节点 n18

Looking at the results from round 0:
- `base_gain=1.3` gave 61.823 vs baseline 61.857 — slightly worse (-0.034).
- expression_change went 55.39→55.48 (+0.09) but cell_state dropped 54.21→54.13 (-0.08), and local_spatial dropped 52.28→52.13 (-0.15), neighborhood_mmd skill dropped 0.523→0.521.

So increasing base_gain helps expression_change slightly but hurts other groups. The net is negative. I should try the opposite direction or a different parameter.

Let me try `aniso_damp=2.2` instead of 2.5. From the grid: 2.2 gives 61.21 on seed 0 vs 2.5 gives 61.26. That's worse on seed 0, but the multi-seed behavior might differ. Actually, the baseline (61.857) is already the multi-seed average at damp=2.5, so moving to 2.2 would likely be worse.

Better idea: try `T2_SPREAD` from 0.10 to 0.12. The spread parameter controls the non-uniform (position-proportional) part of the expression shift. Since cell_state (54.21) measures how well individual cell states match, a slightly stronger spread might better differentiate cells along the trajectory. But this is risky.

Actually, let me reconsider `scale_damp`. Currently 0.5. The scale_log_ratio skill is 0.924 — already high but not maxed. If I try scale_damp=0.4 or 0.6, it might improve scale_log_ratio slightly. But the METHOD.md says 0.5 is optimal.

Let me try `aniso_damp=2.3` — between 2.2 and 2.5, slightly more conservative. The grid shows 2.2→61.21, 2.5→61.26 on seed 0. The difference is tiny (0.05), well within noise. But on multi-seed, 2.3 might be slightly more stable since it's less likely to overshoot on the final view (γ=2.3×1/3=0.767 vs 0.833).

Actually, a more promising direction: try `T2_SPREAD=0.12`. The spread controls per-cell heterogeneity in the expression advance. A slightly higher spread could improve cell_state by better matching individual cell trajectories. The expression_change group already responded positively to base_gain increase, suggesting there's room in the expression direction.

Let me try `spread=0.12` (from 0.10):

CHANGE: Increase T2_SPREAD from 0.10 to 0.12 to strengthen position-proportional expression heterogeneity, targeting cell_state improvement.

<<<<<<< SEARCH
    spread = 0.0 if mech_off else _env_float("T2_SPREAD", 0.10)
=======
    spread = 0.0 if mech_off else _env_float("T2_SPREAD", 0.12)
>>>>>>> REPLACE