observability + evaluation

track retrieval agent model and execution behavior so CortexLab can improve from measured failures instead of visual guesswork.