节点 n21

1. 以 diff 为准而非 PLAN：本节点 PLAN/Engineer 描述的是中位数 VAR_FLOOR=0.1 门控，实际落地的是均值 sqrt(var) clip[0.25,2.0] cap 1.0 的权重，且所谓回退在 diff 中不存在——文档与代码严重脱节。
2. 在已有 SNR 收缩之上再乘一个仅作用于低方差基因、且上限为 1.0 的弱权重，对全部指标影响 <0.05，完全淹没在 ~2 分噪声里；这类小幅乘法扰动不足以移动榜分。
3. 方差门控针对的 covariation（父节点最弱 46.18）只涨 0.05，说明单靠压缩低方差基因位移并不能改善协方差结构。

1. 针对 covariation/X3：不要再叠乘弱权重，改测能改变位移量级的正交维度——对 GENE_SHRINK_POWER 做网格 {1.0,1.5,2.0,2.5}，单独跑并对比 X3 与 covariation。
2. 针对 covariation：把收缩信号从『低方差少动』换成显式协方差保持（如仅在 delta 与基线协方差主方向对齐时保留位移），先小样验证再全量。
3. 针对流程：后续节点务必让落地代码与 PLAN 描述一致，并在提交前用 diff 自检，避免出现本节点这种文档/代码不符且效果被噪声淹没的无效改动。

ANALYSIS.json written

Looking at the results across rounds:

**Parent (baseline):** score 48.83, X3 46.39, covariation 46.18, de_recovery 49.15
**Round 0:** score 48.68, X3 45.25, covariation 45.55, de_recovery 48.83
**Round 1:** score 48.56, X3 44.86, covariation 46.89, de_recovery 47.24

Key observations:
1. **covariation improved** in Round 1 (46.89 vs parent 46.18) - the anti-clipping logic helps!
2. **X3 keeps dropping** (46.39 → 45.25 → 44.86) - the time_factor and anti-clipping are hurting X3
3. **de_recovery dropped significantly** (49.15 → 47.24) - the anti-clipping reduces the effective delta, losing DE signal
4. **proxy/proxy2 stable at 50.40** (slight improvement from parent's 50.04, likely noise)

The anti-clipping helps covariation but hurts de_recovery and X3. The problem is that scaling down negative deltas reduces the overall magnitude of changes, which hurts DE recovery.

**Strategy for Round 2:** Remove the anti-clipping entirely (it hurts more than it helps overall), remove time_factor (it's neutral for T1 and harmful for X3), and instead focus on what actually improved: the proxy score went up slightly. 

Actually, let me reconsider. The best approach is to get back to the parent's behavior (which scored 48.83) and make a targeted improvement. The parent's weakest group is covariation (46.18). 

The key insight: the anti-clipping helped covariation (+0.71) but hurt de_recovery (-1.91) and X3 (-1.53). A better approach: apply anti-clipping only when it doesn't reduce the delta too much. Set a minimum scale threshold - only clip-protect when the scale factor is above some threshold (e.g., 0.5), otherwise allow the clip.

Actually, a simpler approach: revert to parent behavior exactly (remove time_factor, remove anti-clipping), and instead try a different covariation improvement: **variance preservation**. After shifting, rescale each gene's within-type variance to match the original variance. This preserves the spread of the distribution without preventing clipping.

Wait - but the shift is constant per type, so within-type variance is already preserved! The issue must be about between-type covariance structure or the clipping effect.

Let me think about what actually affects the covariation metric. It likely measures gene-gene correlations across the whole population. When we shift different types by different amounts, we change the between-type component of covariance. The clipping at 0 creates artificial correlations (genes that are both pushed to 0 become correlated).

Better idea: **remove the anti-clipping but add a soft floor instead of hard clip**. Use `softplus`-like behavior near zero: `x = log1p(exp(x) - 1)` won't work since we're already in log1p space. Instead, use a gentle floor: `X = np.maximum(X, 0.0)` is what we have. The issue is the hard boundary.

Actually, the simplest winning change: revert everything to parent and just reduce ALPHA_MAX slightly to 0.65. The parent already has good scores; the issue is that aggressive shifts hurt covariation. A slightly lower alpha should help covariation without hurting de_recovery much.

But wait - I need to think about what's different between proxy (single stage, copy_last) and the two-stage case. For proxy, the code just does copy_last and gets 50.04-50.40. For X3 (two stages), the parent gets 46.39. The delta application is hurting X3 relative to what pure copy would give.

Let me try: revert to parent exactly, but with ALPHA_MAX=0.6 (more conservative). This should help covariation and X3 by not over-shifting, at small cost to direction.

```diff
<<<<<<< SEARCH
ALPHA_MAX = 0.75
ALPHA_MIN = 0.1
SIGN_THRESHOLD = 0.05
GENE_SHRINK_POWER = 1.5

（只摘前 40 行）