feat(narration): add FishAudio as a selectable TTS provider

[fancyboi999]

Closes #53

What

Adds FishAudio alongside MiniMax as a narration (voiceover) backend, selectable per workspace in Studio → Settings → Audio. Background music stays MiniMax-only (FishAudio has no music generation).

Why

We use FishAudio TTS heavily (custom reference_id voices), but the studio's narration pipeline was MiniMax-only. See #53 for motivation.

Changes

• core: new fishaudio.ts — resolveFishAudioCredentials / generateFishTts / listFishVoices — plus a shared TtsAudioResult type and 16 unit tests.
• config: MediaConfigStore gains FishAudio creds + a narrationProvider setting; the speech model is env-controlled (FISH_AUDIO_MODEL, default s1).
• server: narration routes by provider; new /api/config/fishaudio, /api/config/narration-provider, /api/fishaudio/voices (the key stays server-side — the browser never sees it).
• studio UI: provider toggle + FishAudio key pane (single host, no region); a searchable voice picker (reference_id) with sample preview; en + zh strings.

Design doc: research/2026-06-15-spec-10-fishaudio-tts-provider.md (RFC-10).

Provider differences absorbed

	MiniMax	FishAudio
Model select	body field	model header (env-driven)
Response	JSON + hex envelope	raw binary stream
Voices	6 fixed voice_id	account reference_id (searchable)
Region	intl / CN split	single global host
Music	yes	none (stays MiniMax)

Screenshots

Verification (real, not mocks)

• Unit: pnpm --filter @html-video/core test → 16/16.
• Real API: generateFishTts → real MP3 (ffprobe-valid); listFishVoices → account models.
• Studio E2E: configure key → switch provider → search + pick a voice → synthesize → asset created; export muxes the FishAudio narration into the MP4 — ffprobe shows h264 video + aac audio, and ffmpeg -f null - decodes clean.
• Regression: cli 14/14, runtime 4/4, adapter-remotion 8/8, smoke all green; the MiniMax narration path is unchanged (only a creds variable rename).

Notes

• A pre-existing adapter-hyperframes test script points at a missing test/ dir (unrelated to this change; present on main).
• MiniMax actual generation wasn't re-run here (no key on hand); its routing + the graceful "key not configured" message were verified.

[lefarcen]

Hey @fancyboi999! 👋

Nice work on the FishAudio integration — the design doc + comprehensive testing (16 unit tests + real API verification + E2E export validation) gives a clear picture of what's changing. The provider abstraction looks clean, and keeping the server-side key handling is the right security move.

I've assigned this to @mrcfps for code review. A few things that look promising from the structure:

• The credential resolution pattern mirrors MiniMax (env fallback + server-side only)
• Provider differences are absorbed cleanly (model header vs body, binary stream vs JSON envelope)
• The searchable voice picker is a nice UX touch for reference_id accounts

Related context: #45 discusses provider selection patterns, and #38 (SenseAudio) has overlapping config/UI seams worth checking for consistency.

Once @mrcfps reviews, we'll move this forward. Let me know if anything's unclear!

💡 To drive this PR to merge hands-free, paste this to your AI coding agent (Claude Code / Codex / opencode / Cursor …):
Take over nexu-io/html-video#54 until it merges — read https://raw.githubusercontent.com/nexu-io/looper/main/skills/pr-takeover/SKILL.md and follow it.