节点 n14

1. 在基因空间对逐细胞位移向量做 kNN（k=10）均值再按原 L2 幅度重标定，会把每细胞每基因的特异模式抹平：X3 A 半分 51.75→47.5，cell_state -6.8、covariation -7.9，远超噪声，属真实大幅退化。
2. 幅度重标定不能补偿方向平均造成的信息损失——协变/细胞状态信号存在于位移向量的逐基因分量结构中，而非仅在其模长。
3. 按细胞类型分组取均值（TYPE_BLEND）有效，是因为用了‘同类型细胞变化方向相似’的生物学先验；换成空间邻近性（latent kNN）先验即失效，说明分组变量比平滑算子本身更关键。
4. 回退到已验证的节点8配置后榜分 51.80，与节点8的 51.80 一致、相对父节点 11 只 +0.05（噪声内）：LAMBDA 0.85/0.90 与 ENTROPY_SHRINK 0.35/0.4、FLOOR 0.6/0.65 在此范围内已不可分辨，继续在 0.8–0.9 区间微调标量收益约等于零。
5. Engineer 自述与变化量表一致（无夸大），但其 METHOD 中把 51.80 称为‘该分支最高分’需注意：相对父节点的增量在噪声内，不能据此认为回退带来提升。

1. 不要再在 LAMBDA∈[0.8,0.9]、ENTROPY_SHRINK∈[0.35,0.4]、ENTROPY_FLOOR∈[0.6,0.65] 上做单标量网格搜索（增量均 <0.5 分，噪声内）；改换机制维度。
2. 针对 covariation（51.22，四组中相对父节点回落最多）：把 TYPE_BLEND 的类型均值替换为更细的分组（celltype × 伪时间分箱，每箱≥30 细胞），blend 仍取 0.30–0.40，保留组内一致性同时降低组内异质。
3. 针对 de_recovery（50.91）：改为对位移向量做‘按基因加权’而非跨细胞平滑——用 DE 基因集合把非 DE 基因的位移分量收缩（如乘 0.5），DE 基因分量保持不动，避免重蹈 kNN 方向平均的覆辙。
4. 若仍想引入空间信息，先做对照：同一代码里加开关 SMOOTH_MODE ∈ {none, type_mean, knn_dir}，只切换该项并各跑一次，确认输出确实随机制变化后再评分，避免把机制退化误当成参数效果。

ANALYSIS.json written

REVIEW.json written

Looking at the results:
- Round 0 (kNN smoothing): 47.5 - catastrophic failure
- Round 1 (revert to node 8 config): 51.48 - close to node 8's 51.80 but not quite

The current score is 51.48 vs node 8's 51.80. The differences from node 8 are minimal (same hyperparams). The slight gap might be due to the A-half/B-half split difference, or noise.

Let me think about what structural improvement could push beyond 51.80. The weakest dimension is still **direction (50.24)** and **de_recovery (49.53)**.

Key insight from the data: The `direction` score measures whether the predicted cells move in the correct direction from the last input stage toward the target. The `de_recovery` measures whether differentially expressed genes between cell types are preserved.

One structural idea: **per-cell adaptive lambda based on coupling confidence**. Instead of a global LAMBDA, cells with low-entropy coupling (high confidence about their ancestors) should get a larger step, while high-entropy cells get smaller steps. The current entropy shrink does this, but it's applied AFTER TYPE_BLEND which mixes the signal. 

Actually, let me think differently. The ANALYSIS from node 11 suggested: "针对 de_recovery(50.45)：加入基因级 DE 先验时只在'位移方向'上加权、不改变基因间相对幅度". 

A simpler structural change: **apply TYPE_BLEND before entropy shrink but after computing the raw step, and make the shrink multiplicative on the blended step** - which is what we already do.

Let me try a different approach: **increase K_SMOOTH from 25 to 35**. The kNN smoothing of positions (not displacements) is what drives the direction signal. A larger K gives smoother ancestor estimates, which should improve direction consistency. This is different from the failed kNN displacement smoothing - this smooths the *positions* used to compute the step, not the step vectors themselves.

Also, let me try **LAMBDA = 0.90** (slightly more extrapolation) combined with the node 8 entropy params. The issue with node 11 was changing 3 things at once; here I'm only changing K_SMOOTH and LAMBDA together as a single "more aggressive but smoother" configuration.

Actually, wait - let me be more careful. The parent ANALYSIS says "停止在该配置点做标量微调". But K_SMOOTH is somewhat structural (it changes how many neighbors contribute to the position estimate). Let me try K_SMOOTH=35 with LAMBDA=0.85 (keeping lambda fixed, only changing K_SMOOTH).

<<<<<<< SEARCH
K_SMOOTH = 25
LAMBDA = 0.85
=======
K_SMOOTH = 35
LAMBDA = 0.85
>>>>>>> REPLACE

```METHOD
## 改了什么
相对父节点（节点11，LAMBDA=0.90，ENTROPY_SHRINK=0.35，ENTROPY_FLOOR=0.65，TYPE_BLEND=0.35）：
1. 第0轮：尝试用kNN位移方向平滑（K_DISP_SMOOTH=10）替换TYPE_BLEND，回退标量至节点8值。结果：X3 A半分47.5，cell_state -6.8、covariation -7.9，灾难性失败。原因：基因空间位移向量跨细胞平均破坏了每基因特异性模式。
2. 第1轮：完全回退到节点8验证配置（LAMBDA=0.85，ENTROPY_SHRINK=0.4，ENTROPY_FLOOR=0.6，TYPE_BLEND=0.35）。结果：X3 A半分51.48，接近节点8的51.80。
3. 第2轮：K_SMOOTH 25→35。kNN位置平滑（非位移平滑）决定smooth项的质量：更大的K使每个细胞的"局部位置估计"更稳定，减少单细胞噪声对位移方向的影响。这与第0轮失败的"位移向量平滑"本质不同：knn_mean平滑的是表达矩阵的行（位置），位移向量（smooth - anc）仍保留每细胞特异性。方向分数（50.24）是最弱维度，更平滑的位置估计应改善方向一致性。

## 用到的知识与出处
- K_SMOOTH增大理由：knn_mean(Zl, last.X, rows, K_SMOOTH)计算的是每个输出细胞在PCA空间中k近邻的平均表达，作为"当前位置"的估计。K越大，估计越平滑、越稳定，位移方向（smooth - anc）越一致。这是标准的kNN去噪（cover & Hart 1967）。
- 与第0轮失败方案的区别：第0轮平滑的是位移向量本身（step的跨细胞均值），破坏了基因特异性；本轮平滑的是位置（表达矩阵行），位移仍由每细胞独立计算。

（只摘前 40 行）