节点 n33

1. 四项指标与榜分全部 Δ=0.00（连小数位都相同）说明输出与父节点逐元素一致，即加权补齐分支在本任务里从未被触发：分型抽样已能凑满 n_target，改动是死代码，不产生任何信号。
2. 在改抽样配额逻辑前，应先离线统计 `len(rows)` 与 `n_target` 的实际差额（哪怕只打印一行），确认分支会被命中再动代码，否则一整轮预算浪费在等价重写上。
3. 回退到父节点已知最优参数（权重 1.6/0.15/0.9、无 EB 位移）能精确复现 52.11，这反证了打分是确定性的：Δ=0.00 可判为‘代码等价’而非‘噪声抵消’。
4. 把‘与父节点等价的重写’当作改动主体时，剩余的有效探索空间为零；每轮至少应包含一个能实际改变被选细胞集合或表达值的改动。

1. 针对全组：先在 run.py 里加一次性诊断输出（各型抽中数/目标配额、rows 总数 vs n_target、补齐名额数），确认哪一型是配额瓶颈，再决定是调权重还是调分层策略，避免继续改不生效的分支。
2. 针对 covariation（52.19，四组中偏低且对位移最敏感）：保持无位移，改试细胞内组成层面的调控，例如按 stage/供体做分层内配额（每个供体最少 k 个细胞），减少同一供体过度重复带来的伪相关，k 先试 2 和 5。
3. 针对 direction（54.01，已是最强组）与 de_recovery（51.33）：不要再动家族权重（已两轮证明 1.75/0.22/0.95 损害 cell_state、EB 位移损害 covariation），改为在固定权重下扫 n_target 规模（如 ±10%、±20%）看是否单调，以定位当前配额是否已在最优点。

ANALYSIS.json written

Looking at the round 1 results: score 51.97 vs parent 52.11. The weight-aware remainder filling didn't help (covariation dropped from 52.19 to 51.31, proxy dropped from 53.19 to 52.99). The direction improved slightly (54.12 vs 54.01) and cell_state improved marginally (51.19 vs 51.14).

Key observations:
- covariation 51.31 vs parent 52.19: the weighted remainder filling may be introducing composition noise
- The parent's simple uniform remainder filling was better for covariation
- direction 54.12 is the best we've seen

Strategy: Revert remainder filling to uniform (like parent), but add a new idea: **within-type expression-aware sampling**. Instead of uniform random sampling within each type, preferentially sample cells that are closer to the type's expression centroid (denser regions of the manifold). This should improve covariation by preserving the core distribution structure of each type, and may help cell_state by selecting more "typical" cells.

This is different from Program 2's distance-based binning (which was complex and didn't help). Here I use a simpler approach: sample with probability proportional to local density (inverse distance to centroid), which is a soft version that doesn't require binning.

Actually, wait - Program 2 tried distance-based binning and got worse covariation (50.79). Let me think differently.

Alternative approach: The issue might be that `include_external=False` causes X3 to fall back to `inputs_by_time(manifest)` which may include external stages. Let me look at the X3 score: 49.94 vs parent's 49.96 - essentially the same.

Let me think about what's actually different between this and the parent that could explain the covariation drop. The only change from parent is the weighted remainder filling. So I should revert that to uniform.

But I need to add something new to improve. Let me try: **soft temperature-based sampling within types**. Instead of uniform within-type sampling, use a Gaussian kernel density estimate approach - sample cells with probability proportional to how many neighbors they have (local density). This preserves the shape of each type's distribution better than uniform sampling, especially for small types where uniform sampling can miss dense regions.

Actually, a simpler and more targeted approach: increase `n_target` slightly. The `target_n_cells` function determines the target, but if we could output more cells (within the allowed range), we'd preserve more of the original distribution. But we don't control that.

Let me try another angle: **stratified sampling by expression quantiles within each type**. For each type, instead of pure random sampling, ensure we sample proportionally from different expression levels (e.g., split by median of top variable genes). This preserves the within-type distribution shape better.

Actually, the simplest improvement that's different from what's been tried: **remove the minimum clamp on n_per_type**. Currently `n_per_type = np.minimum(n_per_type, counts)` means if a type has fewer cells than its target allocation, we take all of them. But then the remainder is filled from remaining cells. What if instead, we redistribute the excess from over-allocated types to under-allocated ones proportionally? This is already partially done by the remainder step, but the current code just takes min and fills uniformly.

Let me try a cleaner approach: revert remainder to uniform (matching parent), and add **density-weighted within-type sampling** using a simple kNN density estimate for types with enough cells.

Actually, let me be more careful. The parent scored 52.11 with uniform remainder. My round 1 scored 51.97 with weighted remainder. The difference is small (0.14, within noise). But covariation dropped notably (52.19→51.31). Let me revert to uniform remainder and try something else entirely.

New idea: **gene-level variance-preserving subsampling**. For each type, instead of sampling cells uniformly, sample to preserve the gene-gene covariance structure. A simple way: use a farthest-point-like greedy selection or stratify by PCA. But this might be too complex.

Simplest safe improvement: revert remainder to uniform, and try adjusting the `n_target` calculation. If `target_n_cells` returns a value less than `n_total`, we're subsampling. What if we use a slightly different target that's closer to `n_total`? We can't override it, but we can check if `n_target >= n_total` and just output everything.

Actually, let me look at this from a different angle. The best scores in the table are around 52.1-52.5. The parent is at 52.11. The key insight from the experiment table is that node 31 (child of 11) got proxy 54.13 - the highest proxy score. But its overall was 51.98.

Let me try: revert to parent's exact behavior (uniform remainder), but add a **small per-cell noise injection** to break ties and avoid exact duplicates in the output. Wait, we're doing without-replace sampling so there are no duplicates.

OK, let me try a genuinely different approach: **adaptive family weights based on type size**. Small types get a boost (they're more likely to be undersampled), large types get slightly reduced. This is a form of rarefaction correction that should help cell_state for rare types without hurting common types.

<<<<<<< SEARCH

（只摘前 40 行）