fix: cover CJK transcripts instead of rendering tofu

[natyang1234]

The defect

Every library recipe declares a Latin webfont stack (Archivo, Playfair Display, …) with Latin system tails, and the engine resolves only the webfonts it fetches — there is no per-glyph system-font fallback. A Chinese, Japanese or Korean transcript therefore renders as tofu boxes on every caption, and probe-qa cannot tell: tofu has ink and contrast.

Found on the first real Chinese video this pipeline was given — a zh-TW birthday clip through the custom whisper provider (mlx-whisper JSON). Every caption drew as empty rectangles while lint, --verify and probe-qa all reported clean.

The fix

generate-recipe.ts now runs a deterministic font-coverage pass inside the one writer of the final document (pipeline/scripts/ensure-font-coverage.ts):

• when the transcript contains CJK and the document's font stacks name no covering family, the matching Noto family (TC / JP / KR by script; Han without kana or hangul defaults to Traditional) is added to the existing Google Fonts css2 request and appended to each font-family list ahead of the generic keyword — the recipe's webfonts keep styling Latin, the Noto face catches only the glyphs they lack;
• a Latin transcript, or a recipe that already names a covering family (Noto CJK / PingFang / Hiragino / Source Han / …), passes through byte-identical, keeping recipe runs deterministic for the content they were authored against;
• no library recipe is edited.

Verification

• Same run (hook-107-peak, zh-TW transcript): tofu before, correct Traditional Chinese after; lint + --verify + probe-qa clean (0 warn).
• Unit tests (node --import tsx --test tests/ensure-font-coverage.test.ts, 8 passing) cover script detection, css2 URL extension, stack placement before the generic keyword, the serif/sans split, already-covered pass-through, no-link documents, and the Latin byte-identical guarantee.

Notes for reviewers

• The TC default for bare Han is undecidable from codepoints; a recipe (or operator --module copy) that names its own CJK family always wins, so an SC-preferring deployment can override per recipe.
• A longer-term companion would be a probe-qa tofu check (e.g. glyph-coverage assert at lint time); happy to attempt that separately if wanted.

🤖 Generated with Claude Code

[jakubtyrcha]

Thanks for the contribution. We plan to address fallback fonts in the coming weeks.