Add pair-isolated P1 v2.2 geometry experiments, fine-type baseline and reproducible inputs

.gitattributes

@@ -58,3 +58,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+p1_v2_fine_types/figures/spatial_fusion_ablation.pdf filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,3 +1,12 @@
+---
+language:
+- en
+- zh
+tags:
+- neuroscience
+- connectomics
+- reproducibility
+---
 # NeuronMimic reusable data (2026-10-30)
 
 This dataset contains compact, source-bound tables and a deterministic passive
@@ -11,13 +20,11 @@ BioFly project at `artifacts/wu_20261030_science_gate/WU_REQUIREMENTS_STATUS_202
 
 ## Contents
 
-- `physics_teacher/`: deterministic 12-compartment Teacher stimulus panel,
-  train/held-out arrays, model comparison, finite-difference gradient check,
-  and Teacher-rechecked active-search results. Every file is
-  `FIXTURE_TEST_ONLY`.
-- `malecns/`: normalized MaleCNS v1.0 directed aggregate edges and annotations,
-  with source hashes. These are `VERIFIED_REAL_DATA` structural tables, while
-  contact-to-branch mapping is explicitly blocked.
+- `physics_teacher/`: historical fixture archive. Its old evaluation metrics are
+  **INVALIDATED**; use `p0_corrected_20261030/fixture_v2/` instead.
+- `malecns/`: historical normalized source-bound aggregate tables. Endpoint annotation
+  joins are **BLOCKED**: 31,355 edge endpoint IDs do not match even the full source id
+  column. These tables are not a validated complete CNS or simulation-ready network.
 - `flywire_kc_atlas/`: compact all-5,177-cell structural profiles and the
   source-bound 20-unit entry registry. This is single-connectome structural
   evidence, not independent biological replication.
@@ -45,3 +52,18 @@ and held-out physiological recordings.
 ## Corrected P0 release
 
 Use `p0_corrected_20261030/fixture_v2/` for reproducible Teacher data, four model baselines, and corrected results. Historical metric files are invalidated; see `HISTORICAL_METRICS_INVALIDATED.md`. All cable experiments are FIXTURE_TEST_ONLY. Code and scientific metric tests: https://github.com/ydchen0806/neuronmimic . Known-equation PhysicsOnly is more accurate than the trained Students; no biological or speed-accuracy advantage is established.
+
+## P1-v2.2 geometry and structure release
+
+`p1_v2_fine_types/` contains the 5,177-cell identity registry, CM/SU crosswalk,
+all directed pair split assignments, contact-free tree descriptors, six genuine spatial
+fusion ablations, full-parent child membership, model predictions, figures and audit receipts.
+See its P1_REPORT.md for conclusions and limits. The 18-cell left-ap1 residual subset
+is exploratory, not a validated functional unit. Original stable8 remains unchanged.
+
+Reproduction instructions: https://github.com/ydchen0806/neuronmimic/blob/master/docs/P1_REPRODUCE.md .
+The full soma distance matrix is exactly regenerated by the script rather than downloaded.
+Raw-source audits and geometry extraction are not rerun by the compact replay; their
+original hashes and receipts are provided. Do not treat imported audit receipts as a fresh audit.
+Upstream data retain their original source terms; this release does not grant a new license
+for the original connectomes. No independent physiology or wet-lab calibration data are included.

SHA256SUMS.json

@@ -1,6 +1,6 @@
 {
   "dataset": "neuronmimic-data",
-  "version": "2026-10-30-P0-corrected",
+  "version": "2026-10-30-P0-corrected-P1-v2.2",
   "files": [
     {
       "path": "HISTORICAL_METRICS_INVALIDATED.md",
@@ -9,8 +9,8 @@
     },
     {
       "path": "README.md",
-      "bytes": 2469,
-      "sha256": "66e2123904f71908b3392df85e221a556f4ae962edffc5e60b23a6a4f6e824f2"
+      "bytes": 3561,
+      "sha256": "1e8c148e210a4f55de23f63d5e8c729f1936f1932df5f914bd24571ced21ef12"
     },
     {
       "path": "flywire_kc_atlas/all5177_cells.parquet",
@@ -167,6 +167,211 @@
       "bytes": 1273,
       "sha256": "131f2141c7aba2ea9f191c4010ef3adaf52fce5978e0762d7cd7d35e836dc9de"
     },
+    {
+      "path": "p1_v2_fine_types/EXECUTION_LOG.md",
+      "bytes": 879,
+      "sha256": "ac509cc278388d0f5584b271ff770c4897c03dbd3dc84216f05a1ffe31a08261"
+    },
+    {
+      "path": "p1_v2_fine_types/MODEL_COMPARISON.csv",
+      "bytes": 1675,
+      "sha256": "09e61fceb6fd41a3382365bcfa7a50778405146b2436525ae605d0f010cd46b7"
+    },
+    {
+      "path": "p1_v2_fine_types/P1_REPORT.md",
+      "bytes": 5383,
+      "sha256": "6f10d61b8368076b9dcd33a9687fac5a9047d5b1630a75e1382f0b113aef0413"
+    },
+    {
+      "path": "p1_v2_fine_types/P1_STATUS.json",
+      "bytes": 447,
+      "sha256": "bef95431eb5e195542c61d52c96a50c5e1ab5c4ee732605fd867d0b540d5ce23"
+    },
+    {
+      "path": "p1_v2_fine_types/analysis/CM019_residual_child_overlap.csv",
+      "bytes": 1178,
+      "sha256": "9e1ac1914f315777e0b2eebb843e050908469d94527a0aa746b698ebfbd743d1"
+    },
+    {
+      "path": "p1_v2_fine_types/analysis/complete_parent_splits.csv",
+      "bytes": 3092,
+      "sha256": "e74154274ebc04b7dbbef0d21d8367e8ab9c6bf55e0ff763e8781f9533433677"
+    },
+    {
+      "path": "p1_v2_fine_types/analysis/discovery_ablation.csv",
+      "bytes": 1273,
+      "sha256": "6064151513ce785c9a0158599abb2988ee4a0dc41a681659c31a308635129905"
+    },
+    {
+      "path": "p1_v2_fine_types/analysis/parent_candidate_counts.csv",
+      "bytes": 408,
+      "sha256": "48c730da0cfeb9c910dfe698fc52174386f673e65fd669643344cd2a10e21ae5"
+    },
+    {
+      "path": "p1_v2_fine_types/analysis/vertex_bootstrap_exploratory.csv",
+      "bytes": 1100,
+      "sha256": "15ac2bf43bd6f40356ffc5a94e1b1680f6cbf98d84c19a608cebe196f6ba49cb"
+    },
+    {
+      "path": "p1_v2_fine_types/audit/competition_literature_lookup.json",
+      "bytes": 275,
+      "sha256": "8a3dc3475c48a4f3b1ca1ea444850980560ce0aee153a13a318ee96ede568208"
+    },
+    {
+      "path": "p1_v2_fine_types/audit/environment.json",
+      "bytes": 432,
+      "sha256": "8801289aed1fcba7bb0fac98a994978da446d483828805a9a59677c06cffd0c9"
+    },
+    {
+      "path": "p1_v2_fine_types/audit/finalizer_snapshot.py",
+      "bytes": 12959,
+      "sha256": "acf21a9a721de1104da75075290cd16472e72bcb9e778e53d555c6d0c4c0e270"
+    },
+    {
+      "path": "p1_v2_fine_types/audit/malecns_id_QC.json",
+      "bytes": 397,
+      "sha256": "9e849f7412be762f50e290c044175076063caa6d10be32b4e26797ac71fa97c9"
+    },
+    {
+      "path": "p1_v2_fine_types/audit/pair_split_QC.json",
+      "bytes": 277,
+      "sha256": "f3794a661cad820c0bad80ca122db5eb369ef66aca2a02d52827a003ddf4392e"
+    },
+    {
+      "path": "p1_v2_fine_types/audit/primitives_snapshot.py",
+      "bytes": 2788,
+      "sha256": "5b5917ee598dc56bbe4cfbcfb24638dfcafd7438deb5d891d8f87e8d121d1bf4"
+    },
+    {
+      "path": "p1_v2_fine_types/audit/protocol.json",
+      "bytes": 2408,
+      "sha256": "e53795499f7d9297a692525a2468c6387728b52f4bb4f37267b28e9c638374cd"
+    },
+    {
+      "path": "p1_v2_fine_types/audit/release_source_manifest.json",
+      "bytes": 342,
+      "sha256": "5ad79bcc5d7993f660162368b8d4b13efaa65241e5f97c6b626074cb1d605fec"
+    },
+    {
+      "path": "p1_v2_fine_types/audit/run.log",
+      "bytes": 840,
+      "sha256": "431554d7cf91c92ad0d1d212f219de8d9da0ea7653334e2cc9bc7f9fcbe867fd"
+    },
+    {
+      "path": "p1_v2_fine_types/audit/runner_snapshot.py",
+      "bytes": 16302,
+      "sha256": "3880e3fd0c529a43c6e714584e9aa98fbfa08d1940980900644ff0baa88628ef"
+    },
+    {
+      "path": "p1_v2_fine_types/audit/source_manifest.json",
+      "bytes": 1117176,
+      "sha256": "8052f87e7da7a68ed86ce9c6cdb1ebfdf5863e10761aded89360b207cb61c3b6"
+    },
+    {
+      "path": "p1_v2_fine_types/audit/standalone_replay_checks.json",
+      "bytes": 257,
+      "sha256": "9d439a0c7ae1c3069611c611152dca3e82169463ed91ef78e5261236129fe4ee"
+    },
+    {
+      "path": "p1_v2_fine_types/audit/structural_checks.csv",
+      "bytes": 422,
+      "sha256": "6a76d649524a320cecbdbe20303af33a452716dc5e14af969cda4176b277d4db"
+    },
+    {
+      "path": "p1_v2_fine_types/data/contact_free_tree_features.parquet",
+      "bytes": 769954,
+      "sha256": "02faead833dd5b846efb11e749b0ced53252176e664d8adfb34897f7db4da2d4"
+    },
+    {
+      "path": "p1_v2_fine_types/data/directed_pair_split.parquet",
+      "bytes": 735097,
+      "sha256": "ac28bbbd928cc8f4459850ea35a8732fd6191228cc7fac5db048676d9f115891"
+    },
+    {
+      "path": "p1_v2_fine_types/data/feature_schema.json",
+      "bytes": 785,
+      "sha256": "1a2cb48fe28d9363b5f892faca5887ab6693a67d76ec517b958ce8c13b13d7ff"
+    },
+    {
+      "path": "p1_v2_fine_types/data/pair_grouped_train.npz",
+      "bytes": 465705,
+      "sha256": "c987937c78c13c5f3859859aafa0cf6ccf4ce28a8ca073e9a6273b9a46d590a8"
+    },
+    {
+      "path": "p1_v2_fine_types/data/prediction_pairs.parquet",
+      "bytes": 437528,
+      "sha256": "0011019ed8394c276f4311f136a87a00d7f3c2f7c774d3280e43c306031cf455"
+    },
+    {
+      "path": "p1_v2_fine_types/data/test_predictions.parquet",
+      "bytes": 847321,
+      "sha256": "d32a46516440e0ec07e1675e9e795c2409e9f86c1e241fe7f030f4806cc09c50"
+    },
+    {
+      "path": "p1_v2_fine_types/data/train_only_embedding.npy",
+      "bytes": 2650752,
+      "sha256": "06e47ce94f182181ef8989e371e1c56e88bb6ea5bc84edc0d0c8d1e28e39ace7"
+    },
+    {
+      "path": "p1_v2_fine_types/figure_source_ledger.csv",
+      "bytes": 1653,
+      "sha256": "7fe1089cf57c317a56d2abbce3407b2306377e9638cace3bbe012bd9f2d7c42c"
+    },
+    {
+      "path": "p1_v2_fine_types/figures/CM019_full_parent_somata.pdf",
+      "bytes": 52765,
+      "sha256": "b04e072d6163ea5a5ebefb3c0e46770ca8a56854bdf5271b461250c13749a241"
+    },
+    {
+      "path": "p1_v2_fine_types/figures/CM019_full_parent_somata.svg",
+      "bytes": 166014,
+      "sha256": "7fa152470f5b47fb64b8a7500f46a7eb8cce7e1777e9a7b18a446acbbd9bf691"
+    },
+    {
+      "path": "p1_v2_fine_types/figures/spatial_fusion_ablation.pdf",
+      "bytes": 245509,
+      "sha256": "866d72796bffd3cb07e65e8570a4445872c1af8748514a292d205a99a59f14f7"
+    },
+    {
+      "path": "p1_v2_fine_types/figures/spatial_fusion_ablation.svg",
+      "bytes": 397732,
+      "sha256": "181ea1daf6ff9dc4d5be30bab1b609fe0f05a8db76e71281718d1ad8a81b3fae"
+    },
+    {
+      "path": "p1_v2_fine_types/registry/CM_SU_crosswalk.csv",
+      "bytes": 783,
+      "sha256": "aae457b66ff7c470c7fd74065658d01d137e4954028aa7a8b9d7b04cb56c06fa"
+    },
+    {
+      "path": "p1_v2_fine_types/registry/P1_exploratory_assignments.parquet",
+      "bytes": 84120,
+      "sha256": "eab3c8ccdaf056fcd085820158de322da0c74f37ddb97b98bd246b073442bebd"
+    },
+    {
+      "path": "p1_v2_fine_types/registry/all5177_identity.parquet",
+      "bytes": 231062,
+      "sha256": "0440f0c0fa77f547df027c929b6b120d8e8121acc59fafe78e880fb8ece4e630"
+    },
+    {
+      "path": "p1_v2_fine_types/registry/ap1L_18_exploratory_members.parquet",
+      "bytes": 31509,
+      "sha256": "6dc7405e9240d0435d2d275f1c0d5a94acebda518ff06a7451620aab8babe24d"
+    },
+    {
+      "path": "p1_v2_fine_types/registry/malecns_unmatched_endpoints.csv",
+      "bytes": 313511,
+      "sha256": "489f6a1882b5167a5a00620c98fe431bad3ab14bf5609a6b41dff9de5b375a1a"
+    },
+    {
+      "path": "p1_v2_fine_types/registry/parent_candidate_residual_atlas.parquet",
+      "bytes": 17946,
+      "sha256": "1e7607070cc9fb9608ee2ec92502c24ea43190a2b1f2bd88114d62d5b8c532f4"
+    },
+    {
+      "path": "p1_v2_fine_types/registry/parent_child_members.parquet",
+      "bytes": 22544,
+      "sha256": "75bf98bd0695832b307ab192adcc09e82a9cb81a63a0a43c80d3d013eb7f32ce"
+    },
     {
       "path": "physics_teacher/active_search_teacher_rechecked.csv",
       "bytes": 2010,

p1_v2_fine_types/EXECUTION_LOG.md

@@ -0,0 +1,10 @@
+# Execution log
+
+- P1-v2.1: original coarse-type pilot completed; preserved in ../p1_v2. Identified omitted KCab-s fine parent and weaker M0.
+- P1-v2.2: protocol amended before rerun; P1A/B/C and descriptor-level D pilot completed. See audit/run.log and source_manifest.json for actual timings, seed and source hashes.
+- Command: PYTHONPATH=simulation/src python simulation/scripts/run_kc_p1_v2.py
+- Validation: PYTHONPATH=simulation/src pytest -q simulation/tests/test_p1_structure.py
+- Posthoc report: python simulation/scripts/finalize_kc_p1_v2.py
+- MaleCNS mapping / independent-connectome confirmation: BLOCKED. Exact missing endpoints are in registry/malecns_unmatched_endpoints.csv.
+- Literature primary-source retrieval: BLOCKED_NETWORK for Crossref/bioRxiv; HF proxy transport is working.
+- NeuronMimic architecture optimization: not part of this structural experiment.

p1_v2_fine_types/MODEL_COMPARISON.csv

@@ -0,0 +1,6 @@
+model,test_log_loss,test_AUROC,validation_log_loss,selected_max_leaves,n_features,n_train,n_validation,n_test,training_budget,training_wall_clock_s,evidence,uncertainty,sampling
+M0,0.2606936352617916,0.946740722974307,0.2608141723135891,31,6,83858,18168,17974,2 settings x100 iterations,0.7510168552398682,RETROSPECTIVE_DESCRIPTOR_PILOT,NOT_ESTIMATED_NO_SIGNIFICANCE_CLAIM,case-control 60000 positive/60000 negative; probabilities not natural prevalence calibrated
+M1,0.25956746261839025,0.9510622572987102,0.2597962231987038,31,20,83858,18168,17974,2 settings x100 iterations,1.0860908031463623,RETROSPECTIVE_DESCRIPTOR_PILOT,NOT_ESTIMATED_NO_SIGNIFICANCE_CLAIM,case-control 60000 positive/60000 negative; probabilities not natural prevalence calibrated
+M2_descriptor,0.24339326296360414,0.956133251125601,0.24257004114678182,31,71,83858,18168,17974,2 settings x100 iterations,2.4745774269104004,RETROSPECTIVE_DESCRIPTOR_PILOT,NOT_ESTIMATED_NO_SIGNIFICANCE_CLAIM,case-control 60000 positive/60000 negative; probabilities not natural prevalence calibrated
+M3_descriptor_discrete,0.23674357076420421,0.9592581579385098,0.2352991587642097,31,74,83858,18168,17974,2 settings x100 iterations,2.54158353805542,RETROSPECTIVE_DESCRIPTOR_PILOT,NOT_ESTIMATED_NO_SIGNIFICANCE_CLAIM,case-control 60000 positive/60000 negative; probabilities not natural prevalence calibrated
+M4_descriptor_continuous,0.2212159097412558,0.9653408044489263,0.2234753210121425,31,263,83858,18168,17974,2 settings x100 iterations,7.409607648849487,RETROSPECTIVE_DESCRIPTOR_PILOT,NOT_ESTIMATED_NO_SIGNIFICANCE_CLAIM,case-control 60000 positive/60000 negative; probabilities not natural prevalence calibrated

p1_v2_fine_types/P1_REPORT.md

@@ -0,0 +1,41 @@
+# P1-v2.2：身份、双向隔离与真实空间融合试验
+
+本轮已完成 P1A–P1C 的可执行试验，以及 P1D 的描述量级模型比较。原 stable8 仍为 8 个、965 个成员；新增稳定功能单元数为 0。全部结果来自实际运行，未用 fixture 替代结构数据。
+
+## 当前结论
+
+1. 胞体与树几何已真正进入发现目标。全体 5,177 KC 的 5,177×5,177 欧氏距离矩阵按 μm 生成；树特征从真实骨架计算，未读取接点标签。六个固定消融方案、每方案五个初始化及三次 20% 无序对删边测试均已运行。
+2. 空间更紧凑不等于更稳定。纯图的平均簇胞体半径为 38.16 μm；大胞体权重降至 26.86 μm，但删边 ARI 为 0.638，未超过纯图的 0.646。图+胞体+树的删边 ARI 为 0.817，而平均半径仍为 37.78 μm。这里的 ARI 是整体分区重现性，不是旧 frozen S/coverage，也不代表个别新簇通过稳定门槛。
+3. 几何描述之后，离散标签还有预测增量，但连续训练图表征更强。测试 log loss：M0（已知粗/细类型+侧别）0.26069；M1（另加胞体/距离/训练图强度）0.25957；M2_descriptor（另加树描述量）0.24339；M3（另加离散标签）0.23674；M4（另加连续图嵌入）0.22122。这不等价于“连续几何胜出”：M4 使用的是连续图表征；M2 还没有精细区域树接近性和分枝接点，M4 特征维数也更高。
+
+4. CM019 residual 有具体追踪线索：融合胞体与树几何后，K=3 与 K=4 得到完全相同的18细胞子簇，含全部16个原 residual 与2个原CM019成员；K=2时它们进入20细胞子簇。成员见 registry/ap1L_18_exploratory_members.parquet。这是融合特征空间中的分离，不能直接推断其胞体构成单一连续空间团；尚需扰动、空间匹配零模型、输入输出留出与跨脑检验。
+
+## 数据与协议
+
+FlyWire v783，293,762 条有向神经元对、379,338 个聚合突触计数。无序 pair 哈希决定训练/验证/测试：205,483 / 44,335 / 43,944 条有向边；三个集合之间同向和反向直接交集均为 0。所有图嵌入与度/强度仅来自训练图。接点派生特征在本轮完全排除；这不是已经完成接点级训练测试映射。
+
+预测样本为 60,000 正例和 60,000 非边；训练/验证/测试样本 83,858 / 18,168 / 17,974。各模型同为两个超参设置×100轮，验证集选择，测试集只评估最终设置。这是 case-control 采样，预测值不是自然患率下的连接概率，更不是突触传递概率。原全图与历史测试数据曾参与开发，本次所有切分均标记回顾性，不能称为盲测。
+
+旧 CM、SU、新 P1 分区独立命名。SU019 与 CM019 为 104/132 重叠。P1 全局 K=30 仅为固定分辨率消融，并非新增 30 个稳定单元。局部试验覆盖完整细类型 parent：左右 KCab-s 311/310、ap1 148/133、KCab-p 67/61；K=2/3/4 是探索尺度，全部子簇成员与残余可追溯。CM019 的 16 个 residual 在各子簇中的去向见 CM019_residual_child_overlap.csv；没有匹配零模型和独立验证前，不称为一致补集。
+
+## 修正与失败项
+
+首轮 p1_v2 使用粗粒度 official_KC_subtype，漏掉 KCab-s parent。该试跑原样保留；当前 p1_v2_fine_types 使用 frozen_detailed_subtype 定义完整 parent，M0 同时保留粗/细标签。右 ap1 的粗标签计数132、细标签133不同，未静默覆盖。
+
+MaleCNS 的完整原始 id 列有211,577个唯一ID，body列187,016个；已存在 body 的行两者一致。但换用完整 id 后仍有31,355个边端点未匹配，不能靠填充body解决。跨版本/来源/范围需继续查原始连接表生产链。本轮不宣称跨脑复现。
+
+本地旧跨连接组比较的 provenance 指向名含 invalid_seed_schedule 的历史benchmark，不能升级为独立复现。BANC 与 FlyWire 的样本来源关系亦须核实，数据集名称不同不能自动算不同动物。
+
+用户提供 Pleijzier 2026 预印本已登记为竞争基线，当前代理访问 Crossref/bioRxiv 失败，题目、日期与具体主张仍待原文核查；不把转述当已核实文献结论。
+
+## 剩余裁决工作
+
+P1D 仍需区域树/分枝接点的训练隔离、模型容量与调参预算敏感性、空间/细胞阻断、类型内困难负例、匹配零模型、PN→KC 与 KC→MBON 的预留终点。附加的1000次顶点权重bootstrap仅为固定模型、同脑内的探索性区间，未覆盖模型重拟合、空间相关和动物变异。
+
+P1E 须先解决 MaleCNS 来源映射、BANC 生物独立性与其他数据版本，再比较可迁移组织规律及实验可定位性。目前不应指定“已验证的2–3个功能微回路”。不继续优化 NeuronMimic 架构。
+
+## 复现与产物
+
+仓库命令：`PYTHONPATH=simulation/src python simulation/scripts/run_kc_p1_v2.py`（已完成目录拒绝覆盖）。公开紧凑数据复现命令见 GitHub 的 docs/P1_REPRODUCE.md。发布包含双向切分、身份与父子登记、树描述量、冻结协议、比较表、图及原始输入哈希；完整骨架须向原数据源获取。
+
+关键文件：MODEL_COMPARISON.csv、P1_STATUS.json、analysis/discovery_ablation.csv、analysis/complete_parent_splits.csv、registry/parent_child_members.parquet、audit/structural_checks.csv、audit/source_manifest.json。

p1_v2_fine_types/P1_STATUS.json

@@ -0,0 +1,11 @@
+{
+  "P1A": "DONE_WITH_MALECNS_BLOCKER",
+  "P1B": "DONE_RETROSPECTIVE_PAIR_ISOLATION",
+  "P1C": "DONE_EXPLORATORY_GEOMETRY_DISCOVERY",
+  "P1D": "PARTIAL_DESCRIPTOR_PILOT_NOT_FULL_BRANCH_CONTACT_MODEL",
+  "P1E": "BLOCKED_INDEPENDENCE_AND_ID_QC",
+  "stable8_unchanged": true,
+  "new_stable_candidates": 0,
+  "scientific_conclusion": "No discrete/continuous verdict yet; full independent competition remains open",
+  "wall_clock_s": 55.4731810092926
+}

p1_v2_fine_types/analysis/CM019_residual_child_overlap.csv

@@ -0,0 +1,19 @@
+arm,k,child,child_n,CM019_residual_n,residual_recall,residual_fraction_of_child,evidence
+graph,2,0,22,16,1.0,0.7272727272727273,DESCRIPTIVE_NO_MATCHED_NULL
+graph,2,1,126,0,0.0,0.0,DESCRIPTIVE_NO_MATCHED_NULL
+graph,3,0,16,15,0.9375,0.9375,DESCRIPTIVE_NO_MATCHED_NULL
+graph,3,1,107,0,0.0,0.0,DESCRIPTIVE_NO_MATCHED_NULL
+graph,3,2,25,1,0.0625,0.04,DESCRIPTIVE_NO_MATCHED_NULL
+graph,4,0,94,0,0.0,0.0,DESCRIPTIVE_NO_MATCHED_NULL
+graph,4,1,12,12,0.75,1.0,DESCRIPTIVE_NO_MATCHED_NULL
+graph,4,2,30,0,0.0,0.0,DESCRIPTIVE_NO_MATCHED_NULL
+graph,4,3,12,4,0.25,0.3333333333333333,DESCRIPTIVE_NO_MATCHED_NULL
+graph_soma_tree,2,0,128,0,0.0,0.0,DESCRIPTIVE_NO_MATCHED_NULL
+graph_soma_tree,2,1,20,16,1.0,0.8,DESCRIPTIVE_NO_MATCHED_NULL
+graph_soma_tree,3,0,61,0,0.0,0.0,DESCRIPTIVE_NO_MATCHED_NULL
+graph_soma_tree,3,1,69,0,0.0,0.0,DESCRIPTIVE_NO_MATCHED_NULL
+graph_soma_tree,3,2,18,16,1.0,0.8888888888888888,DESCRIPTIVE_NO_MATCHED_NULL
+graph_soma_tree,4,0,18,16,1.0,0.8888888888888888,DESCRIPTIVE_NO_MATCHED_NULL
+graph_soma_tree,4,1,55,0,0.0,0.0,DESCRIPTIVE_NO_MATCHED_NULL
+graph_soma_tree,4,2,50,0,0.0,0.0,DESCRIPTIVE_NO_MATCHED_NULL
+graph_soma_tree,4,3,25,0,0.0,0.0,DESCRIPTIVE_NO_MATCHED_NULL

p1_v2_fine_types/analysis/complete_parent_splits.csv

@@ -0,0 +1,37 @@
+parent_type,side,arm,k,parent_n,stable8_selected_n,residual_n,seed_ARI_mean,child_sizes,status
+KCa'b'-ap1,left,graph,2,148,132,16,1.0,"[22, 126]",DESCRIPTIVE_ONLY
+KCa'b'-ap1,left,graph,3,148,132,16,0.968223188463474,"[16, 107, 25]",DESCRIPTIVE_ONLY
+KCa'b'-ap1,left,graph,4,148,132,16,0.8707303917534227,"[94, 12, 30, 12]",DESCRIPTIVE_ONLY
+KCa'b'-ap1,left,graph_soma_tree,2,148,132,16,0.9615650002178049,"[128, 20]",DESCRIPTIVE_ONLY
+KCa'b'-ap1,left,graph_soma_tree,3,148,132,16,0.9876301854004486,"[61, 69, 18]",DESCRIPTIVE_ONLY
+KCa'b'-ap1,left,graph_soma_tree,4,148,132,16,0.9884163797635841,"[18, 55, 50, 25]",DESCRIPTIVE_ONLY
+KCa'b'-ap1,right,graph,2,133,122,11,1.0,"[119, 14]",DESCRIPTIVE_ONLY
+KCa'b'-ap1,right,graph,3,133,122,11,0.9594150796389369,"[101, 21, 11]",DESCRIPTIVE_ONLY
+KCa'b'-ap1,right,graph,4,133,122,11,0.41797314466676044,"[52, 11, 19, 51]",DESCRIPTIVE_ONLY
+KCa'b'-ap1,right,graph_soma_tree,2,133,122,11,0.3692553698353845,"[46, 87]",DESCRIPTIVE_ONLY
+KCa'b'-ap1,right,graph_soma_tree,3,133,122,11,1.0,"[34, 88, 11]",DESCRIPTIVE_ONLY
+KCa'b'-ap1,right,graph_soma_tree,4,133,122,11,0.9940813339598745,"[34, 55, 11, 33]",DESCRIPTIVE_ONLY
+KCab-p,left,graph,2,67,67,0,0.9475021369550763,"[54, 13]",DESCRIPTIVE_ONLY
+KCab-p,left,graph,3,67,67,0,0.9385605512109283,"[32, 24, 11]",DESCRIPTIVE_ONLY
+KCab-p,left,graph,4,67,67,0,0.955884484116566,"[21, 23, 10, 13]",DESCRIPTIVE_ONLY
+KCab-p,left,graph_soma_tree,2,67,67,0,0.7626155489348526,"[21, 46]",DESCRIPTIVE_ONLY
+KCab-p,left,graph_soma_tree,3,67,67,0,0.8578810221084865,"[21, 20, 26]",DESCRIPTIVE_ONLY
+KCab-p,left,graph_soma_tree,4,67,67,0,0.8215200624822263,"[16, 16, 17, 18]",DESCRIPTIVE_ONLY
+KCab-p,right,graph,2,61,61,0,1.0,"[38, 23]",DESCRIPTIVE_ONLY
+KCab-p,right,graph,3,61,61,0,0.9853779455614005,"[20, 9, 32]",DESCRIPTIVE_ONLY
+KCab-p,right,graph,4,61,61,0,0.9194497888847548,"[15, 17, 21, 8]",DESCRIPTIVE_ONLY
+KCab-p,right,graph_soma_tree,2,61,61,0,1.0,"[46, 15]",DESCRIPTIVE_ONLY
+KCab-p,right,graph_soma_tree,3,61,61,0,1.0,"[32, 15, 14]",DESCRIPTIVE_ONLY
+KCab-p,right,graph_soma_tree,4,61,61,0,0.7971718697419794,"[32, 15, 9, 5]",DESCRIPTIVE_ONLY
+KCab-s,left,graph,2,311,114,197,1.0,"[197, 114]",DESCRIPTIVE_ONLY
+KCab-s,left,graph,3,311,114,197,1.0,"[65, 107, 139]",DESCRIPTIVE_ONLY
+KCab-s,left,graph,4,311,114,197,1.0,"[94, 68, 96, 53]",DESCRIPTIVE_ONLY
+KCab-s,left,graph_soma_tree,2,311,114,197,0.9711148060401841,"[190, 121]",DESCRIPTIVE_ONLY
+KCab-s,left,graph_soma_tree,3,311,114,197,1.0,"[62, 141, 108]",DESCRIPTIVE_ONLY
+KCab-s,left,graph_soma_tree,4,311,114,197,0.98192628602563,"[107, 60, 84, 60]",DESCRIPTIVE_ONLY
+KCab-s,right,graph,2,310,147,163,1.0,"[166, 144]",DESCRIPTIVE_ONLY
+KCab-s,right,graph,3,310,147,163,0.9110285018082322,"[105, 126, 79]",DESCRIPTIVE_ONLY
+KCab-s,right,graph,4,310,147,163,1.0,"[81, 109, 32, 88]",DESCRIPTIVE_ONLY
+KCab-s,right,graph_soma_tree,2,310,147,163,1.0,"[165, 145]",DESCRIPTIVE_ONLY
+KCab-s,right,graph_soma_tree,3,310,147,163,0.7885022028960358,"[103, 121, 86]",DESCRIPTIVE_ONLY
+KCab-s,right,graph_soma_tree,4,310,147,163,0.9611244854895695,"[100, 101, 36, 73]",DESCRIPTIVE_ONLY

p1_v2_fine_types/analysis/discovery_ablation.csv

@@ -0,0 +1,7 @@
+arm,soma_weight,tree_weight,mean_cluster_soma_radius_um,seed_ARI_mean,n_clusters,n_cells,status,pair_dropout_ARI_mean,pair_dropout_ARI_values
+graph,0.0,0,38.160028098101485,0.7910987806109194,30,5177,EXPLORATORY_PARTITION_NOT_STABLE_CANDIDATES,0.6459521233688187,"[0.6208263398071404, 0.668976844363313, 0.6480531859360028]"
+graph_soma025,0.25,0,37.52554105336682,0.7467089658094822,30,5177,EXPLORATORY_PARTITION_NOT_STABLE_CANDIDATES,0.6489918988132075,"[0.6421944233864323, 0.6683592191030845, 0.6364220539501056]"
+graph_soma1,1.0,0,35.99740289884258,0.8560880348882907,30,5177,EXPLORATORY_PARTITION_NOT_STABLE_CANDIDATES,0.7033316798759836,"[0.71021080877357, 0.7392016583909505, 0.6605825724634301]"
+graph_soma4,4.0,0,26.85618125502845,0.6615240099659837,30,5177,EXPLORATORY_PARTITION_NOT_STABLE_CANDIDATES,0.6383389635863637,"[0.6231532312096515, 0.6399065723482902, 0.6519570872011492]"
+graph_tree,0.0,1,39.06687797043532,0.7208696248764215,30,5177,EXPLORATORY_PARTITION_NOT_STABLE_CANDIDATES,0.6310862442567852,"[0.6136046808745598, 0.6390119388207619, 0.6406421130750336]"
+graph_soma_tree,1.0,1,37.77500035468918,0.7667214093585388,30,5177,EXPLORATORY_PARTITION_NOT_STABLE_CANDIDATES,0.8172704565561865,"[0.7988362034706604, 0.820391705352876, 0.8325834608450231]"

p1_v2_fine_types/analysis/parent_candidate_counts.csv

@@ -0,0 +1,9 @@
+candidate,parent_fine_type,side,candidate_total,candidate_within_parent,parent_total,residual,minority_outside_parent
+CM012,KCa'b'-m,right,163,160,174,14,3
+CM014,KCa'b'-m,left,152,152,164,12,0
+CM016,KCab-s,right,147,147,310,163,0
+CM019,KCa'b'-ap1,left,132,132,148,16,0
+CM020,KCa'b'-ap1,right,129,122,133,11,7
+CM022,KCab-s,left,114,114,311,197,0
+CM031,KCab-p,left,67,67,67,0,0
+CM033,KCab-p,right,61,61,61,0,0

p1_v2_fine_types/analysis/vertex_bootstrap_exploratory.csv

@@ -0,0 +1,5 @@
+base,augmented,delta_log_loss,interval_low,interval_high,method,n_resamples,seed,limitation
+M1,M2_descriptor,0.016174199654786097,0.010736531871536276,0.021509855136446408,POSTHOC_FIXED_MODEL_VERTEX_POISSON_BOOTSTRAP,1000,20261104,"within-brain endpoint dependence only; no refit, spatial block, tuning or animal uncertainty"
+M2_descriptor,M3_descriptor_discrete,0.006649692199399904,0.002663759778787991,0.010732576210627868,POSTHOC_FIXED_MODEL_VERTEX_POISSON_BOOTSTRAP,1000,20261104,"within-brain endpoint dependence only; no refit, spatial block, tuning or animal uncertainty"
+M2_descriptor,M4_descriptor_continuous,0.02217735322234832,0.016656227046957003,0.027979004103687897,POSTHOC_FIXED_MODEL_VERTEX_POISSON_BOOTSTRAP,1000,20261104,"within-brain endpoint dependence only; no refit, spatial block, tuning or animal uncertainty"
+M3_descriptor_discrete,M4_descriptor_continuous,0.015527661022948416,0.010114642731116131,0.020770471504391454,POSTHOC_FIXED_MODEL_VERTEX_POISSON_BOOTSTRAP,1000,20261104,"within-brain endpoint dependence only; no refit, spatial block, tuning or animal uncertainty"

p1_v2_fine_types/audit/competition_literature_lookup.json

@@ -0,0 +1,12 @@
+[
+  {
+    "url": "https://api.crossref.org/works/10.64898/2026.09.06.749654",
+    "status": "UNVERIFIED",
+    "error": "ProxyError"
+  },
+  {
+    "url": "https://www.biorxiv.org/content/10.64898/2026.09.06.749654v1",
+    "status": "UNVERIFIED",
+    "error": "ProxyError"
+  }
+]

p1_v2_fine_types/audit/environment.json

@@ -0,0 +1,15 @@
+{
+  "python": "3.11.11",
+  "packages": {
+    "numpy": "1.26.4",
+    "scipy": "1.16.3",
+    "scikit-learn": "1.8.0",
+    "pandas": "2.3.3",
+    "pyarrow": "22.0.0",
+    "torch": "2.8.0+cu129",
+    "matplotlib": "3.10.8",
+    "threadpoolctl": "3.6.0"
+  },
+  "threads": 4,
+  "note": "Figure caption corrected from obsolete P1-v2.1 to reference actual protocol.json. Previous PDFs retained in audit/figures_before_caption_correction."
+}

p1_v2_fine_types/audit/finalizer_snapshot.py

@@ -0,0 +1,125 @@
+"""Derive P1 reports and data-quality checks from actual run artifacts."""
+from pathlib import Path
+import json,hashlib
+import numpy as np
+import pandas as pd
+
+ROOT=Path(__file__).resolve().parents[2]
+OUT=ROOT/'artifacts/nature_kc_20261030/p1_v2_fine_types'
+
+def main():
+    reg=pd.read_parquet(OUT/'registry/all5177_identity.parquet')
+    pairs=pd.read_parquet(OUT/'data/directed_pair_split.parquet')
+    models=pd.read_csv(OUT/'MODEL_COMPARISON.csv').set_index('model')
+    ab=pd.read_csv(OUT/'analysis/discovery_ablation.csv').set_index('arm')
+    parent=pd.read_csv(OUT/'analysis/complete_parent_splits.csv')
+    child=pd.read_parquet(OUT/'registry/parent_child_members.parquet')
+    checks=[]
+    def check(name,ok,actual):checks.append({'check':name,'passed':bool(ok),'actual':str(actual)})
+    check('all_KC_unique_strings',len(reg)==5177 and reg.root_id.nunique()==5177 and reg.root_id.map(lambda x:isinstance(x,str)).all(),len(reg))
+    check('stable8_union',reg.stable_backbone.sum()==965,reg.stable_backbone.sum())
+    for side,n in [('left',311),('right',310)]:
+        got=parent.query('parent_type == "KCab-s" and side == @side').parent_n.unique()
+        check('full_KCab-s_'+side,len(got)==1 and got[0]==n,got)
+    check('synapse_conservation',pairs.synapses.sum()==379338,pairs.synapses.sum())
+    sets={s:set(zip(d.pre_id,d.post_id)) for s,d in pairs.groupby('split')}
+    for a,b in [('train','validation'),('train','test'),('validation','test')]:
+        overlap=sets[a]&sets[b];reverse=sets[a]&{(v,u) for u,v in sets[b]}
+        check(a+'_'+b+'_direct_and_reverse_isolation',not overlap and not reverse,[len(overlap),len(reverse)])
+    check('node_JSON_string_roundtrip',json.loads(json.dumps(reg.root_id.tolist()))==reg.root_id.tolist(),len(reg))
+    su=set(reg.loc[reg.unit_id.eq('SU019'),'root_id']);cm=set(reg.loc[reg.candidate_id.eq('CM019')&reg.stable_backbone,'root_id'])
+    check('SU019_CM019_not_aliased',len(su&cm)==104 and len(cm)==132,[len(su),len(cm),len(su&cm)])
+    pd.DataFrame(checks).to_csv(OUT/'audit/structural_checks.csv',index=False)
+    assert all(c['passed'] for c in checks)
+    # Complete parent membership, including minority candidate members outside dominant parent.
+    summaries=[];members=[]
+    for cmid in ['CM012','CM014','CM016','CM019','CM020','CM022','CM031','CM033']:
+        cand=reg.candidate_id.eq(cmid)&reg.stable_backbone.astype(bool)
+        typ,side=reg.loc[cand].groupby(['frozen_detailed_subtype','official_hemisphere']).size().idxmax()
+        par=reg.frozen_detailed_subtype.eq(typ)&reg.official_hemisphere.eq(side)
+        summaries.append({'candidate':cmid,'parent_fine_type':typ,'side':side,'candidate_total':int(cand.sum()),'candidate_within_parent':int((cand&par).sum()),'parent_total':int(par.sum()),'residual':int((par&~cand).sum()),'minority_outside_parent':int((cand&~par).sum())})
+        for _,r in reg.loc[par|cand].iterrows():
+            i=r.name;members.append({'root_id':r.root_id,'candidate':cmid,'parent_type':typ,'side':side,'member_state':'selected_parent' if cand[i] and par[i] else 'minority_outside_parent' if cand[i] else 'parent_residual'})
+    pd.DataFrame(summaries).to_csv(OUT/'analysis/parent_candidate_counts.csv',index=False)
+    pd.DataFrame(members).to_parquet(OUT/'registry/parent_candidate_residual_atlas.parquet',index=False)
+    # Describe where the 16 legacy CM019 residual cells land, without null-based claims.
+    residual=set(reg.loc[reg.frozen_detailed_subtype.eq("KCa'b'-ap1")&reg.official_hemisphere.eq('left')&~reg.root_id.isin(cm),'root_id'])
+    rr=[]
+    for (arm,k),g in child.loc[child.parent_type.eq("KCa'b'-ap1")&child.side.eq('left')].groupby(['arm','k']):
+        for c,cg in g.groupby('child'):
+            nres=int(cg.root_id.isin(residual).sum())
+            rr.append({'arm':arm,'k':k,'child':c,'child_n':len(cg),'CM019_residual_n':nres,'residual_recall':nres/len(residual),'residual_fraction_of_child':nres/len(cg),'evidence':'DESCRIPTIVE_NO_MATCHED_NULL'})
+    pd.DataFrame(rr).to_csv(OUT/'analysis/CM019_residual_child_overlap.csv',index=False)
+    sub=child.loc[child.parent_type.eq("KCa'b'-ap1")&child.side.eq('left')&child.arm.eq('graph_soma_tree')]
+    groups={k:[set(g.root_id) for _,g in sub.loc[sub.k.eq(k)].groupby('child') if residual.issubset(set(g.root_id))] for k in [3,4]}
+    assert len(groups[3])==len(groups[4])==1 and groups[3][0]==groups[4][0]
+    small=reg.loc[reg.root_id.isin(groups[3][0])].copy()
+    small['P1_candidate_id']='P1_ap1L_18_EXPLORATORY';small['status']='NOT_FROZEN_STABLE'
+    small.to_parquet(OUT/'registry/ap1L_18_exploratory_members.parquet',index=False)
+    # Fixed-model vertex-weighted bootstrap: accounts for shared endpoints, not animals.
+    pred=pd.read_parquet(OUT/'data/test_predictions.parquet'); ids=sorted(set(pred.pre_id)|set(pred.post_id)); idx={r:i for i,r in enumerate(ids)}
+    a=pred.pre_id.map(idx).to_numpy();b=pred.post_id.map(idx).to_numpy();y=pred.edge_present.to_numpy()
+    losses={name:-(y*np.log(np.clip(pred[name],1e-12,1-1e-12))+(1-y)*np.log(np.clip(1-pred[name],1e-12,1-1e-12))) for name in models.index}
+    contrasts=[('M1','M2_descriptor'),('M2_descriptor','M3_descriptor_discrete'),('M2_descriptor','M4_descriptor_continuous'),('M3_descriptor_discrete','M4_descriptor_continuous')]
+    diffs=np.column_stack([losses[x]-losses[z] for x,z in contrasts]);rng=np.random.default_rng(20261104);boot=[]
+    for _ in range(1000):
+        w=rng.poisson(1,len(ids));dw=w[a]*w[b];boot.append(dw@diffs/dw.sum())
+    ci=np.quantile(boot,[.025,.975],axis=0)
+    pd.DataFrame([{'base':x,'augmented':z,'delta_log_loss':float(diffs[:,j].mean()),'interval_low':ci[0,j],'interval_high':ci[1,j],'method':'POSTHOC_FIXED_MODEL_VERTEX_POISSON_BOOTSTRAP','n_resamples':1000,'seed':20261104,'limitation':'within-brain endpoint dependence only; no refit, spatial block, tuning or animal uncertainty'} for j,(x,z) in enumerate(contrasts)]).to_csv(OUT/'analysis/vertex_bootstrap_exploratory.csv',index=False)
+    report=f'''# P1-v2.2：身份、双向隔离与真实空间融合试验
+
+本轮已完成 P1A–P1C 的可执行试验，以及 P1D 的描述量级模型比较。原 stable8 仍为 8 个、965 个成员；新增稳定功能单元数为 0。全部结果来自实际运行，未用 fixture 替代结构数据。
+
+## 当前结论
+
+1. 胞体与树几何已真正进入发现目标。全体 5,177 KC 的 5,177×5,177 欧氏距离矩阵按 μm 生成；树特征从真实骨架计算，未读取接点标签。六个固定消融方案、每方案五个初始化及三次 20% 无序对删边测试均已运行。
+2. 空间更紧凑不等于更稳定。纯图的平均簇胞体半径为 {ab.loc['graph','mean_cluster_soma_radius_um']:.2f} μm；大胞体权重降至 {ab.loc['graph_soma4','mean_cluster_soma_radius_um']:.2f} μm，但删边 ARI 为 {ab.loc['graph_soma4','pair_dropout_ARI_mean']:.3f}，未超过纯图的 {ab.loc['graph','pair_dropout_ARI_mean']:.3f}。图+胞体+树的删边 ARI 为 {ab.loc['graph_soma_tree','pair_dropout_ARI_mean']:.3f}，而平均半径仍为 {ab.loc['graph_soma_tree','mean_cluster_soma_radius_um']:.2f} μm。这里的 ARI 是整体分区重现性，不是旧 frozen S/coverage，也不代表个别新簇通过稳定门槛。
+3. 几何描述之后，离散标签还有预测增量，但连续训练图表征更强。测试 log loss：M0（已知粗/细类型+侧别）{models.loc['M0','test_log_loss']:.5f}；M1（另加胞体/距离/训练图强度）{models.loc['M1','test_log_loss']:.5f}；M2_descriptor（另加树描述量）{models.loc['M2_descriptor','test_log_loss']:.5f}；M3（另加离散标签）{models.loc['M3_descriptor_discrete','test_log_loss']:.5f}；M4（另加连续图嵌入）{models.loc['M4_descriptor_continuous','test_log_loss']:.5f}。这不等价于“连续几何胜出”：M4 使用的是连续图表征；M2 还没有精细区域树接近性和分枝接点，M4 特征维数也更高。
+
+4. CM019 residual 有具体追踪线索：融合胞体与树几何后，K=3 与 K=4 得到完全相同的18细胞子簇，含全部16个原 residual 与2个原CM019成员；K=2时它们进入20细胞子簇。成员见 registry/ap1L_18_exploratory_members.parquet。这是融合特征空间中的分离，不能直接推断其胞体构成单一连续空间团；尚需扰动、空间匹配零模型、输入输出留出与跨脑检验。
+
+## 数据与协议
+
+FlyWire v783，293,762 条有向神经元对、379,338 个聚合突触计数。无序 pair 哈希决定训练/验证/测试：205,483 / 44,335 / 43,944 条有向边；三个集合之间同向和反向直接交集均为 0。所有图嵌入与度/强度仅来自训练图。接点派生特征在本轮完全排除；这不是已经完成接点级训练测试映射。
+
+预测样本为 60,000 正例和 60,000 非边；训练/验证/测试样本 83,858 / 18,168 / 17,974。各模型同为两个超参设置×100轮，验证集选择，测试集只评估最终设置。这是 case-control 采样，预测值不是自然患率下的连接概率，更不是突触传递概率。原全图与历史测试数据曾参与开发，本次所有切分均标记回顾性，不能称为盲测。
+
+旧 CM、SU、新 P1 分区独立命名。SU019 与 CM019 为 104/132 重叠。P1 全局 K=30 仅为固定分辨率消融，并非新增 30 个稳定单元。局部试验覆盖完整细类型 parent：左右 KCab-s 311/310、ap1 148/133、KCab-p 67/61；K=2/3/4 是探索尺度，全部子簇成员与残余可追溯。CM019 的 16 个 residual 在各子簇中的去向见 CM019_residual_child_overlap.csv；没有匹配零模型和独立验证前，不称为一致补集。
+
+## 修正与失败项
+
+首轮 p1_v2 使用粗粒度 official_KC_subtype，漏掉 KCab-s parent。该试跑原样保留；当前 p1_v2_fine_types 使用 frozen_detailed_subtype 定义完整 parent，M0 同时保留粗/细标签。右 ap1 的粗标签计数132、细标签133不同，未静默覆盖。
+
+MaleCNS 的完整原始 id 列有211,577个唯一ID，body列187,016个；已存在 body 的行两者一致。但换用完整 id 后仍有31,355个边端点未匹配，不能靠填充body解决。跨版本/来源/范围需继续查原始连接表生产链。本轮不宣称跨脑复现。
+
+本地旧跨连接组比较的 provenance 指向名含 invalid_seed_schedule 的历史benchmark，不能升级为独立复现。BANC 与 FlyWire 的样本来源关系亦须核实，数据集名称不同不能自动算不同动物。
+
+用户提供 Pleijzier 2026 预印本已登记为竞争基线，当前代理访问 Crossref/bioRxiv 失败，题目、日期与具体主张仍待原文核查；不把转述当已核实文献结论。
+
+## 剩余裁决工作
+
+P1D 仍需区域树/分枝接点的训练隔离、模型容量与调参预算敏感性、空间/细胞阻断、类型内困难负例、匹配零模型、PN→KC 与 KC→MBON 的预留终点。附加的1000次顶点权重bootstrap仅为固定模型、同脑内的探索性区间，未覆盖模型重拟合、空间相关和动物变异。
+
+P1E 须先解决 MaleCNS 来源映射、BANC 生物独立性与其他数据版本，再比较可迁移组织规律及实验可定位性。目前不应指定“已验证的2–3个功能微回路”。不继续优化 NeuronMimic 架构。
+
+## 复现与产物
+
+仓库命令：`PYTHONPATH=simulation/src python simulation/scripts/run_kc_p1_v2.py`（已完成目录拒绝覆盖）。公开紧凑数据复现命令见 GitHub 的 docs/P1_REPRODUCE.md。发布包含双向切分、身份与父子登记、树描述量、冻结协议、比较表、图及原始输入哈希；完整骨架须向原数据源获取。
+
+关键文件：MODEL_COMPARISON.csv、P1_STATUS.json、analysis/discovery_ablation.csv、analysis/complete_parent_splits.csv、registry/parent_child_members.parquet、audit/structural_checks.csv、audit/source_manifest.json。
+'''
+    (OUT/'P1_REPORT.md').write_text(report)
+    (OUT/'EXECUTION_LOG.md').write_text('''# Execution log
+
+- P1-v2.1: original coarse-type pilot completed; preserved in ../p1_v2. Identified omitted KCab-s fine parent and weaker M0.
+- P1-v2.2: protocol amended before rerun; P1A/B/C and descriptor-level D pilot completed. See audit/run.log and source_manifest.json for actual timings, seed and source hashes.
+- Command: PYTHONPATH=simulation/src python simulation/scripts/run_kc_p1_v2.py
+- Validation: PYTHONPATH=simulation/src pytest -q simulation/tests/test_p1_structure.py
+- Posthoc report: python simulation/scripts/finalize_kc_p1_v2.py
+- MaleCNS mapping / independent-connectome confirmation: BLOCKED. Exact missing endpoints are in registry/malecns_unmatched_endpoints.csv.
+- Literature primary-source retrieval: BLOCKED_NETWORK for Crossref/bioRxiv; HF proxy transport is working.
+- NeuronMimic architecture optimization: not part of this structural experiment.
+''')
+    print('structural checks',len(checks),'passed; report written')
+
+if __name__=='__main__':main()

p1_v2_fine_types/audit/malecns_id_QC.json

@@ -0,0 +1,12 @@
+{
+  "annotation_rows": 211577,
+  "body_valid_unique": 187016,
+  "id_valid_unique": 211577,
+  "edge_endpoints": 41105,
+  "body_unmatched": 31355,
+  "id_unmatched": 31355,
+  "recovered_by_id": 0,
+  "body_vs_id_disagreement": 0,
+  "status": "BLOCKED_SOURCE_VERSION_OR_SCOPE_UNRESOLVED",
+  "note": "Full source id column does not resolve missing edge endpoints; no cross-version identity inference."
+}

p1_v2_fine_types/audit/pair_split_QC.json

@@ -0,0 +1,11 @@
+{
+  "n_directed_pairs": 293762,
+  "synapses_total": 379338.0,
+  "train_pairs": 205483,
+  "validation_pairs": 44335,
+  "test_pairs": 43944,
+  "heldout_direct_overlap": 0,
+  "heldout_reverse_overlap": 0,
+  "contact_features_used": false,
+  "exposure": "RETROSPECTIVE_NOT_BLIND"
+}

p1_v2_fine_types/audit/primitives_snapshot.py

@@ -0,0 +1,58 @@
+"""P1-v2 primitives: pair isolation and contact-free geometry, no label promotion."""
+from __future__ import annotations
+import hashlib
+import numpy as np
+
+
+def pair_uniform(ids, pre, post, seed=20261030):
+    """Order-invariant identity hash; every direction/contact of a pair shares a draw."""
+    tokens = np.array([int.from_bytes(hashlib.blake2b(str(i).encode(), digest_size=8).digest(), 'little') for i in ids], dtype=np.uint64)
+    if len(np.unique(tokens)) != len(tokens):
+        raise ValueError('duplicate ID or hash collision')
+    a, b = tokens[np.asarray(pre)], tokens[np.asarray(post)]
+    lo, hi = np.minimum(a,b), np.maximum(a,b)
+    with np.errstate(over='ignore'):
+        x = lo ^ (hi * np.uint64(0x9E3779B97F4A7C15)) ^ np.uint64(seed)
+        x = (x ^ (x >> np.uint64(30))) * np.uint64(0xBF58476D1CE4E5B9)
+        x = (x ^ (x >> np.uint64(27))) * np.uint64(0x94D049BB133111EB)
+        x ^= x >> np.uint64(31)
+    return (x >> np.uint64(11)).astype(np.float64) / 2**53
+
+
+def pair_split(ids, pre, post, seed=20261030):
+    u=pair_uniform(ids,pre,post,seed)
+    return np.where(u<.70,0,np.where(u<.85,1,2)).astype(np.int8)
+
+
+def tree_features(d):
+    """Only vertices, tree edges, branch IDs and skeleton paths; ignore contacts."""
+    v=np.asarray(d['vertices'],float); e=np.asarray(d['edges'],int)
+    if not np.isfinite(v).all() or v.shape[1]!=3:
+        raise ValueError('invalid geometry')
+    length=np.linalg.norm(v[e[:,0]]-v[e[:,1]],axis=1)
+    if length.sum()<=0: raise ValueError('zero cable length')
+    mid=(v[e[:,0]]+v[e[:,1]])/2
+    center=np.average(mid,axis=0,weights=length)
+    delta=mid-center
+    cov=(delta*length[:,None]).T@delta/length.sum()
+    eigen=np.maximum(np.linalg.eigvalsh(cov),0)
+    degree=np.bincount(e.ravel(),minlength=len(v))
+    path=np.asarray(d['path_from_soma_nearest_um'],float)
+    finite=path[np.isfinite(path)]
+    out={'cable_length_um':length.sum(),'n_branch_ids':len(set(d['branch_per_edge'])),
+         'n_terminals':int((degree==1).sum()),'n_branchpoints':int((degree>2).sum()),
+         'edge_length_median_um':np.median(length),'edge_length_p95_um':np.quantile(length,.95),
+         'path_median_um':np.median(finite),'path_p95_um':np.quantile(finite,.95)}
+    for j,axis in enumerate('xyz'):
+        out[f'centroid_{axis}_um']=center[j]
+        out[f'extent_{axis}_um']=np.ptp(v[:,j])
+        out[f'principal_sd_{j}_um']=np.sqrt(eigen[j])
+    return out
+
+
+def fuse_blocks(graph, soma, tree, soma_weight, tree_weight):
+    """Squared distance = graph distance + weighted soma/tree distances."""
+    blocks=[graph/np.sqrt(graph.shape[1])]
+    if soma_weight: blocks.append(np.sqrt(soma_weight)*soma)
+    if tree_weight: blocks.append(np.sqrt(tree_weight)*tree/np.sqrt(tree.shape[1]))
+    return np.column_stack(blocks)

p1_v2_fine_types/audit/protocol.json

@@ -0,0 +1,69 @@
+{
+  "protocol_version": "P1-v2.2-retrospective-fine-type-geometry-pilot",
+  "frozen_at_utc": "2026-10-30T08:29:23.509181+00:00",
+  "code_commit": "f15ab12e1988000f71ec5e0ab138ab116f150e7d",
+  "seed": 20261030,
+  "scope": "all 5177 FlyWire v783 KC; complete official ap1/KCab-s/KCab-p parents",
+  "split": "deterministic unordered neuron pair 70/15/15 train/validation/test; same rule for all nonedges; both directions share split",
+  "exposure": "RETROSPECTIVE: full graph and historical test outcomes already used in project development; not a blind test",
+  "features": "train-only directed spectral32; actual soma Euclidean distances (um); contact-free skeleton geometry; no held-out synapse labels or contact statistics",
+  "contact_features": "excluded entirely in this pilot; no claims of completed branch-contact M2",
+  "discovery_arms": {
+    "graph": [
+      0,
+      0
+    ],
+    "graph_soma025": [
+      0.25,
+      0
+    ],
+    "graph_soma1": [
+      1,
+      0
+    ],
+    "graph_soma4": [
+      4,
+      0
+    ],
+    "graph_tree": [
+      0,
+      1
+    ],
+    "graph_soma_tree": [
+      1,
+      1
+    ]
+  },
+  "global_k": 30,
+  "parent_k": [
+    2,
+    3,
+    4
+  ],
+  "init_seeds": [
+    20261030,
+    20261031,
+    20261032,
+    20261033,
+    20261034
+  ],
+  "pair_dropout_seeds": [
+    20261101,
+    20261102,
+    20261103
+  ],
+  "pair_dropout_fraction": 0.2,
+  "n_init": 5,
+  "no_candidate_promotion": true,
+  "model_comparison": "exploratory M0-M4 descriptor pilot only; equal two-setting validation budget, max_iter100, leaf15/31, l2=1, learning_rate=.08; final fit train-only",
+  "sample_budget": "at most60000 observed directed positives and60000 nonedges, fixed RNG; no test-driven choice",
+  "uncertainty": "edge-IID significance prohibited; confirmatory cell/spatial blocks and cross-brain not complete",
+  "outputs_label": "DESCRIPTIVE_ONLY, COMPUTED_MODEL_RESULT (retrospective structural prediction); no biological functional validation",
+  "amendment": "Previous pilot used coarser official_KC_subtype and omitted KCab-s parents. Preserved p1_v2; this revision uses both coarse and frozen fine known-type categories in M0 and frozen_detailed_subtype for parent selection. No tuning changes.",
+  "parent_scope": [
+    "KCab-s",
+    "KCab-p",
+    "KCa'b'-ap1"
+  ],
+  "fine_label_source": "frozen_detailed_subtype in audited full registry; retains official_KC_subtype separately"
+}

p1_v2_fine_types/audit/release_source_manifest.json

@@ -0,0 +1,5 @@
+{
+  "postrun_runner_sha256": "486f03df01c93923a74775fd69edfb5fb19dd391809f0bba32502b6f98fc2704",
+  "finalizer_sha256": "acf21a9a721de1104da75075290cd16472e72bcb9e778e53d555c6d0c4c0e270",
+  "scope": "Added compact-bundle replay and corrected figure protocol caption after the raw-source run; original runner preserved in runner_snapshot.py."
+}

p1_v2_fine_types/audit/run.log

@@ -0,0 +1,14 @@
+P1A: 5177 identities / frozen stable8 verified
+P1B: unordered pair split / training-only embedding complete
+geometry 1000
+geometry 2000
+geometry 3000
+geometry 4000
+geometry 5000
+P1C: six geometry ablations / full-parent child registries complete
+M0 0.2606936352617916
+M1 0.25956746261839025
+M2_descriptor 0.24339326296360414
+M3_descriptor_discrete 0.23674357076420421
+M4_descriptor_continuous 0.2212159097412558
+{"P1A": "DONE_WITH_MALECNS_BLOCKER", "P1B": "DONE_RETROSPECTIVE_PAIR_ISOLATION", "P1C": "DONE_EXPLORATORY_GEOMETRY_DISCOVERY", "P1D": "PARTIAL_DESCRIPTOR_PILOT_NOT_FULL_BRANCH_CONTACT_MODEL", "P1E": "BLOCKED_INDEPENDENCE_AND_ID_QC", "stable8_unchanged": true, "new_stable_candidates": 0, "scientific_conclusion": "No discrete/continuous verdict yet; full independent competition remains open", "wall_clock_s": 55.4731810092926}

p1_v2_fine_types/audit/runner_snapshot.py

@@ -0,0 +1,197 @@
+"""Run the frozen retrospective P1A-C and bounded M0-M4 descriptor pilot.
+
+Run from repository root: PYTHONPATH=simulation/src python simulation/scripts/run_kc_p1_v2.py
+"""
+from __future__ import annotations
+import hashlib,json,time,subprocess
+from pathlib import Path
+import numpy as np
+import pandas as pd
+from scipy import sparse
+from scipy.spatial.distance import cdist
+from sklearn.preprocessing import StandardScaler
+from sklearn.cluster import KMeans
+from sklearn.metrics import adjusted_rand_score,roc_auc_score,log_loss
+from sklearn.ensemble import HistGradientBoostingClassifier
+from threadpoolctl import threadpool_limits
+from bio_fly.p1_structure import pair_split,pair_uniform,tree_features,fuse_blocks
+from bio_fly.kc_weighted_structure import _embedding
+
+ROOT=Path(__file__).resolve().parents[2]
+OUT=ROOT/'artifacts/nature_kc_20261030/p1_v2_fine_types'
+ATLAS=ROOT/'artifacts/wu_20261030_all5177_structural_atlas'
+GRAPH=ROOT/'nature_methods_connectome/next_stage_structure_first_20260912/artifacts/flywire_all_kc_partner_pilot_20260912'
+
+def sha(p):
+    h=hashlib.sha256()
+    with p.open('rb') as f:
+        for b in iter(lambda:f.read(8*1024*1024),b''):h.update(b)
+    return h.hexdigest()
+
+def dump(p,d):p.write_text(json.dumps(d,indent=2,ensure_ascii=False,default=lambda x:x.item() if hasattr(x,'item') else str(x)))
+
+def main():
+    start=time.time()
+    if (OUT/'P1_STATUS.json').exists():raise RuntimeError('completed run exists; use a separately versioned protocol/output')
+    for n in ['registry','data','analysis','figures','audit']:(OUT/n).mkdir(exist_ok=True,parents=True)
+    cfg=json.loads((OUT/'audit/protocol.json').read_text()); seed=cfg['seed']; sources=[]
+    def source(p):
+        p=Path(p);sources.append({'path':str(p.relative_to(ROOT)),'bytes':p.stat().st_size,'sha256':sha(p)});return p
+    ids=np.array(json.loads(source(GRAPH/'node_order.json').read_text()),dtype=str)
+    reg=pd.read_parquet(source(ROOT/'artifacts/wu_20261019/registry/kc_all.parquet'))
+    reg.root_id=reg.root_id.astype(str);reg=reg.set_index('root_id').loc[ids].reset_index()
+    su=pd.read_parquet(source(ROOT/'artifacts/wu_20261030_fine_units_integrated_atlas/registry/fine_unit_members.parquet'))
+    su.root_id=su.root_id.astype(str)
+    reg=reg.merge(su[['root_id','unit_id']],on='root_id',how='left',validate='one_to_one')
+    assert len(reg)==len(set(ids))==5177
+    expected={'CM012':163,'CM014':152,'CM016':147,'CM019':132,'CM020':129,'CM022':114,'CM031':67,'CM033':61}
+    counts=reg.loc[reg.stable_backbone.astype(bool),'candidate_id'].value_counts().to_dict()
+    assert counts==expected
+    reg['legacy_CM_namespace']='v9_frozen';reg['SU_namespace']='20261030_scale_conditional';reg['P1_role']='full_registry_no_promotion'
+    reg.to_parquet(OUT/'registry/all5177_identity.parquet',index=False)
+    cross=reg.groupby(['candidate_id','unit_id'],dropna=False).size().reset_index(name='n_shared')
+    cross.to_csv(OUT/'registry/CM_SU_crosswalk.csv',index=False)
+    print('P1A: 5177 identities / frozen stable8 verified',flush=True)
+    # Audit raw annotation alternatives, never silently repair historical normalized tables.
+    a=pd.read_parquet(source(ROOT/'audit/meeting_revision_20261027/reproduced_structure/MaleCNS_v1_0_source_annotations.parquet'))
+    me=pd.read_parquet(source(ROOT/'audit/meeting_revision_20261027/reproduced_structure/MaleCNS_v1_0_annotated_input_edges.parquet'))
+    ends=set(me.pre.astype(str))|set(me.post.astype(str)); body=set(a.body.dropna().astype('int64').astype(str)); aid=set(a.id.astype(str))
+    qc={'annotation_rows':len(a),'body_valid_unique':len(body),'id_valid_unique':len(aid),'edge_endpoints':len(ends),
+        'body_unmatched':len(ends-body),'id_unmatched':len(ends-aid),'recovered_by_id':len((ends-body)&aid),
+        'body_vs_id_disagreement':int((a.loc[a.body.notna(),'id'].astype(str)!=a.loc[a.body.notna(),'body'].astype('int64').astype(str)).sum()),
+        'status':'BLOCKED_SOURCE_VERSION_OR_SCOPE_UNRESOLVED','note':'Full source id column does not resolve missing edge endpoints; no cross-version identity inference.'}
+    dump(OUT/'audit/malecns_id_QC.json',qc)
+    pd.DataFrame({'unmatched_endpoint_id':sorted(ends-aid)}).to_csv(OUT/'registry/malecns_unmatched_endpoints.csv',index=False)
+    # Split all observed pairs BEFORE embeddings, features or clustering.
+    full=sparse.load_npz(source(GRAPH/'data/kc_graph.npz')).astype(float).tocsr();co=full.tocoo()
+    split=pair_split(ids,co.row,co.col,seed)
+    train=sparse.csr_matrix((co.data[split==0],(co.row[split==0],co.col[split==0])),shape=full.shape)
+    sparse.save_npz(OUT/'data/pair_grouped_train.npz',train)
+    pairs=pd.DataFrame({'pre_id':ids[co.row],'post_id':ids[co.col],'synapses':co.data.astype(int),'split':np.array(['train','validation','test'])[split]})
+    pairs.to_parquet(OUT/'data/directed_pair_split.parquet',index=False)
+    tr=set(zip(co.row[split==0],co.col[split==0]));he=set(zip(co.row[split!=0],co.col[split!=0]))
+    assert not tr&he and not tr&{(j,i) for i,j in he}
+    splitqc={'n_directed_pairs':full.nnz,'synapses_total':float(full.sum()),'train_pairs':int((split==0).sum()),'validation_pairs':int((split==1).sum()),'test_pairs':int((split==2).sum()),'heldout_direct_overlap':0,'heldout_reverse_overlap':0,'contact_features_used':False,'exposure':'RETROSPECTIVE_NOT_BLIND'}
+    dump(OUT/'audit/pair_split_QC.json',splitqc)
+    emb=_embedding(train,32,seed);np.save(OUT/'data/train_only_embedding.npy',emb)
+    print('P1B: unordered pair split / training-only embedding complete',flush=True)
+    # Tree-only geometry, with a hash for every original morphology.
+    features=[]
+    for i,rid in enumerate(ids):
+        p=source(ATLAS/'geometry'/f'{rid}.json');d=json.loads(p.read_text())
+        assert str(d['id'])==rid and d['dataset']=='FlyWire v783'
+        features.append({'root_id':rid,**tree_features(d)})
+        if (i+1)%1000==0:print('geometry',i+1,flush=True)
+    geo=pd.DataFrame(features);geo.to_parquet(OUT/'data/contact_free_tree_features.parquet',index=False)
+    soma=reg[['soma_x_um','soma_y_um','soma_z_um']].to_numpy(float)
+    assert np.isfinite(soma).all(),'missing soma prevents full distance matrix'
+    matrix=np.lib.format.open_memmap(OUT/'data/soma_distance_um.npy',mode='w+',dtype='float32',shape=(len(ids),len(ids)))
+    for i in range(0,len(ids),256):matrix[i:i+256]=cdist(soma[i:i+256],soma)
+    matrix.flush()
+    soma_scale=float(np.sqrt(np.mean(np.sum((soma-soma.mean(0))**2,axis=1))))
+    S=(soma-soma.mean(0))/soma_scale
+    T=StandardScaler().fit_transform(geo.drop(columns='root_id'))
+    dump(OUT/'data/feature_schema.json',{'soma_unit':'um','soma_scale_um':soma_scale,'tree_columns':geo.columns[1:].tolist(),'tree_source':'skeleton geometry only; no contacts, partner labels, old CM/SU labels','fusion':'concatenate E/sqrt(dE), sqrt(ws)*(soma-mean)/scale, sqrt(wt)*z(tree)/sqrt(dT)','embedding_source':'pair_grouped_train.npz','normalization':'transductive unlabeled full registry; not inductive new-cell evaluation'})
+    arms=cfg['discovery_arms']; assignments=reg[['root_id','candidate_id','unit_id','official_KC_subtype','official_hemisphere']].copy()
+    rows=[]; labelsets={}
+    def fit(x,k,rs):return KMeans(n_clusters=k,n_init=cfg['n_init'],random_state=rs).fit_predict(x)
+    for name,(ws,wt) in arms.items():
+        x=fuse_blocks(emb,S,T,ws,wt);labs=[fit(x,30,rs) for rs in cfg['init_seeds']]
+        labelsets[name]=labs[0];assignments[name]=labs[0]
+        compact=np.mean([np.linalg.norm(soma[labs[0]==c]-soma[labs[0]==c].mean(0),axis=1).mean() for c in np.unique(labs[0])])
+        rows.append({'arm':name,'soma_weight':ws,'tree_weight':wt,'mean_cluster_soma_radius_um':compact,'seed_ARI_mean':np.mean([adjusted_rand_score(labs[0],l) for l in labs[1:]]),'n_clusters':30,'n_cells':len(ids),'status':'EXPLORATORY_PARTITION_NOT_STABLE_CANDIDATES'})
+    dropout={name:[] for name in arms}
+    tc=train.tocoo()
+    for rs in cfg['pair_dropout_seeds']:
+        keep=pair_uniform(ids,tc.row,tc.col,rs)>=.2
+        pert=sparse.csr_matrix((tc.data[keep],(tc.row[keep],tc.col[keep])),shape=train.shape)
+        pe=_embedding(pert,32,seed)
+        for name,(ws,wt) in arms.items():
+            lab=fit(fuse_blocks(pe,S,T,ws,wt),30,seed)
+            dropout[name].append(adjusted_rand_score(labelsets[name],lab))
+    for row in rows:row['pair_dropout_ARI_mean']=float(np.mean(dropout[row['arm']]));row['pair_dropout_ARI_values']=json.dumps(dropout[row['arm']])
+    pd.DataFrame(rows).to_csv(OUT/'analysis/discovery_ablation.csv',index=False)
+    assignments.to_parquet(OUT/'registry/P1_exploratory_assignments.parquet',index=False)
+    parent_rows=[];parent_members=[]
+    # Complete official parents, including all residuals, never candidate-only reclustering.
+    for (typ,side),rr in reg.groupby(['frozen_detailed_subtype','official_hemisphere']):
+        if typ not in cfg['parent_scope']:continue
+        idx=rr.index.to_numpy()
+        for name in ['graph','graph_soma_tree']:
+            ws,wt=arms[name];x=fuse_blocks(emb,S,T,ws,wt)[idx]
+            for k in cfg['parent_k']:
+                labs=[fit(x,k,rs) for rs in cfg['init_seeds']]
+                parent_rows.append({'parent_type':typ,'side':side,'arm':name,'k':k,'parent_n':len(idx),'stable8_selected_n':int(rr.stable_backbone.sum()),'residual_n':int((~rr.stable_backbone.astype(bool)).sum()),'seed_ARI_mean':np.mean([adjusted_rand_score(labs[0],l) for l in labs[1:]]),'child_sizes':json.dumps(np.bincount(labs[0]).tolist()),'status':'DESCRIPTIVE_ONLY'})
+                parent_members.extend({'root_id':ids[j],'parent_type':typ,'side':side,'arm':name,'k':k,'child':int(labs[0][ii])} for ii,j in enumerate(idx))
+    pd.DataFrame(parent_rows).to_csv(OUT/'analysis/complete_parent_splits.csv',index=False)
+    pd.DataFrame(parent_members).to_parquet(OUT/'registry/parent_child_members.parquet',index=False)
+    print('P1C: six geometry ablations / full-parent child registries complete',flush=True)
+    comparison(full,train,ids,reg,geo,soma,emb,labelsets['graph_soma_tree'],seed,source)
+    figures(reg,soma,labelsets,rows)
+    dump(OUT/'audit/source_manifest.json',{'sources':sources,'runner_sha256':sha(Path(__file__)),'primitives_sha256':sha(ROOT/'simulation/src/bio_fly/p1_structure.py'),'protocol_sha256':sha(OUT/'audit/protocol.json'),'code_commit':subprocess.check_output(['git','rev-parse','HEAD'],cwd=ROOT,text=True).strip(),'seed':seed,'dataset_version':'FlyWire v783 / local MaleCNS-v1.0 named source','run_id':cfg['protocol_version'],'wall_clock_s':time.time()-start})
+    status={'P1A':'DONE_WITH_MALECNS_BLOCKER','P1B':'DONE_RETROSPECTIVE_PAIR_ISOLATION','P1C':'DONE_EXPLORATORY_GEOMETRY_DISCOVERY','P1D':'PARTIAL_DESCRIPTOR_PILOT_NOT_FULL_BRANCH_CONTACT_MODEL','P1E':'BLOCKED_INDEPENDENCE_AND_ID_QC','stable8_unchanged':True,'new_stable_candidates':0,'scientific_conclusion':'No discrete/continuous verdict yet; full independent competition remains open','wall_clock_s':time.time()-start}
+    dump(OUT/'P1_STATUS.json',status)
+    print(json.dumps(status),flush=True)
+
+def comparison(full,train,ids,reg,geo,soma,emb,labels,seed,source):
+    rng=np.random.default_rng(seed);co=full.tocoo();ix=rng.choice(len(co.data),min(60000,len(co.data)),replace=False)
+    a=list(co.row[ix]);b=list(co.col[ix]);positive=set(zip(co.row,co.col));neg=set()
+    while len(neg)<60000:
+        q=rng.integers(0,len(ids),size=(10000,2))
+        for i,j in q:
+            if i!=j and (i,j) not in positive:neg.add((int(i),int(j)))
+            if len(neg)==60000:break
+    na,nb=np.array(sorted(neg)).T;y=np.r_[np.ones(len(a)),np.zeros(len(na))].astype(int)
+    a=np.r_[a,na].astype(int);b=np.r_[b,nb].astype(int);sp=pair_split(ids,a,b,seed)
+    pd.DataFrame({'pre_id':ids[a],'post_id':ids[b],'edge_present':y,'split':np.array(['train','validation','test'])[sp]}).to_parquet(OUT/'data/prediction_pairs.parquet',index=False)
+    codes=pd.Categorical(reg.official_KC_subtype.fillna('unknown')).codes
+    fine=pd.Categorical(reg.frozen_detailed_subtype.fillna('unknown')).codes
+    sides=pd.Categorical(reg.official_hemisphere.fillna('unknown')).codes
+    base=np.column_stack([codes[a],codes[b],fine[a],fine[b],sides[a],sides[b]])
+    deg=np.column_stack([np.asarray(train.sum(1)).ravel(),np.asarray(train.sum(0)).ravel()])
+    spatial=np.column_stack([soma[a],soma[b],np.abs(soma[a]-soma[b]),np.linalg.norm(soma[a]-soma[b],axis=1),np.log1p(deg[a]),np.log1p(deg[b])])
+    t=geo.drop(columns='root_id').to_numpy(float)
+    tree=np.column_stack([t[a],t[b],np.abs(t[a]-t[b])])
+    m1=np.column_stack([base,spatial]);m2=np.column_stack([m1,tree])
+    matrices={'M0':base,'M1':m1,'M2_descriptor':m2,'M3_descriptor_discrete':np.column_stack([m2,labels[a],labels[b],labels[a]==labels[b]]),'M4_descriptor_continuous':np.column_stack([m2,emb[a],emb[b],np.abs(emb[a]-emb[b])])}
+    records=[];preds=pd.DataFrame({'pre_id':ids[a[sp==2]],'post_id':ids[b[sp==2]],'edge_present':y[sp==2]})
+    for name,x in matrices.items():
+        cat=np.zeros(x.shape[1],dtype=bool);cat[:6]=True
+        if name.startswith('M3'):cat[-3:-1]=True
+        fits=[]
+        for leaves in [15,31]:
+            tm=time.time();model=HistGradientBoostingClassifier(max_iter=100,max_leaf_nodes=leaves,learning_rate=.08,l2_regularization=1,categorical_features=cat,early_stopping=False,random_state=seed)
+            model.fit(x[sp==0],y[sp==0]);vl=log_loss(y[sp==1],model.predict_proba(x[sp==1])[:,1]);fits.append((vl,leaves,model,time.time()-tm))
+        vl,leaves,model,_=min(fits,key=lambda z:z[0]);p=model.predict_proba(x[sp==2])[:,1];preds[name]=p
+        records.append({'model':name,'test_log_loss':log_loss(y[sp==2],p),'test_AUROC':roc_auc_score(y[sp==2],p),'validation_log_loss':vl,'selected_max_leaves':leaves,'n_features':x.shape[1],'n_train':int((sp==0).sum()),'n_validation':int((sp==1).sum()),'n_test':int((sp==2).sum()),'training_budget':'2 settings x100 iterations','training_wall_clock_s':sum(z[3] for z in fits),'evidence':'RETROSPECTIVE_DESCRIPTOR_PILOT','uncertainty':'NOT_ESTIMATED_NO_SIGNIFICANCE_CLAIM','sampling':'case-control 60000 positive/60000 negative; probabilities not natural prevalence calibrated'})
+        print(name,records[-1]['test_log_loss'],flush=True)
+    pd.DataFrame(records).to_csv(OUT/'MODEL_COMPARISON.csv',index=False);preds.to_parquet(OUT/'data/test_predictions.parquet',index=False)
+
+def figures(reg,soma,labels,rows):
+    import matplotlib
+    matplotlib.use('Agg')
+    import matplotlib.pyplot as plt
+    selected=reg.candidate_id.eq('CM019') & reg.stable_backbone.astype(bool)
+    parent=reg.frozen_detailed_subtype.eq("KCa'b'-ap1") & reg.official_hemisphere.eq('left')
+    fig,axes=plt.subplots(1,3,figsize=(12,4))
+    for ax,(i,j) in zip(axes,[(0,1),(0,2),(1,2)]):
+        ax.scatter(soma[:,i],soma[:,j],s=1,c='#dddddd',rasterized=True,label='all KC')
+        ax.scatter(soma[parent&~selected,i],soma[parent&~selected,j],s=16,facecolors='none',edgecolors='#d95f02',label='parent residual')
+        ax.scatter(soma[selected,i],soma[selected,j],s=5,c='#147d92',label='CM019')
+        ax.set(xlabel='XYZ'[i]+' (um)',ylabel='XYZ'[j]+' (um)');ax.set_aspect('equal')
+    axes[-1].legend(fontsize=7);fig.suptitle(f'CM019: selected {selected.sum()}, left ap1 parent {parent.sum()}, residual {(parent&~selected).sum()}')
+    fig.text(.01,.01,'FlyWire v783; all5177_identity.parquet; P1-v2.1; DESCRIPTIVE_ONLY; soma positions, no dynamic validation',fontsize=7)
+    fig.tight_layout(rect=(0,.04,1,.93))
+    for ext in ['pdf','svg']:fig.savefig(OUT/'figures'/f'CM019_full_parent_somata.{ext}')
+    plt.close(fig)
+    fig,axes=plt.subplots(2,3,figsize=(12,8))
+    for ax,(name,lab) in zip(axes.ravel(),labels.items()):
+        ax.scatter(soma[:,0],soma[:,1],s=2,c=lab,cmap='turbo',rasterized=True);ax.set_title(name);ax.set_aspect('equal');ax.set(xlabel='X (um)',ylabel='Y (um)')
+    fig.suptitle('Actual spatial fusion: same K=30, six frozen feature arms')
+    fig.text(.01,.01,'FlyWire v783; P1_exploratory_assignments.parquet; n=5177 each; P1-v2.1; exploratory partitions, not new stable units',fontsize=7)
+    fig.tight_layout(rect=(0,.04,1,.95))
+    for ext in ['pdf','svg']:fig.savefig(OUT/'figures'/f'spatial_fusion_ablation.{ext}')
+    plt.close(fig)
+
+if __name__=='__main__':
+    with threadpool_limits(limits=4):main()

p1_v2_fine_types/audit/standalone_replay_checks.json

@@ -0,0 +1,7 @@
+{
+  "models_exact": true,
+  "assignments_exact": true,
+  "parents_exact": true,
+  "source_audit_repeated": false,
+  "note": "Compact bundle replay reconstructs graph and recalculates discovery/model results; imports source audit and precomputed geometry."
+}

p1_v2_fine_types/audit/structural_checks.csv

@@ -0,0 +1,11 @@
+check,passed,actual
+all_KC_unique_strings,True,5177
+stable8_union,True,965
+full_KCab-s_left,True,[311]
+full_KCab-s_right,True,[310]
+synapse_conservation,True,379338
+train_validation_direct_and_reverse_isolation,True,"[0, 0]"
+train_test_direct_and_reverse_isolation,True,"[0, 0]"
+validation_test_direct_and_reverse_isolation,True,"[0, 0]"
+node_JSON_string_roundtrip,True,5177
+SU019_CM019_not_aliased,True,"[104, 132, 104]"

p1_v2_fine_types/data/contact_free_tree_features.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:02faead833dd5b846efb11e749b0ced53252176e664d8adfb34897f7db4da2d4
+size 769954

p1_v2_fine_types/data/directed_pair_split.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ac28bbbd928cc8f4459850ea35a8732fd6191228cc7fac5db048676d9f115891
+size 735097

p1_v2_fine_types/data/feature_schema.json

@@ -0,0 +1,27 @@
+{
+  "soma_unit": "um",
+  "soma_scale_um": 104.1533268297617,
+  "tree_columns": [
+    "cable_length_um",
+    "n_branch_ids",
+    "n_terminals",
+    "n_branchpoints",
+    "edge_length_median_um",
+    "edge_length_p95_um",
+    "path_median_um",
+    "path_p95_um",
+    "centroid_x_um",
+    "extent_x_um",
+    "principal_sd_0_um",
+    "centroid_y_um",
+    "extent_y_um",
+    "principal_sd_1_um",
+    "centroid_z_um",
+    "extent_z_um",
+    "principal_sd_2_um"
+  ],
+  "tree_source": "skeleton geometry only; no contacts, partner labels, old CM/SU labels",
+  "fusion": "concatenate E/sqrt(dE), sqrt(ws)*(soma-mean)/scale, sqrt(wt)*z(tree)/sqrt(dT)",
+  "embedding_source": "pair_grouped_train.npz",
+  "normalization": "transductive unlabeled full registry; not inductive new-cell evaluation"
+}

p1_v2_fine_types/data/pair_grouped_train.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c987937c78c13c5f3859859aafa0cf6ccf4ce28a8ca073e9a6273b9a46d590a8
+size 465705

p1_v2_fine_types/data/prediction_pairs.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0011019ed8394c276f4311f136a87a00d7f3c2f7c774d3280e43c306031cf455
+size 437528

p1_v2_fine_types/data/test_predictions.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d32a46516440e0ec07e1675e9e795c2409e9f86c1e241fe7f030f4806cc09c50
+size 847321

p1_v2_fine_types/data/train_only_embedding.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:06e47ce94f182181ef8989e371e1c56e88bb6ea5bc84edc0d0c8d1e28e39ace7
+size 2650752

p1_v2_fine_types/figure_source_ledger.csv

@@ -0,0 +1,5 @@
+figure,source_file,dataset_version,protocol,n,denominator,metric_definition,simulation_vs_biology_label,run_id
+figures/CM019_full_parent_somata.svg,registry/all5177_identity.parquet;registry/P1_exploratory_assignments.parquet,FlyWire v783,P1-v2.2-retrospective-fine-type-geometry-pilot,5177,all registered KC,real soma coordinate in um; categorical discovered partition or CM019/parent/background,DESCRIPTIVE_REAL_STRUCTURE_NOT_FUNCTIONAL_VALIDATION,P1-v2.2-retrospective-fine-type-geometry-pilot
+figures/CM019_full_parent_somata.pdf,registry/all5177_identity.parquet;registry/P1_exploratory_assignments.parquet,FlyWire v783,P1-v2.2-retrospective-fine-type-geometry-pilot,5177,all registered KC,real soma coordinate in um; categorical discovered partition or CM019/parent/background,DESCRIPTIVE_REAL_STRUCTURE_NOT_FUNCTIONAL_VALIDATION,P1-v2.2-retrospective-fine-type-geometry-pilot
+figures/spatial_fusion_ablation.svg,registry/all5177_identity.parquet;registry/P1_exploratory_assignments.parquet,FlyWire v783,P1-v2.2-retrospective-fine-type-geometry-pilot,5177,all registered KC,real soma coordinate in um; categorical discovered partition or CM019/parent/background,DESCRIPTIVE_REAL_STRUCTURE_NOT_FUNCTIONAL_VALIDATION,P1-v2.2-retrospective-fine-type-geometry-pilot
+figures/spatial_fusion_ablation.pdf,registry/all5177_identity.parquet;registry/P1_exploratory_assignments.parquet,FlyWire v783,P1-v2.2-retrospective-fine-type-geometry-pilot,5177,all registered KC,real soma coordinate in um; categorical discovered partition or CM019/parent/background,DESCRIPTIVE_REAL_STRUCTURE_NOT_FUNCTIONAL_VALIDATION,P1-v2.2-retrospective-fine-type-geometry-pilot

p1_v2_fine_types/figures/spatial_fusion_ablation.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:866d72796bffd3cb07e65e8570a4445872c1af8748514a292d205a99a59f14f7
+size 245509

p1_v2_fine_types/registry/CM_SU_crosswalk.csv

@@ -0,0 +1,64 @@
+candidate_id,unit_id,n_shared
+CM001,,390
+CM002,,307
+CM003,,304
+CM004,,252
+CM005,,241
+CM006,,220
+CM007,SU002,206
+CM008,SU003,180
+CM009,SU004,171
+CM010,SU005,167
+CM011,,164
+CM012,SU017,116
+CM012,,47
+CM013,,158
+CM014,SU022,97
+CM014,,55
+CM015,,150
+CM016,SU021,98
+CM016,,49
+CM017,SU011,144
+CM018,SU013,132
+CM018,,6
+CM019,SU019,104
+CM019,,28
+CM020,SU018,108
+CM020,,21
+CM021,,129
+CM022,SU027,84
+CM022,,30
+CM023,,94
+CM024,SU036,55
+CM024,,38
+CM025,SU037,50
+CM025,,39
+CM026,SU024,88
+CM027,SU033,68
+CM027,,14
+CM028,,75
+CM029,SU039,41
+CM029,,30
+CM030,SU030,70
+CM031,SU034,67
+CM032,,62
+CM033,SU035,61
+CM034,,46
+CM035,,41
+CM036,,13
+CM037,,11
+CM038,,8
+CM039,SU030,7
+CM040,,6
+CM041,SU003,5
+CM042,,5
+CM043,SU024,5
+CM044,,4
+CM045,,3
+CM046,,3
+CM047,,3
+CM048,,2
+CM049,,2
+CM050,,1
+CM051,,1
+CM052,SU005,1

p1_v2_fine_types/registry/P1_exploratory_assignments.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eab3c8ccdaf056fcd085820158de322da0c74f37ddb97b98bd246b073442bebd
+size 84120

p1_v2_fine_types/registry/all5177_identity.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0440f0c0fa77f547df027c929b6b120d8e8121acc59fafe78e880fb8ece4e630
+size 231062

p1_v2_fine_types/registry/ap1L_18_exploratory_members.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6dc7405e9240d0435d2d275f1c0d5a94acebda518ff06a7451620aab8babe24d
+size 31509

p1_v2_fine_types/registry/parent_candidate_residual_atlas.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1e7607070cc9fb9608ee2ec92502c24ea43190a2b1f2bd88114d62d5b8c532f4
+size 17946

p1_v2_fine_types/registry/parent_child_members.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:75bf98bd0695832b307ab192adcc09e82a9cb81a63a0a43c80d3d013eb7f32ce
+size 22544