实验定位 carrier package 正向 / 路由混合
A51 解决 A50 后的下一层问题：有了 10-case face/vertex masks，并不等于所有 case 都可以直接跑 depth-sign residual。A51 把 A23 的 repair_hybrid 作为 10-case topology carrier candidate 打包，并用无 GT 的 topology + mask evidence 判断每个 case 是否 ready、是否值得 residual、是否应 defer。
结果：completed=10/10，carrier/mask 都持久化成功；topology valid=9/10，mask shape match=10/10，carrier-ready=8/10。路由为：high-value depth-sign=2，standard depth-sign=1，low-conflict defer=5，not-ready defer=2。
实验设计（Image-2 风格）
模块设计（Image-2 风格）
No-GT Readiness Rule
- 拓扑：水密、winding consistent、单组件、无 boundary edges、无 nonmanifold edges。
- mask：A50 face confidence ≥ 0.85，p95 sample-to-face transfer ≤ 0.025。
- activity：conflict face ratio ≥ 0.20 标记为 high-value residual；低于 0.05 标记 carrier-only/defer。
- GT Chamfer/F@5 只用于 retrospective audit，不参与路由。
Why This Matters
A51 把“能跑 residual”和“应该跑 residual”分开。gso_006 虽然 conflict 很高，但 carrier 有 204 个 components，必须 defer；gso_005 conflict 近 0，强行移动只会制造假收益风险。这种可拒绝机制比工程式全量后处理更接近论文里的可靠生成控制。
实验结果（表格）
| case | route | reasons | topology valid | mask match | components | faces | face ratio | face conf | p95 transfer | conflict face | Δ Chamfer | Δ F@5 | carrier OBJ |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| gso_000_input4 | ready standard depth sign | ['moderate_conflict_activity_region'] | True | True | 1 | 2734 | 0.2026 | 0.9110 | 0.01337 | 0.1306 | -0.00160 | 0.0202 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a51_10case_topology_carrier_readiness/gso_000_input4/carrier_repair_hybrid.obj |
| gso_001_input4 | ready but defer low conflict | ['low_conflict_carrier_only_preferred'] | True | True | 1 | 449604 | 1.0068 | 0.9697 | 0.01112 | 0.0290 | -0.00105 | 0.0082 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a51_10case_topology_carrier_readiness/gso_001_input4/carrier_repair_hybrid.obj |
| gso_002_input4 | ready high value depth sign | ['high_conflict_activity_region'] | True | True | 1 | 234960 | 0.1607 | 0.9010 | 0.01703 | 0.2018 | -0.00377 | 0.0502 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a51_10case_topology_carrier_readiness/gso_002_input4/carrier_repair_hybrid.obj |
| gso_003_input4 | defer carrier not ready | ['low_face_vote_confidence'] | True | True | 1 | 295076 | 0.1124 | 0.8489 | 0.01880 | 0.0897 | -0.00005 | -0.0041 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a51_10case_topology_carrier_readiness/gso_003_input4/carrier_repair_hybrid.obj |
| gso_004_input4 | ready but defer low conflict | ['low_conflict_carrier_only_preferred'] | True | True | 1 | 100120 | 0.0779 | 0.9895 | 0.01353 | 0.0038 | -0.00041 | 0.0032 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a51_10case_topology_carrier_readiness/gso_004_input4/carrier_repair_hybrid.obj |
| gso_005_input4 | ready but defer low conflict | ['low_conflict_carrier_only_preferred'] | True | True | 1 | 328064 | 1.0284 | 1.0000 | 0.00589 | 0.0000 | 0.00119 | -0.0079 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a51_10case_topology_carrier_readiness/gso_005_input4/carrier_repair_hybrid.obj |
| gso_006_input4 | defer carrier not ready | ['carrier_topology_not_valid'] | False | True | 204 | 1832108 | 1.2150 | 0.8698 | 0.01430 | 0.4054 | 0.00144 | 0.0204 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a51_10case_topology_carrier_readiness/gso_006_input4/carrier_repair_hybrid.obj |
| gso_007_input4 | ready but defer low conflict | ['low_conflict_carrier_only_preferred'] | True | True | 1 | 278432 | 0.5190 | 0.9925 | 0.00771 | 0.0264 | 0.00281 | -0.0188 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a51_10case_topology_carrier_readiness/gso_007_input4/carrier_repair_hybrid.obj |
| gso_008_input4 | ready high value depth sign | ['high_conflict_activity_region'] | True | True | 1 | 827994 | 1.5079 | 0.9431 | 0.00802 | 0.7050 | 0.00474 | -0.0168 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a51_10case_topology_carrier_readiness/gso_008_input4/carrier_repair_hybrid.obj |
| gso_009_input4 | ready but defer low conflict | ['low_conflict_carrier_only_preferred'] | True | True | 1 | 83166 | 0.1094 | 0.9784 | 0.00889 | 0.0176 | -0.00006 | -0.0029 | /home/jiachen/TRELLIS/experiments/nonorthogonal_gpt_orthoview/outputs/data_survey/vggt_spatial_adapter/support_guidance_a51_10case_topology_carrier_readiness/gso_009_input4/carrier_repair_hybrid.obj |
可视化结果
实验结论
- A51 成功打包 10/10 carrier OBJ + mask NPZ + manifest，为 10-case A47/A48 rerun 提供了统一入口。
- 8/10 case 通过 carrier readiness；
gso_003因 face vote confidence 低 defer，gso_006因 carrier topology 异常 defer。 - 真正适合马上做 depth-sign 的是
gso_002_input4, gso_008_input4, gso_000_input4；其中gso_002_input4, gso_008_input4是高 conflict/high-value residual target。 - 5 个 low-conflict case 被路由到 carrier-only/defer：这避免了把无冲突区域也强行 residual，正好回应 A44-A46 里“过度投影会伤 GT”的问题。
- retrospective GT 显示 carrier 相比 baseline 平均 Δ Chamfer=0.00032、Δ F@5=0.0052：topology carrier 的几何收益是混合的，所以后续 residual 必须只在高价值 conflict 区活动。
下一步想法
- A52：只对 A51 route=ready_high_value/ready_standard 的 case 跑 10-case depth-sign residual，先覆盖 gso_000/gso_002/gso_008。
- A53：为 gso_006 做 carrier repair fallback：从 204 components 中选择 visible/mask 支持最大的单组件或重新 voxel close，再重新评估 readiness。
- A54：把 A51 route 写入 A47/A48 selector，形成 no-GT pipeline：carrier package -> readiness -> depth sign -> confidence selector -> defer。
- A55：对 low-conflict defer case 做消融，证明“强行 residual”不会带来稳定收益，支撑论文里的 reject option。