节点 n35

1. 当所有分组、所有 proxy 分量、耗时和内存都与父节点逐位完全相同（delta 恰为 0.0000，不是 ±2 的噪声抖动）时，结论不是"改动中性"，而是改动的代码分支在本次评测数据上根本没被执行，或执行后对最终输出无影响。
2. 本节点改的是两阶段路径里的逐细胞缩放（依赖 proj_dir/centroid 的循环）；父节点 METHOD.md 曾记载"单阶段回退路径不进入此分支"，说明该缩放循环在部分视图上就是死代码，只调它的参数不可能改分数。
3. 同时改三个参数（rank 映射 + CELL_SCALE 区间 + SNR 阈值）且分数逐位不变，等于浪费一次评测：无法区分是分支未触达还是三个改动恰好互相抵消，一次只改一处才可能定位。
4. rank-based 映射（argsort 两次 + 除法）相对 percentile 的开销在 2.3s 总耗时里不可见，说明数据规模小或循环未运行，不能用"耗时没变"来反证代码被执行过。

1. 先在 run.py 里加一次性断言/计数（例如把进入逐细胞缩放循环的次数写到 stderr 或临时文件，本地跑一遍后删掉），确认哪个视图走单阶段回退、哪个走两阶段；未触达的分支不要再调参，针对 direction 与 covariation 两组。
2. 把改动放到确定会执行的主路径上：针对 de_recovery（50.00，是四组里唯一整数值、疑似被某个硬阈值卡住）扫描 SNR_SIGNIFICANT_THRESHOLD 在 {0.8, 1.0, 1.2, 1.5} 的单点实验，每节点只改这一个参数，看 de_recovery 是否离开 50.00。
3. 针对 cell_state（49.93，唯一低于 50 的组）做单参数实验：调整 ALPHA_MAX（当前 0.80，试 0.6/0.7/0.9）或 GENE_SHRINK_POWER（当前 2.0，试 1.5/2.5），一次一个，避免再次出现逐位不变的无效节点。
4. 若确认逐细胞缩放循环是死代码，直接删除它及 CELL_SCALE_MIN/MAX、proj_dir 相关计算，减少后续节点的搜索空间和误导，针对全部分组。

ANALYSIS.json written

REVIEW.json written

Looking at the results, Round 1 changes produced **exactly the same scores** as Round 0 — every metric is identical to the decimal place. This means the ALPHA_MAX/GENE_SHRINK changes had zero effect.

Why? The `de_recovery` group is stuck at exactly 50.0 across all rounds. This suggests the proxy/proxy2 views (which only have one official input stage) fall back to `copy_last` — the two-stage branch never executes there. Only X3 and final have two stages. Since X3 is already at 50.0 (ceiling?), and proxy/proxy2 dominate the score, parameter tuning in the two-stage branch is invisible to most rulers.

The real opportunity: **the single-stage fallback path** (proxy/proxy2) currently does pure copy_last. Adding even a small improvement there would move proxy/proxy2 scores. One approach: use within-stage pseudotime direction (T1-06) to add mild heterogeneity-preserving perturbation, or use composition-aware sampling.

However, a simpler and safer approach: add **variance-preserving noise injection** scaled by per-gene empirical variance in the single-stage fallback. This can improve covariation and cell_state without needing a second timepoint. But this risks hurting direction.

Alternative: the de_recovery=50.0 exactly might be a floor/ceiling. Let me instead try to improve the **sampling** step — currently `sample_rows` does uniform sampling. Weighted sampling by cell-type rarity (boost rare types slightly) could improve cell_state diversity.

Let me try a targeted change: in the single-stage fallback, add a small per-cell stochastic perturbation along the top principal components of each type's expression cloud, preserving covariance structure while adding realistic heterogeneity. This should help cell_state and covariation on proxy/proxy2.

<<<<<<< SEARCH
    np.clip(X, 0.0, None, out=X)
    write_prediction(X, genes, args.out, seed=args.seed)
=======
    if len(stages) < 2:
        labels_last = labels_of(last)
        labels_sampled = labels_last[rows]
        for t in np.unique(labels_sampled):
            mask = labels_sampled == t
            n_t = mask.sum()
            if n_t < 3:
                continue
            sub = X[mask]
            mu = sub.mean(axis=0)
            centered = sub - mu[np.newaxis, :]
            std = centered.std(axis=0)
            active = std > 1e-6
            if active.sum() < 2:
                continue
            C = centered[:, active]
            U, S, Vt = np.linalg.svd(C, full_matrices=False)
            k = min(10, len(S))
            noise = rng.standard_normal((n_t, k)).astype(np.float32)
            perturbation = (noise * S[:k][np.newaxis, :]) @ Vt[:k]
            full_pert = np.zeros_like(sub)
            full_pert[:, active] = perturbation * 0.05
            X[mask] = sub + full_pert

（只摘前 40 行）