节点 n52

1. 在 T1 native 路线上，任何非零的伪批量 delta 位移（ALPHA=0.03 甚至 0.008）都会破坏类型内基因共变结构：把 ALPHA 归零后 covariation +1.11、X3 +0.68，direction 不受影响，验证了“delta 位移收益为噪声级、代价落在 covariation”的判断。
2. 以实验表里的 Program 1（无 delta）分数作为上界先验，能准确预测归零后的分组落点（covariation 52.19 预测 → 实测 51.91；X3 49.96 → 实测 50.00），先做小规模对照实验再改主程序是有效的。
3. 单组 +1.11 仍小于 T1 约 2 分的噪声，只有当多个相关组同向移动且与实验预测吻合时才可判定为真实改进；单看一个组的小幅变化不能下“有效”结论。
4. 参数关闭后遗留的死代码（size_factor 分母 50→120）不影响分数，但会让 diff 与 PLAN 描述互相矛盾；后续改动应同步说明父节点真实取值，避免 Researcher 误读谱系。

1. 针对 covariation（51.91，仍未到 Program 1 的 52.19 且是最弱组之一）：在 ALPHA=0 基线上做保结构的生成，如“真实细胞锚 + 类型内低秩残差”（方向库 T1-12），残差秩取 5~10 并按类型内方差缩放，避免逐基因不均匀缩放。
2. 针对 X3（50.00）与 cell_state（51.20）：在 N_BINS=12、MIN_CELLS_FOR_BINNING=15 上做网格（N_BINS ∈ {12,16}，阈值 ∈ {10,15}），只对细胞数≥阈值的类型启用分箱分层，小类型回退到按类型权重整体重采样。
3. 针对 direction（54.04，已是最强组）：不要再用 delta 位移去追它（收益在噪声内且损 covariation）；改为在 fw() 类型权重上做 ±20% 的灵敏度试验，确认其对 direction 的边际贡献，若小于 0.5 分则固定现值并把预算全部投给 covariation/X3。

ANALYSIS.json written

Looking at Round 1 results: ALPHA=0 gives covariation 51.10 (vs Program 1's 52.19 with replacement sampling). The stratified binning without replacement distorts within-type gene-gene correlations. Program 1's simple replacement sampling preserves the empirical joint distribution exactly in expectation.

Strategy: switch to within-type replacement sampling (preserves covariance) while keeping type-level proportion control via fw() weights (preserves direction).

<<<<<<< SEARCH
def stratified_sample(labels: np.ndarray, X, n_target: int, rng: np.random.Generator, tw: np.ndarray | None = None) -> np.ndarray:
    unique, counts = np.unique(labels, return_counts=True)
    n_total = len(labels)
    if n_target >= n_total:
        return np.arange(n_total)
    if tw is not None:
        wc = counts.astype(np.float64) * tw
        props = wc / wc.sum() if wc.sum() > 0 else counts / n_total
    else:
        props = counts / n_total
    n_per_type = np.floor(props * n_target).astype(int)
    remainder = n_target - n_per_type.sum()
    frac = (props * n_target) - n_per_type
    top_idx = np.argsort(-frac)[:remainder]
    n_per_type[top_idx] += 1
    n_per_type = np.minimum(n_per_type, counts)
    indices = []
    for u, n_take in zip(unique, n_per_type):
        if n_take <= 0:
            continue
        type_idx = np.where(labels == u)[0]
        if n_take >= len(type_idx):
            indices.append(type_idx)
            continue
        if len(type_idx) > MIN_CELLS_FOR_BINNING and n_take >= 5:
            sub = _submatrix(X, type_idx)
            centroid = sub.mean(axis=0)
            dists = np.linalg.norm(sub - centroid, axis=1)
            bin_edges = np.quantile(dists, np.linspace(0, 1, N_BINS + 1)[1:-1])
            bin_ids = np.searchsorted(bin_edges, dists)
            bin_props = np.bincount(bin_ids, minlength=N_BINS) / len(type_idx)
            per_bin = np.floor(bin_props * n_take).astype(int)
            rem = n_take - per_bin.sum()
            fr = (bin_props * n_take) - per_bin
            for i in np.argsort(-fr)[:rem]:

（只摘前 40 行）