feat(remotion-composer): CJK-aware captions and caption styling props

[sakuraozation]

Summary

CaptionOverlay assumes space-delimited text, which breaks captions for CJK languages in three ways:

1. A hardcoded " " is rendered between words — Japanese/Chinese don't use inter-word spaces, so captions show spurious gaps.
2. Lines can wrap anywhere inside a word (every CJK character is a CSS line-break opportunity), producing mid-word breaks and punctuation stranded at line starts.
3. Pages are cut every wordsPerPage words regardless of sentence boundaries, so a clause from the next sentence can leak onto the current page.

This PR makes the caption system CJK-capable while keeping the default rendering for space-delimited languages pixel-identical, and exposes the caption colors that TalkingHead previously hardcoded.

Before / after (same props, Japanese captions — the caption text itself explains the fix: "Japanese captions don't put spaces between words!"):

before	after

Before: spurious spaces, a break in the middle of 単語のあいだに, and a stranded ん！ at a line start. After: no spaces, wraps only at word (clause) boundaries.

Related issue

None filed — happy to open one first if you prefer issue-first workflow.

Changes

• CaptionOverlay: new wordSeparator prop (default " " keeps current behavior; CJK callers pass "").
• CaptionOverlay: each word renders as an unbreakable inline block, so lines wrap only at word boundaries. For space-delimited text this matches the previous behavior (browsers already broke at the spaces); for CJK it prevents mid-word breaks.
• WordCaption: optional pageBreakAfter flag; buildPages flushes the page early when set, letting callers align pages with sentence/scene boundaries.
• TalkingHead: exposes captionColor / captionBackgroundColor (previously hardcoded inline), captionFontFamily, and captionWordSeparator. All default to the current values, so existing props files render identically.

One known limitation, by design: a single word/chunk wider than the caption box will overflow rather than break. Callers control chunk granularity (for Japanese, clause-level chunks of 2–10 characters are natural), so this hasn't been a problem in practice.

Testing

• Stills rendered with the same Japanese props before/after (screenshots above). Rendered with the resolveAsset fix (fix(remotion-composer): resolve videoSrc through resolveAsset in TalkingHead #506) applied so the demo clip loads; the caption diff is exactly this branch.
• Default-props render of English captions confirmed visually unchanged (separator defaults to " ", and word-boundary wrapping matches the browser's existing space-based breaks).
• tsc --noEmit clean.
• In production use for Japanese vertical (1080x1920) ad videos — clause-chunked captions with wordSeparator: "", pageBreakAfter on scene boundaries, and a Japanese rounded-gothic captionFontFamily.

Checklist

• The change is focused on a single logical concern (caption i18n/styling surface of TalkingHead).
• I ran the relevant tests locally where applicable — no TS test infra for remotion-composer; verified manually as above.
• I updated docs/README if behavior or usage changed — props are documented inline; happy to add a README section if you'd like one.
• No unrelated files are included in the diff.

---

Disclosure: this patch was developed with AI assistance, then human-reviewed and verified in production use. Happy to adjust anything to match project conventions.