实验定位 诊断正 / 未执行 mesh edit
A33 继续推进 A32 的下一步：如果 case-level 三路路由成立，那么真正的方法应在 surface patch 层面执行局部编辑。由于 A30/A31 没有保存逐点 surface labels，本轮不伪装成真实 mesh edit，而是用已有 surface ratios 做 local edit feasibility diagnosis：判断当前证据是否足够支持 repair-lock、baseline-revert、completion、defer 四类局部状态。
结果：10-case 中 executable local edit cases=4，neutral defer cases=5，mean direct editable ratio=0.3847，mean defer ratio=0.6153。这说明 A33 的最大贡献是指出下一步必须保存 per-sample labels 并做 patch clustering。
实验设计（Image-2 风格）
模块设计（Image-2 风格）
Local Edit States
- repair_lock：surface_improved regions should keep repair geometry and can receive consistency locks
- baseline_revert：surface_conflict regions should be replaced by baseline or penalized during generation
- completion：unobserved_closable regions can be completed/watertight-filled with weak constraints
- defer：surface_neutral regions lack evidence and should avoid hard acceptance
Evidence Boundary
A33 的输入是聚合比例，不是逐点 mesh surface label。因此它只能判断“是否值得做 patch-level edit”，不能输出真实编辑后的 mesh。这个限制很关键：下一轮 A34 必须修改 A30/A31 的数据保存流程，把每个 surface sample 的位置、标签、残差和投影视角持久化。
实验结果（表格）
| case | A32 decision | local edit plan | repair lock | baseline revert | completion | defer | direct editable | edit confidence | GT repair Chamfer Δ | GT repair F@5 Δ | 原因 |
|---|---|---|---|---|---|---|---|---|---|---|---|
| gso_000_input4 | accept | lock_repair_dominant | 0.4801 | 0.1295 | 0.2552 | 0.1351 | 0.8649 | 0.7973 | -0.00160 | 0.0202 | ['repair-improved surface dominates with bounded conflict'] |
| gso_001_input4 | defer | defer_neutral_dominant | 0.0294 | 0.0470 | 0.0000 | 0.9236 | 0.0764 | 0.0000 | -0.00105 | 0.0082 | ['neutral evidence dominates; local edit lacks positive support'] |
| gso_002_input4 | accept | lock_repair_dominant | 0.6640 | 0.1862 | 0.0119 | 0.1379 | 0.8621 | 0.7932 | -0.00377 | 0.0502 | ['repair-improved surface dominates with bounded conflict'] |
| gso_003_input4 | defer | ambiguous_patch_route | 0.1505 | 0.1767 | 0.0000 | 0.6727 | 0.3272 | 0.0000 | -0.00005 | -0.0041 | ['mixed local evidence; requires patch clustering before editing'] |
| gso_004_input4 | defer | defer_neutral_dominant | 0.0106 | 0.0194 | 0.0000 | 0.9700 | 0.0300 | 0.0000 | -0.00041 | 0.0032 | ['neutral evidence dominates; local edit lacks positive support'] |
| gso_005_input4 | defer | defer_neutral_dominant | 0.0000 | 0.0000 | 0.0000 | 1.0000 | 0.0000 | 0.0000 | 0.00119 | -0.0079 | ['neutral evidence dominates; local edit lacks positive support'] |
| gso_006_input4 | accept | hybrid_lock_and_complete | 0.2412 | 0.2970 | 0.1941 | 0.2676 | 0.7324 | 0.5958 | 0.00144 | 0.0204 | ['moderate repair evidence plus unobserved completion budget'] |
| gso_007_input4 | defer | defer_neutral_dominant | 0.0055 | 0.0275 | 0.0000 | 0.9670 | 0.0330 | 0.0000 | 0.00281 | -0.0188 | ['neutral evidence dominates; local edit lacks positive support'] |
| gso_008_input4 | reject | revert_conflict_dominant | 0.1739 | 0.7091 | 0.0000 | 0.1170 | 0.8830 | 0.8091 | 0.00474 | -0.0168 | ['visible conflict dominates; use baseline for conflict regions'] |
| gso_009_input4 | defer | defer_neutral_dominant | 0.0152 | 0.0225 | 0.0000 | 0.9623 | 0.0377 | 0.0000 | -0.00006 | -0.0029 | ['neutral evidence dominates; local edit lacks positive support'] |
可视化结果
实验结论
- A33 把论文方法边界推进了一步：问题不应再被表述为“选 baseline 还是 repair”，而应表述为“每个 observable surface patch 应该 lock repair、revert baseline、complete 还是 defer”。
- 当前证据还不足以执行真实局部编辑：mean defer ratio=0.6153，明显高于 mean completion budget=0.0461，说明大部分表面是中性/证据不足，而不是可直接补全。
- A33 保留了 A32 的关键发现：neutral false accept risk cases=['gso_005_input4', 'gso_007_input4']，这些 case 不应该被 repair 接受，而应进入 defer。
- 坏 repair 的平均 conflict ratio=0.2455，好 repair 的平均 improved ratio=0.2273；这支持使用 surface state 作为局部约束，但也说明仅靠聚合均值还不够。
- 最重要的工程/研究缺口是数据产物：A30 只保存了 summary 和 PNG，没有保存 per-sample labels。要做真正 patch edit，必须先保存 sample-level evidence。
下一步想法
- A34：重跑或改造 A30，保存 per-sample surface evidence：xyz、normal、view ids、depth residual delta、improved/conflict/neutral/unobserved 标签。
- 用 surface sample graph 或 geodesic/Euclidean clustering 形成 patch，再对 patch 执行 repair-lock / baseline-revert / completion / defer。
- 实现最小 mesh edit 原型：不追求完美重拓扑，先输出 patch colored mesh 或 vertex mask，证明局部决策可以落到 mesh 上。
- 把 patch labels 映射回 TRELLIS.2 sparse/SLat token，设计 generation-time observable-state guidance。