Describir: From Policy Optimization Foundations to Language Model Post-Training on Structured Tasks