节点 n10

Looking at the results:
- Round 0 (TOP_K=20): 51.002 (worse)
- Round 1 (alpha=2.0): 51.118 (worse than baseline 51.221)
- Baseline (alpha=1.6, TOP_K=10): 51.221

The current code is the parent's code with TOP_K=10 and ALPHA=1.6 (the baseline). Both my changes were harmful. Let me think about what else to try.

Key observations from the component scores:
- alpha=1.6: expr=50.54, state=49.22, shape=50.78, local=53.57
- alpha=2.0: expr=50.14, state=49.22, shape=50.78, local=54.33
- alpha=0.8: expr=50.11, state=50.43, shape=51.03, local=51.63

The cell_state component is the bottleneck at higher alpha. The variogram worsens with displacement. 

One idea: try alpha=1.2 (between 0.8 and 1.6). At alpha=0.8, cell_state was 50.43 (good), but local_spatial was only 51.63 (bad). At alpha=1.6, cell_state dropped to 49.22 but local_spatial jumped to 53.57. Maybe alpha=1.2 gives a better balance.

Another idea: try alpha=1.4, which is closer to 1.6 but might slightly improve cell_state.

Actually, let me think differently. The problem is that the displacement hurts cell_state (variogram). What if I try a smaller alpha like 1.2 that preserves more of the original expression while still getting some local_spatial gain?

From the data points:
- alpha=0.8: overall 50.80, local=51.63, state=50.43
- alpha=1.6: overall 51.09, local=53.57, state=49.22

The jump in local_spatial from 0.8→1.6 is +1.94, while state drops -1.21. The net is positive. Going from 1.6→2.0, local gains +0.76 but state stays same and expr drops -0.40. Net slightly negative.

Let me try alpha=1.2 as a middle ground. If local_spatial scales roughly linearly, at 1.2 we'd expect local ≈ 52.6, state ≈ 49.8, which might give a better overall.

Actually wait - let me reconsider. The baseline is 51.221 and I need to beat it by 1 point (52.221). That's a big jump. The alpha grid showed max 51.09. I need something more creative.

What about MIN_CELLS? Currently 20. If I lower it to 10, more types get displaced. But looking at the stderr, the skipped types are "no_orient" not below MIN_CELLS. So that won't help.

What about trying a different approach: use `t` (the interpolation parameter) to scale alpha? Currently t=0.4. What if alpha_effective = alpha * t or alpha * (1+t)?

Actually, looking more carefully at the code, `t` is read but only used in the stderr report. What if I scale alpha by some function of t? Like alpha * (1 + t) = 1.6 * 1.4 = 2.24? That's close to 2.4 which was worse.

Or alpha / (1+t) = 1.6/1.4 = 1.14? That's close to 1.2.

Let me try alpha=1.2. It's the most promising untested point between 0.8 and 1.6.

（只摘前 40 行）