CALM Before the STORM: Unlocking Native Reasoning for Optimization Modeling