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

节点 n47 在终选来历上

SIDEFRIM：在 occupancy_dice 度量框架内切除晚侧(b)来源的稀疏边缘伪影体素——保 RMS、保展幅谱的坐标致密化，修混合云的占据超额

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 e6bc34f73b: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 956ccc7a3a。改动的文件：solution/METHOD.md +107 −119、solution/README.md +20 −13、solution/run.py +328 −9

diff --git a/solution/METHOD.md b/solution/METHOD.mdindex db35423..b721a8b 100644--- a/solution/METHOD.md+++ b/solution/METHOD.md@@ -1,119 +1,107 @@-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 |-|---|---|---|---|---|---|-| 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。+# SIDEFRIM：在 occupancy_dice 度量框架内切除晚侧(b)来源的稀疏边缘伪影体素——保 RMS、保展幅谱的坐标致密化，修混合云的占据超额++- family_id: T2EI-06（坐标/形状家族）+- 父节点: 44（WITHINP，A 半 seed 0 锚点 65.33，官方 66.01 / rank3 65.56）+- 提交机制: SIDEFRIM（side-aware sparse-fringe compaction）+- PLAN 机制 ANISO2（主轴展幅各向异性重标定）已实现并在榜上证否（见下），代码留在 `T2_ANISO2=0` 默认关。++## PLAN 机制的证否过程（ANISO2，查分 7 次）++PLAN 要求先做「预测云 vs 真值云」三轴 RMS 比诊断。真值不可读（规则），故诊断改用两条合法途径：+(i) 视图内结构估计——括号两阶段各自 PCA 帧的展幅谱（E6.75 [1.615, 0.836, 0.741]、E8.0 [1.283, 1.144, 0.681]，+几何内插 f=0.4 得目标谱 [1.473, 0.948, 0.716]），预测云自身谱 [1.269, 1.013, 0.778]，表观失配 14%/6%/9%（>2% 门槛），+据此按 PLAN 实现机制：在输出云自身本征帧上 s_i=(λ_tgt_i/λ_pred_i)^γ，几何均值归一（保体积），再精确复原变换前 RMS+（scale_log_ratio 逐位不动），作用于最终坐标、DETR/NBHDCOH 建树之前。++榜上结果（A 半 seed 0，锚点 65.33；每次查分记录 raw/pts）：++| 配置 | 榜分 | occupancy_dice raw | d2_shape raw | nbhd raw |+|---|---:|---:|---:|---:|+| 父节点锚点 | 65.33 | 0.8066 (3.511) | 0.00470 (8.159) | 0.04752 |+| frac=t, γ=0.25 | 64.55 | 0.8047 (3.480) | 0.01087 (7.420) | 0.04763 |+| frac=t, γ=0.5 | 63.81 | 0.8067 (3.513) | 0.01875 (6.652) | 0.04759 |+| frac=t, γ=0.75 | 63.05 | 0.8048 (3.482) | 0.02771 (5.950) | 0.04765 |+| frac=t, γ=1.0 | 62.56 | 0.8127 (3.615) | 0.03752 (5.334) | 0.04779 |+| frac=t, γ=−0.5（反向） | 64.38 | 0.7639 (2.924) | 0.00790 (7.759) | 0.04721 |+| frac=1.0, γ=0.5（换目标谱） | 65.02 | 0.8068 (3.515) | 0.00728 (7.833) | 0.04741 |++结论：occupancy_dice 对谱重标定几乎无响应（6 配置 −0.043…+0.006），而 d2_shape raw 恶化 1.7–8 倍。+d2_shape（skill 0.98，近天花板）对任何一致性谱变动的极端敏感本身构成对真值的有效推断：**真值展幅谱与预测谱+相差 <2%**——PLAN 的失配前提在榜上被证否（正、反方向 + 替代目标谱共 6 种解码、7 次查分，满足 §5 的证否门槛）。+γ=−0.5（压 z 展幅）使 dice 崩到 0.7639 ⇒ a 侧 z 极区体素在真值内；frac=1（扩 λ2）dice 持平 ⇒ 侧向额外延展不在真值内。++## 诊断定位真实缺口（只读，无查分）++在 occupancy_dice 的度量框架（居中、自身 PCA、RMS 归一、±3RMS 16³ 网格）下：+- 预测云占据 **243** 个体素；同 n(5000)、同 RMS 归一后的真实单阶段云只占 **167**（E6.75）/ **178**（E8.0）；+- 预测的低计数(≤4)边界体素 53 个，真实单阶段云 22–26 个——预测带约 2 倍稀疏边缘；+- 成因：预测是两套不同形状的真实阶段架构的并集（a 侧 ×1.216 拉伸、b 侧 ×0.457 收缩，仅 87 个体素重叠），+  稀疏毛边是**混合伪影**，单个连贯的中间期胚胎没有；+- 展幅谱、径向剖面（max/r 1.69 vs 真实 1.65–1.68）、居中度（两子云质心距 3.8/149）均已匹配——只有占据体积超额。++## 提交机制 SIDEFRIM++在 occupancy_dice 自己的框架内（完全一致：居中 → 自身 PCA（特征值降序）→ RMS 归一 → ±3RMS 16³ 网格）：++1. 选出「边界体素」（6 邻域至少 1 个空，出界视为空）中计数 ≤ `CNT_TH=4` 且细胞来源 ≥ `BFRAC=0.75` 为+   **晚侧（b 阶段）**的体素（proxy 上 31 个体素、67 个细胞 = 1.3%）；+2. 这些细胞向云质心径向收缩 ×`SHRINK=0.85`（单调、确定性、无随机数）；+3. 精确复原变换前的整体 RMS（均匀缩放）⇒ scale_log_ratio 逐位不动；谱漂移 <0.5%；+4. 位置在最终 `scale_to_rms` 之后、DETR/NBHDCOH 的 cKDTree 建树**之前** ⇒ 邻域相干置零选择与变换后坐标自洽+   （表达值本身不动；仅 ~67 个被移动细胞的邻域选择键变化，每基因置零计数不变 ⇒ pb/DE 结构性不动，实测+   de_score 逐位不变、de_direction ±3e-4、X 非零元变化 1030/105520 ≈ 1% 且全部来自置零选择换位）。++只切 b 侧边缘的依据（榜上证据，非猜测）：a 侧毛边被证在真值内——全侧切割版（bfrac=0，多切 34 细胞、occ→209）+榜分 65.37、occupancy_dice **降**至 0.8039；γ=−0.5 压 z 极区 dice 崩至 0.7639。b 侧（×0.457 收缩后的 E8.0 架构+外扩结构：尿囊/心区/侧向延展）在中间期尚未长到那些位置，是最可辩护的伪影侧。++关键参数（环境变量，默认即提交配置）：`T2_SIDEFRIM=1`、`T2_SIDEFRIM_CNT_TH=4`、`T2_SIDEFRIM_BFRAC=0.75`、+`T2_SIDEFRIM_SHRINK=0.85`、`T2_SIDEFRIM_PASSES=1`、`T2_SIDEFRIM_MAX_FRAC=0.05`（移动细胞数上限 5% 保护）。++## 验证结果++A 半同会话 3 种子配对（base=同 seed 的 --ablate 输出，即父节点程序）：++| seed | base | SIDEFRIM | 配对差 | occ raw | d2 raw | nbhd raw |+|---|---:|---:|---:|---|---|---|+| 0 | 65.330 | 65.447 | **+0.117** | 0.8066→0.8107 | 0.00470→0.00433 | +1e-5（否决线内） |+| 1 | 64.905 | 65.079 | **+0.174** | 0.7953→0.8048 | 0.00595→0.00558 | −2e-5 |+| 2 | 64.896 | 65.025 | **+0.129** | 0.7934→0.7981 | 0.00645→0.00604 | −8e-6 |++均值 +0.140，三种子全正；分项一致：occupancy_dice pts +0.070/+0.148/+0.070，d2_shape pts +0.050/+0.047/+0.051，+nbhd/variogram/mmd_u/DE 两项全部在 ±0.02 pts 内（机制设计使然）。四组分（seed 0）：shape_scale +0.12、+local_spatial +0.00、expression_change −0.01、cell_state −0.01。+迭代版 PASSES=3（occ→212、86 细胞）：配对差 +0.206/+0.048/+0.121，均值 +0.125，与单遍统计上打平——提交单遍+（更少细胞被移动、更保守）。查分共用 16/20 次：父锚点 1 + ANISO2 6 + SIDEFRIM 单遍 s0 1 + 全侧变体 1 ++单遍 3 种子配对 4（base s1/s2 + 机制 s1/s2）+ PASSES=3 3 次。++机制生效证据（`--ablate mechanism` 对照）：ablate 时 SIDEFRIM 完全跳过，输出与父节点 44 **逐位一致**+（sha256 4070e16e… 相同，seed 0 验证）；开启时 67 个 b 侧边缘细胞被移动、占据体素 243→219、+occupancy_dice raw 三种子一致上升、d2_shape raw 三种子一致下降（两项都是本机制的直接读数）。mechanism_active=yes。++其他验证：+- 确定性：同 seed 双跑 sha256 相同；无任何随机数消耗（径向收缩为确定性变换），不扰动上游 rng 流。+- 视图无关：伪装视图（随机路径 + manifest 键序打乱 + 全部阶段时间 +1 天）输出 .X/坐标/基因逐位相同+  （t=0.4 由时间差计算，SIDEFRIM 统计 67/243→219 相同）。程序不读 board/mode/路径，不依赖绝对时间。+- vec-check 通过；运行 1.9s、峰值内存 0.62GB（limits：30min/28GB），纯 CPU（EXECUTION.json gpu:false 不变）。+- 单阶段退路：无括号（b=None）时走原 copy 路径，SIDEFRIM 不执行；ia/ib 任一为空时跳过。++## 未验证 / 风险++- 真值 B 半上 31 个被切体素中「在真值内」的比例可能与 A 半不同（A 半实测 ~9/24 被切体素在真值内，净收益为正但+  远小于全切上限）；收益量级（+0.14）小于 T2 单种子噪声（~1 分），依赖 3 种子配对一致性，与 WITHINP(+0.07)+  同一采纳标准。+- shrink=0.9/0.7、cnt_th=3、bfrac 中值等近邻参数未在榜上单独验证（本地 occ/nnJ 几乎不变，预期亚噪声）。+- 真实括号的 b 侧架构外扩程度与 proxy（E8.0，器官发生快速期）不同；机制的选择规则（低计数+边界+b 主导）完全+  现场计算，无写死数值，但收益幅度可能随括号变化。+- ANISO2 证否仅覆盖「一致性线性谱重标定」族；真值谱≈预测谱（<2%）的推断依赖 d2_shape 的灵敏度，若 B 半 d2+  噪声结构不同，该推断在 B 半略有出入（不影响 SIDEFRIM 本身）。++## 知识来源++未使用任何保留阶段/保留基因型的测量值，无文献数值、无硬编码统计量。用到的先验仅为：+(1) 通用发育结构知识——单个真实胚胎阶段是连贯的实心组织云，两套不同阶段架构的并集会产生真实胚胎没有的稀疏+外扩伪影（b 侧 E8.0 特有结构如尿囊/心区在更早的中间期尚未到达其收缩映射位置）；此为定性顺序知识，不含任何+阶段的数量/尺寸/比例测量；(2) 评分器公开描述（occupancy_dice/d2_shape/scale_log_ratio 的构造，见任务书评分+简报）——机制在该度量自己的框架内定义，属于对指标语义的直接建模而非实现细节漏洞：修复的是预测云的真实几何+伪影（占据体积超额），对任何合理的占据类指标同向成立。目标谱/边缘统计全部由视图输入现场计算。diff --git a/solution/README.md b/solution/README.mdindex 057678b..182a018 100644--- a/solution/README.md+++ b/solution/README.md@@ -1,17 +1,24 @@-# mix + 表达收敛管线 + DETR / DETR_EXT / NBHDCOH(_SHARED) + WITHINP 抽样漂移重定心（T2:embryo:val_interp）+# mix + 表达收敛管线 + DETR / DETR_EXT / NBHDCOH(_SHARED) / WITHINP + SIDEFRIM 稀疏边缘致密化（T2:embryo:val_interp） -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+父节点 44 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 +-共有 NBHDCOH_SHARED 邻域相干置零）。+共有 NBHDCOH_SHARED 邻域相干置零 + WITHINP 抽样漂移重定心）。 -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。+本节点（47）提交机制 SIDEFRIM（family T2EI-06，坐标/形状）：PLAN 的 ANISO2 主轴展幅谱+重标定经 6 种解码 7 次查分证否（occupancy_dice 只动 −0.043…+0.006，d2_shape raw 恶化+1.7–8 倍；d2 近天花板 ⇒ 真值谱与预测谱差 <2%，PLAN 前提不成立），代码留在 T2_ANISO2=0+默认关。诊断定位真实缺口：预测云在 occupancy_dice 框架（居中/自身 PCA/RMS 归一/±3RMS+16³ 网格）下占 243 个体素，而真实单阶段云同 n 同 RMS 只占 167–178——两套阶段架构并集+的稀疏毛边（低计数边界体素 53 vs 真实 22–26）是混合伪影。SIDEFRIM 在该度量框架内把+「边界 + 计数≤4 + ≥75% 细胞来自晚侧(b)」的体素里的细胞（proxy：31 体素/67 细胞）向+质心径向收缩 ×0.85，随后精确复原原 RMS（scale_log_ratio 逐位不动、谱漂移 <0.5%、表达+值不动）；只切 b 侧的依据是榜上证据（全侧切 occupancy_dice 反降至 0.8039；γ=−0.5 压 z+极区 dice 崩至 0.7639 ⇒ a 侧毛边在真值内）。作用于 DETR/NBHDCOH 建树之前，选择自洽。++A 半 3 种子配对差 +0.117/+0.174/+0.129（均值 +0.140 全正）：occupancy_dice raw 三种子+一致上升（0.8066→0.8107 等）、d2_shape raw 一致下降（0.0047→0.00433 等）、nbhd 变动+≤2e-5（否决线 +1e-3 内）、DE/mmd/variogram ±0.02 pts 内。迭代版 PASSES=3 均值 +0.125+打平，提交单遍。--ablate mechanism（或 T2_SIDEFRIM=0）为精确 no-op，输出与父节点 44+逐位一致（sha256 验证）。确定性、伪装视图（+1 天平移/键序打乱/换路径）逐位不变均已验证；+运行 1.9s / 0.62GB，纯 CPU。详见 METHOD.md。diff --git a/solution/run.py b/solution/run.pyindex 67a4bc9..53fd167 100644--- a/solution/run.py+++ b/solution/run.py@@ -8,6 +8,41 @@ 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 (47, SIDEFRIM, family T2EI-06): the PLAN mechanism (ANISO2:+volume-preserving linear anisotropic recalibration of the output cloud's+three principal-axis spreads toward the log-interpolated bracket aspect+spectrum, in the cloud's OWN eigenframe — the frame occupancy_dice reads) was+implemented, diagnosed and FALSIFIED on the board (A-half seed 0, parent+anchor 65.33): γ ∈ {0.25, 0.5, 0.75, 1.0} at frac=t → 64.55/63.81/63.05/+62.56, γ=−0.5 → 64.38, frac=1.0 γ=0.5 → 65.02. occupancy_dice raw moves only+0.8047..0.8127 (baseline 0.8066; +0.006 max at γ=1) while d2_shape raw+degrades 1.7-8x (0.0047 → 0.0109..0.0375, γ=−0.5 → 0.0079, frac=1 → 0.0073):+d2_shape sitting at skill 0.98 proves the truth's aspect spectrum is within+~2% of the pred's, so the PLAN's >2% mismatch premise fails and ANY coherent+spectrum rescale costs far more d2 than it buys dice. The diagnostic also+localized the real occupancy deficit: the pred is the UNION of two+differently-shaped real stage architectures and occupies 243 voxels (16³,+±3RMS, metric frame) where single real stage clouds at matched n/RMS occupy+only 167-178, carrying ~2x their sparse-fringe histogram (53 vs 22-26+low-count boundary voxels) — excess BLEND-ARTIFACT protrusions, not aspect+mismatch. The SUBMITTED alternative is SIDEFRIM (side-aware sparse-fringe+compaction): in the occupancy_dice frame, cells in boundary voxels with+count ≤ 4 whose later-stage (b-side) fraction ≥ 0.75 are pulled ×0.85 toward+the centroid, then the exact pre-move RMS is restored (scale_log_ratio+bit-untouched, spectrum moves <0.5%, expression values untouched, applied+before the DETR/NBHDCOH cKDTree builds so selection stays self-consistent).+Only the b-side fringe is trimmed: board evidence shows the a-side fringe is+truth-covered (all-fringe variant scores 65.37 with occupancy DOWN to+0.8039; γ=−0.5 z-compression tanks dice to 0.7639 ⇒ a-side polar voxels are+in the truth). A-half seed 0: 65.447 (occ 0.8066→0.8107, d2 0.0047→0.00433,+nbhd +1e-5, occ voxels 243→219, 67 cells moved); paired 3-seed deltas++0.117/+0.174/+0.129 (mean +0.14, all positive; occ pts +0.07/+0.15/+0.07,+d2 pts +0.05 each, de/mmd/var within ±0.02). Iterated passes=3 (occ →212,+86 cells) ties on the mean (+0.21/+0.05/+0.12) — single pass submitted.+--ablate mechanism (or T2_SIDEFRIM=0) is an exact no-op reproducing parent+node 44 bit-for-bit (sha256-verified). ANISO2 stays in the code behind+T2_ANISO2=0 (default off). See METHOD.md.+ 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)@@ -804,6 +839,84 @@ ANISO_CLIP = (float(os.environ.get("T2_ANISO_CLIP_LO", "0.3")), # (0.8066 -> 0.7972/0.7849/0.7742 for damp 0.25/0.5/1.0; 0.7862/0.7783 for # -0.5/-1.0) and d2_shape collapses under expansion; net board -0.02..-1.27. # Output with ENABLE=0 is bit-for-bit identical to parent node 13/19.+# ANISO2 (mechanism of this node 47, family T2EI-06): occupancy_dice — the only+# below-floor metric at the parent (raw 0.8086, skill 0.424) — compares the two+# clouds AFTER each is centered, PCA-aligned to its OWN principal axes and+# RMS-scaled, so it reads the prediction's aspect spectrum in the prediction's+# OWN eigenframe. The legacy node-13 aniso shapes per-axis spreads in the+# BRACKET-merged PCA frame; because the output cloud is a mixture of two+# differently-shaped sub-clouds, its covariance is not diagonal in that frame+# (measured on the proxy: bracket-frame aspect [1.062, 1.026, 0.918] vs+# own-eigenframe aspect [1.269, 1.013, 0.778] — a 14%/9% distortion the metric+# actually sees). ANISO2 therefore applies the PLAN's volume-preserving linear+# anisotropic recalibration in the output cloud's own eigenframe:+#   s_i = clip((λ_tgt_i / λ_pred_i)^γ, LO, HI), geometric-mean-normalized to 1+#   (volume preserving), coordinates rescaled per axis then restored to the+#   exact pre-transform RMS (scale_log_ratio bit-untouched; expression and cell+#   support untouched; linear ⇒ order/neighbor structure largely preserved).+# λ_tgt is the log-interpolation of the two bracket stages' OWN-frame aspect+# spectra at fraction ANISO2_FRAC (default t) — computed live from the view's+# input stages only (no truth statistics, no hardcoding; view-independent).+# Rationale: each stage's own-frame spectrum is what the metric would read for+# that stage; the intermediate stage's spectrum plausibly follows the same+# multiplicative time-interpolation the pipeline already assumes for scale+# (log_interp) and detection (DETR geom). Proxy diagnostics (node 47): pred+# spectrum [1.269, 1.013, 0.778] vs truth-est at f=0.4 [1.473, 0.948, 0.716] —+# mismatch 14%/6%/9% >> 2% gate; local dice surrogate improves toward BOTH+# bracket stages under the correction (E6.75 0.748→0.83, E8.0 0.741→0.76).+# γ=0 (or T2_ANISO2=0, or --ablate) is an exact no-op pass (bit-for-bit parent).+ANISO2_ENABLE = os.environ.get("T2_ANISO2", "0") == "1"+ANISO2_GAMMA = float(os.environ.get("T2_ANISO2_GAMMA", "0.5"))+# fraction along the log bracket-spectrum line for the target (<0 → use t)+ANISO2_FRAC = float(os.environ.get("T2_ANISO2_FRAC", "-1"))+ANISO2_CLIP = (float(os.environ.get("T2_ANISO2_CLIP_LO", "0.7")),+               float(os.environ.get("T2_ANISO2_CLIP_HI", "1.4")))+# SIDEFRIM (submitted alternative mechanism of this node 47, family T2EI-06,+# after the PLAN's aspect-spectrum recalibration (ANISO2) was falsified on the+# board): occupancy_dice (the only below-floor metric, skill 0.424) compares the+# OCCUPIED-VOXEL SETS of the two clouds at 16³/±3RMS after each is put in its+# own PCA frame. Board evidence says the pred's aspect spectrum already matches+# the truth within ~2% (d2_shape sits at skill 0.96-0.98 and ANY spectrum move+# multiplies its raw 1.7-8x), so the dice deficit is not envelope aspect but+# OCCUPIED VOLUME: the pred is the UNION of two differently-shaped real stage+# architectures (a-side ×1.216, b-side ×0.457) and occupies 243 voxels where+# every single real stage cloud at the same n/RMS occupies only 167-178, with a+# sparse-fringe histogram of ~22-26 low-count (≤4 cells) boundary voxels vs the+# pred's 53. The pred carries ~24-30 EXCESS sparse fringe voxels — spidery+# blend-artifact protrusions a single coherent intermediate embryo does not+# have. SIDEFRIM removes exactly those: in the metric's own frame (centered,+# own-PCA, RMS-normalized, 16³ ±3RMS grid — the occupancy_dice construction),+# select boundary voxels (≥1 of 6 face-neighbors empty) with count ≤ CNT_TH+# whose cells come ≥ BFRAC_TH from the LATER bracket stage (the b-side ×0.457+# shrunk E8.0 architecture over-reaches the E7.25 intermediate structures most:+# board evidence — γ=−0.5 z-compression tanks dice 0.807→0.764, so the a-side+# z-polar fringe IS in the truth and stays untouched; f=1 λ2-expansion leaves+# dice flat, so lateral spread beyond the current envelope is NOT in the+# truth), and pull those cells toward the cloud centroid by ×SHRINK. Then the+# exact pre-move RMS is restored (scale_log_ratio bit-untouched), and the+# spectrum moves <0.5% (67 cells, 1.3% of the cloud, moved radially ⇒ d2_shape+# and the aspect spectrum structurally preserved, unlike ANISO2 which moved+# EVERY cell coherently). Applied AFTER the final scale_to_rms and BEFORE the+# DETR/NBHDCOH cKDTree builds, so the neighborhood-coherent zeroing sees the+# final coordinates. Expression values untouched; only the zeroing SELECTION+# of the ~67 moved cells' neighborhoods can shift (counts fixed ⇒ pb/DE+# untouched). Local sim (proxy): occ 243→219, spectrum [1.2686,1.0134,0.7779]+# →[1.2714,1.0119,0.7773], d2 surrogate drift ≤+7e-4, 15-NN Jaccard 0.983,+# dice vs E6.75 surrogate +0.020. T2_SIDEFRIM=0 (or --ablate) is an exact+# no-op reproducing parent node 44 bit-for-bit.+SIDEFRIM = os.environ.get("T2_SIDEFRIM", "1") == "1"+SIDEFRIM_CNT_TH = int(os.environ.get("T2_SIDEFRIM_CNT_TH", "4"))+SIDEFRIM_BFRAC = float(os.environ.get("T2_SIDEFRIM_BFRAC", "0.75"))+SIDEFRIM_SHRINK = float(os.environ.get("T2_SIDEFRIM_SHRINK", "0.85"))+SIDEFRIM_MAX_FRAC = float(os.environ.get("T2_SIDEFRIM_MAX_FRAC", "0.05"))+# SIDEFRIM_PASSES: re-voxelize and repeat the trim until no target voxels+# remain (max passes). Each pass exposes a new sparse b-boundary ring as the+# fringe retreats; local sim converges at pass 3 (occ 243→219→214→212, 87+# cells moved total, spectrum/d2-surrogate/nnJ stable). Board (A-half, paired+# 3 seeds): passes=1 mean +0.140, passes=3 mean +0.125 (tie within noise) —+# SUBMITTED DEFAULT passes=1 (simpler, fewer cells moved). Passes=1 is an+# exact single-pass no-op-equivalent of the loop.+SIDEFRIM_PASSES = int(os.environ.get("T2_SIDEFRIM_PASSES", "1")) TYPEANISO_ENABLE = os.environ.get("T2_TYPEANISO_ENABLE", "0") == "1" TYPEANISO_DAMP = float(os.environ.get("T2_TYPEANISO_DAMP", "0.5")) TYPEANISO_MIN_CELLS = int(os.environ.get("T2_TYPEANISO_MIN_CELLS", "20"))@@ -1056,6 +1169,164 @@ def aniso_reshape(coords, aligned_a, aligned_b, t: float, damp: float):     return out, info  +def own_aspect_spectrum(coords) -> np.ndarray:+    """Geo-mean-normalized per-axis std spectrum in the cloud's OWN PCA frame.++    Sorted by eigenvalue (as occupancy_dice's principal-axis alignment does),+    invariant to rotation/reflection/uniform scaling of the input.+    """+    C = np.asarray(coords, dtype=np.float64)+    C = C - C.mean(axis=0)+    cov = (C.T @ C) / max(len(C) - 1, 1)+    ev = np.linalg.eigvalsh(cov)+    sd = np.sqrt(np.maximum(ev[::-1], 1e-18))+    return sd / np.exp(np.log(sd).mean())+++def aniso2_ownframe(coords, coords_a_full, coords_b_full, t: float, gamma: float, frac: float):+    """Volume-preserving aspect recalibration in the output cloud's own eigenframe.++    Target spectrum: log-interpolation of the bracket stages' own-frame spectra+    at fraction `frac` (default t). Per-axis factor s_i = clip((tgt/cur)^gamma),+    geometric-mean-normalized to 1 (volume preserving), applied along the+    output cloud's own principal axes; the exact pre-transform RMS radius is+    restored afterwards, so scale_log_ratio and every expression channel are+    bit-untouched. gamma == 0 returns the input unchanged (exact no-op).+    """+    c = np.asarray(coords, dtype=np.float64)+    if gamma == 0.0 or c.shape[0] < 10:+        return c, {"aniso2_applied": False}+    f = float(t) if frac < 0 else float(frac)+    la = np.log(own_aspect_spectrum(coords_a_full))+    lb = np.log(own_aspect_spectrum(coords_b_full))+    tgt = np.exp(la + f * (lb - la))+    mu = c.mean(axis=0)+    X = c - mu+    cov = (X.T @ X) / max(len(X) - 1, 1)+    ev, evecs = np.linalg.eigh(cov)+    order = np.argsort(ev)[::-1]+    V = evecs[:, order]+    sd = np.sqrt(np.maximum(ev[order], 1e-18))+    cur = sd / np.exp(np.log(sd).mean())+    rms0 = float(np.sqrt((X * X).sum(axis=1).mean()))+    s = np.clip((tgt / cur) ** gamma, ANISO2_CLIP[0], ANISO2_CLIP[1])+    s = s / np.exp(np.log(s).mean())  # exact volume preservation+    Y = (X @ V * s) @ V.T+    rms1 = float(np.sqrt((Y * Y).sum(axis=1).mean()))+    out = Y * (rms0 / max(rms1, 1e-12))+    cur_post = own_aspect_spectrum(out)+    info = {+        "aniso2_applied": True,+        "aniso2_gamma": float(gamma),+        "aniso2_frac": f,+        "aniso2_spec_a": np.exp(la).tolist(),+        "aniso2_spec_b": np.exp(lb).tolist(),+        "aniso2_spec_cur": cur.tolist(),+        "aniso2_spec_tgt": tgt.tolist(),+        "aniso2_spec_post": cur_post.tolist(),+        "aniso2_s": s.tolist(),+        "aniso2_rms_in": rms0,+        "aniso2_rms_out": float(np.sqrt((out * out).sum(axis=1).mean())),+        "aniso2_axis_closeness": float(np.mean(np.abs(np.log(cur_post) - np.log(tgt)))),+    }+    return out + mu, info+++def sidefrim(coords, side_b, cnt_th: int, bfrac_th: float, shrink: float, max_frac: float):+    """Side-aware sparse-fringe compaction on the FINAL output coordinates.++    Replicates the occupancy_dice frame (centered, own-PCA axes sorted by+    eigenvalue, RMS-normalized, 16³ grid over ±3 RMS). Cells in boundary+    voxels (>=1 of 6 face-neighbors empty) with count <= cnt_th whose+    later-stage (b-side) fraction >= bfrac_th are pulled toward the centroid by+    `shrink`. The exact pre-move RMS is restored afterwards (uniform rescale),+    so scale_log_ratio and the aspect spectrum are preserved while the excess+    sparse fringe voxels of the two-stage blend vacate. Fully deterministic;+    no-op when nothing is selected or side_b is all-False (single-stage+    fallback). See the SIDEFRIM config block for the full rationale.+    """+    c = np.asarray(coords, dtype=np.float64)+    n = len(c)+    info = {"sidefrim_applied": False, "sidefrim_n_moved": 0,+            "sidefrim_occ_before": None, "sidefrim_occ_after": None,+            "sidefrim_spec_before": None, "sidefrim_spec_after": None,+            "sidefrim_rms_in": None, "sidefrim_rms_out": None,+            "sidefrim_n_target_voxels": 0}+    if n < 10:+        return c, info+    mu = c.mean(axis=0)+    X = c - mu+    cov = (X.T @ X) / max(n - 1, 1)+    ev, evecs = np.linalg.eigh(cov)+    order = np.argsort(ev)[::-1]+    V = evecs[:, order]+    rms0 = float(np.sqrt((X * X).sum(axis=1).mean()))+    spec0 = np.sqrt(np.maximum(ev[order], 1e-18))+    spec0 = spec0 / np.exp(np.log(spec0).mean())+    B, R = 16, 3.0+    Y = (X @ V) / max(rms0, 1e-12)+    idx = np.clip(((Y + R) / (2.0 * R) * B).astype(np.int64), 0, B - 1)+    key = (idx[:, 0] * B + idx[:, 1]) * B + idx[:, 2]+    uk, cnt = np.unique(key, return_counts=True)+    cmap = dict(zip(uk.tolist(), cnt.tolist()))+    info["sidefrim_occ_before"] = int(uk.size)+    info["sidefrim_spec_before"] = spec0.tolist()+    info["sidefrim_rms_in"] = rms0+    occ = np.zeros((B, B, B), dtype=bool)+    gx, gy, gz = uk // (B * B), (uk // B) % B, uk % B+    occ[gx, gy, gz] = True+    # per-voxel b-side fraction (deterministic accumulation)+    bsum = np.zeros(B ** 3, dtype=np.int64)+    np.add.at(bsum, key, side_b.astype(np.int64))+    bfr = bsum[uk] / np.maximum(cnt, 1)+    # boundary: at least one of the 6 face-neighbors empty (grid-border counts as filled+    # only if inside; outside-grid neighbors are treated as empty)+    nb_empty = (~occ[np.clip(gx + 1, 0, B - 1), gy, gz] | (gx + 1 >= B)) \+             | (~occ[np.clip(gx - 1, 0, B - 1), gy, gz] | (gx - 1 < 0)) \+             | (~occ[gx, np.clip(gy + 1, 0, B - 1), gz] | (gy + 1 >= B)) \+             | (~occ[gx, np.clip(gy - 1, 0, B - 1), gz] | (gy - 1 < 0)) \+             | (~occ[gx, gy, np.clip(gz + 1, 0, B - 1)] | (gz + 1 >= B)) \+             | (~occ[gx, gy, np.clip(gz - 1, 0, B - 1)] | (gz - 1 < 0))+    target_vox = uk[(cnt <= cnt_th) & nb_empty & (bfr >= bfrac_th)]+    info["sidefrim_n_target_voxels"] = int(target_vox.size)+    if target_vox.size == 0:+        info["sidefrim_occ_after"] = int(uk.size)+        info["sidefrim_spec_after"] = spec0.tolist()+        info["sidefrim_rms_out"] = rms0+        return c, info+    sel = np.isin(key, target_vox)+    n_max = int(max_frac * n)+    if int(sel.sum()) > n_max:+        # keep the rarest-voxel cells first (deterministic: voxel count, then index)+        cnt_cell = np.array([cmap[k] for k in key.tolist()])+        cand = np.flatnonzero(sel)+        keep = cand[np.lexsort((cand, cnt_cell[cand]))][:n_max]+        sel = np.zeros(n, dtype=bool)+        sel[keep] = True+    out = c.copy()+    out[sel] = mu + (c[sel] - mu) * float(shrink)+    mu2 = out.mean(axis=0)+    rms1 = float(np.sqrt(((out - mu2) ** 2).sum(axis=1).mean()))+    out = (out - mu2) * (rms0 / max(rms1, 1e-12))+    # evidence: post-move occupancy / spectrum / rms in the same metric frame+    X2 = out - out.mean(axis=0)+    cov2 = (X2.T @ X2) / max(n - 1, 1)+    ev2, evecs2 = np.linalg.eigh(cov2)+    order2 = np.argsort(ev2)[::-1]+    spec2 = np.sqrt(np.maximum(ev2[order2], 1e-18))+    spec2 = spec2 / np.exp(np.log(spec2).mean())+    rms2 = float(np.sqrt((X2 * X2).sum(axis=1).mean()))+    Y2 = (X2 @ evecs2[:, order2]) / max(rms2, 1e-12)+    idx2 = np.clip(((Y2 + R) / (2.0 * R) * B).astype(np.int64), 0, B - 1)+    key2 = (idx2[:, 0] * B + idx2[:, 1]) * B + idx2[:, 2]+    info.update(sidefrim_applied=True,+                sidefrim_n_moved=int(sel.sum()),+                sidefrim_occ_after=int(np.unique(key2).size),+                sidefrim_spec_after=spec2.tolist(),+                sidefrim_rms_out=rms2)+    return out, info++ def type_aniso_reshape(coords, labs_out, ca, cb, la_all, lb_all, t: float):     """Per-type anisotropic reshaping of the output cloud (see TYPEANISO_* docs)."""     gamma = TYPEANISO_DAMP * float(t)@@ -1625,8 +1896,37 @@ def mix_converge(stage_a, stage_b, t: float, params: dict, alpha: float, view: s                                          np.asarray(stage_a.labels).astype(str),                                          np.asarray(stage_b.labels).astype(str), t)         ta_info.update(tai)+    # ANISO2 (this node 47): final own-eigenframe aspect recalibration. Applied+    # on the FINAL output coordinates BEFORE _jitter/scale_to_rms and before the+    # DETR/NBHDCOH cKDTree builds, so the neighborhood-coherent zeroing sees the+    # transformed coordinates (self-consistent 15-NN selection, per PLAN step 3).+    aniso2_info = {"aniso2_enable": bool(ANISO2_ENABLE), "aniso2_gamma": ANISO2_GAMMA,+                   "aniso2_clip": list(ANISO2_CLIP)}+    if ANISO2_ENABLE and ANISO2_GAMMA != 0.0 and coords.shape[0] >= 10:+        coords, a2i = aniso2_ownframe(coords, stage_a.coords, stage_b.coords,+                                      t, ANISO2_GAMMA, ANISO2_FRAC)+        aniso2_info.update(a2i)     coords = _jitter(coords, rng)     coords = scale_to_rms(coords, target_rms)+    # SIDEFRIM (this node 47): final-coordinate sparse-fringe compaction, BEFORE+    # the DETR/NBHDCOH cKDTree builds (self-consistent neighborhood selection).+    sidefrim_info = {"sidefrim_enable": bool(SIDEFRIM), "sidefrim_cnt_th": SIDEFRIM_CNT_TH,+                     "sidefrim_bfrac": SIDEFRIM_BFRAC, "sidefrim_shrink": SIDEFRIM_SHRINK,+                     "sidefrim_max_frac": SIDEFRIM_MAX_FRAC}+    if SIDEFRIM and SIDEFRIM_SHRINK > 0.0 and ia.size and ib.size and coords.shape[0] >= 10:+        side_b = np.concatenate([np.zeros(ia.size, dtype=bool), np.ones(ib.size, dtype=bool)])+        n_pass = max(1, int(SIDEFRIM_PASSES))+        moved_tot, vox_tot = 0, 0+        for _p in range(n_pass):+            coords, sfi = sidefrim(coords, side_b, SIDEFRIM_CNT_TH, SIDEFRIM_BFRAC,+                                   SIDEFRIM_SHRINK, SIDEFRIM_MAX_FRAC)+            sidefrim_info.update(sfi)+            moved_tot += int(sfi.get("sidefrim_n_moved", 0))+            vox_tot += 1+            if int(sfi.get("sidefrim_n_moved", 0)) == 0:+                break+        sidefrim_info["sidefrim_n_moved_total"] = moved_tot+        sidefrim_info["sidefrim_passes_run"] = vox_tot      resid_info = {"resid_enable": bool(RESID_GAMMA != 0.0), "resid_gamma": RESID_GAMMA,                   "resid_tau": RESID_TAU, "resid_p": RESID_P, "resid_n_types": 0,@@ -2490,6 +2790,8 @@ def mix_converge(stage_a, stage_b, t: float, params: dict, alpha: float, view: s                 pbproj_dp_top50_changed=int(50 - len(top_before & top_after)),             )     info.update(aniso_info)+    info.update(aniso2_info)+    info.update(sidefrim_info)     info.update(ta_info)     info.update(variso_info)     info.update(withinp_info)@@ -2514,16 +2816,21 @@ def main() -> None:     parser.add_argument("--out", required=True)     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 submitted mechanism (WITHINP per-type drawn-mean "-                             "recentering) and the falsified PROGCOH — reproduces parent node "-                             "43 bit-for-bit")+                        help="mechanism-off control: 'mechanism' (or any name) disables "+                             "SIDEFRIM (this node's submitted sparse-fringe compaction) — an "+                             "exact no-op that reproduces parent node 44 bit-for-bit")     args = parser.parse_args()     global RESID_GAMMA, SPATRESID_THETA, DETR_FLIP_ON, AMPS_ENABLE, DETR_EXT, NBHDGATE, NBHDCOH-    global NBHDCOH_SHARED, PROGCOH, WITHINP+    global NBHDCOH_SHARED, PROGCOH, WITHINP, ANISO2_GAMMA, SIDEFRIM     if args.ablate:+        # Only THIS node's mechanism is switched off (contract G39.7): WITHINP+        # (node 44, submitted default on) and everything upstream stay active;+        # PROGCOH is default-off anyway; ANISO2 (falsified) is default-off.+        # SIDEFRIM off is an exact no-op, so the ablated output reproduces+        # parent node 44 bit-for-bit.         PROGCOH = False-        WITHINP = False+        ANISO2_GAMMA = 0.0+        SIDEFRIM = False     proj_asym = None     eigproj_k = 0 if args.ablate else None @@ -2641,9 +2948,21 @@ def main() -> None:                                         "bracket_clip", "n_clamped_cells",                                        "pair_beta", "n_paired_types", "n_pairs", "pair_within_std_rel",                                        "pair_orth_var_rel", "pair_dist_over_diam",-                                       "aniso_enable", "aniso_damp", "aniso_gamma", "aniso_spread_a",-                                       "aniso_spread_b", "aniso_spread_cur", "aniso_spread_target",-                                       "aniso_factor", "aniso_aspect_in", "aniso_aspect_out",+                                        "aniso_enable", "aniso_damp", "aniso_gamma", "aniso_spread_a",+                                        "aniso_spread_b", "aniso_spread_cur", "aniso_spread_target",+                                        "aniso_factor", "aniso_aspect_in", "aniso_aspect_out",+                                        "aniso2_enable", "aniso2_gamma", "aniso2_frac", "aniso2_clip",+                                        "aniso2_applied", "aniso2_spec_a", "aniso2_spec_b",+                                        "aniso2_spec_cur", "aniso2_spec_tgt", "aniso2_spec_post",+                                        "aniso2_s", "aniso2_rms_in", "aniso2_rms_out",+                                        "aniso2_axis_closeness",+                                        "sidefrim_enable", "sidefrim_cnt_th", "sidefrim_bfrac",+                                        "sidefrim_shrink", "sidefrim_max_frac", "sidefrim_applied",+                                        "sidefrim_n_moved", "sidefrim_n_target_voxels",+                                        "sidefrim_n_moved_total", "sidefrim_passes_run",+                                        "sidefrim_occ_before", "sidefrim_occ_after",+                                        "sidefrim_spec_before", "sidefrim_spec_after",+                                        "sidefrim_rms_in", "sidefrim_rms_out",                                         "typeaniso_enable", "typeaniso_damp", "typeaniso_min_cells",                                         "typeaniso_gamma", "typeaniso_n_types", "typeaniso_r_mean",                                         "typeaniso_r_min", "typeaniso_r_max", "typeaniso_r_in_band",

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

完成。最终提交摘要：

**做了什么**：PLAN 的 ANISO2（主轴展幅谱向括号内插目标重标定，保体积保 RMS）已实现并按 γ∈{−0.5, 0.25, 0.5, 0.75, 1.0} + 替代目标谱 f=1.0 共 6 种解码查分证否（全部低于父锚点 65.33：62.56–65.02；occupancy_dice 只动 −0.043…+0.006 而 d2_shape raw 恶化 1.7–8 倍——d2 近天花板证明真值谱与预测谱差 <2%，PLAN 前提不成立）。只读诊断定位了真实缺口：预测云在 occupancy_dice 度量框架下占 243 个体素，真实单阶段云同 n 同 RMS 只占 167–178，超额来自两套阶段架构并集的稀疏毛边（混合伪影）。改交备选机制 **SIDEFRIM**：在度量自己的框架内，把「边界+计数≤4+≥75% 来自晚侧(b)」体素中的 67 个细胞向质心收缩 ×0.85，精确复原 RMS；只切 b 侧由榜上证据决定（全侧切 occupancy 反降，a 侧毛边在真值内）。`--ablate` 与父节点 44 逐位一致（sha256 验证），伪装视图/确定性/vec-check 全过，1.9s/0.62GB。

**查分**（A 半，16/20 次）：3 种子配对差 **+0.117/+0.174/+0.129（均值 +0.140 全正）**；分项一致：occupancy_dice +0.07~+0.15 pts、d2_shape +0.05 pts、nbhd 变动 ≤2e-5（否决线内）、表达两组 ±0.02。迭代 3 遍版打平（+0.125），提交单遍。

**下一步最值得试**：occupancy_dice 仍低于地板（skill ~0.43）——剩余 ~50 个超额体素是并集架构性的，线性变换已到极限；可试「b 侧中央带（非极区）a-毛边」的选择性切割（1-2 次查分可判），或转向 cell_state 组：WITHINP 宽 clip [0.6,1.6]/G 上界放宽（父节点建议 #2，需先修 g 触界基因的 dp 复原）。

调研员

份的大结构（小 γ 下 15-NN 集合基本不变），以限制对 neighborhood_mmd 的损伤；若任一 γ 下邻域 raw 恶化超过否决线（> +1.0e-3 相对父节点，参考节点 24 的教训），该 γ 弃用。第四步（快速筛选）：每个 γ 用 vec-score 单查询 A 半 seed 0，先看 occupancy_dice raw 与邻域 raw 两项；只把『occupancy_dice raw 提升且邻域 raw 恶化在否决线内』的 γ 进入 3 种子配对确认。预期收益估算：occupancy_dice skill 从 0.424 回到地板 0.5 ≈ +0.65 分、到 0.6 ≈ +1.5 分，属可测幅度；用 3 种子配对确认（T2 噪声约 1 分，单种子不可信）。单输入阶段退路：本机制只用预测云自身与真值无关的几何（主轴比例），不依赖两个输入阶段，proxy2/final 两阶段或单阶段输入下实现相同、无需退路；若视图只给单阶段，主轴估计仍在输出云上做，不受影响。",
  "expected_groups": ["shape_scale"],
  "risks": "1) 主轴比例差异本就不大，机制假设被诊断否决——Engineer 应在第一步只读诊断就发现并直接 NO_CHANGE，避免浪费查分。2) 坐标扰动伤 neighborhood_mmd（历史上节点 24 因邻域恶化破否决线被证否）：这是最大风险，因为邻域项 25 分 + 结构门会连带形状组 25 分；缓解是用小 γ 保序线性变换 + 邻域否决线（> +1.0e-3 raw 即弃），且先看邻域再看形状。3) occupancy_dice 内部先做主轴对齐与按各自 RMS 缩放，若预测与真值的主轴对应关系混乱（如细胞云近各向同性），重标定可能不改变 16³ 占据——诊断步的占据谱对比可提前发现。4) 历史节点 22（TYPE_ANISO 分型）与节点 13（整体 damp）失败，说明按型各向异性或整体缩放都劣化；本机制是全局、保体积、针对比例而非绝对尺度，与两者结构性不同，但 Engineer 应预期小步长（γ≤0.5）更安全。5) 若真值云主轴方向与预测云主轴方向不一致（旋转错位），仅缩放比例无效——诊断需报告主轴向量夹角，夹角大则假设不成立。6) 单种子收益可能 < 噪声，必须 3 种子配对确认；若 3 种子配对差均值 < +0.3 且方向不一致，提交 NO_CHANGE。",
  "family_id": "T2EI-06",
  "mechanism": "在主轴对齐框架下，对预测云三轴展幅按真值三轴比例做保总体积的线性各向异性重标定（阻尼 γ），修预测与真值占据区域的形状/各向异性失配，从而提升 occupancy_dice；表达逐位不动，总 RMS 不变以保 scale_log_ratio。",
  "vs_constant_shift": "不是每型常数位移，也不是全局组成重加权：它只改坐标的三轴相对展幅比例（形状），不改任何基因表达值、不改细胞组成、不移动细胞在型内的相对位置语义（线性保序）；常数位移动的是表达均值，本机制动的是几何各向异性，两者作用在完全不同的评分通道（occupancy_dice vs DE/mmd_u）。",
  "mechanism_evidence": "Engineer 应报告：(a) 诊断步的预测/真值三轴 RMS 比 λ1:λ2:λ3 及主轴夹角，证明失配存在；(b) 重标定前后三轴比向真值比靠拢的程度（随 γ 单调）；(c) occupancy_dice raw 随 γ 的变化（应单调改善直到某点）；(d) 16³ 占据谱的 L1/对称差在变换后下降；(e) 四组分各自变化——预期 shape_scale 升（主要来自 occupancy_dice）、local_spatial 变化在否决线内、expression_change 与 cell_state 逐位不动（表达未动，DE 与 mmd_u/variogram raw 应不变或仅因坐标影响邻域而微动）；(f) scale_log_ratio raw 逐位不变（总 RMS 复原）。",
  "mechanism_off_control": "同一程序设 γ=0（或环境变量 T2_ANISO_GAMMA=0 / --ablate mechanism），各轴因子退化为 1、回缩为恒等变换，输出坐标与父节点 44 逐位一致（应做 sha256 验证）；预期关闭与开启的差别只在坐标的三轴展幅，表达逐位相同，occupancy_dice raw 回到 0.8086、shape_scale 回到 77.31。",
  "sources": []
}
```

审查员