节点 n6

Looking at the results, SHIFT=0.5 gave only +0.093 over baseline — within noise. Let me analyze what happened:

- expression_change dropped from 64.96 to 64.48 (lost DE benefit)
- cell_state improved from 63.14 to 63.93 (variogram improved: 0.03012 vs 0.03211)
- local_spatial slightly improved (52.32 → 52.39)

The variogram improved significantly with SHIFT=0.5 (skill 0.650 vs 0.635). The pure mix (SHIFT=0) scored 58.90 in the parent's testing. So the trend suggests going even lower on SHIFT might help, but we need to beat the parent baseline (58.495) by more than 1 point.

Looking more carefully: the parent baseline is 58.495 (which is the SHIFT=1.0 configuration averaged over scorer seeds). Pure mix scored 58.90 on seed 0. So SHIFT=0 might give ~58.9 average, which is only +0.4 over baseline — still not enough.

A different approach: increase MIN_TYPE_CELLS to filter out noisy type-specific deltas. With only 5 shared types and MIN_TYPE_CELLS=30, some types might have poorly estimated deltas. Raising to 50 or 100 would restrict shifts to only the best-estimated types, potentially keeping DE gains while reducing variogram damage.

But actually, looking at the numbers more carefully: the pure mix (SHIFT=0) already scores higher than any shifted version. The mechanism hurts more than it helps overall. The best strategy might be to set SHIFT=0 (pure mix) which scored 58.90 on s0. But that's only +0.4 over baseline 58.495.

Wait — I need to reconsider. The parent node's official score is 58.495. I need >59.495. Pure mix got 58.90 on a single seed. That's probably ~58.8-58.9 averaged, still not enough.

Let me think differently: what about the `nonzero` mask with SHIFT=0.5? The parent tested `add + 只移非零元` at SHIFT=1.0 and got 58.75. With SHIFT=0.5, the nonzero mask might preserve zeros better (helping variogram) while still providing some DE benefit.

CHANGE: T2HI_SHIFT from 0.5 to 0.0 (pure mix), since pure mix scored highest (58.90) in parent's tests and all shift variants degrade variogram/neighborhood more than they improve DE

<<<<<<< SEARCH
SHIFT = float(os.environ.get("T2HI_SHIFT", "0.5"))
=======
SHIFT = float(os.environ.get("T2HI_SHIFT", "0.0"))
>>>>>>> REPLACE

This reverts to pure mix which scored 58.90 on seed 0 in the parent's analysis, the highest of any configuration tested. The expected average should be ~58.8-58.9, which is still likely below the +1 threshold but represents the best known configuration. If this doesn't clear the bar, the node will be voided.