📉 Phase 4 — MuSGD

WP-033 — toy convergence golden

MuSGD reaches the 0.1 MSE threshold in 47 steps against momentum-SGD's 90 on a fully seeded micro-CNN regression task.

The tolerance is the interesting part: step counts are pinned at ±15, final losses at 0.005. Cross-platform torch float drift shifts trajectories slightly (Float reproducibility), so a tight pin would be a golden about the machine. The 43-step separation is what keeps the directional claim decisive inside a tolerance that loose — the golden asserts the ordering it cares about rather than the number it happened to get.