节点 n11

Looking at the results from Round 1, increasing clip-c to 0.75 and lowering match-rmin to 0.5 *hurt* the score (-0.55). The expression_change dropped from 68.64 to 66.60, and cell_state from 52.58 to 51.83. This suggests the current parameters are already well-tuned and larger deltas or more matches introduce noise.

Let me think about what else to try. The shape_scale is the weakest component (47.74). The METHOD.md notes that comp resampling changes the coordinate cloud, costing shape_scale. One parameter-level adjustment: reduce the composition interpolation weight `w_c` by scaling it down. Currently `w_c = clip((t_tgt - t_a) / (t_o - t_a), 0, 1)`. If I multiply by a factor < 1 (e.g., 0.75), the composition shift is less aggressive, which should preserve more of the anchor's spatial structure (helping shape_scale and local_spatial) while still getting some composition benefit.

The comp mechanism gave a huge boost (52.05 → 56.51), so I don't want to disable it, but a gentler blend might recover some shape_scale while keeping most of the expression/local gains.

CHANGE: comp weight scaling factor 1.0→0.75, reducing composition interpolation aggressiveness to recover shape_scale while retaining most expression/local gains.

<<<<<<< SEARCH
            w_c = float(np.clip((t_tgt - t_a) / (t_o - t_a), 0.0, 1.0)) if t_o != t_a else 0.0
=======
            w_c = float(np.clip((t_tgt - t_a) / (t_o - t_a), 0.0, 1.0)) * 0.75 if t_o != t_a else 0.0
>>>>>>> REPLACE