节点 n14

Looking at the results:
- Round 0: fallback_scale=0.6 hurt (-0.41), mainly expression_change dropped (60.80→59.52) and local_spatial dropped (54.03→52.97)
- Round 1: MIN_TYPE_CELLS=5 had zero effect (typed stayed at 5, meaning no additional types had ≥5 cells in both stages)

The parent's alpha=t=0.4 is already optimal per the METHOD.md analysis. The scores show alpha=0.3 gave 53.72 vs alpha=0.4 gave 53.75 — very close. Let me try a slightly different angle: increase alpha modestly to 0.5, which is between the tested 0.4 and 0.6. The METHOD.md shows alpha=0.6 gives expression_change 61.36 (higher than 60.80 at alpha=0.4) while cell_state drops to 48.31. At alpha=0.5 we might get some expression_change gain with less cell_state penalty.

Alternatively, I could try a blend: use a slightly higher alpha for the type-specific deltas while keeping the global fallback at the original alpha. This could improve expression_change (which benefits from larger displacement) while the fallback (which affects cell_state/local_spatial) stays moderate.

Let me try: scale the type-specific delta by 1.25x (effectively alpha=0.5 for those 5 types) while keeping fallback at alpha=0.4. The 5 typed cells get more precise displacement, potentially improving expression_change without hurting other metrics much since only 5/50 types are affected.

CHANGE: Scale type-specific delta by 1.25x (from alpha*t to 1.25*alpha*t) while keeping global fallback unchanged, to boost expression_change for the 5 well-characterized types without affecting the 45 fallback types that drive cell_state/local_spatial.

<<<<<<< SEARCH
        d = np.where(ok[:, None], alpha * (m_hi - m_lo), fb[None, :]).astype(np.float32)
=======
        d = np.where(ok[:, None], 1.25 * alpha * (m_hi - m_lo), fb[None, :]).astype(np.float32)
>>>>>>> REPLACE