Few-shot optimization closeout: prototype+combo heads; WM gain confirmed K<=5; honest K>=10 tree recovery; docs+logs updated

4B Flow zero-shot v2 (direction-cosine) confirmed to fail -> insight: transfer identity-invariance (WM), not transformation (flow)

Full autonomous run: 8 proposals with real results (few-shot WM wins K<=10; zero-shot via multi-LoD 0.67; noisy/cross-LoD solved by baseline) + all code, results, derived data