What simulation fidelity level (agent realism, dialogue authenticity, decision distributions) accurately predicts outcomes for specific deliberative formats under real-world constraints, and where does increased fidelity stop improving predictive value?