Implement real-time content moderation for narration output using a
local safety model (Llama Guard 3 via Ollama). The pipeline is designed
to run in parallel with TTS preprocessing for near-zero added latency.
Key components:
- ContentModerator singleton with async check() method
- Game-context profiles (Morrowind vocabulary whitelist, fallback narrations)
- Configurable fail-open/fail-closed degradation when model unavailable
- Llama Guard response parsing (safe/unsafe with category codes)
- 40 unit tests covering profiles, parsing, whitelist, and async checks
Config settings: moderation_enabled, moderation_model, moderation_timeout_ms,
moderation_fail_open, moderation_game_profile
Fixes#987
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
3e5a3ac05f