Does eliminating the torus leak (diffused shadows → max filter, no x-wrap) break the memory-specificity trade-off?
Finding: Direct-material co-presence (max filter, no x-wrapping) eliminates the torus leak that created sim12's false boundaries. The initial 1-seed co-presence is <1% of the 2-seed value. But the max filter's hard cutoff creates a narrower boundary band than diffusion — at small radii (8-12), the boundary is too narrow to prevent merging. At medium radii (15-20), agent-wander material creates false positives. At radius=30, the b_scale normalization effect produces clean composition (2-seed coexist, 1-seed does not) — the first clean composition where the 1-seed control does not fire.
| Radius | 2-seed outcome | 2-seed stable | 2-seed cells | 1-seed outcome | Clean? |
|---|
| Mode | Seeds | L2 crossed | Coexist | Stable | H7 | Clean |
|---|