeps ja/ko out of the picker and out of System detection, so Japanese and Korean browser users keep the English UI instead of boxes. Nothing is fetched from the network.
- **Docs:** localization-parity (rows, CJK UI fonts section, totals, revision), translating (catalog list, CJK conventions, fonts, plural note), target-app-parity, ROADMAP, gaps (gap 5), roadmap (focus 4), mcp.md. Localization 30% -> 35% (graded low: unreviewed drafts), 40-65 h -> 32-55 h; ready for real work ~52% -> ~53% (52.9 -> 53.1 by the dimension table); full parity ~284-554 h -> ~276-544 h; beta ~162-292 h -> ~155-280 h.
159de3ce43