Summary

Stop shipping every romanization engine in the worker and the main bundle. Romanization now loads the engine for the current song on demand.

Closes #396.

Changes

• Set Vite worker.format to es so lyric-romanizer's per-engine import() stays dynamic.
• Call pinned romanizeLines(lines, options) once for the whole array. The library already keeps Latin lines and isolates engine failures.
• Bump lyric-romanizer to ^0.3.1. The dict plugin copies kuromoji files through lyric-romanizer/dict and no longer depends on [@sglkc](/sglkc)/kuromoji directly.
• Replace the main-thread pinyin-pro import in the alphabet rail with a compact Han first-letter table.
• Drop the main-thread import("lyric-romanizer") fallback. Tauri webviews have Worker. Without a worker the original lines are returned.

Acceptance criteria

From #396:

• vite.config.ts sets worker: { format: "es" } — tests/romanize-packaging.test.ts
• Production romanize.worker-*.js is the orchestrator only and keeps dynamic import()
• First Mandarin romanize does not load kuroshiro / to-jyutping / sanscript
• Pinned SongLanguage uses one romanizeLines(lines, options) call — src/lib/romanize-options.test.ts
• Dict plugin uses lyric-romanizer/dict; [@sglkc](/sglkc)/kuromoji removed from OpenKara devDependencies
• pinyin-pro is not a direct main-thread import (alphabet-index included)
• lyric-romanizer is ^0.3.1
• pnpm vitest run + pnpm build green; worker still serves /dict/ for Japanese

Idle warmup("japanese") is still out of this change. The first Japanese overlay still pays dictionary parse.

Production build evidence

Local pnpm build after this change:

Worker orchestrator dynamic imports (from dist-kcbpQyEL.js):

import(\./dist-yO-POjyd.js`), import(`./dist-Ck1oGm7X.js`), import(`./dist-CUAsjn7q.js`), import(`./lib-BXhtaYAn.js`), import(`./tamil-romanizer-Dx1k4D2F.js`), import(`./sanscript-BfGYTtJQ.js`), import(`./bundle.esm.min-B4fP4wtX.js`), import(`./CyrillicToTranslit-aHZG3v3q.js`), import(`./src-BfiIdoAE.js`), import(`./src-CXAzzDWw.js`)`

Japanese dictionary: dist/dict/ still has the 12 kuromoji files (base.dat.gz, cc.dat.gz, check.dat.gz, tid.dat.gz, tid_map.dat.gz, tid_pos.dat.gz, unk.dat.gz, unk_char.dat.gz, unk_compat.dat.gz, unk_invoke.dat.gz, unk_map.dat.gz, unk_pos.dat.gz). The worker creates the romanizer with japaneseDictPath: "/dict/".

A Mandarin song therefore loads the pinyin chunk only. Kuroshiro, Jyutping, and Sanscript stay on disk until first use.

Test plan

• pnpm lint and pnpm build
• pnpm vitest run (follow-up commit patch coverage 100%)
• cd src-tauri && cargo clippy / cargo nextest — skipped; no Rust change
• node --run check:i18n — skipped; no copy change
• IPC contracts — no public command, payload, or event change
• Product-standard profiles reviewed:
  • Lifecycle, quality, and testing — ISO/IEC 25010 performance and maintainability. Evidence: worker size, batched call tests, packaging test.
  • Security, privacy, and release — dependency change only. lyric-romanizer@0.3.1 from npm; pinyin-pro and [@sglkc](/sglkc)/kuromoji remain transitive. Flatpak node-sources.0.json regenerated.
  • Interaction, language, and IPC profiles do not apply. Alphabet rail buckets for Han titles stay A–Z first letters.

Residual risk

A rare Han character outside the generated first-letter table falls into #. The table covers 20,935 mapped initials across CJK Ext A, Unified, and Compatibility. Supplementary-plane characters such as 𠀀 stay #, matching the previous unmapped path.

Related issues

Closes #396