Across 160 sampled CPU payloads, width stayed fixed at 48 while runs split among bottleneck 16 versus none and ortho lambda 0.0 versus 0.1. Bottlenecked runs usually pushed the bottleneck effective rank down near about 11.8 to 12.9, but their OOD accuracy still commonly sat only around 0.24 to 0.45 with gaps around 0.50 to 0.68; orthogonality-only runs without a bottleneck often kept much larger ranks around 33 to 35 and usually still showed large gaps around 0.65 to 0.73, with only scattered better runs near gap 0.42 to 0.50.