Every render measurement so far has used a 2-unit training scenario, and a 2-unit scenario says a match performs well. It does not scale like one. Three scenarios holding everything still except unit count and board area - 2v2 on two map sheets, 4v4 and 4x4 (16 units) on four - plus scripts/perf/ stats.py, which summarises Suramadu's per-frame stats out of the container log with percentiles rather than medians. Percentiles are the point. Across the three, total bytes for the same scripted sweep move 1%, and the p90 frame goes 20KB -> 197KB -> 299KB. A median hides that completely.