Remaining cells are still being collected or verified.
Protocol & interpretation
¹ Paper imbalance bounds are [(0.5 − p)W, (0.5 + p)W], with p = 0.02 / 0.10 (KaHyPar epsilon 0.04 / 0.20). Each candidate receives one native V-cycle. Improvement is 100 × (hMETIS − MagicPart) / hMETIS; the headline mean is the arithmetic mean across 18 instances, not a ratio of cut sums. hMETIS is a published reference, not a rerun on this machine. Local runs used 12 parallel workers; no speed comparison is claimed.
Excluded finals remain recorded and never compete for a winner.