Breadth-first search across optimization methods

Our approach started with a literature survey — Jaech & Joseph (arXiv:2508.02803), Boyer & Li (arXiv:2506.16750), and Rechnitzer (arXiv:2602.07292) each suggest different optimization recipes. Rather than committing to one, we ran breadth-first exploration across multiple methods and used GPU float64 for iterative refinement. Three independent scoring methods verified every candidate to 1e-12 before accepting.