实验定位 V2 完成 / 数据覆盖审计
A79 是三版路线中的第 2 版：它不重新生成 mesh，而是把 A78 的 3-case observation-state 扩展到 A51/A70/A71 的 10-case route/evaluator 证据。目标是回答：A80 要学习的状态类型是否真实存在，哪些状态还缺完整 sampler trajectory。
结果：10 个 case 覆盖 5/5 个必要 proxy 状态；但完整 A65 sampler trajectory 只覆盖 3 类 active 状态，defer/topology-risk 仍是 proxy-only。active 3 case 上 A70 的 mean ΔChamfer=-0.0004057，mean ΔF@5=0.002796。
Only 3 cases have full sampler trajectory labels from A65; the other 7 cases use proxy labels derived from readiness, surface masks, active/defer route and same-evaluator metrics.
实验设计（Image-2 风格）
模块设计（Image-2 风格）
State Taxonomy
- late-safe global gain：中等 conflict + 大 unobserved/completion room。
- visible-region recovery：visible surface improvement 很强且置信度高。
- saturated local damage：高 conflict 且几乎没有 closure room。
- defer low-conflict：大部分表面 neutral，编辑收益低。
- carrier topology failure：carrier 未 ready 或面片/组件异常。
Training Implication
- A78 的失败不是单纯阈值没调好，而是 held-out train split 缺状态类型。
- A79 给 A80 指定 state-balanced 采样，避免只学 active case。
- proxy-only 状态必须补 full sampler trajectory，否则模型只能学到 route，而不是 step-level budget。
实验结果（表格）
state summaries
| state | cases | active | conflict | unobserved | improved | A70 ΔChamfer | A70 ΔF@5 | case ids |
|---|---|---|---|---|---|---|---|---|
| carrier_topology_failure | 1 | 0 | 0.4054 | 0.0010 | 0.2711 | 0.0000000 | 0.000000 | ['gso_006_input4'] |
| defer_low_conflict | 5 | 0 | 0.0153 | 0.0000 | 0.0083 | 0.0000000 | 0.000000 | ['gso_001_input4', 'gso_004_input4', 'gso_005_input4', 'gso_007_input4', 'gso_009_input4'] |
| defer_not_ready | 1 | 0 | 0.0897 | 0.0000 | 0.0895 | 0.0000000 | 0.000000 | ['gso_003_input4'] |
| late_safe_global_gain | 1 | 1 | 0.1306 | 0.2359 | 0.4952 | -0.0001331 | -0.004499 | ['gso_000_input4'] |
| saturated_local_damage | 1 | 1 | 0.7050 | 0.0002 | 0.1707 | 0.0002912 | -0.000896 | ['gso_008_input4'] |
| visible_region_recovery | 1 | 1 | 0.2018 | 0.0265 | 0.6134 | -0.0013750 | 0.013782 | ['gso_002_input4'] |
case labels
| case | state | A65 trajectory | route | conflict | improved | unobserved | carrier faces | reason |
|---|---|---|---|---|---|---|---|---|
| gso_000_input4 | late_safe_global_gain | 是 | active | 0.1306 | 0.4952 | 0.2359 | 2734 | large unobserved/completion room with moderate conflict |
| gso_001_input4 | defer_low_conflict | 否 | defer | 0.0290 | 0.0156 | 0.0000 | 449604 | low conflict and mostly neutral observed surface |
| gso_002_input4 | visible_region_recovery | 是 | active | 0.2018 | 0.6134 | 0.0265 | 234960 | visible surface has strong improvement evidence |
| gso_003_input4 | defer_not_ready | 否 | defer | 0.0897 | 0.0895 | 0.0000 | 295076 | carrier not ready, so sampler should not edit this case |
| gso_004_input4 | defer_low_conflict | 否 | defer | 0.0038 | 0.0038 | 0.0000 | 100120 | low conflict and mostly neutral observed surface |
| gso_005_input4 | defer_low_conflict | 否 | defer | 0.0000 | 0.0000 | 0.0000 | 328064 | low conflict and mostly neutral observed surface |
| gso_006_input4 | carrier_topology_failure | 否 | defer | 0.4054 | 0.2711 | 0.0010 | 1832108 | carrier not ready and carrier topology is fragmented/high-face |
| gso_007_input4 | defer_low_conflict | 否 | defer | 0.0264 | 0.0014 | 0.0000 | 278432 | low conflict and mostly neutral observed surface |
| gso_008_input4 | saturated_local_damage | 是 | active | 0.7050 | 0.1707 | 0.0002 | 827994 | saturated visible conflict with near-zero closure room |
| gso_009_input4 | defer_low_conflict | 否 | defer | 0.0176 | 0.0205 | 0.0000 | 83166 | low conflict and mostly neutral observed surface |
可视化结果
实验结论
- A79 给出了 A80 可以学习的状态空间：10-case proxy 覆盖了 active、defer、topology-risk，不再只有 A65 的 3 个 active case。
- 但 A79 也暴露了训练数据缺口：defer_low_conflict、defer_not_ready、carrier_topology_failure 没有完整 sampler trajectory。
- 因此下一步的“大提升架构”不能再是手写阈值，而应是 state-balanced region-budget head，并继续补采 trajectory。
- 这一步的正面进展是：A78 的泛化失败现在被解释为状态覆盖问题，而不是 observation-state 思路本身无效。
下一步想法
- A80：把状态标签变成模型结构，训练 VGGT Observation Encoder + Region Budget Head + TRELLIS.2 sampler controller。
- 优先补 proxy-only 状态的 sparse/SLat trajectory，尤其是 topology-risk 和 low-conflict defer。
- 训练时联合 Chamfer、F@5/coverage、topology-risk 和 face-budget，避免 A66 的 Chamfer 正向但 recall 受损。