feat(plugins): pluggable TTS backends + Volcengine (Doubao) reference provider

[Hypnus-Yuan]

Summary

Mirrors the image_gen plugin refactor for TTS:

• New TtsProvider ABC + tts_registry in agent/
• PluginContext.register_tts_provider() hook
• Prepend-style dispatcher in text_to_speech tool (mirror of image_generation_tool._dispatch_to_plugin_provider)
• Picker integration (hermes_cli/tools_config.py), setup-wizard status probe (hermes_cli/setup.py)
• First reference provider: Volcengine (Doubao seed-tts-2.0) for Chinese-native TTS — plugins/voice/volcengine/

Why

Adding a new TTS backend today requires editing tools/tts_tool.py (1500+ LOC). This PR makes that a plugin drop, matching how image_gen already works. The dispatcher is prepend-style so all existing users see zero change — only tts.provider: <non-legacy-name> routes through the plugin path.

What's NOT in this PR

• STT — separate follow-up PR (Volcengine STT already exists locally, will ship after this merges)
• Porting existing hardcoded TTS providers (edge/openai/elevenlabs/...) to the plugin interface — one at a time in future PRs
• Streaming TTS via plugins — stays ElevenLabs-only for now (regression test tests/tools/test_tts_streaming_untouched.py locks this in)

Behavior preservation

• Users with tts.provider: edge / elevenlabs / openai / minimax / mistral / gemini / xai / kittentts / neutts → zero change
• Dispatcher fires only when tts.provider is set to a non-legacy name
• Existing test suite: all ~17k tests pass
• New test coverage: 8 test files, ~55 test cases covering ABC contract, registry, plugin hook, dispatcher, requirements probe, streaming regression guard, picker injection, setup status probe, Volcengine synthesis (mocked)

How to test locally

uv pip install -e ".[all,dev,voice-volcengine]"
echo 'VOLCENGINE_APP_ID=your_app_id' >> ~/.hermes/.env
echo 'VOLCENGINE_ACCESS_TOKEN=your_token' >> ~/.hermes/.env
# Set tts.provider: volcengine in ~/.hermes/config.yaml
hermes -q "use text_to_speech to say 你好"

Platforms tested

• macOS 15.x (Apple Silicon)
• CI: Ubuntu 24.04 + Python 3.11 (via this PR's tests.yml run)

Design notes

1. TtsProvider.synthesize() returns a Dict[str, Any] (mirror of ImageGenProvider.generate()) rather than writing and raising. This lets providers carry native_opus / voice_compatible metadata forward without a hardcoded name list in _finalize_tts_response.

2. tts_registry.get_active_provider() returns None when tts.provider is unset — deliberate divergence from image_gen_registry (which auto-selects FAL). The divergence preserves the existing unset-config → Edge-default flow; documented in the registry module docstring.

3. Volcengine's client.py WebSocket streaming code is lazily imported inside synthesize() so machines without the optional voice-volcengine extra installed can still discover the plugin without ImportError.

4. Output path derivation in Volcengine provider: when audio_format=ogg_opus, the provider writes to a .ogg path (not the caller-suggested .mp3) and returns the actual path in file_path.

Closes

(no open issue — happy to split / revert / rework if you prefer a different architecture)

[Hypnus-Yuan]

Closing this PR — saw that #17843 (command-type provider registry) shipped
the same day from a different angle, and it subsumes most of what this PR
was trying to solve for TTS extensibility.

Migration path for the Volcengine work:

I've reshaped the 868-line Volcengine WebSocket-streaming client into a
thin CLI that plays nicely with the #17843 config-only interface:

tts:
  provider: volcengine
  providers:
    volcengine:
      type: command
      command: 'python ~/.hermes/scripts/volcengine_tts.py
                --text-file {input_path} --out {output_path}
                --voice {voice} --speed {speed} --format {format}'
      output_format: mp3
      max_text_length: 1024
      timeout: 30

This works for non-streaming use cases (cold-start ~1-2s per call to open
the WebSocket). If a native Python-ABC plugin path is ever wanted for
low-latency streaming backends (Volcengine bidi-stream, MiniMax realtime,
Azure Speech WS, etc.), the TtsProvider + tts_registry + PluginContext
scaffolding in this branch is 100% mirrored on the existing image_gen
plugin system and can be revisited then. No hard feelings closing it — just
wanted to note the scaffolding is available if it becomes useful.

Branch stays on my fork at Hypnus-Yuan/hermes-agent:feat/pluggable-tts-backends
for reference.