总览 · ← 返回运行 20261003-171955-search-t2-embryo-interp-chain-12h

节点 n44 终选程序?按该运行锁定的规则最终选出的程序；可能是候选节点，也可能由护栏回退到基线。在终选来历上

WITHINP：按类型×侧把抽样漂移的 drawn 均值用有界乘性因子回正到全阶段均值（支持集/坐标不动，池化伪批量中和保 DE 通道），覆盖共有与单侧类型；PLAN 的 PROGCOH 共表达相关置零经 5 种解码证否。

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 956ccc7a3a: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：父节点版本 2fded41e1e。改动的文件：solution/METHOD.md +118 −67、solution/README.md +14 −17、solution/run.py +320 −9

diff --git a/solution/METHOD.md b/solution/METHOD.mdindex ad067ac..db35423 100644--- a/solution/METHOD.md+++ b/solution/METHOD.md@@ -1,68 +1,119 @@-NBHDCOH-SHARED：把节点41的邻域相干OFF选择从单侧类型推广到共有类型DETR——每基因移除数不变，只把"选哪些细胞置零"从随机键u改为"15-NN邻域该基因平均表达(排除自身)最低的k个"，针对local_spatial(neighborhood_mmd)。--## 方法（family T2EI-01，improve on node 41）--父节点 41 的 NBHDCOH 证明了：在 DETR_EXT 支持集修复上，**保持每基因移除数不变、只改选哪些细胞置零**（邻域相干选择）能同时改善 nbhd 与 variogram，三种子均值 +0.72。但那套相干选择只作用于**单侧独有类型**（DETR_EXT，~37% 输出细胞）；**共有类型 DETR 置零**（~63% 输出细胞、强度 1.5、更大的支持集修复）仍用节点 37 的**逐细胞随机键 u** 选择。--本节点 **NBHDCOH_SHARED** 把与单侧完全相同的相干选择推广到共有类型 DETR：--- 对每个共有类型的每个激活基因 g、每一侧（a/b），在该型的非零细胞中，按**15-NN 邻域内基因 g 的平均表达**（排除自身）排序，置零**最低的 k 个**。-- 邻域场 NS 在**置零前的输出云**上算一次：cKDTree 建在最终输出坐标 `coords[:n_out]`，query k=15（含自身），确定性 tie-break `(dist, index)`；`NS' = (k·NS − x_i)/(k−1)` 排除自身（`NBHDCOH_SELF=0`，沿用父默认）。-- 确定性选择键：`np.lexsort((nz, u[nz], NS'))[:k]`（主键 NS'，次键遗留随机键 u，三键 cell 局部索引），与单侧 NBHDCOH 逐字一致。-- **k（每基因移除数）完全不变**：仍由 DETR 的 `p_t = p_a^(1−t)·p_b^t`（geom 模式）、强度 s=1.5、公式 `k = round((p_t − p_cur)·n·s)` 决定。所以 marginal 检出率 / variogram 的目标不动，只把移除搬到**微环境本就缺该转录本**的位置，最小化对 neighborhood_mmd 所读的"邻域平均表达场"的扰动（空间自相关：相干移除像真实深度差异，孤立随机 dropout 才是 nbhd 惩罚的对象）。--机理不是常数位移、也不是组成重加权：同一共有类型内、不同空间位置的细胞有不同的 15-NN 邻域表达谱，因此获得不同的置零决策；不改任何基因表达值、不改细胞总数、不改坐标，只改**哪些细胞的哪些基因被置零**。--单输入阶段退路不变：`b is None` → 分层照抄，DETR/NBHDCOH_SHARED 自然不激活。--## 关键参数（提交默认）--- `T2_NBHDCOH_SHARED=1`（机制开）-- `T2_NBHDCOH_SHARED_K=15`（邻域尺度 = neighborhood_mmd 的 15-NN）-- `T2_NBHDCOH_SELF=0`（排除自身，沿用父）-- `T2_DETR_STRENGTH=1.5`（父强度，**不动**——纯位置改变，最干净的归因）-- `T2_DETR_EXT_STRENGTH=0.45`、单侧 `T2_NBHDCOH=1`（父节点 41 原样保留）--## 验证过什么--所有分数为 A 半、**同一评分会话**（父教训：本地锚点跨会话不可比），proxy 视图 `T2:embryo:val_interp`。--**1) 关闭机制对照（--ablate mechanism / T2_NBHDCOH_SHARED=0）**：共有类型 DETR 回退随机键 u，单侧 NBHDCOH 保留。输出与父节点 41 **逐位一致**（seed 0 sha256 相同，`bae7f50e…`）——机制归因干净，ablate = 父。--**2) 同强度机制归因（coh − nocoh，matched strength，seed 0）**：--| s | coh board | nocoh board | Δ机制 | coh nbhd | nocoh nbhd |+WITHINP：按类型×侧把抽样漂移的 drawn 均值用有界乘性因子回正到全阶段均值（支持集/坐标不动，池化伪批量中和保 DE 通道），覆盖共有与单侧类型；PLAN 的 PROGCOH 共表达相关置零经 5 种解码证否。++## 节点 44（op=improve，父节点 43，board T2:embryo:val_interp，family T2EI-01）++### 提交机制：WITHINP（draw-noise recentering，作用于 cell_state 组）++**动机（只读诊断，全部由视图数据现场计算）**：分层无放回抽样使每个类型的 drawn 细胞均值+偏离该类型全阶段均值，偏离是纯抽样噪声——代理括号上测得小样本侧（b 侧 drawn 7–38 细胞的+类型）漂移达 0.15–0.17 倍类型内 std，且同一类型 a 侧与 b 侧漂移的相关系数仅 0.0–0.08+（若是真实生物信号应同向）。真实中间阶段不带我们的抽样噪声，因此每个输出类型子云在+mmd_u 所读的真值 PCA 空间里整体偏移。现有通道都碰不到这一项：位移 Δ 用全阶段类型均值+计算（无法感知 drawn 子集已实现的偏移）；DETR/NBHDCOH 只修支持集；pair-β 只压缩型内散布。++**实现**（`mix_converge`，插在 VARISO 之后、坐标/RESID/DETR 之前；`T2_WITHINP` 块）：++1. 对每个输出类型 × 每侧（`WITHINP_ALLTYPES=1`：不限于共有类型——单侧公式只需该侧自己的+   全阶段均值，故单侧独有类型（约 37% 输出细胞）同样校正；门槛：该侧全阶段 ≥10 细胞、+   drawn ≥5 细胞）：逐基因乘性因子+   `f_g = clip(((μ_full_g + ε)/(μ_pre_g + ε))^w, LO, HI)`，+   μ_full = 该类型该侧全阶段均值，μ_pre = 该侧 drawn 细胞**位移前**的均值（抽样噪声比经+   nnz-only 加性位移近似乘性传递）。提交参数：w=1.0、ε=0.03、clip=[0.8, 1.25]。+   乘性 ⇒ 零支持集逐位保留（无 ON 注入，遵守节点 37 教训；全零列自动 no-op），坐标不动。+2. 池化伪批量中和（`WITHINP_PBNEUTRAL=1`）：所有类型处理完后，对每基因整列乘+   `g_g = clip(pb0_g/pb1_g, 0.9, 1.1)`（pb 为全体输出细胞的池化均值，pb0 为 WITHINP 前的值），+   把每基因池化 pb 复原 ⇒ dp（de_score/de_direction 的输入）结构上不动，机制的可见效应+   纯粹是基因质量在类型/细胞之间的再分配（cell-state 通道）。++**机制生效证据**（seed 0，A 半）：43 个类型×侧被校正；|f−1| 均值 0.14；输出相对父节点+107,258 个矩阵元素改变（4.3%，其中支持集差异来自下游 NBHDCOH 的 NS 场在重定心后的值上+重算，选择随之微移）；坐标逐位相同（shape_scale 三项 raw 三种子全部逐位不变）。+四组分变化（对父节点，A 半同会话）：local_spatial +0.13/+0.25/+0.14（seed 0/1/2，主收益，+nbhd raw 三种子一致下降 −2.6e-4/−5.0e-4/−2.7e-4）；cell_state +0.11/−0.10/+0.10+（mmd_u raw 2/3 种子下降）；expression_change +0.04/+0.16/−0.09（de_direction +1.5e-3/++7.5e-3/−4.0e-3，中和后不再系统性受损）；shape_scale 0/0/0。++**查分记录（A 半，同会话锚点 parent seed 0 = 65.2615）**：++| 配置 | 榜分 | de_score | de_dir | mmd_u | variogram | nbhd | 形状三项 |+|---|---|---|---|---|---|---|---|+| parent（锚点） | 65.2615 | 7.621 | 7.921 | 7.986 | 6.89 | 15.405 | 8.159/3.511/7.767 |+| 共有类型 w1 ε.03 clip[.8,1.25] | 65.3012 | 7.622 | 7.903 | 8.006 | 6.906 | 15.427 | 同上 |+| 共有类型 w0.5 ε.03 | 65.2767 | 7.621 | 7.911 | 7.994 | 6.901 | 15.410 | 同上 |+| 共有类型 w1 ε.1 | 65.3045 | 7.621 | 7.909 | 8.004 | 6.906 | 15.427 | 同上 |+| 共有类型 w1 clip[.9,1.12] | 65.3127 | 7.622 | 7.918 | 8.005 | 6.900 | 15.427 | 同上 |+| 共有+中和 w1 ε.03 | 65.3167 | 7.622 | 7.930 | 7.995 | 6.888 | 15.445 | 同上 |+| 共有+中和 w1 clip[.9,1.12] | 65.2935 | 7.622 | 7.926 | 7.986 | 6.886 | 15.435 | 同上 |+| **全类型+中和 w1 ε.03（提交）** | **65.3318** | 7.622 | 7.931 | 8.003 | 6.900 | 15.438 | 同上 |+| 全类型+中和 w1 clip[.9,1.12] | 65.3067 | 7.622 | 7.928 | 7.998 | 6.897 | 15.425 | 同上 |+| 全类型+中和 宽clip[.6,1.6]/G[.75,1.35] | 65.3555 | 7.622 | 7.938 | 8.004 | 6.906 | 15.450 | 同上 |++（mmd_u/variogram/nbhd 列为 points；raw 见下。宽 clip 仅 seed 0 领先 +0.024，在配对噪声内，+且 g 触到 1.35 上界（dp 复原不完全）、|f−1| 均值升到 0.24（幅值通道激进，有 AMPSHRINK 式+流形风险），额度已尽无法做 3 种子确认，**不取**。）++**3 种子配对确认（提交配置 vs 父节点，A 半同会话）**：++| seed | 提交 | 父 | 差 | nbhd raw 提交/父 | mmd_u raw | variogram raw | de_dir raw |+|---|---|---|---|---|---|---|---|+| 0 | 65.3318 | 65.2615 | **+0.070** | 0.04752/0.04778 | 0.01010/0.01017 | 0.007098/0.007124 | 0.3773/0.3758 |+| 1 | 64.9060 | 64.8268 | **+0.079** | 0.04793/0.04843 | 0.00978/0.00972 | 0.007289/0.007272 | 0.3699/0.3624 |+| 2 | 64.8927 | 64.8570 | **+0.036** | 0.04762/0.04789 | 0.01011/0.01019 | 0.007020/0.007023 | 0.3720/0.3760 |++三种子配对差全正（均值 +0.062）；nbhd raw 三种子一致下降（与节点 41/43 的采纳标准一致）。+幅度小于父节点当时的 +0.21，属于小机制，但方向在全部三种子一致。++**对照**：`--ablate mechanism`（或 `T2_WITHINP=0`）关闭 WITHINP（PROGCOH 默认已关），+输出与父节点 43 逐位一致（sha256 264727…b327，seed 0 验证）。默认输出对 seed 确定+（同 seed 两次运行 sha256 相同）；`vec-check` 通过；程序纯 CPU（EXECUTION.json gpu=false），+seed 0 全程 2.0 s / 峰值 0.62 GB（限额 30 min / 28 GB）。++### PLAN 机制 PROGCOH：已实现并证否（5 种解码/幅度组合，5 次查分）++按 PLAN 实现：输出云（DETR 前）上算基因-基因 Pearson 相关，每基因取 |r|≥RMIN 的前 P 个+伙伴（伙伴候选限检出率>DET 的基因），共有类型 DETR 的置零选择键从 NS′(g) 改为+JNS = NS′(g) + λ·mean_h NS′(h)（h∈N(g)，同一 15-NN 场，确定性 lexsort (JNS,u,index)），+仅在有重排自由的调用（0<k<非零数）生效，每基因移除数不变。Step-0 门槛通过：853 个自由+调用中 44.1%（376 个）有 ≥1 伙伴（PLAN 门槛 40%），伙伴 |r| 中位 0.36；置零确实变得+程序相关（top-50 共表达对置零指示相关 +3e-4…+1.4e-3，随解码增强）。但真 variogram 不动：++| 配置（seed 0） | 榜分 | variogram raw | nbhd raw | mmd_u raw | de_dir raw | |---|---|---|---|---|---|-| 1.5 | 65.262 | 65.043 | **+0.218** | 0.04778 | 0.04929 |-| 1.75 | 65.165 | 64.940 | **+0.225** | 0.04825 | 0.04972 |-| 2.0 | 65.090 | 64.862 | **+0.228** | 0.04870 | 0.05023 |--机制在**每个强度**都稳定值 ~+0.22 榜分，nbhd raw 每档降 ~1.5e-3。--**3) 三种子配对差（提交配置 s=1.5/k=15，coh − parent）**：--| seed | coh | parent(ablate) | Δboard | coh nbhd | parent nbhd | Δnbhd |-|---|---|---|---|---|---|---|-| 0 | 65.262 | 65.043 | +0.218 | 0.04778 | 0.04929 | −1.51e-3 |-| 1 | 64.827 | 64.644 | +0.183 | 0.04843 | 0.04995 | −1.52e-3 |-| 2 | 64.857 | 64.681 | +0.176 | 0.04789 | 0.04944 | −1.55e-3 |--三种子 board 差全正（+0.18~+0.22），nbhd raw **每种降 1.5e-3**（远超 1e-4 判无效门槛）。--**4) 四组分分解（seed 0，coh vs parent）**：local_spatial 改善（nbhd 0.04929→0.04778，skill 0.609→0.616，+0.19 分，本机制主目标）；cell_state 微动（mmd_u 0.01028→0.01016 +0.03、variogram 0.007107→0.007124 −0.01，因共有 DETR 移除数不变，per-gene 检出率不变，variogram 读的是 on/off 结构 → 基本不动，符合预期）；expression_change 不变（de_score raw 逐位相同 0.3214）；shape_scale **逐位不变**（d2_shape 0.0047、occupancy_dice 0.8066、scale_log_ratio 0.0162 三项 raw 完全相同，坐标未动）。--**5) 机制生效证据**：共有类型中被 NBHDCOH_SHARED 改变置零决策的细胞比例 = **13.8%**（2888 个 gene×side 调用，18600 次移除）。低的原因：**70.5%** 的共有 DETR 基因调用是**全移除**（k == nz.size，无重排自由）；在有自由的基因上 move_frac = **46.7%**（实质重排）。NS' 区分度良好：中位 0.398、IQR 1.377、零占比 0.74% —— PLAN risk #2（邻域同质、NS' 区分度不足）**未发生**。--**6) 确定性**：同 seed 重跑逐位一致（sha256 相同）。**vec-check ok**。**运行时**：seed 0 约 3–4s（+cKDTree 15-NN 约 0.5s），内存 <1GB，远在 limits 内。**k 敏感性**：k=10/15/20 board 65.262/65.262/65.270、nbhd 0.0479/0.04778/0.04775，全在噪声内；保留 k=15（对齐 neighborhood_mmd 尺度，最有原则）。--## 没验证 / 已知弱点--- **增益幅度 +0.19（3 种子均值）远小于父节点 41 的 +0.72**，也小于 T2 ~1 分噪声；但配对差三种子全正、机制归因（matched-strength coh−nocoh ~+0.22）稳定、nbhd raw 每种一致降 1.5e-3，方向可信（沿用父节点的归因逻辑：信配对差与 matched-strength 对照，非信单一 board）。增益小是 PLAN risk #1 预判的：共有 DETR 的 nbhd 影响本就接近平，70% 基因全移除锁死了大部分重排自由。-- **未降 s 到 <1.5**：s=1.25 board 65.290（略高但噪声内），且降 s = 减少移除总量，正是节点 41 明确证否的方向（预算削减 variogram 损失 ~3× nbhd 收益，1:3 兑换率）；s=1.25 的 variogram 已劣化到 0.007226（vs 父 0.007107）。**未升 s 到 2.0**（PLAN step 5）：board 65.090 < 65.262，更差。故 s 保持父的 1.5。-- **未在 B 半 / 官方尺子验证**（无权限）；A 半增益是否兑现到官方分未知，但配对一致性是父节点采用的同类证据。-- **未做值通道改动**（node 17/19/37/39 教训：值通道一律伤 nbhd/mmd），本机制只动支持集（哪些细胞哪些基因置零）。-- 视图无关：机制只依赖坐标、表达、类型标签、相对时间差，不读视图路径 / 绝对时间 / manifest 排版；伪装视图平移时间不影响。--## 知识来源--方法为纯算法（空间自相关 / 邻域相干 dropout 选择），不引入任何保留阶段或保留基因型的测量值、细胞类型清单、比例、表达量或形态尺寸。检出率插值目标 `p_t = p_a^(1−t)·p_b^t` 由程序从**输入括号阶段**（view 内 E6.75/E8.0，均非 T2 全胚禁窗 (7.25,8.0) 内的保留阶段）现场计算；邻域场由程序从输出坐标与表达现场计算。无外部数据集、无 prior 资源、无文献数值。（沿用父节点 41 的合规基线。）+| parent 锚点 | 65.2615 | 0.007124 | 0.04778 | 0.01017 | 0.3758 |+| nbhd λ=1 P5 R.25 D.05（PLAN 原样） | 65.2422 | 0.007115 | 0.04793 | 0.01017 | 0.3755 |+| self λ=2（伙伴用自身表达值） | 65.2056 | 0.007131 | 0.04815 | 0.01017 | 0.3752 |+| both λ=1（nbhd+self） | 65.2262 | 0.007118 | 0.04811 | 0.01014 | 0.3754 |+| nbhd-cov λ=1 P10 R.20 D.02（覆盖扩展） | 65.2513 | 0.007122 | 0.04786 | 0.01016 | 0.3756 |+| nbhd λ=0.5（弱耦合） | 65.2588 | 0.007108 | 0.04786 | 0.01016 | 0.3758 |++（points 分解见上表前半；各配置形状三项与 de_score 逐位不变，符合"只动置零选择"的设计。）+证否结论：variogram raw 只在 ±1.6e-5（≈±0.02 分）内摆动、无剂量-响应；任何偏离纯 NS′(g)+排序的选择都按偏离幅度损失 nbhd raw（+8e-5…+3.7e-4）；总效应上限受"有重排自由的移除+条目数"（约 1–2k / 18.6k 共有移除，70.5% 调用为全移除）硬性封顶，而真 variogram 用 2 万+**随机**基因对，共表达对只占极小份额——PLAN 的效应量估计（variogram skill 0.564→0.60）+高估了约一个数量级。覆盖扩展（更多伙伴、更低检出门槛）不改变结论。代码保留在+`T2_PROGCOH=0`（默认关）之后。++### 备选机制探索中的其他已测方向（未提交，均记录于上表）++- WITHINP 不做池化中和：de_direction 系统性受损（−0.6e-3…−3.4e-3），中和后被消除——+  证明 dp 复原设计有效。+- 只覆盖共有类型：收益约为全类型版的一半（+0.04 vs +0.07），单侧类型（37% 细胞）的+  抽样漂移是真实可校正的。++### 知识来源++未使用任何外部生物学知识或文献数值。WITHINP 与 PROGCOH 的全部量（类型均值、漂移、+基因相关、检出率、移除预算）均在运行时从视图输入现场计算；不读取绝对阶段时间+（只用 t 比例与时间差），不读取 board/mode 字段，满足视图无关要求。++### 未验证 / 风险++- 宽 clip [0.6,1.6] 在 seed 0 略好（+0.024，噪声内）但未做多种子确认，未提交；若后续+  节点额度充足可复核（预期方向不确定，g 上界触界说明 dp 复原在宽 clip 下不完全）。+- WITHINP 与下游 DETR/NBHDCOH 的交互（NS 场在重定心值上重算）包含在提交配置内一并+  验证，未单独拆解。+- 真实括号（final 视图，两个相邻已发布阶段）上 drawn 漂移的幅度可能与代理括号不同：+  细胞数、类型数都从视图现场读取，公式对两种情况同样成立（漂移大小自动由数据决定），+  但收益幅度未在真实括号上验证过（无法验证）。+- proxy2 类视图（若挂载）：单侧输入退化路径（b is None）不经过 WITHINP，行为与父节点一致。diff --git a/solution/README.md b/solution/README.mdindex e2e2214..057678b 100644--- a/solution/README.md+++ b/solution/README.md@@ -1,20 +1,17 @@-# mix + 表达收敛管线 + DETR / DETR_EXT 检出率几何插值 + 邻域相干置零（T2:embryo:val_interp）+# mix + 表达收敛管线 + DETR / DETR_EXT / NBHDCOH(_SHARED) + WITHINP 抽样漂移重定心（T2:embryo:val_interp） -父节点 41 管线原样保留（mix 混抽 + α=5 类型级收敛位移 + 软阈值基因权重 + 相关扩散+父节点 43 管线原样保留（mix 混抽 + α=5 类型级收敛位移 + 软阈值基因权重 + 相关扩散 η=1/asym=0.85 + λ=6 投影加权 + β=0.2 型内配对收缩 + aniso 坐标整形 damp=1.25 + DETR-共有类型检出率 OFF 校正 s=1.5 + DETR_EXT 单侧类型 OFF 校正 s=0.45 + 单侧 NBHDCOH-邻域相干置零；RESID/EIGPROJ/ITERDIFF/VARISO/TYPE_ANISO/PBPROJ/TSHIFT/SPATRESID/-AMPSHRINK/NBHDGATE 默认关闭）。+共有类型检出率 OFF 校正 s=1.5 + DETR_EXT 单侧类型 OFF 校正 s=0.45 + 单侧 NBHDCOH ++共有 NBHDCOH_SHARED 邻域相干置零）。 -本节点（43）提交机制 NBHDCOH_SHARED：把节点 41 的邻域相干 OFF 选择从单侧类型推广到-共有类型 DETR——对每基因每侧，在该型非零细胞中按 15-NN 邻域该基因平均表达（排除-自身，cKDTree 建于最终输出坐标，确定性 lexsort (NS', u, index)）取最低的 k 个置零；-k（每基因移除数）由 DETR 公式与 s=1.5 决定、完全不变。移除搬到微环境本就缺该转录本-的位置，最小化对 neighborhood_mmd 所读邻域平均场的扰动。-A 半同会话：机制在 matched strength 下稳定 +0.22（s=1.5/1.75/2.0）；提交配置-s=1.5/k=15 三种子配对差 +0.218/+0.183/+0.176，nbhd raw 每种降 1.5e-3-（0.0493→0.0478 / 0.0500→0.0484 / 0.0494→0.0479）；shape_scale 与 de_score 逐位不变。-共有类型置零决策改变率 13.8%（70.5% 基因调用为全移除、无选择自由；有自由的基因上-move_frac 46.7%）。--ablate mechanism（或 T2_NBHDCOH_SHARED=0）逐位还原父节点 41-（sha256 已验证）。s=2.0 更差（65.09）、s=1.25 落入已证否的预算削减方向，均不取。-详见 METHOD.md。+本节点（44）提交机制 WITHINP：分层无放回抽样使每类型 drawn 均值带纯抽样噪声漂移+（小样本侧达 0.17 倍类型 std，两侧漂移不相关 r≈0–0.08）。对每个输出类型×侧（含单侧+独有类型），逐基因乘 f_g=clip(((μ_full+ε)/(μ_pre+ε))^w, 0.8, 1.25)（ε=0.03、w=1）把+drift 回正到全阶段均值——乘性 ⇒ 零支持集与坐标逐位不动；再对每基因整列乘+g_g=clip(pb0/pb1, 0.9, 1.1) 复原池化伪批量，dp（DE 两项输入）结构上不动，机制可见效应+纯粹是类型间质量再分配。A 半同会话 3 种子配对差 +0.070/+0.079/+0.036 全正，nbhd raw+三种子一致下降（−2.6e-4/−5.0e-4/−2.7e-4）。PLAN 的 PROGCOH（共表达程序相关置零）已实现+并经 5 种解码证否（variogram 只动 ±1.6e-5、nbhd 按偏离幅度受损、效应量受重排自由调用数+封顶），代码留在 T2_PROGCOH=0（默认关）。--ablate mechanism（或 T2_WITHINP=0）逐位还原+父节点 43（sha256 验证）。详见 METHOD.md。diff --git a/solution/run.py b/solution/run.pyindex f8d3bca..67a4bc9 100644--- a/solution/run.py+++ b/solution/run.py@@ -8,6 +8,42 @@ exp(log r_a + SCALE_DAMP·t·Δlog r), and draws cells stratified by type: round(t·n) from the later stage, the rest from the earlier one. Coordinates travel with the cells. n is log-linear in t, clipped to the board range. +This node (44, WITHINP, family T2EI-01): the PLAN mechanism (PROGCOH:+co-expression-program-correlated DETR zeroing — rank each shared-DETR gene's+candidate cells by a joint neighborhood score JNS = NS'(g) + λ·mean_h NS'(h)+over g's top-|r| correlation partners on the output cloud, so co-expressed+genes are zeroed in the same cells) was implemented and FALSIFIED: partner+coverage passes the PLAN gate (44% of the 853 free shared-DETR calls have ≥1+partner at RMIN=0.25/DET=0.05) and zeroing does become program-correlated+(top-50-pair zero-indicator corr +3e-4..+1.4e-3), but the real variogram moves+only ±1.6e-5 (≈±0.02 pts) across 5 decode/strength combos (nbhd λ=0.5/1,+nbhd-cov P10/R0.2/D0.02, self λ=2, both λ=1; A-half seed 0: 65.26/65.24/+65.25/65.21/65.23 vs parent anchor 65.26) while every deviation from the pure NS'(g) ranking costs nbhd+raw +8e-5..+3.7e-4 — the 20k-RANDOM-pair variogram is dominated by pairs the+program coupling cannot touch, and the reordering freedom (~1-2k entries of+18.6k shared removals) caps the effect size. PROGCOH stays in the code behind+T2_PROGCOH=0 (default off). The SUBMITTED alternative is WITHINP (draw-noise+recentering, targets the same weak group cell_state): the stratified draw+makes each type's DRAWN-cell mean deviate from its FULL-stage mean by pure+sampling noise (measured: |drift| up to 0.17 type-std on small-n sides;+cross-side drift correlation ≈0.0-0.08 → noise, not signal), offsetting every+output type sub-cloud in the 30-d truth PCA that mmd_u reads — a channel no+existing step touches (displacement Δ uses FULL-stage means; DETR/NBHDCOH act+on support). Per type × side (all output types incl. one-side-only, own stage+≥10 full cells, ≥5 drawn), each gene column is rescaled by+f_g = clip(((μ_full+ε)/(μ_pre_drawn+ε))^w, 0.8, 1.25), ε=0.03, w=1 —+multiplicative ⇒ zero support bit-preserved, coordinates untouched; then one+uniform per-gene pooled rescale g_g = clip(pb0/pb1, 0.9, 1.1) restores each+gene's pooled output pseudobulk, so dp (de_score/de_direction input) stays+≈fixed and the scorer-visible effect is a pure between-type mass+redistribution. A-half seeds 0/1/2: 65.33/64.91/64.89 vs parent 65.26/64.83/+64.86 (paired +0.070/+0.079/+0.036, all positive); nbhd raw lower in all 3+seeds (−2.6e-4/−5.0e-4/−2.7e-4), mmd_u lower in 2/3, de_direction +1.5e-3/++7.5e-3/−4.0e-3. Wider clip [0.6,1.6] scores 65.36 at seed 0 (+0.02 over the+submitted clip, unconfirmed on seeds 1/2 — quota exhausted, not taken).+--ablate mechanism (or T2_WITHINP=0) reproduces parent node 43 bit-for-bit+(sha256-verified). See METHOD.md.+ This node (43, NBHDCOH_SHARED, family T2EI-01): node 41's NBHDCOH proved that at matched per-gene removal counts, neighborhood-coherent OFF selection improves nbhd AND variogram together — but only applied it to DETR_EXT@@ -463,6 +499,86 @@ NBHDCOH_SELF = os.environ.get("T2_NBHDCOH_SELF", "0") == "1" # random key u). NBHDCOH_SELF is shared for the self-exclusion choice. NBHDCOH_SHARED = os.environ.get("T2_NBHDCOH_SHARED", "1") == "1" NBHDCOH_SHARED_K = int(os.environ.get("T2_NBHDCOH_SHARED_K", "15"))+# PROGCOH (mechanism of this node 44, family T2EI-01): NBHDCOH_SHARED selects+# the zeroed cells for each gene INDEPENDENTLY, so for co-expressed genes the+# dropout patterns are uncorrelated — artificially inflating E|x_i−x_j|^0.5 for+# those pairs, which is exactly what the variogram reads (skill 0.564, the+# weakest metric above floor). PROGCOH coordinates the selection across each+# gene's co-expression program: on the PRE-removal output cloud, compute the+# gene-gene Pearson correlation, and for each gene g give partner set N(g) =+# top-P genes by |r| with |r| >= RMIN among genes detected in > DET of cells+# (low-detection correlations are noise-gated out). The selection key becomes+# the joint neighborhood score JNS_i(g) = NS'_i(g) + λ·mean_{h∈N(g)} NS'_i(h),+# with NS' the same 15-NN neighborhood average (self-exclusion per+# NBHDCOH_SELF) on the same pre-removal cKDTree field, so co-expressed genes+# are preferentially zeroed in the SAME cells (those whose microenvironment+# lacks the whole program), preserving the pairwise gene-gene covariance+# structure. Applied only where reordering freedom exists (0 < k < n_nonzero);+# full-removal genes and genes without partners fall back to NBHDCOH_SHARED.+# Per-gene removal counts k are untouched → de_score/detection targets and all+# coordinate/shape metrics are structurally unchanged.+# T2_PROGCOH=0 (or --ablate mechanism) disables and reproduces parent node 43+# bit-for-bit.+PROGCOH = os.environ.get("T2_PROGCOH", "0") == "1"+PROGCOH_LAMBDA = float(os.environ.get("T2_PROGCOH_LAMBDA", "1.0"))+PROGCOH_P = int(os.environ.get("T2_PROGCOH_P", "5"))+PROGCOH_RMIN = float(os.environ.get("T2_PROGCOH_RMIN", "0.25"))+PROGCOH_DET = float(os.environ.get("T2_PROGCOH_DET", "0.05"))+# PROGCOH_KEY: which partner signal enters the joint score —+#   "nbhd": mean_h NS'_i(h), the 15-NN neighborhood average of the partner+#          genes (PLAN literal; spatially smoothed, nbhd-protective);+#   "self": mean_h x0_i[h], the cell's OWN pre-removal partner expression+#          (direct co-occurrence signal: zero g where its partners are+#          already low/absent in the same cell, the most direct preservation+#          of the pair's joint on/off pattern);+#   "both": sum of the two partner terms.+PROGCOH_KEY = os.environ.get("T2_PROGCOH_KEY", "nbhd")+# WITHINP (submitted alternative mechanism of this node 44, family T2EI-01,+# after PROGCOH falsification): the stratified WITHOUT-replacement draw makes+# each type's DRAWN-cell mean deviate from its FULL-stage mean by pure+# sampling noise (proxy diagnostics: b-side drawn means of small-n types sit up+# to 0.17 type-std away from the full-stage mean; a-side vs b-side drifts are+# uncorrelated, r ≈ 0.0–0.08 → sampling noise, not a coherent biological+# signal). The truth's intermediate stage carries no such draw noise, so every+# output type sub-cloud sits offset from where that type's intermediate cells+# belong — an mmd_u (cell-state distribution) mismatch that no existing+# channel touches: the displacement Δ is computed from FULL-stage type means+# (so it cannot undo the drawn subset's realized offset), DETR/NBHDCOH repair+# the support channel, pair-β only compresses within-type spread.+# WITHINP recenters each shared type's drawn cells per side by a per-gene+# multiplicative factor f_g = clip(((μ_full_g + ε)/(μ_pre_g + ε))^w, LO, HI),+# where μ_full is the type's FULL-stage mean and μ_pre its PRE-displacement+# drawn mean (the draw-noise ratio, carried multiplicatively through the+# additive nnz-only displacement). Exact pseudobulk recentering; zero support+# bit-preserved (no ON-flip injection — node-37 lesson; all-zero columns are+# automatic no-ops); coordinates untouched; ε pseudocount + clip bound the+# factor for low-expression / small-sample genes. w = 0 (or T2_WITHINP=0, or+# --ablate mechanism) reproduces the parent bit-for-bit.+WITHINP = os.environ.get("T2_WITHINP", "1") == "1"+WITHINP_W = float(os.environ.get("T2_WITHINP_W", "1.0"))+WITHINP_EPS = float(os.environ.get("T2_WITHINP_EPS", "0.03"))+WITHINP_CLIP = (float(os.environ.get("T2_WITHINP_LO", "0.8")),+                float(os.environ.get("T2_WITHINP_HI", "1.25")))+WITHINP_MIN_FULL = int(os.environ.get("T2_WITHINP_MIN_FULL", "10"))+WITHINP_MIN_DRAWN = int(os.environ.get("T2_WITHINP_MIN_DRAWN", "5"))+# WITHINP_ALLTYPES: also recenter one-side-only output types (present at just+# one bracket stage). The recentering for one side only needs that side's own+# FULL-stage mean, so the shared-type gate (both stages >= MIN_FULL) is not+# required — this extends the draw-noise correction to the ~37% of output cells+# in one-side-only types.+WITHINP_ALLTYPES = os.environ.get("T2_WITHINP_ALLTYPES", "1") == "1"+# WITHINP_PBNEUTRAL: after the per-type recentering, rescale each gene's FULL+# output column by g_g = pb0_g / pb1_g (pooled over all output cells) so the+# per-gene pooled pseudobulk — and hence dp, the input of de_score /+# de_direction — is restored bit-for-bit (float32 rounding aside). The scorer-+# visible effect of WITHINP is then purely a redistribution of each gene's+# mass BETWEEN types/cells at fixed pooled mean: exactly the cell-state-+# distribution channel (mmd_u / variogram / neighborhood) with the DE-rank+# channel structurally untouched. The uniform g slightly offsets every type+# mean but preserves the recentered between-type structure.+WITHINP_PBNEUTRAL = os.environ.get("T2_WITHINP_PBNEUTRAL", "1") == "1"+WITHINP_G_CLIP = (float(os.environ.get("T2_WITHINP_GLO", "0.9")),+                  float(os.environ.get("T2_WITHINP_GHI", "1.1"))) # AMPSHRINK (mechanism of this node 39, family T2EI-01): DETR repaired the # zero/nonzero SUPPORT channel; the nnz AMPLITUDE channel is still frozen at # the source stage (a-side cells keep the E_a nnz-value distribution, b-side@@ -786,6 +902,43 @@ def corr_kernel_within(stage_a, stage_b) -> np.ndarray:     return C.astype(np.float32)  +def progcoh_partner_sets(X0: np.ndarray, det_th: float, rmin: float, top_p: int):+    """Per-gene co-expression partner sets on the output cloud (PROGCOH).++    Dense Pearson correlation over all cells of X0; partner candidates are+    gated to genes detected (>0) in more than det_th of cells (low-detection+    correlations are dropout noise). For each gene g: top_p genes by |r| among+    candidates with |r| >= rmin (deterministic: stable argsort on -|r|, ties by+    gene index); None when no candidate passes rmin.+    """+    Xc = X0.astype(np.float64)+    det_c = (Xc > 0).mean(axis=0)+    gate_c = det_c > det_th+    Xc = Xc - Xc.mean(axis=0, keepdims=True)+    sd_c = Xc.std(axis=0)+    sd_c[sd_c < 1e-12] = np.inf+    C = (Xc.T @ Xc) / Xc.shape[0]+    C = C / np.outer(sd_c, sd_c)+    np.fill_diagonal(C, 0.0)+    absC = np.abs(C).astype(np.float32)+    absC[:, ~gate_c] = -1.0+    parts, rvals = [], []+    for g in range(absC.shape[0]):+        r = absC[g].copy()+        r[g] = -1.0+        ok = np.where(r >= rmin)[0]+        if ok.size:+            order = ok[np.argsort(-r[ok], kind="stable")][:top_p]+            parts.append(order)+            rvals.append(np.asarray(r[order], dtype=np.float64))+        else:+            parts.append(None)+    stats = {"gate_frac": float(gate_c.mean()),+             "coverage": float(np.mean([p is not None for p in parts])),+             "r_med": float(np.median(np.concatenate(rvals))) if rvals else None}+    return parts, stats++ def pathway_boost(abs_d: np.ndarray, M, k: int):     """boost_g = (# top-k active sets containing g) / (# all sets containing g)."""     s = (M.astype(np.float64) @ abs_d) / np.maximum(M.sum(axis=1), 1)@@ -1355,6 +1508,99 @@ def mix_converge(stage_a, stage_b, t: float, params: dict, alpha: float, view: s                 variso_pb_maxshift=float(np.max(pbshifts)),             ) +    # WITHINP (this node 44, alternative mechanism): recenter each shared+    # type's drawn cells per side onto the full-stage type mean by a bounded+    # per-gene multiplicative factor (draw-noise removal; support preserved).+    # See the config block for the full rationale.+    withinp_info = {"withinp_enable": bool(WITHINP), "withinp_w": WITHINP_W,+                    "withinp_eps": WITHINP_EPS, "withinp_clip": list(WITHINP_CLIP),+                    "withinp_n_sides": 0, "withinp_f_dev_mean": None,+                    "withinp_f_dev_max": None, "withinp_pb_maxshift": None,+                    "withinp_drift_rel_med": None, "withinp_alltypes": bool(WITHINP_ALLTYPES),+                    "withinp_pbneutral": None,+                    "withinp_g_min": None, "withinp_g_max": None,+                    "withinp_g_dev_mean": None}+    if WITHINP and WITHINP_W != 0.0 and ia.size and ib.size:+        la_w = np.asarray(stage_a.labels).astype(str)+        lb_w = np.asarray(stage_b.labels).astype(str)+        lao_w = la_w[ia]+        lbo_w = lb_w[ib]+        cnt_af = {k: int((la_w == k).sum()) for k in set(la_w.tolist())}+        cnt_bf = {k: int((lb_w == k).sum()) for k in set(lb_w.tolist())}+        pre_a = as_dense(stage_a.X, ia)+        pre_b = as_dense(stage_b.X, ib)+        n_out_w = int(xa.shape[0] + xb.shape[0])+        pb_pool0 = ((xa.astype(np.float64).sum(axis=0) + xb.astype(np.float64).sum(axis=0))+                    / n_out_w) if WITHINP_PBNEUTRAL else None+        fdev, drifts, pbsh = [], [], []+        # WITHINP_ALLTYPES: the recentering for one side only needs that side's+        # FULL-stage mean of the type — no cross-side information. Restricting+        # to types present at BOTH stages (the displacement gate) would skip+        # the one-side-only types (~37% of output cells), whose drawn subsets+        # carry the same sampling drift. With ALLTYPES, every output type whose+        # own source stage has >= MIN_FULL cells of it is recentered on that+        # side.+        if WITHINP_ALLTYPES:+            labs_w = sorted(set(lao_w.tolist()) | set(lbo_w.tolist()))+        else:+            labs_w = sorted(set(lao_w.tolist()) & set(lbo_w.tolist()))+        for lab in labs_w:+            for side in ("a", "b"):+                lo_side = lao_w if side == "a" else lbo_w+                cnt_full = cnt_af if side == "a" else cnt_bf+                mo = lo_side == lab+                if int(mo.sum()) < WITHINP_MIN_DRAWN:+                    continue+                if cnt_full.get(lab, 0) < WITHINP_MIN_FULL:+                    continue+                X = xa if side == "a" else xb+                pre = pre_a if side == "a" else pre_b+                st = stage_a if side == "a" else stage_b+                lfull = la_w if side == "a" else lb_w+                Xfd = np.asarray(st.X[lfull == lab].toarray(), dtype=np.float64)+                mu_full = Xfd.mean(axis=0)+                tstd = Xfd.std(axis=0)+                mu_pre = pre[mo].astype(np.float64).mean(axis=0)+                ratio = (mu_full + WITHINP_EPS) / (mu_pre + WITHINP_EPS)+                f = np.clip(ratio ** WITHINP_W, WITHINP_CLIP[0], WITHINP_CLIP[1]).astype(np.float32)+                pb0 = X[mo].astype(np.float64).mean(axis=0)+                X[mo] = X[mo] * f[None, :]+                pb1 = X[mo].astype(np.float64).mean(axis=0)+                fdev.append(np.abs(f - 1.0))+                drifts.append(np.abs(mu_full - mu_pre) / np.maximum(tstd, 1e-6))+                pbsh.append(float(np.abs(pb1 - pb0).max()))+                withinp_info["withinp_n_sides"] += 1+        if fdev:+            fall = np.concatenate(fdev)+            withinp_info["withinp_f_dev_mean"] = float(fall.mean())+            withinp_info["withinp_f_dev_max"] = float(fall.max())+            withinp_info["withinp_pb_maxshift"] = float(np.max(pbsh))+            withinp_info["withinp_drift_rel_med"] = float(np.median(np.concatenate(drifts)))+        if WITHINP_PBNEUTRAL and pb_pool0 is not None and n_out_w:+            # Restore each gene's POOLED output pseudobulk (hence dp, the input+            # of de_score / de_direction) to its pre-WITHINP value with one+            # uniform per-gene rescale of all output cells. WITHINP's+            # scorer-visible effect becomes purely a between-type redistribution+            # of each gene's mass at fixed pooled mean — the cell-state channel+            # (mmd_u / variogram / neighborhood) with the DE-rank channel+            # structurally untouched.+            pb_pool1 = (xa.astype(np.float64).sum(axis=0) + xb.astype(np.float64).sum(axis=0)) / n_out_w+            g = np.ones_like(pb_pool0)+            ok = (pb_pool1 > 1e-9) & (pb_pool0 > 1e-9)+            g[ok] = np.clip(pb_pool0[ok] / pb_pool1[ok], WITHINP_G_CLIP[0], WITHINP_G_CLIP[1])+            gf32 = g.astype(np.float32)+            if ia.size:+                xa *= gf32[None, :]+            if ib.size:+                xb *= gf32[None, :]+            withinp_info["withinp_pbneutral"] = True+            withinp_info["withinp_g_min"] = float(g.min())+            withinp_info["withinp_g_max"] = float(g.max())+            withinp_info["withinp_g_dev_mean"] = float(np.abs(g - 1.0).mean())+        else:+            withinp_info["withinp_pbneutral"] = False+        del pre_a, pre_b+     # Final coordinates are computed BEFORE the residual-collapse steps so that     # SPATRESID can gate on the exact output coordinate space. The coordinate     # pipeline does not read xa/xb and consumes no rng between mix_indices and@@ -1618,8 +1864,10 @@ def mix_converge(stage_a, stage_b, t: float, params: dict, alpha: float, view: s                  "detr_coh_shared_move_frac": None, "detr_coh_shared_ns_zero_frac": None,                  "detr_coh_shared_ns_med": None, "detr_coh_shared_ns_iqr": None}         NS_d = None+        X0_d = None         kk_d = 0         move_fr_d, nsz_fr_d, ns_all_d = [], [], []+        pc_move_d = []         if NBHDCOH_SHARED and shared_d:             n_out_d = int(ia.size + ib.size)             kk_d = int(min(NBHDCOH_SHARED_K, n_out_d))@@ -1641,7 +1889,27 @@ def mix_converge(stage_a, stage_b, t: float, params: dict, alpha: float, view: s                                       (rows_d, nbr_d.ravel())), shape=(n_out_d, n_out_d))                 Xfull_d = np.vstack([xa, xb]).astype(np.float32)                 NS_d = np.asarray(W_d @ Xfull_d, dtype=np.float32)-                del W_d, Xfull_d, rows_d, nbr_d+                del W_d, rows_d, nbr_d+                if PROGCOH and DETR_KEY != "gene":+                    X0_d = Xfull_d+                else:+                    del Xfull_d+        # PROGCOH (this node 44): co-expression partner sets on the PRE-removal+        # output cloud (same cloud NS_d was built from). Dense Pearson over all+        # output cells; partner candidates gated to genes detected in > DET of+        # cells (low-detection correlations are dominated by dropout noise).+        PART_pc = None+        pc = {"progcoh_enable": bool(PROGCOH and X0_d is not None),+              "progcoh_lambda": PROGCOH_LAMBDA, "progcoh_p": PROGCOH_P,+              "progcoh_rmin": PROGCOH_RMIN, "progcoh_det": PROGCOH_DET, "progcoh_key": PROGCOH_KEY,+              "progcoh_gate_frac": None, "progcoh_coverage_all": None,+              "progcoh_n_free": 0, "progcoh_n_free_part": 0, "progcoh_n_nopart": 0,+              "progcoh_move_vs_coh": None, "progcoh_r_med": None}+        if X0_d is not None:+            PART_pc, pc_stats = progcoh_partner_sets(X0_d, PROGCOH_DET, PROGCOH_RMIN, PROGCOH_P)+            pc["progcoh_gate_frac"] = pc_stats["gate_frac"]+            pc["progcoh_coverage_all"] = pc_stats["coverage"]+            pc["progcoh_r_med"] = pc_stats["r_med"]         for lab in shared_d:             ma = la_o == lab             mb = lb_o == lab@@ -1704,10 +1972,38 @@ def mix_converge(stage_a, stage_b, t: float, params: dict, alpha: float, view: s                             ns_g = NS_d[gidx[nz], int(g)].astype(np.float64)                             if not NBHDCOH_SELF:                                 ns_g = (ns_g * kk_d - col[nz].astype(np.float64)) / (kk_d - 1.0)-                            ord_ = np.lexsort((nz, u[nz], ns_g))[:k]+                            # PROGCOH (this node 44): joint neighborhood score+                            # over g's co-expression partners — only where+                            # reordering freedom exists (k < nz.size) and the+                            # partner set is non-empty; else NBHDCOH_SHARED key.+                            key = ns_g+                            part = PART_pc[int(g)] if PART_pc is not None else None+                            free = k < nz.size+                            if free:+                                pc["progcoh_n_free"] += 1+                                if part is not None and part.size:+                                    pc["progcoh_n_free_part"] += 1+                                else:+                                    pc["progcoh_n_nopart"] += 1+                            if part is not None and part.size and free:+                                rows_pc = gidx[nz]+                                X0p = X0_d[np.ix_(rows_pc, part)].astype(np.float64)+                                pterm = np.zeros(rows_pc.size)+                                if PROGCOH_KEY in ("nbhd", "both"):+                                    NSp = NS_d[np.ix_(rows_pc, part)].astype(np.float64)+                                    if not NBHDCOH_SELF:+                                        NSp = (NSp * kk_d - X0p) / (kk_d - 1.0)+                                    pterm = pterm + NSp.mean(axis=1)+                                if PROGCOH_KEY in ("self", "both"):+                                    pterm = pterm + X0p.mean(axis=1)+                                key = ns_g + PROGCOH_LAMBDA * pterm+                            ord_ = np.lexsort((nz, u[nz], key))[:k]                             sel = nz[ord_]                             leg = nz[np.argsort(u[nz], kind="stable")[:k]]                             move_fr_d.append(float(np.setdiff1d(sel, leg).size) / k)+                            if part is not None and part.size and free:+                                selc = nz[np.lexsort((nz, u[nz], ns_g))[:k]]+                                pc_move_d.append(float(np.setdiff1d(sel, selc).size) / k)                             nsz_fr_d.append(float((ns_g <= 0.0).mean()))                             ns_all_d.append(ns_g)                             coh_d["detr_coh_shared_n_genes"] += 1@@ -1744,7 +2040,10 @@ def mix_converge(stage_a, stage_b, t: float, params: dict, alpha: float, view: s             coh_d["detr_coh_shared_ns_med"] = float(np.median(ns_cat))             coh_d["detr_coh_shared_ns_iqr"] = float(                 np.subtract(*np.percentile(ns_cat, [75, 25])))+        if pc_move_d:+            pc["progcoh_move_vs_coh"] = float(np.mean(pc_move_d))         det_info.update(coh_d)+        det_info.update(pc)         if n_genes_l:             det_info["detr_n_genes"] = n_genes_l             det_info["detr_rate_gap_before"] = float(np.mean(gaps_before))@@ -2193,6 +2492,7 @@ def mix_converge(stage_a, stage_b, t: float, params: dict, alpha: float, view: s     info.update(aniso_info)     info.update(ta_info)     info.update(variso_info)+    info.update(withinp_info)     info.update(pbproj_info)     info.update(tshift_info)     info.update(clamp_info)@@ -2215,14 +2515,15 @@ def main() -> None:     parser.add_argument("--seed", type=int, default=0)     parser.add_argument("--ablate", default=None,                         help="mechanism-off control: 'mechanism' (or any name) disables this "-                             "node's mechanism (NBHDCOH_SHARED neighborhood-coherent OFF "-                             "selection for the shared-type DETR), keeping the one-side NBHDCOH "-                             "of parent node 41 — reproduces parent node 41 bit-for-bit")+                             "node's submitted mechanism (WITHINP per-type drawn-mean "+                             "recentering) and the falsified PROGCOH — reproduces parent node "+                             "43 bit-for-bit")     args = parser.parse_args()     global RESID_GAMMA, SPATRESID_THETA, DETR_FLIP_ON, AMPS_ENABLE, DETR_EXT, NBHDGATE, NBHDCOH-    global NBHDCOH_SHARED+    global NBHDCOH_SHARED, PROGCOH, WITHINP     if args.ablate:-        NBHDCOH_SHARED = False+        PROGCOH = False+        WITHINP = False     proj_asym = None     eigproj_k = 0 if args.ablate else None @@ -2314,7 +2615,12 @@ def main() -> None:                                            "detr_coh_shared_enable", "detr_coh_shared_k",                                            "detr_coh_shared_n_genes", "detr_coh_shared_n_off",                                            "detr_coh_shared_move_frac", "detr_coh_shared_ns_zero_frac",-                                           "detr_coh_shared_ns_med", "detr_coh_shared_ns_iqr",+                                            "detr_coh_shared_ns_med", "detr_coh_shared_ns_iqr",+                                            "progcoh_enable", "progcoh_lambda", "progcoh_p",+                                            "progcoh_rmin", "progcoh_det", "progcoh_key", "progcoh_gate_frac",+                                            "progcoh_coverage_all", "progcoh_n_free",+                                            "progcoh_n_free_part", "progcoh_n_nopart",+                                            "progcoh_move_vs_coh", "progcoh_r_med",                                            "ampshrink_enable", "ampshrink_strength", "ampshrink_tau", "ampshrink_decode",                                            "ampshrink_n_types", "ampshrink_n_pairs", "ampshrink_n_shrunk",                                            "ampshrink_frac_shrunk", "ampshrink_f_mean", "ampshrink_f_min",@@ -2344,7 +2650,12 @@ def main() -> None:                                         "typeaniso_clip_frac", "typeaniso_tgt_over_cur_mean", "typeaniso_rs",                                         "variso_enable", "variso_damp", "variso_n_types", "variso_f_mean",                                        "variso_f_in_band", "variso_f_min", "variso_f_max",-                                       "variso_cur_over_tgt", "variso_pb_maxshift")}+                                        "variso_cur_over_tgt", "variso_pb_maxshift",+                                        "withinp_enable", "withinp_w", "withinp_eps", "withinp_clip",+                                        "withinp_n_sides", "withinp_f_dev_mean", "withinp_f_dev_max",+                                        "withinp_pb_maxshift", "withinp_drift_rel_med", "withinp_alltypes",+                                        "withinp_pbneutral", "withinp_g_min", "withinp_g_max",+                                        "withinp_g_dev_mean")}     print(json.dumps({"bracket": [a["stage"], b["stage"]], **keep}, default=float), file=sys.stderr)     write_t2(args.out, expr, coords, genes, seed=args.seed) 

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 summary:

**做了什么**：按 PLAN 实现 PROGCOH（共表达程序相关的 DETR 置零选择）——Step-0 门槛通过（44% 自由调用有伙伴），置零确实变得程序相关，但 5 种解码/幅度（nbhd λ=0.5/1、nbhd-cov、self λ=2、both λ=1）全部 ≤ 父节点（65.21–65.26 vs 锚点 65.2615）：真 variogram 只动 ±1.6e-5，nbhd 按偏离幅度受损，效应量被"有重排自由的移除条目"（~1–2k/18.6k）硬性封顶 → 证否（代码留在 `T2_PROGCOH=0` 后）。改交备选机制 **WITHINP**（同一弱项 cell_state）：诊断发现分层抽样使每类型 drawn 均值带纯噪声漂移（小样本侧达 0.17 倍类型 std，两侧漂移不相关），逐基因乘性因子把每类型×侧回正到全阶段均值（支持集/坐标逐位不动），再用整列 g=pb0/pb1 复原池化伪批量使 DE 通道结构性不动；覆盖单侧独有类型（37% 细胞）。

**查分**（A 半同会话，额度 20 用满）：提交配置 3 种子配对差 **+0.070/+0.079/+0.036 全正**（65.33/64.91/64.89 vs 65.26/64.83/64.86），nbhd raw 三种子一致下降（−2.6e-4/−5.0e-4/−2.7e-4），shape 三项逐位不变。`--ablate mechanism` 逐位还原父节点 43（sha256 验证）；vec-check 通过；seed 确定；2.0 s / 0.62 GB。

**下一步最值得试**：① WITHINP 宽 clip [0.6,1.6]（seed 0 达 65.3555，+0.024 未做多种子确认，需先解决 g 触界导致的 dp 复原不完全）；② occupancy_dice 是唯一低于地板的指标（skill 0.42，差地板 −0.65 分、满配 +4.8），先做只读诊断（三轴 RMS 比、16³ 占据谱）再决定是否动坐标；③ 共有 DETR 的 70.5% 全移除调用可试"保留最低 NS′ 的少量细胞"，但需 matched 对照盯 variogram 兑换率。

调研员

oint neighborhood score that averages 15-NN expression across the gene's top co-expression partners, so co-expressed genes are preferentially zeroed in the same cells, preserving the pairwise gene-gene covariance structure that variogram measures.",
  "vs_constant_shift": "No expression value is changed, no cell is added/removed, no coordinate moves. The mechanism changes only WHICH cells lose which transcripts, and does so in a gene-correlated, spatially-conditioned manner: two cells in the same type at different spatial positions have different 15-NN neighborhoods and thus different JNS rankings, producing different zeroing outcomes. A constant shift or global composition reweighting would move all cells' values uniformly and cannot create gene-pair-specific zeroing correlation.",
  "mechanism_evidence": "1) Partner coverage: fraction of non-full-removal DETR genes with ≥1 partner (expect >50%). 2) Zeroing correlation: for the top-20 co-expressed gene pairs, compute the Pearson correlation of their binary zeroing indicator vectors across cells—PROGCOH should show significantly higher correlation than parent (where zeroing is independent). 3) variogram raw must decrease (primary target). 4) Per-gene removal counts must be identical to parent (verify sum of zeros per gene unchanged). 5) Four-group decomposition: cell_state (variogram) should improve; expression_change (de_score raw) must be unchanged; shape_scale must be bit-identical (coordinates untouched); local_spatial (nbhd raw) should be neutral or slightly better.",
  "mechanism_off_control": "Set T2_PROGCOH=0 (or pass --ablate mechanism): the shared-type DETR selection reverts to per-gene independent NBHDCOH_SHARED (NS'_i(g) only, no partner averaging). Output must reproduce parent node 43 bit-for-bit (sha256-verified on seed 0). Expected difference: variogram raw should worsen back to ~0.007073, nbhd raw unchanged, board drops by the mechanism's contribution (~0.3–0.5 if successful).",
  "sources": []
}
```

审查员

REVIEW.json written