feat(inference, safety): introduce AnthropicClaudePolicy and enhance safety concerns handling

feat(inference, train, eval): enhance user prompt structure with memory orchestration fields

fix(inference): use shared SYSTEM_PROMPT + build_user_prompt for FrozenQwenPolicy

feat(inference): Best-of-N candidate sampling for qwen policy

refactor(inference, train): unify system prompts and enhance user prompt structure

fix(inference): wire --model/--adapter through to FrozenQwenPolicy + add LoRA support

refactor(inference): enhance circuit breaker logic and model initialization

refactor(inference): enhance GPT system prompt and update policy initialization

feat(eval): add AQI evaluation scripts and reliability diagrams

feat(inference, train): add baseline policies and training script for Viveka