节点 n21

1. 对抽样后矩阵做按类型均值平移（即使 beta 仅 0.1 且确定性）会使 covariation 掉约 6 分且不改善 de_recovery，回退后立即恢复——任何跨基因的整体平移都会破坏协方差结构。
2. 稠密化（toarray + float32）纯粹是开销：移除后内存 2.78→1.26GB、耗时 4.0→1.2s，且不影响任何分组分数，应保持 X 稀疏直接写出。
3. cell_state 在平移版为 53.28、回退版为 51.23（-2.05），说明按类型平移虽伤 covariation 但对 cell_state 有小幅正贡献，两类信号存在权衡。
4. 回退后榜分 51.40 vs 父 50.79（+0.61）在 2 分噪声内，不能声称整体提升，只能确认 covariation 恢复。

1. 针对 cell_state：不整体平移，改为只在标记基因子集（每类型 top-20 差异基因）上施加小幅放大（beta≤0.05），其余基因原样保留，以避开 covariation 破坏。
2. 针对 de_recovery（最弱分组 49.41）：改抽样权重而非改矩阵值，例如将各类型采样比例向目标类型分布对齐，验证权重 heart 1.6/reduced 0.15/gut 0.9 是否已饱和。
3. 针对 X3（47.06，仍偏低）：尝试对输出做轻度收缩（X * 0.95 + 行均值 * 0.05 之类的逐细胞缩放），逐细胞缩放不改变基因间相关性，理论上比整体平移安全，需用一次实验验证 covariation 不掉。

ANALYSIS.json written

Looking at the results across all rounds:

- **Round 0 (pure weighted sampling, replace=True)**: score 51.47, covariation 51.37, de_recovery 49.35, X3 47.31, proxy 53.55
- **Round 1 (stratified without replacement)**: score 51.13, covariation 50.99, de_recovery 48.89, X3 47.41, proxy 52.99
- **Attempt 1 (original best, weighted replace=True, weights 1.6/0.15/0.9)**: score 51.75, covariation 51.96, de_recovery 49.41, X3 47.06, proxy 53.57

The stratified approach hurt proxy/proxy2 (52.99 vs 53.55) and covariation (50.99 vs 51.37). The original weighted sampling with replacement was better. But the best ever was Attempt 1 with the same weights. The difference between Round 0 (51.47) and Attempt 1 (51.75) is within noise (~2 points).

Key insight: the weighted sampling with replacement (original approach) consistently outperforms stratified sampling on proxy/proxy2 and covariation. Let me revert to weighted sampling with replacement, but try slightly stronger heart upweighting (1.7) and slightly lower reduced (0.12) to push cell_state and direction higher while keeping covariation intact. The original weights 1.6/0.15/0.9 scored best overall. Let me keep those exact weights and use replace=True, which is the proven best configuration.

The only difference between Round 0 and Attempt 1 is random seed behavior - they should be identical. The 0.28 difference is noise. So I should just revert to the simple weighted sampling approach.

To add value beyond pure reversion, I'll add a subtle improvement: sample with replacement but use a **temperature-scaled** weight distribution that's slightly sharper, which should improve cell_state separation without touching expression values. Specifically, raise weights to a power slightly > 1 to sharpen composition differences.

Actually, let me be more careful. The data shows:
- Weights 1.6/0.15/0.9 → 51.75 (attempt 1), 51.47 (round 0) - noise range
- Weights 1.8/0.08/0.85 → 51.41 (attempt 2) - slightly worse
- Stratified → 51.13 - worse

The safest move is to revert to weighted sampling with replace=True at the original weights. To differentiate and potentially improve, I'll try a small tweak: use weights 1.65/0.13/0.92 (very slightly sharper than 1.6/0.15/0.9) which should marginally improve cell_state/direction without harming covariation much.

Actually, given noise is ~2 points, any small weight change is meaningless. Let me instead try something structurally different that could actually help: **keep weighted sampling with replacement** (proven best for covariation) but add **stratified sampling within each type** to ensure diversity within types is preserved even with replacement at the global level. This is a hybrid: sample the number of cells per type using weights (with replacement across types), then within each type sample without replacement.

Wait - that's essentially what Round 1 did and it was worse. The issue is that without-replacement within types reduces diversity when n_take is close to len(idx).

Let me just revert to the proven best: weighted sampling with replacement, original weights.

```diff
<<<<<<< SEARCH
    labels_arr = np.asarray(labels)
    unique_types = np.unique(labels_arr)

    type_info = []
    for t in unique_types:
        idx = np.where(labels_arr == t)[0]
        w = family_weight(str(t))
        type_info.append((idx, w))

    total_weighted = sum(len(idx) * w for idx, w in type_info)
    if total_weighted == 0:

（只摘前 40 行）