AbstractVoice v0.10.3

Added

• Expanded the AbstractCore capability plugin discovery surface around the
  provider/model/voice abstraction: list_models(kind=..., provider=...),
  list_tts_models(provider=...), list_stt_models(provider=...),
  list_tts_voices(provider=..., model=...), list_cloned_voices(...), and
  richer voice_catalog() provider entries via tts_catalog_by_provider and
  stt_catalog_by_provider.
• Added combined provider:model selector support for both TTS and STT through
  the AbstractCore plugin, so callers can pass either split fields such as
  provider="openai", model="tts-1" or combined values such as
  openai:tts-1, faster-whisper:large, and
  transformers-asr:Qwen/Qwen3-ASR-1.7B.

Changed

• Clarified the public AbstractCore integration contract in the README, API
  docs, llms.txt, and llms-full.txt, including how to back clean
  OpenAI-compatible /v1/audio/voices, /v1/audio/speech/models, and
  /v1/audio/transcriptions/models discovery endpoints from the plugin
  surface.

Fixed

• Provider-scoped TTS model catalogs no longer treat synthetic catalog/profile
  ids as model ids, so AbstractCore selectors now expose only real provider
  models.
• Provider/model-specific TTS and STT requests now reuse a cached
  VoiceManager only when the active model actually matches the requested one,
  preventing model bleed between requests in AbstractCore/Gateway usage.
• Deduplicated provider voice/profile catalog rows before exposing them through
  the AbstractCore plugin, keeping UI consumers and OpenAI-compatible discovery
  endpoints simpler.