p20030920p · Oct 9, 2026 · Oct 9, 2026 · Oct 9, 2026 · Oct 9, 2026 · Oct 9, 2026
diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
@@ -22,8 +22,11 @@ jobs:
       - uses: astral-sh/setup-uv@v6
         with:
           version: '0.9.2'
-      - run: uv sync --frozen --python 3.10 --extra dev
+      - run: uv sync --frozen --python 3.10 --extra dev --extra audit
       - run: uv run ruff check src tests scripts
+      - run: uv run --no-sync deptry src
+      - run: uv run --no-sync lint-imports --no-cache
+      - run: uv run --no-sync vulture
       - run: uv run pytest -q
       - run: uv run python scripts/verify_evidence.py
       - run: uv run python scripts/verify_paired_evidence.py

diff --git a/README.md b/README.md
diff --git a/README.zh-CN.md b/README.zh-CN.md
diff --git a/docs/AUDIT.md b/docs/AUDIT.md
@@ -2,6 +2,8 @@
 
 English | [中文](AUDIT.zh-CN.md)
 
+> Historical measurements/analysis; current question and next decision: [STUDY](STUDY.md) · [PLAN](PLAN.md).
+
 Baseline inspected: `af1e58b`. Original files remain recoverable in Git; a local sibling backup was preserved. Historical scores are **unverified inherited claims**, not inputs to the new ledger.
 
 | Finding | Consequence | Replacement |

diff --git a/docs/AUDIT.zh-CN.md b/docs/AUDIT.zh-CN.md
@@ -2,6 +2,8 @@
 
 [English](AUDIT.md) | 中文
 
+> 历史测量／分析；当前问题与下一项决策统一见 [STUDY](STUDY.zh-CN.md) · [PLAN](PLAN.zh-CN.md)。
+
 检查基线为 `af1e58b`，原文件可从 Git 恢复，也保留了本机相邻备份。历史分数是未核实的继承记录，不参与新汇总。
 
 | 发现 | 影响 | 替换 |

diff --git a/docs/DELAYED_RESULTS.md b/docs/DELAYED_RESULTS.md
@@ -2,6 +2,8 @@
 
 English | [中文](DELAYED_RESULTS.zh-CN.md)
 
+> Historical measurements/analysis; current question and next decision: [STUDY](STUDY.md) · [PLAN](PLAN.md).
+
 **The frozen experiment passes its historical-decision prerequisite, but a simpler support-gate control removes the selected target deficit without reassociation. H1 remains a candidate; a bounded replay prototype is not justified yet.** This is a new-scene ConceptGraphs diagnostic, not a SLAM, navigation or four-method benchmark.
 
 ## What was frozen and checked

diff --git a/docs/DELAYED_RESULTS.zh-CN.md b/docs/DELAYED_RESULTS.zh-CN.md
@@ -2,6 +2,8 @@
 
 [English](DELAYED_RESULTS.md) | 中文
 
+> 历史测量／分析；当前问题与下一项决策统一见 [STUDY](STUDY.zh-CN.md) · [PLAN](PLAN.zh-CN.md)。
+
 **冻结主实验通过了历史决策效应的前提检验，但更简单的支持门槛对照无需重新关联，就消除了所选目标的恢复缺口。H1 仍是候选，暂不足以支持实现有界回放原型。** 这是新场景上的 ConceptGraphs 诊断，不是完整 SLAM、导航或四方法基准。
 
 ## 冻结内容与核验

diff --git a/docs/DISCLOSURE.md b/docs/DISCLOSURE.md
@@ -2,8 +2,8 @@
 
 English | [中文](DISCLOSURE.zh-CN.md)
 
-AI contributed to literature discovery and analysis, implementation, debugging, annotation assistance and writing. Measurements come from executed runs; generated explanations do not replace experiments. Room0 and room1 partial-surface annotations were visually checked but await independent human review. Room1 references are excluded from mapping; its six support-gate controls are explicitly post-hoc and separate from 35 frozen cells.
+AI assisted literature analysis, implementation, debugging, annotation and writing. Numbers come from recorded execution; generated explanations do not replace experiments. Room0/room1 partial-surface labels lack independent human review. Room2 uses an explicitly AI-only exploratory protocol after the owner waived manual review; this does not constitute confirmation.
 
-Measurements, exploratory interpretations and candidate mechanisms are labeled separately. The applicant should read sources, review annotations and personally walk through at least one reproduction. This repository does not claim independent authorship or complete applicant understanding of every implementation.
+Primary protocols, post-hoc controls and candidate mechanisms are distinguished. Room1 references are held out from mapping; its six support controls are separate from 35 frozen cells. Room2's 28 mapping cells ran, but analysis remains pending. Original labels and evidence are preserved.
 
-All current mapping runs use supplied poses. Complete SLAM trajectories, semantic benchmarks, navigation and physical validation remain unfinished. Failures and adverse evidence are retained in public records.
+The four research cores use supplied poses. Later original-code runs include scoped semantic scoring; the full benchmarks, complete SLAM, navigation and physical H1 validation remain incomplete. Failures and adverse evidence are public. The applicant should personally read sources and run a reproduction; the repository does not claim independent authorship or complete understanding of every implementation.
diff --git a/docs/DISCLOSURE.zh-CN.md b/docs/DISCLOSURE.zh-CN.md
@@ -2,8 +2,8 @@
 
 [English](DISCLOSURE.md) | 中文
 
-AI 参与文献发现与分析、实现、调试、标注辅助和写作。原始运行产生测量值；AI 生成的解释不代替实验。Room0 与 room1 部分表面标注经视觉检查，但尚未独立人工审核。Room1 参考视图不进入建图；6 个支持门槛对照明确属于事后探索，与 35 个冻结单元分开。
+AI 辅助文献分析、实现、调试、标注和写作。数字来自运行记录，生成解释不代替实验。Room0／room1 部分表面标注尚未独立人审；room2 在所有者豁免人工复核后使用明确的 AI-only 探索协议，不构成确认性证据。
 
-实测、探索性解释和候选机制分别标明。申请人应亲自阅读来源、检查标注并走通至少一个复现流程；仓库不声称申请人独立写成或已理解全部实现。
+原协议、事后对照和候选机制分别标明。Room1 参考帧不参与建图，六个支持门槛对照与 35 个冻结单元分开。Room2 的 28 个建图单元已运行，分析待完成。原标注及证据保留。
 
-所有当前建图运行使用给定位姿。完整 SLAM 轨迹、完整语义 benchmark、导航与实物验证尚未完成。失败与不利证据保留在公开记录中。
+四个研究核心使用给定位姿。后续作者原流程有范围限定的语义评分；完整 benchmark、完整 SLAM、导航和实物 H1 验证仍未完成。失败与不利证据公开。申请人应亲自读来源并运行复现；仓库不声称独立写成或已理解全部实现。
diff --git a/docs/LITERATURE.md b/docs/LITERATURE.md
@@ -2,7 +2,7 @@
 
 English | [中文](LITERATURE.zh-CN.md)
 
-Literature cutoff: **7 October 2026**. “Active direction” means several recent representative works pursue it; this is not a bibliometric popularity ranking. The eight core works are chosen for comparable map-update decisions, not a leaderboard across incompatible tasks.
+Literature cutoff: **7 October 2026**. “Active direction” means several recent representative works pursue it, not a popularity ranking. Four works were reproduced; four additional papers constrain the interpretation and novelty claim.
 
 ## Active directions within fields 1–2
 
@@ -15,25 +15,25 @@ Literature cutoff: **7 October 2026**. “Active direction” means several rece
 
 [AnyLoc](https://arxiv.org/abs/2308.00688) and [Revisit Anything](https://arxiv.org/abs/2409.18049) are adjacent visual place recognition examples. Recall depends on dataset and positive-match definitions; historical scores are not reused. No fresh VPR run is claimed.
 
-## Eight core works
+## Four reproduced works and four context papers
 
 | Work | Paper evidence | Structural implication (our inference) | Existing protection / boundary |
 | --- | --- | --- | --- |
 | [DUFOMap, 2024](https://arxiv.org/html/2403.01449v1) | III-A/B models observed free space and sensor/pose margins; IV-C specifies `d_p=1` | Cleaning uses supplied poses; a margin is not shared-residual covariance | Already addresses pose errors and compares pose sources; “noise unaware” would be wrong |
 | [BeautyMap, 2024](https://arxiv.org/html/2405.07283v1) | III uses binary occupancy and ray-based refinement/protection; Table I reports HA | Inconsistency after registration can reflect pose error or changes | Already protects against visibility mistakes; acknowledges ground-level limitations |
 | [ConceptGraphs, ICRA 2024](https://arxiv.org/html/2309.16650v1) | Object association combines geometric overlap and semantics, then fuses points/features | Incorrect world alignment may fragment or merge instances before reasoning | Semantic association helps; this is not evidence that all semantic maps assume static worlds |
-| [HOV-SG, RSS 2024](https://arxiv.org/html/2403.17846v2) | III-A uses accurate odometry for projection/fusion; IV-C uses external LiDAR SLAM | Graph/navigation geometry inherits the pose interface | External odometry and loop closure are safeguards; no fresh sensitivity run here |
+| [HOV-SG, RSS 2024](https://arxiv.org/html/2403.17846v2) | III-A uses accurate odometry for projection/fusion; IV-C uses external LiDAR SLAM | Graph/navigation geometry inherits the pose interface | External odometry and loop closure are safeguards; paired subset tests do not establish delayed recovery |
 | [Khronos, 2024](https://arxiv.org/html/2402.13817v2) | IV discusses local consistency; V-B jointly optimizes poses/fragments/background; V-C distinguishes absence of evidence | Candidate generation and local consistency still matter before corrections | **Counterexample:** already has joint optimization, history and visibility; cannot be grouped as fixed-pose independent cleaning |
 | [DovSG, 2024 preprint / 2025 paper](https://arxiv.org/html/2410.11989v2) | III-D relocalizes using ACE, matching and ICP before local updates | Wrong relocalization may affect associations and obsolete-voxel decisions | Explicitly supports dynamic local updates; refutes “all scene graphs are static” |
 | [SuperMap, August 2026](https://arxiv.org/html/2608.22896v1) | III conditions on pose; IV uses depth observation states and recursive evidence | Displayed model motivates checking shared pose covariance and calibration | Already separates observable, unobservable and disappeared; an omitted covariance formula does not prove a code defect |
 | [PerSeM, September 2026 preprint](https://arxiv.org/html/2609.19542v2) | 3.1/3.2 projects into persistent world voxels; 5 discusses drift/correlated errors | Label stability does not guarantee correct geometric evidence assignment | Memory/refinement already exists; correlated geometry is an acknowledged limitation |
 
-The recurring interface is narrow: establish spatial correspondence, then interpret map evidence. These papers do not implement one algorithm, and several already mitigate errors. The candidate question concerns residual shared geometric uncertainty in destructive-update confidence; an empirical cross-field claim still needs annotated pose-error controls beyond the first semantic subset.
+The recurring interface is narrow: establish spatial correspondence, then interpret map evidence. Several papers already mitigate errors. The current question tests recovery after late correction; a common dominant shared-uncertainty defect is unestablished. [Current analysis](STUDY.md).
 
 ## Code accessibility and scope
 
-DUFOMap and BeautyMap have executable author implementations and small labeled public data. Their exact commits are in [upstreams.json](../configs/upstreams.json). A class-agnostic ConceptGraphs frontend now executed on 40 posed Replica observations; [source/settings](../configs/semantic.json), [scope](SEMANTIC.md). HOV-SG now executes its 8-observation segment feature-map core; [paper card](papers/hovsg.md). The remaining four works were read, not freshly reproduced. [Four-paper structural analysis](STUDY.md).
+The [four core snapshots](papers/README.md) pin author commits, inputs and adaptations. [Later original-code runs at 535a278](https://github.com/p20030920p/SLAM_Learning/blob/535a2780af7ca7eb3aa02f722e1340fe90bc2dcf/docs/SCOPE.md) add paper-table checks and scoped semantic scoring. The other four papers are read context, not freshly reproduced baselines.
 
 The [SuperMap repository](https://github.com/superxslam/SuperMap), inspected at `ec95b1d50a458645b2669836e2c431f1e957bbc6`, contained a README, paper and teaser, but no implementation, requirements file or runnable examples. Its setup text is not a runnable recipe at this snapshot. PerSeM is treated as a recent preprint, not an independently validated baseline.
 
-A focused update interface permits author-code measurements and negative controls before the deadline. The next step is the same intervention in a real semantic frontend at matched coverage, rather than presenting synthetic object positions as semantic-navigation results.
+Next, finish the frozen room2 equal-cap analysis before proposing bounded replay. Earlier synthetic mechanisms are exploratory context. [Decision gate](PLAN.md).
diff --git a/docs/LITERATURE.zh-CN.md b/docs/LITERATURE.zh-CN.md
@@ -2,7 +2,7 @@
 
 [English](LITERATURE.md) | 中文
 
-文献截止 2026 年 10 月 7 日。“活跃方向”表示有多项近期代表工作，不是引用量热度排名。核心工作按地图更新决策选择，不能对不同任务直接排排行榜。
+文献截止 2026 年 10 月 7 日。“活跃方向”表示有多项近期代表工作，不是热度排名。四篇已复现，另外四篇用于约束解释与新颖性声明。
 
 ## 方向 1–2 的活跃主题
 
@@ -15,27 +15,25 @@
 
 [AnyLoc](https://arxiv.org/abs/2308.00688)、[Revisit Anything](https://arxiv.org/abs/2409.18049) 是相关视觉地点识别例子。召回依赖数据集与正样本定义，本库不沿用历史分数，未做新的 VPR 运行。
 
-## 八篇核心工作
+## 四篇复现工作与四篇背景文献
 
 | 工作 | 论文证据 | 结构推断 | 已有保护／边界 |
 | --- | --- | --- | --- |
 | [DUFOMap，2024](https://arxiv.org/html/2403.01449v1) | III-A/B：观测自由空间与传感器／位姿容差；IV-C：`d_p=1` | 使用给定扫描位姿清图，容差不等于共享残差协方差 | 已考虑位姿误差并比较位姿来源，不能说它“不管噪声” |
 | [BeautyMap，2024](https://arxiv.org/html/2405.07283v1) | III：二进制占据与射线修正／保护；表 I 用 HA | 配准后不一致可以来自位姿误差或环境变化 | 已保护可见性错误，承认地面层级局限 |
 | [ConceptGraphs，ICRA 2024](https://arxiv.org/html/2309.16650v1) | 几何重叠与语义相似度关联对象，然后融合点和特征 | 世界坐标对齐错误可能先造成实例碎片化或合并 | 语义帮助关联，不能据此说所有语义地图假设静态世界 |
-| [HOV-SG，RSS 2024](https://arxiv.org/html/2403.17846v2) | III-A：准确里程计投影／融合；IV-C：外部 LiDAR SLAM | 图和导航几何继承位姿接口质量 | 外部里程计／回环是保护，本库未实测其敏感性 |
+| [HOV-SG，RSS 2024](https://arxiv.org/html/2403.17846v2) | III-A：准确里程计投影／融合；IV-C：外部 LiDAR SLAM | 图和导航几何继承位姿接口质量 | 外部里程计／回环是保护；配对子集不代表迟到修正恢复 |
 | [Khronos，2024](https://arxiv.org/html/2402.13817v2) | IV：局部一致性；V-B：联合位姿／片段／背景优化；V-C：区分缺少证据 | 修正前的候选生成与局部一致性仍有影响 | **反例：** 已有联合优化、历史和可见性，不能归为固定姿态独立清图 |
 | [DovSG，2024 预印本／2025 论文](https://arxiv.org/html/2410.11989v2) | III-D：ACE、匹配和 ICP 重定位后局部更新 | 错误重定位可能影响关联与过期体素处理 | 明确支持动态局部更新，反驳“场景图都静态” |
 | [SuperMap，2026 年 8 月](https://arxiv.org/html/2608.22896v1) | III：条件于位姿；IV：深度观测状态与递归证据 | 展示的模型值得检查共享位姿协方差与校准 | 已区分可观测、不可观测、消失；未展示某公式不等于代码缺陷 |
 | [PerSeM，2026 年 9 月预印本](https://arxiv.org/html/2609.19542v2) | 3.1/3.2：长期世界体素；5：漂移／相关误差 | 标签稳定不保证几何证据分配正确 | 已有记忆和修正；相关几何是作者承认的局限 |
 
-共同接口只限于“先建立空间对应，再解释地图证据”。这些论文不是一个算法，多个工作已缓解相关问题。候选问题关注剩余共享几何不确定性与不可逆更新的置信度。首个语义子集已执行；实测共性仍需标注和位姿误差对照。
+共同接口只限于“先建立空间对应，再解释地图证据”，多个工作已有保护。当前问题检验迟到修正后的恢复；尚未证明共享不确定性是共同主导缺陷。[当前分析](STUDY.zh-CN.md)。
 
 ## 代码可得性与复现范围
 
-DUFOMap、BeautyMap 有可执行作者实现与小型公开标注数据，固定提交见 [upstreams.json](../configs/upstreams.json)。ConceptGraphs 的 class-agnostic 前端已处理 40 次提供位姿的 Replica 观测，HOV-SG 已完成 8 观测分段特征建图核心，见[边界](SEMANTIC.zh-CN.md)。其余四篇工作做了阅读，未新近复现。
+[四篇核心快照](papers/README.zh-CN.md)固定作者提交、输入与适配。[535a278 的后续作者原流程](https://github.com/p20030920p/SLAM_Learning/blob/535a2780af7ca7eb3aa02f722e1340fe90bc2dcf/docs/SCOPE.zh-CN.md)补充论文表格核查与限定语义评分。其余四篇作为已读背景，不计作新复现基线。
 
 [SuperMap 仓库](https://github.com/superxslam/SuperMap)在检查提交 `ec95b1d50a458645b2669836e2c431f1e957bbc6` 时，仅有 README、论文、演示，没有实现、依赖文件与可运行例程，安装文字不构成该快照下的复现配方。PerSeM 是近期预印本，不当作独立验证过的成熟基线。
 
-先复现可检查的接口和负面对照，再进入真实语义前端；不能把合成对象位置包装成语义导航结果。[阶段安排](PLAN.zh-CN.md)。
-
-本次主选四篇为 DUFOMap、BeautyMap、ConceptGraphs、HOV-SG；前两篇完整 teaser，后两篇语义核心子集。HOV-SG 已执行 8 观测分段特征建图，其他四篇只作已读比较与反例。[四篇研究分析](STUDY.zh-CN.md)、[实际范围索引](papers/README.zh-CN.md)。
+先完成冻结的 room2 相同候选上限分析，再决定有界重放。早期合成机制只作探索背景。[决策门槛](PLAN.zh-CN.md)。
diff --git a/docs/PAIRED_RESULTS.md b/docs/PAIRED_RESULTS.md
@@ -2,6 +2,8 @@
 
 English | [中文](PAIRED_RESULTS.zh-CN.md)
 
+> Historical measurements/analysis; current question and next decision: [STUDY](STUDY.md) · [PLAN](PLAN.md).
+
 **Decision: keep H1 as a candidate and narrow the claim.** The completed exploratory study tests downstream map decisions under controlled pose errors. It does not establish a universal failure caused specifically by shared uncertainty, and it does not validate a new method. Temporal ordering can help one map objective and hurt another; simple parameter changes already explain part of the apparent improvement opportunity.
 
 ## Evidence and scope

diff --git a/docs/PAIRED_RESULTS.zh-CN.md b/docs/PAIRED_RESULTS.zh-CN.md
@@ -2,6 +2,8 @@
 
 [English](PAIRED_RESULTS.md) | 中文
 
+> 历史测量／分析；当前问题与下一项决策统一见 [STUDY](STUDY.zh-CN.md) · [PLAN](PLAN.zh-CN.md)。
+
 **结论：保留候选 H1，收窄主张。** 本轮探索检查受控位姿误差对地图决策的影响，尚未证明“错误处理共享不确定性”是四篇工作的共同主导瓶颈，也未验证新方法。同一种时间结构可能帮助某个地图目标，却损害另一个目标；简单参数变化已经解释了部分看起来值得改进的空间。
 
 ## 证据与范围