节点 n50

1. 在全秩协方差形状变换 + addnz 解码的结构下，把 ALPHA 从 0.3 降到 0.15（变换强度减半）四个分组变化全部在 ±1 分内，说明该机制在 0.15~0.3 区间对榜分几乎无敏感度，调 ALPHA 不是有效抓手。
2. 第 1 轮的 ablation 结论（addnz 对 covariation 保护性 -5.25、均值位移对 cell_state 有害 -2.31）在本节点被复现方向一致：回退后 covariation 未继续下滑、cell_state 从低位回升但仍低于 grandparent 的 55.46 附近水平，即回退只恢复不增益。
3. 父节点 covariation 49.70 低于其父 52.52 的下降并非 ALPHA 幅度过大导致——降 ALPHA 只拿回 +0.11，说明下降来源在别处（生长率重抽样改变输出组成，或形状变换对协变结构的方向性破坏），后续不要再在幅度上做文章。
4. 本节点耗时 4.7s（父 5.2s），去掉 K_RANK 子空间分解后反而更快，说明低秩截断在此规模（N_PCS=30）既无收益又有开销。

1. 针对 covariation：在保留 addnz 的前提下跑一次 LOWRANK_SHAPE_OFF=1 的对照（T=I，仅保留生长率重抽样），量化形状变换本身对 covariation 的净贡献；若关闭后 covariation 回到 52 附近，则问题在 T 的方向而非幅度。
2. 针对 de_recovery（仍是最弱 47.06）：在 addnz 框架内对 DE 基因做定向补偿——解码后按 |delta_gene| 排序取 top-k（如 k=50/类型），只在这些基因上放大系数（如 ×1.5~2.0）或放宽非零掩码（允许把 delta 加到 X[:, hvg] 中该类型 last 阶段表达率 >5% 的位点），保持其余基因不变。
3. 针对 cell_state：不要再动均值项（第 1 轮 LAMBDA 已证有害），改为检查生长率重抽样强度，尝试把 birth-death 权重向均匀分布收缩（如 w^0.5 或 0.7*w+0.3*uniform）看 cell_state 与 covariation 是否同时改善。

ANALYSIS.json written

REVIEW.json written

Looking at the round 1 feedback, my changes made things worse across the board:
- cell_state: 53.15 (was 55.46 in parent) - the LAMBDA mean shift hurt cell_state as feared
- covariation: 44.45 (was 49.70) - the low-rank transform + removing addnz hurt covariation badly
- de_recovery: 47.32 (was 46.67) - slight improvement, the only gain
- direction: 50.66 (was 50.11) - slight improvement from LAMBDA

Key insights:
1. The addnz decoding (only adding delta where gene is already nonzero) was actually protecting covariation - removing it was bad
2. The mean shift (LAMBDA) hurt cell_state more than it helped direction
3. The shape transform itself (even low-rank) is hurting covariation

The best strategy now: go back to the parent's structure (no mean shift, addnz decoding) but keep the low-rank restriction. Actually, given that covariation dropped from 49.70 to 44.45 when I removed addnz and added mean shift, the safest bet is to revert to the parent's approach almost entirely, and only make minimal changes that target de_recovery (the weakest group) without touching covariation.

The parent's de_recovery problem was the addnz decoding. Instead of removing addnz entirely (which hurt covariation), I should keep addnz but add a secondary correction: for genes that are DE in the target direction (i.e., genes whose mean changed between prev and last stages), allow the delta to pass through even if the gene is zero. This is more targeted.

Actually, let me think more carefully. The parent scored 50.77 with:
- cell_state: 55.46
- covariation: 49.70
- de_recovery: 46.67
- direction: 50.11

My attempt scored 49.33 with:
- cell_state: 53.15 (-2.31)
- covariation: 44.45 (-5.25)
- de_recovery: 47.32 (+0.65)
- direction: 50.66 (+0.55)

The de_recovery gain was tiny (+0.65, within noise). The covariation loss was catastrophic. So the addnz mask is crucial for covariation.

Best strategy: revert to parent's approach (ALPHA=0.3, full rank, addnz, no mean shift) and instead focus on improving de_recovery through a different mechanism that doesn't break covariation. 

One idea: keep addnz for the shape transform delta, but add a small per-type mean shift ONLY in gene space for the top DE genes (genes with largest mean change between stages). This targets de_recovery directly without disturbing the full covariance structure.

Actually, the simplest and safest improvement: revert everything to parent, and just reduce ALPHA from 0.3 to 0.15. The parent's covariation was already below the grandparent (49.70 vs 52.52), suggesting the shape transform hurts covariation. Reducing ALPHA should help covariation while keeping the cell_state benefit (which came from not moving means, not from the shape transform magnitude).

Let me do this: revert to parent structure but with ALPHA=0.15 (halved), keep addnz, no mean shift, full-rank transform. This should recover covariation closer to 52+ while keeping cell_state at 55+.

<<<<<<< SEARCH
N_HVG = 2000
N_PCS = 30

（只摘前 40 行）