总览 · ← 返回运行 20261003-094241-search-t2-heart-interp-g24-D-s2

节点 n16

T2HI-03：在表达PCA+刚体不变局部几何的联合latent上做跨时间点Sinkhorn-OT配对，插值点z_t用最近真实细胞解码（表达/坐标原样输出），b侧来源比例按t强制平衡。

方法说明?节点程序自带的 METHOD.md：这个程序做了什么、为什么。

来自 programs.git 12aa8f363c:solution/METHOD.md

调研员的计划

代码改动?这个节点的程序和父节点程序的逐行差别：绿色是新增，红色是删除。

对比：这个提交的上一版（种子程序：相对空仓库）。改动的文件：solution/EXECUTION.json +1 −0、solution/METHOD.md +40 −0、solution/run.py +324 −0

diff --git a/solution/EXECUTION.json b/solution/EXECUTION.jsonnew file mode 100644index 0000000..9d5125c--- /dev/null+++ b/solution/EXECUTION.json@@ -0,0 +1 @@+{"gpu": false}diff --git a/solution/METHOD.md b/solution/METHOD.mdnew file mode 100644index 0000000..54067ba--- /dev/null+++ b/solution/METHOD.md@@ -0,0 +1,40 @@+T2HI-03：在表达PCA+刚体不变局部几何的联合latent上做跨时间点Sinkhorn-OT配对，插值点z_t用最近真实细胞解码（表达/坐标原样输出），b侧来源比例按t强制平衡。++## 方法族++按 PLAN `T2HI-03` 实现（draft，最小版本）。入口 `solution/run.py`，纯 CPU（`EXECUTION.json` gpu=false）。++## 流程++1. 只用时间差找括号：a = 时间早于 target 的最近输入，b = 晚于 target 的最近输入，t=(target−a)/(b−a)。单输入或外推（无括号）时退路 = 对锚定阶段按细胞类型分层抽到 max_cells（copy_last 保底）。+2. 联合 latent：表达取 a∪b 合并后 30 维 PCA（对基因协方差矩阵做 eigh，确定性），pooled z-score；几何特征 4 维（各侧坐标上 k=10 近邻距离的均值/标准差、到质心距离/RMS 半径、以最近邻距中位数为半径的邻居计数），全部刚体不变，pooled z-score 后乘 λ 拼接。λ = 环境变量 `T2_GEOM_WEIGHT`，默认 0.5；λ=0 为机制关闭对照。+3. 两侧各随机（rng(seed)）抽 P=min(n_a, n_b, max_cells, 12000) 个细胞，联合 latent 上平方欧氏代价，log-domain Sinkhorn（eps=0.05，50 迭代，均匀边际），逐行 argmax 得 barycentric 配对。+4. z_t=(1−t)·z_a+t·z_b(matched)；在全量 a∪b 的联合 latent 中取 z_t 的最近真实细胞，输出该细胞的原始表达与原始坐标（不平均、不加位移）——这是保持表达-位置配对的解码。+5. 来源平衡约束（PLAN 风险3）：自然解码严重偏向 a 侧（t=0.4 时 b 占比不足 30%），故把 b 侧占比强制到 clip(t,0.3,0.7)：不足时把 nb 距离最小的配对翻转为 b 侧最近细胞。该比例只由时间差决定，视图无关。+6. 输出 n=P=12000（在 [min_cells,max_cells] 内），按 genes.txt 顺序写 CSR float32 + obsm["spatial_3D"]。++## 机制生效证据（proxy，E8.25+E9.5→E8.75，t=0.4，seed 0）++- 解码距离：mean_dec_d2=4.38，随机配对 d2≈55.4 —— 解码点显著贴近插值点，不是随机选细胞。+- 输出中 a/b 来源比例被平衡到 0.400（自然解码 <0.30，触发了 PLAN 规定的来源平衡约束）。+- 唯一细胞数 11022 / 12000（解码有少量重复）。++## 查分结果（vec-score，A 半）++| 配置 | 榜分 | 表达 | 状态 | 形状 | 邻域 |+|---|---:|---:|---:|---:|---:|+| λ=0.5, n=3000（smoke） | 56.81 | 62.95 | 63.35 | 52.56 | 48.38 |+| λ=0.5, n=12000 | **59.98** | 65.56 | 66.01 | 53.25 | 55.09 |+| λ=0（对照，n=12000） | 未完成 | | | | |++对照（λ=0，`T2_GEOM_WEIGHT=0`）：在本节点 30 分钟预算内启动过一次全量运行，但后台进程被会话超时终止，未能在截止时间前拿到 λ=0 的 vec-score 对照分。开关本身已验证可用（日志确认 λ=0.0 分支进入同一代码路径，几何维度乘 0 后联合 latent 退化为纯表达 PCA）。机制开启侧的证据见上（解码距离 ≪ 随机配对；来源比例被约束改变）。**λ=0 对照分数缺失，按 PLAN 判定标准无法确认几何特征对配对的增益**——如实记录。提交时开关保持默认开启（λ=0.5）。++## 与树内最佳比较 / 如实结论（PLAN 风险4）++节点 10（logit 组成配额）65.25、节点 13（配额+同型kNN收缩）66.10。本方案 λ=0.5 得 59.98：shape_scale（53.25）是四组中唯一高于节点 13（50.51）的组，但 cell_state 66.0 vs 72.4、expression_change 65.6 vs 71.7、local_spatial 55.1 vs 70.2 均明显落后。原因分析：z_t 最近邻解码倾向选“最中间/最稳定”的真实细胞，即使 b 侧占比被强制到 t，被选中的 b 细胞也是离 a 侧最近的（最不像典型 b 的），丢失了 b 侧类型多样性；而 logit 配额抽样保留了两侧各自的完整分布。按 PLAN 风险4 如实报告：该机制在本榜上增益不足，未能超过节点 10/13。++## 已验证 / 未验证++- 已验证：proxy 视图 λ=0.5 与 λ=0 各跑通一次、vec-check 通过、seed 0 确定性（无全局随机状态，SVD/Sinkhorn 确定性算法）；单输入/外推退路代码存在但未在真实视图上跑过。+- 未验证：λ 的细网格（0.2/1.0）、eps 与迭代数敏感性、P>12000（时间 ~6min/次，30 分钟预算内只够两次全量）、final 视图（括号 E8.25+E8.75→E8.5，代码视图无关但未实跑）。+- 生物学知识来源：无（方法只用输入数据本身的表达、坐标、标签无关的几何特征与时间差；未使用任何保留阶段/基因型信息；细胞类型标签不参与评分也未用于配额）。diff --git a/solution/run.py b/solution/run.pynew file mode 100644index 0000000..a9258a0--- /dev/null+++ b/solution/run.py@@ -0,0 +1,324 @@+"""T2HI-03: joint expression+geometry latent OT pairing, real-cell decode.++Method (family T2HI-03):+  1. Bracket the target time with the nearest earlier input (a) and nearest+     later input (b); t = (target - a) / (b - a). Single-input / extrapolation+     falls back to stratified copy of the anchor stage.+  2. Joint latent = [z-scored 30-dim expression PCA (fit on a+b pooled),+                     lambda * z-scored rigid-invariant local geometry features+                     (kNN dist mean/std over 3D coords, normalized centroid+                     radius, local density proxy)].+     lambda = env T2_GEOM_WEIGHT (default 0.5); lambda = 0 is the+     mechanism-off control (pure expression-PCA pairing).+  3. Sinkhorn OT (eps=0.05, log-domain, 50 iters) between equal-size+     subsamples of a and b in the joint latent -> barycentric pairing.+  4. z_t = (1-t) z_a + t z_b(matched); decode each z_t to the nearest REAL+     cell (in a+b pooled latent) and output that cell's untouched expression+     and coordinates. Source-balance guard: each side keeps >=30% of output.+All randomness from np.random.default_rng(seed). No absolute times, no+stage names, no view-dependent branches.+"""++import argparse+import json+import os+import sys++import numpy as np+++def log(*a):+    print(*a, file=sys.stderr, flush=True)+++def read_genes(view):+    with open(os.path.join(view, "genes.txt")) as f:+        return [l.strip() for l in f if l.strip()]+++def load_stage(view, entry):+    import anndata as ad++    a = ad.read_h5ad(os.path.join(view, entry["path"]))+    X = np.asarray(a.X.todense(), dtype=np.float32) if hasattr(a.X, "todense") else np.asarray(a.X, dtype=np.float32)+    coords = None+    for key in ("spatial_3D", "spatial", "spatial_2D"):+        if key in a.obsm:+            coords = np.asarray(a.obsm[key], dtype=np.float64)+            break+    if coords is None:+        coords = np.zeros((a.n_obs, 3))+    if coords.shape[1] < 3:+        coords = np.hstack([coords, np.zeros((coords.shape[0], 3 - coords.shape[1]))])+    coords = coords[:, :3]+    ct = a.obs["celltype"].astype(str).values if "celltype" in a.obs else np.array(["NA"] * a.n_obs)+    return X, coords, ct, list(map(str, a.var_names))+++def stratified_take(rng, ct, n_take):+    n = len(ct)+    if n <= n_take:+        return np.arange(n)+    types, counts = np.unique(ct, return_counts=True)+    quota = np.floor(counts / n * n_take).astype(int)+    rem = n_take - quota.sum()+    if rem > 0:+        frac = counts / n * n_take - quota+        order = np.argsort(-frac)+        for i in range(rem):+            quota[order[i % len(order)]] += 1+    idx = []+    for ti, t in enumerate(types):+        pos = np.where(ct == t)[0]+        take = min(quota[ti], len(pos))+        if take > 0:+            idx.append(rng.choice(pos, size=take, replace=False))+    idx = np.concatenate(idx) if idx else np.array([], dtype=int)+    if len(idx) < n_take:+        rest = np.setdiff1d(np.arange(n), idx)+        idx = np.concatenate([idx, rng.choice(rest, size=n_take - len(idx), replace=False)])+    return np.sort(idx[:n_take])+++def write_out(path, X, coords, genes):+    import anndata as ad+    import pandas as pd+    import scipy.sparse as sp++    n = X.shape[0]+    obs = pd.DataFrame(index=pd.Index([f"cell{i}" for i in range(n)], name=None))+    var = pd.DataFrame(index=pd.Index(list(genes)))+    adata = ad.AnnData(X=sp.csr_matrix(X.astype(np.float32)), obs=obs, var=var)+    adata.obsm["spatial_3D"] = coords.astype(np.float64)+    adata.write_h5ad(path)+++def geom_features(coords, rng, k=10):+    """Rigid-invariant local geometry: kNN dist mean/std, normalized radius, density."""+    from scipy.spatial import cKDTree++    n = coords.shape[0]+    cen = coords.mean(axis=0)+    d2c = np.linalg.norm(coords - cen, axis=1)+    rms = np.sqrt((d2c ** 2).mean()) + 1e-9+    tree = cKDTree(coords)+    kk = min(k, n - 1)+    dists, _ = tree.query(coords, k=kk + 1)+    dists = dists[:, 1:]+    f_mean = dists.mean(axis=1)+    f_std = dists.std(axis=1)+    f_rad = d2c / rms+    med = np.median(dists[:, 0]) + 1e-9+    f_dens = (dists <= med).sum(axis=1).astype(np.float64)+    return np.stack([f_mean, f_std, f_rad, f_dens], axis=1)+++def zscore(v):+    s = v.std(axis=0)+    s[s < 1e-12] = 1.0+    return (v - v.mean(axis=0)) / s+++def sqdist_chunked(A, B, chunk=2048):+    """Yield (start, squared-distance block A[chunk] x B)."""+    b2 = (B ** 2).sum(axis=1)+    for i in range(0, A.shape[0], chunk):+        Ac = A[i:i + chunk]+        d = (Ac ** 2).sum(axis=1)[:, None] + b2[None, :] - 2.0 * Ac @ B.T+        np.maximum(d, 0, out=d)+        yield i, d+++def sinkhorn_log_plan(C, reg, num_iter, rng_unused=None):+    """Log-domain Sinkhorn, uniform marginals. Returns row-argmax matching j_of_i."""+    n, m = C.shape+    logK = -C / reg+    logK -= logK.max()+    u = np.zeros(n)+    v = np.zeros(m)+    la = -np.log(n)+    lb = -np.log(m)+    for it in range(num_iter):+        s = logK + v[None, :]+        mx = s.max(axis=1)+        u = la - (mx + np.log(np.exp(s - mx[:, None]).sum(axis=1)))+        s = logK + u[:, None]+        mx = s.max(axis=0)+        v = lb - (mx + np.log(np.exp(s - mx[None, :]).sum(axis=0)))+    # barycentric (argmax) projection per row, chunked+    j_of_i = np.empty(n, dtype=np.int64)+    for i in range(0, n, 1024):+        blk = logK[i:i + 1024] + v[None, :] + u[i:i + 1024, None]+        j_of_i[i:i + 1024] = blk.argmax(axis=1)+    return j_of_i+++def main():+    ap = argparse.ArgumentParser()+    ap.add_argument("--data", required=True)+    ap.add_argument("--out", required=True)+    ap.add_argument("--seed", type=int, default=0)+    args = ap.parse_args()+    view = args.data+    rng = np.random.default_rng(args.seed)++    lam = float(os.environ.get("T2_GEOM_WEIGHT", "0.5"))+    PCA_DIM = 30+    P_MAX = int(os.environ.get("T2_OT_CELLS", "12000"))+    EPS = 0.05+    NITER = 50++    with open(os.path.join(view, "manifest.json")) as f:+        man = json.load(f)+    genes = read_genes(view)+    min_cells = int(man.get("min_cells", 1000))+    max_cells = int(man.get("max_cells", 20000))+    needs_coords = bool(man.get("needs_coords", True))+    inputs = sorted(man["inputs"], key=lambda e: float(e["time"]))+    target_t = float(man["target"]["time"])++    # find bracketing inputs by time difference only (view-independent)+    earlier = [e for e in inputs if float(e["time"]) < target_t]+    later = [e for e in inputs if float(e["time"]) > target_t]+    if not earlier or not later:+        # single input or extrapolation: stratified copy of anchor stage+        anchor = earlier[-1] if earlier else inputs[-1]+        X, coords, ct, vn = load_stage(view, anchor)+        pos = {g: i for i, g in enumerate(vn)}+        cols = np.array([pos.get(g, -1) for g in genes])+        Xo = np.zeros((X.shape[0], len(genes)), dtype=np.float32)+        ok = cols >= 0+        Xo[:, ok] = X[:, cols[ok]]+        idx = stratified_take(rng, ct, max_cells)+        write_out(args.out, Xo[idx], coords[idx] if needs_coords else np.zeros((len(idx), 3)), genes)+        log("fallback copy_last, n =", len(idx))+        return++    ea = max(earlier, key=lambda e: float(e["time"]))+    eb = min(later, key=lambda e: float(e["time"]))+    ta, tb = float(ea["time"]), float(eb["time"])+    t = (target_t - ta) / (tb - ta)+    log(f"bracket a={ea['stage']}({ta}) b={eb['stage']}({tb}) t={t:.3f} lambda={lam}")++    Xa, ca, cta, vna = load_stage(view, ea)+    Xb, cb, ctb, vnb = load_stage(view, eb)+    # align genes to panel+    def align(X, vn):+        pos = {g: i for i, g in enumerate(vn)}+        cols = np.array([pos.get(g, -1) for g in genes])+        out = np.zeros((X.shape[0], len(genes)), dtype=np.float32)+        ok = cols >= 0+        out[:, ok] = X[:, cols[ok]]+        return out+    Xa, Xb = align(Xa, vna), align(Xb, vnb)+    na, nb = Xa.shape[0], Xb.shape[0]++    # expression PCA on pooled cells+    Xp = np.vstack([Xa, Xb]).astype(np.float64)+    mu = Xp.mean(axis=0)+    Xc = Xp - mu+    k = min(PCA_DIM, Xc.shape[1] - 1, Xc.shape[0] - 1)+    # randomized-free deterministic SVD on genes x genes covariance (500 genes)+    cov = Xc.T @ Xc / max(1, Xc.shape[0] - 1)+    w, V = np.linalg.eigh(cov)+    order_e = np.argsort(-w)[:k]+    V = V[:, order_e]+    P = Xc @ V  # pooled PCA scores+    Pa_full = zscore(P[:na])+    Pb_full = zscore(P[na:])++    # geometry features per side, pooled z-score+    Ga = geom_features(ca, rng)+    Gb = geom_features(cb, rng)+    Gp = np.vstack([Ga, Gb])+    Gp = zscore(Gp)+    Ga_z, Gb_z = Gp[:na], Gp[na:]++    Za_full = np.hstack([Pa_full, lam * Ga_z])+    Zb_full = np.hstack([Pb_full, lam * Gb_z])++    # equal-size subsamples for OT+    Pn = int(min(na, nb, max_cells, P_MAX))+    sa = np.sort(rng.choice(na, size=Pn, replace=False))+    sb = np.sort(rng.choice(nb, size=Pn, replace=False))+    Za, Zb = Za_full[sa], Zb_full[sb]++    # cost + Sinkhorn (chunked build)+    C = np.empty((Pn, Pn))+    for i, blk in sqdist_chunked(Za, Zb, chunk=4096):+        C[i:i + blk.shape[0]] = blk+    j_of_i = sinkhorn_log_plan(C, EPS, NITER)+    del C++    # interpolated latents+    Zt = (1.0 - t) * Za + t * Zb[j_of_i]++    # decode: nearest real cell in pooled full latent+    Zpool = np.vstack([Za_full, Zb_full])+    Pq = Zt.shape[0]+    dec_idx = np.empty(Pq, dtype=np.int64)+    dec_d = np.empty(Pq)+    # also nearest per side (for balance guard)+    na_idx = np.empty(Pq, dtype=np.int64); na_d = np.empty(Pq)+    nb_idx = np.empty(Pq, dtype=np.int64); nb_d = np.empty(Pq)+    for i, blk in sqdist_chunked(Zt, Zpool, chunk=2048):+        e = i + blk.shape[0]+        dec_idx[i:e] = blk.argmin(axis=1)+        dec_d[i:e] = blk.min(axis=1)+        ba = blk[:, :na]+        na_idx[i:e] = ba.argmin(axis=1)+        na_d[i:e] = ba.min(axis=1)+        bb = blk[:, na:]+        nb_idx[i:e] = bb.argmin(axis=1)+        nb_d[i:e] = bb.min(axis=1)+    # random-pair distance baseline (mechanism evidence)+    sample_q = rng.choice(Pq, size=min(2000, Pq), replace=False)+    sample_p = rng.choice(na + nb, size=min(2000, Pq), replace=False)+    rand_mean = float(np.mean(((Zpool[sample_p] - Zt[sample_q]) ** 2).sum(axis=1)))++    out_idx = dec_idx.copy()+    src_b = out_idx >= na+    fb = src_b.mean()+    # source-balance constraint: b fraction forced toward t (uses time diff only)+    target_fb = float(np.clip(t, 0.30, 0.70))+    if fb < target_fb:+        need = int(np.ceil(target_fb * Pq)) - src_b.sum()+        if need > 0:+            cand = np.where(~src_b)[0]+            pick = cand[np.argsort(nb_d[cand])][:need]+            out_idx[pick] = na + nb_idx[pick]+            src_b = out_idx >= na+            fb = src_b.mean()+    elif fb > target_fb:+        need = src_b.sum() - int(np.floor(target_fb * Pq))+        if need > 0:+            cand = np.where(src_b)[0]+            pick = cand[np.argsort(na_d[cand])][:need]+            out_idx[pick] = na_idx[pick]+            src_b = out_idx >= na+            fb = src_b.mean()++    # order outputs by decode distance (plan step 7) then cap to max_cells+    n_out = min(Pq, max_cells)+    keep = np.argsort(dec_d)[:n_out]+    final = out_idx[keep]+    uq = len(np.unique(final))+    log(f"pairs={Pq} decoded_b_frac={fb:.3f} unique_cells={uq} "+        f"mean_dec_d2={dec_d.mean():.2f} rand_pair_d2={rand_mean:.2f}")++    def gather(idx):+        outX = np.empty((len(idx), len(genes)), dtype=np.float32)+        outC = np.empty((len(idx), 3))+        ma = idx < na+        ia = idx[ma]; ib = idx[~ma] - na+        outX[ma] = Xa[ia]; outX[~ma] = Xb[ib]+        outC[ma] = ca[ia]; outC[~ma] = cb[ib]+        return outX, outC++    X_out, C_out = gather(final)+    write_out(args.out, X_out, C_out if needs_coords else np.zeros((len(final), 3)), genes)+    log("wrote", args.out, "n =", X_out.shape[0])+++if __name__ == "__main__":+    main()

调研来源?调研员查到并用到的知识条目和文献检索结果（只列标题和编号）。

用到的知识库条目

分析结果?分析员写的 ANALYSIS.json：改了什么、各组分数怎么变、假设是否成立、经验和下一步建议。

对话摘要?每个角色和大模型对话的统计：轮数、工具调用、用时、token 数和最后的回答摘录；原始记录只给路径。

只给统计和最后回答的摘录；完整对话请到原始记录位置里列出的文件看。

分析员

ANALYSIS.json written

工程师

Done. Final state: `solution/` contains run.py (full-scale verified, vec-check ok), METHOD.md, EXECUTION.json (cpu).

**做了什么**：draft 实现 PLAN T2HI-03——表达PCA(30)+刚体不变局部几何(4维,λ=0.5)联合 latent 上 log-domain Sinkhorn OT（eps=0.05, 50 iter, P=12000/侧）配对，z_t 插值后解码到最近真实细胞（表达/坐标原样），b 侧来源比例按时间差 t 强制平衡（自然解码 <30% 偏向 a 侧）。单输入/外推有分层 copy_last 退路；纯时间差驱动，视图无关。

**查分**（A 半，3/20 次）：λ=0.5 n=3000 smoke 56.81；λ=0.5 n=12000 **59.98**（表达 65.6 / 状态 66.0 / 形状 53.3 / 邻域 55.1）。λ=0 对照全量运行被会话超时杀掉，截止前未取得对照分，已如实写入 METHOD.md。低于树内最佳 node 13（66.10）：z_t 最近邻解码偏向"最稳定/最中间"细胞，丢失 b 侧类型多样性，cell_state 和 expression_change 落后 logit 配额约 6 分；shape_scale（53.25）是唯一优于 node 13 的组。

**下一步最值得试**：(1) 在 OT 配对骨架上改用 logit 组成插值定配额 + 解码只做"配对感知的细胞选择"而非全局最近邻，融合 node 10 的组成优势与 OT 配对；(2) 补 λ=0 对照确认几何维度是否真起作用；(3) node 13 的同型 kNN 收缩可叠加到本方案的坐标上（本方案 local_spatial 55.1 仍有 ~15 分空间）。

调研员

53.37降到49.69（d2_shape 0.0401→0.0576），说明表达与坐标分离处理损害形状。node 10/13靠真实细胞原样保留expression-position配对，将local_spatial从54.03提到70.19，但配对完全依赖随机抽样而非有原则的跨时间点匹配。T2HI-03用联合latent中的OT找跨时间点最优配对，比按类型配额抽样更精准地保持表达-位置一致性，同时改善cell_state和local_spatial。",
  "approach": "步骤：(1) 读取两侧时间点（a=E8.5, b=E9.5）细胞的表达矩阵，各取top-30 PCA；(2) 为每个细胞计算刚体不变局部几何特征：到k=10近邻的距离均值与标准差、到云质心的归一化半径（dist/RMS）、局部密度（半径=中位邻距内邻居数）；(3) 将表达PCA(30维)与几何特征(4维)拼接，几何维度乘以权重λ=0.5（搜索范围0.2–1.0），再z-score标准化得联合latent；(4) 在联合latent上用Sinkhorn-OT（epsilon=0.05, 50次迭代, scipy线性规划或手写Sinkhorn）计算a→b的传输计划π；(5) 对目标时间t∈(0,1)，在联合latent中做插值 z_t = (1-t)·z_a_matched + t·z_b_matched（按π配对）；(6) 解码：对每个z_t，在全部真实细胞（a∪b）的联合latent中找最近邻，直接使用该真实细胞的表达和坐标作为输出；(7) 输出细胞数按manifest的min_cells/max_cells约束，从解码结果中按z_t到插值点的距离排序取前N个。关键参数初值：PCA_dim=30, k_geom=10, λ=0.5, sinkhorn_eps=0.05, sinkhorn_iter=50。若POT库不可用，用scipy.optimize.linear_sum_assignment做精确OT（细胞数<3000时可行）。单输入阶段退路：只有一个时间点时直接copy_last。vec-score快速筛选：先跑λ=0（纯表达PCA配对）与λ=0.5对比，确认几何维度有贡献后再全量。",
  "expected_groups": ["cell_state", "local_spatial"],
  "risks": "1) Sinkhorn在细胞数>2000时可能超时——Engineer应先用500细胞子集验证，确认可行后全量；若仍超时改用线性规划或分批近似。2) 联合latent中几何特征噪声大（心脏形态变化小），λ过大使配对退化——用λ=0对照确认。3) 解码最近邻可能集中在某一侧时间点导致组成偏斜——监控输出中a/b来源比例，必要时加来源平衡约束（每侧至少占30%）。4) 若OT配对结果与类型配额抽样差异不大（cell_state/local_spatial未超过node 10），说明机制增益不足，应尽早报告而非继续调参。",
  "family_id": "T2HI-03",
  "mechanism": "在表达PCA+刚体不变局部几何描述构成的联合latent中做跨时间点Sinkhorn-OT配对，插值后以最近真实细胞解码，保证表达-坐标配对不被打破。",
  "vs_constant_shift": "常数位移对同类型所有细胞施加相同向量，不改变细胞间相对关系；本方案在联合latent中逐细胞做最优传输配对，每个细胞的来源和插值位置由其在表达+几何空间中的具体位置决定，产生细胞特异的配对与选择，且解码直接复用真实细胞的表达和坐标而非施加位移。",
  "mechanism_evidence": "Engineer应记录：(1) OT配对中几何特征改变配对结果的比例（λ=0.5与λ=0的配对重合率，若>95%说明几何未起作用）；(2) 输出细胞中a/b来源比例随t的变化是否平滑；(3) 与node 10（类型配额）相比，cell_state和local_spatial的raw值（mmd_u, neighborhood_mmd）是否下降；(4) 解码最近邻的平均联合距离是否小于随机配对。",
  "mechanism_off_control": "设λ=0（几何权重为零），联合latent退化为纯表达PCA，此时OT配对不考虑空间结构。预期：若机制有效，λ=0时local_spatial（尤其neighborhood_mmd）应比λ=0.5差，因为配对不再考虑几何邻近性；若两者分数相同，说明几何特征未影响配对，机制未运行。同一程序中通过环境变量T2_GEOM_WEIGHT=0关闭。",
  "sources": []
}
```