Phase 3 Formulator: QLoRA adapter (r=32, alpha=64) trained on 1,215 formulation problems

initial commit