Remove old eval/aggregate_universal_v2.py (renamed to main_benchmark)

Remove old eval/universal_v2_evaluate.py (renamed to main_benchmark)

Rename universal_v2 -> main_benchmark (scripts, bench/ dir, env vars SF_MAIN_BENCH_*, README)

Add universal_v2 benchmark scoring (universal_v2_evaluate + aggregate) + 60-case reference structures/annotations/cases.yaml (paths scrubbed; CPU-only; reproduces hit flags 2594/2594)

Add benchmark reproduction: CPU scoring/eval (evaluate_prediction+batch_eval), reference structures, region manifests, headline prediction sets, reproduce_benchmark.py (reproduces main minority_hit_rate table; 15/15 cells verified by re-scoring 1440 PDBs)