0.4.0: speaking-style instructions for Gemini and OpenAI, carried by avatar.json

[snakajima]

Closes #3.

What changes

• Gemini prompt is always DIRECTOR'S NOTES.

  ### DIRECTOR'S NOTES
  <character instruction (instructions, if any)>
  <the line's emotion style (default or emotionInstructions)>
  
  #### TRANSCRIPT
  <text>
  

  This is MulmoCast's form. It is now used even when no options are set, because sending the text alone to gemini-2.5-flash-preview-tts fails with HTTP 400 "Model tried to generate text" whenever the line is a question (3 of 3; MulmoCast's ttsGeminiAgent fails the same way without an instruction). The mismatch check (re-request, then error) stays as the safety net.
• Shared character + emotion logic. Combining the two and the default emotion styles now live in src/tts/style.ts, used by both OpenAI and Gemini. OpenAI now takes instructions too; it's ignored on tts-1* models, which don't accept instructions.
• avatar.json voice.<provider>.options. A character's speaking style can live in the avatar file, and the CLI passes it to createTts(). The model / voice / options shape is defined once in src/tts/types.ts and shared by the createTts() config and avatar.json.
• trimToVoice extends a held final sound. When the voice runs on without a pause after a phrase's last character, that character is extended over it, up to the next spoken character. A long, sad 「の？」 kept sounding about 0.5 s past its aligned end, so createTts() rejected it as "speech that is not in the text," and the mouth closed early. The early mouth close also happened in compileTimeline, which MulmoCast uses.
• 503 / 429 retry for Gemini. Requests are retried after 1, 2 and 4 s. The default model stays gemini-3.8-flash-tts.

Live test (2026-10-07)

8 cases × 2 models (boy = Puck, マルモ = Autonoe, Japanese and English, neutral/happy/sad/surprised, with and without instructions), every clip transcribed with gpt-4o-transcribe.

	gemini-3.8-flash-tts	gemini-2.5-flash-preview-tts
Passed	8/8 (no re-requests)	7/8
Instructions read aloud	0	0

The one 2.5 failure: the model garbles the end of the line in a child's voice (transcribed as 「何をすれば」「何をするまん」), and the mismatch check correctly rejects it. Before these fixes, 2.5 returned 400 on the question line, and the held-vowel line failed on both models.

Compatibility

• Gemini's prompt changes even without options (the default emotion style is now sent). The cache key doesn't include the prompt, so Gemini audio already in .tts-cache (made from the text alone) is still reused; delete the cache to regenerate it.
• avatar.json voice entries with an empty voice / model string are now rejected, the same as in createTts().

Checks

• npm run format, lint, typecheck and test (71) all pass, and scripts/smoke.sh avatars/ani passed.

🤖 Generated with Claude Code

work in avatarscript