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

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

fix(train): preserve exact backwards compatibility when --resume is not passed

feat(train): add --resume flag to continue from latest checkpoint

fix(train): pass eos_token_id list via GRPOConfig.generation_kwargs (Qwen GRPO termination)

fix(train): pin chat-end token as EOS so Qwen generation stops on turn end

fix(prompts,train): kill multi-step example leakage that made Qwen ramble past EOS

fix(train): teacher uses real reversibility from registry + better last_result check

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

fix(train): revert num_generations to 4 — G=16 was 30hr ETA on T4

fix(train): bump grad_accum to 16 to match num_generations=16

feat(train): teacher-rollout multi-step trajectories + multi-step examples in prompt

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

fix(train): enable KL penalty (beta=0.04) to prevent mode collapse

fix(train+kaggle): warnings_issued patch for transformers 5.x + proper notebook shim

fix(train): filter LLM extras + bound scenario_idx to real per-tier count

fix(train): rewrite reward_func to TRL 0.24 signature + drop mergekit

fix(train): proper llm_blender shim + add mergekit shim for TRL 0.24

fix(train): stub llm_blender + unpin TRL/hub for GRPOConfig access

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

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