Merge branch 'gpt' of git.simba-robotics.com:HelixCopex/mcm-26-b into gpt

2026-09-11 02:34:16 +08:00
parent 12da6aafd2 4f91f242fe
commit 846b45e420
19 changed files with 971 additions and 45 deletions
@@ -0,0 +1,60 @@
+# 修复计划：7 项 P2 问题 + 3 项规划调整
+
+已核实：7 个问题的定位与审核报告一致；另确认关联缺陷 `planner/__init__.py:392` 用 `>` 判断遮挡预算（实际放行 5 个被遮挡探点，注释口径为 4）。用户已确认：环探保留启发式+重写依据；误差场采用逐点白噪声哈希。
+
+## 修改一：模拟器固定误差场（问题1）
+文件 `src/mcm26b/sim/simulator.py`：
+- 新增 `_bearing_error(channel, pos)`：`sha256(f"{error_seed}|{channel}|{round(x,6)}|{round(y,6)}")` 前 8 字节映射到 `[-error_half_deg, +error_half_deg)`。同一（源频道, 位置）读数恒定，不同点独立。
+- `_observe`（L193）改为调用它；删除顺序随机流 `self._rng`（构造参数 `error_seed` 保留）。
+- 模块 docstring 更新：同位置误差固定；分布仍是测试假设。
+- 新增测试：同位置重复测量（不同 request_id）svd_deg 相同；同种子两次构造结果一致。
+
+## 修改二：Q4 对跖探点公式（问题2）
+文件 `src/mcm26b/planner/__init__.py`（L440-447）：
+- L447：`dx / norm * scale` → `dx * scale`（dy 同理），使偏移模长 = `min(norm, safe)`。圆心为原点、原探点 (200,0)、safe≥200 时返回 (-200,0)。
+- L444 退化分支：`min(safe, 1.0)` → `safe`（方向 +x），不再钳到 1m。
+- 新增单测：对跖点坐标（safe≥norm 与 safe<norm 两种）。
+
+## 修改三：环探依据重写 + 预算一致（问题3，保留启发式）
+- `planner/__init__.py` 模块 docstring、`DirectionalExtension` docstring、`_next_probe` docstring（L428-431）：删除"≥180° 弧"声称，改为诚实表述：源位于环上且朝向沿外切线时可见弧可退化为单点，固定 45° 步进无最坏命中保证；梯子只提高提前命中率，完备性完全由后备清除兜底。
+- `docs/协议规则表与状态机.md` §7（L166-172）同步重写。
+- L392 `>` → `>=`，`config.py:85` 注释与文档同步为"4 个被遮挡探点（选点/对跖/环探×2）"。
+- 新增 `scripts/benchmark_q4.py`：批量种子扫描（`--problem 4`，可配源数/定向数/种子范围/输出路径），逐局记录 measures、clears、后备清除格数与 attempts、virtual_time、completed，落盘 `output/q4/benchmark.json` 并打印汇总表——用于验证修复后的后备清除成本（审核"验证后备成本"要求），同时为论文问题4结果表提供可复现数字。
+
+## 修改四：通信失败可靠收尾（问题4）
+文件 `planner/__init__.py` `run()`（L195-205）：
+- `client.enter()` 纳入 try；except 顺序：`DeadlineExceeded` → `(TransportError, ProtocolError, InconsistentBelief)` → `Exception` 兜底；stop_reason 分别记 `real_budget_exhausted` / `transport_failure` / `protocol_error` / `inconsistent_belief` / `unexpected_error`，summary 增 `abort_detail`（异常类型+消息，截断）。
+- `/exit` 改为尽力而为：`client.exit()` 失败不再中断收尾；`_summary` 接受 `info=None`（`exit_reason="exit_unreachable"`，virtual_time 取 `client.state.virtual_time_s`）。
+- summary 增 `channel_status` 快照（每频道状态 + 候选 MEC 半径，约 20 条）——即"保存未完成状态"；`--log` 的 JSONL 本就有全过程留痕。
+- cli.py 无需改（`completed=false` 已映射退出码 1）。
+- 新增测试：注入连续三次传输失败（若 `ScriptedFaultTransport` 不支持持久失败则扩展其故障类型），断言 `run()` 正常返回 summary 不抛异常、有 /exit 尝试留痕、退出码路径为 1。
+
+## 修改五：Q1 近平行退化情形（问题5）
+文件 `src/mcm26b/geometry/localization.py`（L216-226）：
+- 衰退方向判定改为对浮点平面系数的精确有理数运算（`Fraction`）：候选方向仍枚举各平面法线的垂直方向 `(b,-a)/(-b,a)`；先用浮点快速预筛（max > 1e-9 直接排除，浮点误差远小于该阈值），近零候选再用 Fraction 精确检验 `∀(d,e,f): d·dir_x + e·dir_y ≤ 0`。平面系数本身是 float（精确有理数），故结论对量化输入精确，消除 2e-14 阈值与顶点过滤 eps=1e-9 的口径错配导致的漏判。
+- 顶点枚举与 `PARALLEL_EPS` 预筛保持现状。
+- 回归测试：审核复现输入 `[(0,0,1°), (0,0.01,358.99999999999°)]` → `status=unbounded`、`diameter_infinite=true`、不给清除保证；`scripts/benchmark_q1.py` 两档结果不回归。
+
+## 修改六：默认舍入裕量统一（问题6）
+- `config.py:75`：`bearing_rounding_margin_deg` 默认 `0.0` → `0.005`，注释改为"覆盖服务器两位小数舍入偏差（≤0.005°）；库默认与 CLI 默认一致"。
+- `q1` 子命令与 `geometry/__main__.py` 的 `--rounding-margin-deg` 默认保持 0（离线敏感性分析语义不变，审核未要求改）。
+- 新增断言 `SearchConfig().bearing_rounding_margin_deg == 0.005`，并用默认配置复跑"边缘读数保留真源"场景；检查显式传参的现有测试不受影响；README 如有默认值描述一并更新。
+
+## 修改七：后验评分补 near 分支（问题7）
+文件 `planner/__init__.py` `_worst_posterior_radius`（L93-110）：
+- 增加 near 分支：`probe.clip_disk(q, field.near_radius_m, search.polygon_sides)`，多边形非空则纳入 max；direction 端点分支保留，docstring 说明 safe 盘内 no_signal 不可能故不枚举。
+- 调用点（粗评/加密复评两处）传入 `near_radius_m` 与 `polygon_sides`。
+- 新增测试：候选集 `[(-4,0),(4,0)]`（两点退化区域）、q=原点 → 最坏后验 ≈ 4m（修复前报 2m）；`test_loop_q3/q4` 场景回归（选点可能变化，但断言均为清除状态类，预期无碍）。
+
+## 规划调整落地
+- **A（第二问定位）**：`docs/algo-2.md` §13（L231 附近）明确交付为"安全选点的启发式基线"：解析安全盘证书已实现并证明；§10 区间上界未实现，采样最大值仅作粗筛，不得称严格最坏上界。
+- **B（第三问）**：无代码改动；顺序约束（先修改一/四，再官方演练）在 `docs/B题建模分析与实施计划.md` 演练清单处注明。
+- **C（第四问预算口径）**：`_summary` 增 `total_requests`（measures+clears）；`DirectionalExtension._summary` 增 `fallback_clear_attempts`（Σ attempts）；`docs/协议规则表与状态机.md` §7 后备清除段注明"后备 /clear 请求计入总请求与现实预算"。
+
+## 测试与验收
+- 每项修改配单测（新增约 10-12 项），全套 pytest 通过（95 → ~107 项）。
+- 冒烟：`uv run mcm26b sim-check`；`uv run mcm26b run --problem 3/4 --self-test` 若干种子。
+- 运行 `scripts/benchmark_q4.py` 记录修复后后备成本，输出落盘。
+
+## 执行顺序
+修改一（模拟器）→ 修改四（收尾）→ 修改六 → 修改七 → 修改二 → 修改三 → 修改五 → 文档同步 → 全量回归。
@@ -21,8 +21,8 @@ src/mcm26b/
 ├── planner/       # P3 保证接收盘选点器 / P4 全向闭环 / P5 混合源定向闭环（已完工）
 ├── evaluation/    # 日志复算与统计（已完工）
 └── sim/scenes.py  # 自检场景生成器：全向与混合源（测试端专用）
-tests/             # pytest（兼容 unittest 风格），95 项
-scripts/           # 采集、基准、可视化与校核脚本
+tests/             # pytest（兼容 unittest 风格），105 项
+scripts/           # 采集、基准、可视化与校核脚本（含问题四批量仿真 benchmark_q4.py）
 output/ examples/  # 演练证据、算例与生成产物
 ```

@@ -215,11 +215,22 @@ P0–P5（自建部分）与 P2 均已交付：协议层（串行幂等客户端
 全向闭环与 `DirectionalExtension` 混合源定向闭环（49 点覆盖对定向源同样成立、
 no_signal 联合约束基线不删候选、对跖+环探遮挡梯子、`fallback_clear_cells`
 有限圆盘覆盖后备清除，格数与现实预算双硬顶下不盲扫）。自建验收：
-`uv run pytest`（95 项）与 `uv run mcm26b sim-check`；随机场景自检：
+`uv run pytest`（105 项）与 `uv run mcm26b sim-check`；随机场景自检：
 `uv run mcm26b run --robot-id <队号> --self-test`（问题4加 `--problem 4`）。

+评审修订（2026-09-11）：本地模拟器改为逐点固定误差场（同一"源×位置"读数
+恒定，与官方演练实测一致）；对跖复核探点公式修正（此前塌缩到 MEC 圆心 1m
+内，浪费探点预算并推高后备清除）；环探"≥180° 可见弧"依据删除（最坏可见弧
+可退化为单点，梯子只是启发式，完备性由后备清除兜底）；断连/协议错误中断后
+仍尽力 /exit 并产出未完成 summary；库默认舍入裕量统一为 0.005°；后验评分
+补 near 分支；问题一近平行细长区域保留远端顶点。后备清除成本用
+`scripts/benchmark_q4.py` 批量种子扫描量化（结果落盘 `output/q4/`）。
+
 下一步优先把两个闭环接入官方演练（问题3用全向、问题4用混合，先实测单请求
 延迟与总请求数，核对 1200s 现实预算与 2MB 日志体积），再做 P6 消融对照。
+顺序约束：上述模拟器与故障收尾修复先行合入，官方演练在修复后的版本上执行；
+总请求口径必须计入后备清除（summary 的 `total_requests` 与
+`fallback_clear_attempts`），仅统计检测次数会低估现实耗时。
 基线量级供参照：10 源场景约 550–630 次检测、虚拟总时 1.4–1.7 万秒；16 源
 场景约 200–320 次检测即触发数量上界停止。该数字说明共享测点与跳点策略
 （P6）有明确改进空间。不要重复开发几何内核或提前堆叠学习算法；以官方演练
@@ -228,7 +228,7 @@

 性能层应在相同自建源位置与确定的空间误差场上，比较轴向点、两个横向点、有限候选最坏后验策略。误差场在同一位置保持固定；不得让每次调用重新独立生成误差。记录第二次测量类型、后验直径和包围半径、额外移动、计算耗时及是否达到20米清除阈值。至少包含误差边界、源域边界、接收半径1000米、近距离和近共线案例。没有这些实验前，不报告未经验证的性能提升比例。

-目前 [planning.py](C:/Users/HelixCopex/code/MATH/geometry/planning.py) 中的 `first_bearing_candidate_disk` 已实现第5至6节的解析证书；问题一内核可用于纯半平面区域的顶点和最小圆计算。第9至11节描述的是完整选点器的数学方案，包括本次补充的保守区间评估方法，尚未在现有仓库中实现为完整第二问优化器，也没有相应正式成绩。该实现边界不影响解析候选区域的证明。
+目前 [planning.py](C:/Users/HelixCopex/code/MATH/geometry/planning.py) 中的 `first_bearing_candidate_disk` 已实现第5至6节的解析证书；问题一内核可用于纯半平面区域的顶点和最小圆计算。**本仓库第二问交付的定位是"安全选点的启发式基线"**：解析安全候选盘证书已实现并证明（选中点必能收到信号），而选点评分采用边界采样点 × 读数区间端点的有限枚举近似（含 near 分支）。第9至11节描述的完整选点器数学方案，特别是第10节覆盖全部未来读数的保守区间上界，尚未在现有仓库中实现；采样最大值只用于候选点之间的相对比较与粗筛，不得当作严格最坏情形上界报告，也没有相应正式成绩。该实现边界不影响解析候选区域的证明。

 对后续实现而言，应新增接收任意半平面约束的交集入口，或使用独立的多边形裁剪函数，再复用第一问的最小包围圆。不能把圆盘先验伪装成额外测向观测传入现有 `solve(measurements)` 接口。

@@ -165,11 +165,13 @@ no_signal/near/direction（direction 附 svd_deg），`/clear` 仍只看 20m 半
  （`belief.update_measure` 对 no_signal 无操作）；缺席证明仍只由完整覆盖
  给出。唯一启用的推理：若 no_signal 发生在保证接收盘 B(mec, 1000-ρ) 内
  （该盘内任一全向源必然可收），则只能由朝向遮挡解释 [观测推理]——该频道
-  记为"观测到定向"，进入探点升级梯子：对 MEC 圆心对跖复核一次，再绕
-  MEC、半径 = ρ 的圆环按 45° 步进探点（环包围任意候选源，发射半平面在
-  环上截出 ≥180° 弧）。被遮挡探点不消耗定位步数，单轮由
-  `max_blocked_probes`（4 [工程]）限流；梯子不构成接收保证，失败由后备
-  清除兜底；
+  记为"观测到定向"，进入探点升级梯子：对 MEC 圆心取对跖方向复核（偏移
+  裁进保证接收盘），再绕 MEC、半径 = ρ 的圆环按 45° 步进探点。**几何边界
+  必须如实陈述**：环上可见弧没有下界——源可位于环上（|g−c| = ρ）且朝向
+  沿外切线，此时可见集退化为单个切点，固定步进环探可以全部漏过。梯子
+  只提高提前命中概率，不构成接收或命中保证；完备性完全由后备清除兜底。
+  被遮挡探点不消耗定位步数，单轮被遮挡探点（首次选点、对跖复核、环探
+  合计）达 `max_blocked_probes`（4 [工程]）即转后备清除；
 - **后备清除的协议用法**：定位停滞（候选集有界但未达 §6 清除证书）时，
  对候选外包多边形构造 20m 有限圆盘覆盖（`fallback_clear_cells`：全局
  半开网格索引、相交判定带正容差，格内任一点距某格心 ≤ 10√2 < 20m），
@@ -180,12 +182,18 @@ no_signal/near/direction（direction 附 svd_deg），`/clear` 仍只看 20m 半
    (⌈w/20⌉+2)(⌈h/20⌉+2) 做规模检查，宽候选集（如仅一次读数的 750m
    截断角域，数千格）必须继续定位而非硬扫；
  - 现实预算：N × `real_rtt_p95_s` + `reserve_exit_s` ≤ 剩余现实时间。
-  虚拟耗时上界 ≈ 移动 L/5 + 3N（全失败）或 3(N-1)+5（末次成功）；
+  虚拟耗时上界 ≈ 移动 L/5 + 3N（全失败）或 3(N-1)+5（末次成功）。
+  后备 /clear 请求**计入总请求口径**：summary 的 `total_requests`
+  （measures+clears）与 `fallback_clear_attempts`（后备逐格尝试数）
+  如实上报——仅统计检测次数会明显低估现实耗时；
 - **真源不误删**：no_signal、失败的 `/clear`（含后备逐格全失败）均不删
  空间候选；后备全失败按守恒性破坏记异常（格心距格内任一点 < 20m，
  理论上必命中）；
 - **停止条件与预算语义不变**：仍是清除数达 16、每频道均"已清除或覆盖
-  证明不存在"、现实预算耗尽（记未证明完成）三种；
+  证明不存在"、现实预算耗尽（记未证明完成）三种；断连/协议错误/
+  守恒性破坏中断时同样尽力 /exit 并产出未完成 summary（`stop_reason`
+  记 `transport_failure`/`protocol_error`/`inconsistent_belief` 等，
+  附逐频道状态快照），不静默丢失会话；
 - **验收**：`tests/test_loop_q4.py`（15 项：覆盖格无空洞与格线两侧保留、
  退化多边形、格数上界公式；混合 10/16 源全清、朝向背离圆心的近边界源
  不漏、批量种子回归；no_signal 不删候选；格数硬顶与预算跳过；后备
@@ -0,0 +1,288 @@
+{
+  "summary": {
+    "config": {
+      "sources": 10,
+      "directional": 3,
+      "bearing_margin_deg": 0.005
+    },
+    "games": 16,
+    "completed": 16,
+    "all_cleared": 16,
+    "measures_total": 9453,
+    "clears_total": 471,
+    "total_requests_total": 9924,
+    "fallback_attempts_total": 316,
+    "fallback_attempts_max": 165,
+    "virtual_time_s_max": 18756.107,
+    "mismatch_total": 0,
+    "anomaly_total": 0
+  },
+  "games": [
+    {
+      "seed": 0,
+      "sources": 10,
+      "directional": 3,
+      "completed": true,
+      "stop_reason": "all_resolved",
+      "all_cleared": true,
+      "measures": 623,
+      "clears": 10,
+      "total_requests": 633,
+      "fallback_cells": [],
+      "fallback_attempts": 0,
+      "virtual_time_s": 17147.133,
+      "virtual_recompute_mismatches": 0,
+      "anomaly_count": 0
+    },
+    {
+      "seed": 1,
+      "sources": 10,
+      "directional": 3,
+      "completed": true,
+      "stop_reason": "all_resolved",
+      "all_cleared": true,
+      "measures": 591,
+      "clears": 10,
+      "total_requests": 601,
+      "fallback_cells": [],
+      "fallback_attempts": 0,
+      "virtual_time_s": 16383.06,
+      "virtual_recompute_mismatches": 0,
+      "anomaly_count": 0
+    },
+    {
+      "seed": 2,
+      "sources": 10,
+      "directional": 3,
+      "completed": true,
+      "stop_reason": "all_resolved",
+      "all_cleared": true,
+      "measures": 580,
+      "clears": 10,
+      "total_requests": 590,
+      "fallback_cells": [],
+      "fallback_attempts": 0,
+      "virtual_time_s": 16380.242,
+      "virtual_recompute_mismatches": 0,
+      "anomaly_count": 0
+    },
+    {
+      "seed": 3,
+      "sources": 10,
+      "directional": 3,
+      "completed": true,
+      "stop_reason": "all_resolved",
+      "all_cleared": true,
+      "measures": 559,
+      "clears": 10,
+      "total_requests": 569,
+      "fallback_cells": [],
+      "fallback_attempts": 0,
+      "virtual_time_s": 14575.032,
+      "virtual_recompute_mismatches": 0,
+      "anomaly_count": 0
+    },
+    {
+      "seed": 4,
+      "sources": 10,
+      "directional": 3,
+      "completed": true,
+      "stop_reason": "all_resolved",
+      "all_cleared": true,
+      "measures": 576,
+      "clears": 10,
+      "total_requests": 586,
+      "fallback_cells": [],
+      "fallback_attempts": 0,
+      "virtual_time_s": 16128.325,
+      "virtual_recompute_mismatches": 0,
+      "anomaly_count": 0
+    },
+    {
+      "seed": 5,
+      "sources": 10,
+      "directional": 3,
+      "completed": true,
+      "stop_reason": "all_resolved",
+      "all_cleared": true,
+      "measures": 609,
+      "clears": 10,
+      "total_requests": 619,
+      "fallback_cells": [],
+      "fallback_attempts": 0,
+      "virtual_time_s": 15963.724,
+      "virtual_recompute_mismatches": 0,
+      "anomaly_count": 0
+    },
+    {
+      "seed": 6,
+      "sources": 10,
+      "directional": 3,
+      "completed": true,
+      "stop_reason": "all_resolved",
+      "all_cleared": true,
+      "measures": 624,
+      "clears": 10,
+      "total_requests": 634,
+      "fallback_cells": [],
+      "fallback_attempts": 0,
+      "virtual_time_s": 16364.634,
+      "virtual_recompute_mismatches": 0,
+      "anomaly_count": 0
+    },
+    {
+      "seed": 7,
+      "sources": 10,
+      "directional": 3,
+      "completed": true,
+      "stop_reason": "all_resolved",
+      "all_cleared": true,
+      "measures": 569,
+      "clears": 10,
+      "total_requests": 579,
+      "fallback_cells": [],
+      "fallback_attempts": 0,
+      "virtual_time_s": 17010.511,
+      "virtual_recompute_mismatches": 0,
+      "anomaly_count": 0
+    },
+    {
+      "seed": 8,
+      "sources": 10,
+      "directional": 3,
+      "completed": true,
+      "stop_reason": "all_resolved",
+      "all_cleared": true,
+      "measures": 599,
+      "clears": 12,
+      "total_requests": 611,
+      "fallback_cells": [
+        5
+      ],
+      "fallback_attempts": 3,
+      "virtual_time_s": 18756.107,
+      "virtual_recompute_mismatches": 0,
+      "anomaly_count": 0
+    },
+    {
+      "seed": 9,
+      "sources": 10,
+      "directional": 3,
+      "completed": true,
+      "stop_reason": "all_resolved",
+      "all_cleared": true,
+      "measures": 671,
+      "clears": 13,
+      "total_requests": 684,
+      "fallback_cells": [
+        5
+      ],
+      "fallback_attempts": 4,
+      "virtual_time_s": 17391.766,
+      "virtual_recompute_mismatches": 0,
+      "anomaly_count": 0
+    },
+    {
+      "seed": 10,
+      "sources": 10,
+      "directional": 3,
+      "completed": true,
+      "stop_reason": "all_resolved",
+      "all_cleared": true,
+      "measures": 613,
+      "clears": 10,
+      "total_requests": 623,
+      "fallback_cells": [],
+      "fallback_attempts": 0,
+      "virtual_time_s": 17981.333,
+      "virtual_recompute_mismatches": 0,
+      "anomaly_count": 0
+    },
+    {
+      "seed": 11,
+      "sources": 10,
+      "directional": 3,
+      "completed": true,
+      "stop_reason": "all_resolved",
+      "all_cleared": true,
+      "measures": 552,
+      "clears": 149,
+      "total_requests": 701,
+      "fallback_cells": [
+        142
+      ],
+      "fallback_attempts": 140,
+      "virtual_time_s": 17099.436,
+      "virtual_recompute_mismatches": 0,
+      "anomaly_count": 0
+    },
+    {
+      "seed": 12,
+      "sources": 10,
+      "directional": 3,
+      "completed": true,
+      "stop_reason": "all_resolved",
+      "all_cleared": true,
+      "measures": 543,
+      "clears": 174,
+      "total_requests": 717,
+      "fallback_cells": [
+        168
+      ],
+      "fallback_attempts": 165,
+      "virtual_time_s": 17802.514,
+      "virtual_recompute_mismatches": 0,
+      "anomaly_count": 0
+    },
+    {
+      "seed": 13,
+      "sources": 10,
+      "directional": 3,
+      "completed": true,
+      "stop_reason": "all_resolved",
+      "all_cleared": true,
+      "measures": 579,
+      "clears": 10,
+      "total_requests": 589,
+      "fallback_cells": [],
+      "fallback_attempts": 0,
+      "virtual_time_s": 16392.501,
+      "virtual_recompute_mismatches": 0,
+      "anomaly_count": 0
+    },
+    {
+      "seed": 14,
+      "sources": 10,
+      "directional": 3,
+      "completed": true,
+      "stop_reason": "all_resolved",
+      "all_cleared": true,
+      "measures": 587,
+      "clears": 13,
+      "total_requests": 600,
+      "fallback_cells": [
+        5
+      ],
+      "fallback_attempts": 4,
+      "virtual_time_s": 16565.569,
+      "virtual_recompute_mismatches": 0,
+      "anomaly_count": 0
+    },
+    {
+      "seed": 15,
+      "sources": 10,
+      "directional": 3,
+      "completed": true,
+      "stop_reason": "all_resolved",
+      "all_cleared": true,
+      "measures": 578,
+      "clears": 10,
+      "total_requests": 588,
+      "fallback_cells": [],
+      "fallback_attempts": 0,
+      "virtual_time_s": 15288.608,
+      "virtual_recompute_mismatches": 0,
+      "anomaly_count": 0
+    }
+  ]
+}
@@ -0,0 +1,195 @@
+{
+  "summary": {
+    "config": {
+      "sources": 16,
+      "directional": 6,
+      "bearing_margin_deg": 0.005
+    },
+    "games": 10,
+    "completed": 10,
+    "all_cleared": 10,
+    "measures_total": 3222,
+    "clears_total": 1024,
+    "total_requests_total": 4246,
+    "fallback_attempts_total": 872,
+    "fallback_attempts_max": 199,
+    "virtual_time_s_max": 18879.215,
+    "mismatch_total": 0,
+    "anomaly_total": 0
+  },
+  "games": [
+    {
+      "seed": 0,
+      "sources": 16,
+      "directional": 6,
+      "completed": true,
+      "stop_reason": "sixteen_cleared",
+      "all_cleared": true,
+      "measures": 375,
+      "clears": 16,
+      "total_requests": 391,
+      "fallback_cells": [],
+      "fallback_attempts": 0,
+      "virtual_time_s": 17651.199,
+      "virtual_recompute_mismatches": 0,
+      "anomaly_count": 0
+    },
+    {
+      "seed": 1,
+      "sources": 16,
+      "directional": 6,
+      "completed": true,
+      "stop_reason": "sixteen_cleared",
+      "all_cleared": true,
+      "measures": 356,
+      "clears": 16,
+      "total_requests": 372,
+      "fallback_cells": [],
+      "fallback_attempts": 0,
+      "virtual_time_s": 17003.576,
+      "virtual_recompute_mismatches": 0,
+      "anomaly_count": 0
+    },
+    {
+      "seed": 2,
+      "sources": 16,
+      "directional": 6,
+      "completed": true,
+      "stop_reason": "sixteen_cleared",
+      "all_cleared": true,
+      "measures": 297,
+      "clears": 214,
+      "total_requests": 511,
+      "fallback_cells": [
+        204
+      ],
+      "fallback_attempts": 199,
+      "virtual_time_s": 18137.964,
+      "virtual_recompute_mismatches": 0,
+      "anomaly_count": 0
+    },
+    {
+      "seed": 3,
+      "sources": 16,
+      "directional": 6,
+      "completed": true,
+      "stop_reason": "sixteen_cleared",
+      "all_cleared": true,
+      "measures": 346,
+      "clears": 16,
+      "total_requests": 362,
+      "fallback_cells": [],
+      "fallback_attempts": 0,
+      "virtual_time_s": 16885.059,
+      "virtual_recompute_mismatches": 0,
+      "anomaly_count": 0
+    },
+    {
+      "seed": 4,
+      "sources": 16,
+      "directional": 6,
+      "completed": true,
+      "stop_reason": "sixteen_cleared",
+      "all_cleared": true,
+      "measures": 372,
+      "clears": 177,
+      "total_requests": 549,
+      "fallback_cells": [
+        201,
+        172
+      ],
+      "fallback_attempts": 163,
+      "virtual_time_s": 18525.209,
+      "virtual_recompute_mismatches": 0,
+      "anomaly_count": 0
+    },
+    {
+      "seed": 5,
+      "sources": 16,
+      "directional": 6,
+      "completed": true,
+      "stop_reason": "sixteen_cleared",
+      "all_cleared": true,
+      "measures": 330,
+      "clears": 145,
+      "total_requests": 475,
+      "fallback_cells": [
+        130,
+        6
+      ],
+      "fallback_attempts": 131,
+      "virtual_time_s": 18089.132,
+      "virtual_recompute_mismatches": 0,
+      "anomaly_count": 0
+    },
+    {
+      "seed": 6,
+      "sources": 16,
+      "directional": 6,
+      "completed": true,
+      "stop_reason": "sixteen_cleared",
+      "all_cleared": true,
+      "measures": 187,
+      "clears": 209,
+      "total_requests": 396,
+      "fallback_cells": [
+        207
+      ],
+      "fallback_attempts": 194,
+      "virtual_time_s": 15131.949,
+      "virtual_recompute_mismatches": 0,
+      "anomaly_count": 0
+    },
+    {
+      "seed": 7,
+      "sources": 16,
+      "directional": 6,
+      "completed": true,
+      "stop_reason": "sixteen_cleared",
+      "all_cleared": true,
+      "measures": 258,
+      "clears": 199,
+      "total_requests": 457,
+      "fallback_cells": [
+        188,
+        6
+      ],
+      "fallback_attempts": 185,
+      "virtual_time_s": 18879.215,
+      "virtual_recompute_mismatches": 0,
+      "anomaly_count": 0
+    },
+    {
+      "seed": 8,
+      "sources": 16,
+      "directional": 6,
+      "completed": true,
+      "stop_reason": "sixteen_cleared",
+      "all_cleared": true,
+      "measures": 325,
+      "clears": 16,
+      "total_requests": 341,
+      "fallback_cells": [],
+      "fallback_attempts": 0,
+      "virtual_time_s": 16301.93,
+      "virtual_recompute_mismatches": 0,
+      "anomaly_count": 0
+    },
+    {
+      "seed": 9,
+      "sources": 16,
+      "directional": 6,
+      "completed": true,
+      "stop_reason": "sixteen_cleared",
+      "all_cleared": true,
+      "measures": 376,
+      "clears": 16,
+      "total_requests": 392,
+      "fallback_cells": [],
+      "fallback_attempts": 0,
+      "virtual_time_s": 15287.399,
+      "virtual_recompute_mismatches": 0,
+      "anomaly_count": 0
+    }
+  ]
+}
@@ -0,0 +1,124 @@
+"""问题四批量仿真基准：固定误差场下统计后备清除成本与总请求量。
+
+逐局在进程内驱动 DirectionalExtension（本地模拟器；示向度误差场逐点
+固定，与官方"同一地点误差不变"口径一致），记录 completed、measures、
+clears、total_requests、fallback_clear_attempts、virtual_time_s 等指标，
+落盘 JSON 并打印汇总表。用于：
+- 验证后备清除成本（对跖探点修正后梯子命中率的现实校核）；
+- 为论文问题四结果表提供可复现数字（种子、配置、逐局指标全部留档）。
+
+用法示例：
+    uv run python scripts/benchmark_q4.py --seeds 0-15 --sources 10 --directional 3
+    uv run python scripts/benchmark_q4.py --seeds 20260910 --sources 16 --directional 6
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+from dataclasses import replace
+from pathlib import Path
+
+from mcm26b.config import ProblemConfig
+from mcm26b.evaluation import recompute_virtual_time
+from mcm26b.planner import DirectionalExtension
+from mcm26b.protocol.client import SimulatorClient
+from mcm26b.sim import LocalSimTransport, LocalSimulator
+from mcm26b.sim.scenes import mixed_scene
+
+ROOT = Path(__file__).resolve().parents[1]
+
+
+def parse_seeds(spec: str) -> list[int]:
+    """种子列表："0-15" 区间、"0,5,9" 枚举或单个值。"""
+    seeds: list[int] = []
+    for part in spec.split(","):
+        part = part.strip()
+        if "-" in part:
+            lo, hi = part.split("-", 1)
+            seeds.extend(range(int(lo), int(hi) + 1))
+        else:
+            seeds.append(int(part))
+    return seeds
+
+
+def run_one(seed: int, count: int, directional: int, margin: float) -> dict:
+    config = replace(ProblemConfig(),
+                     search=replace(ProblemConfig().search,
+                                    bearing_rounding_margin_deg=margin))
+    sources = mixed_scene(seed=seed, count=count, directional=directional)
+    simulator = LocalSimulator(sources, robot_id="BENCH", config=config)
+    client = SimulatorClient("BENCH", config=config,
+                             transport=LocalSimTransport(simulator))
+    loop = DirectionalExtension(client, config)
+    summary = loop.run()
+    report = recompute_virtual_time(client.log.action_events())
+    truth = sorted(s.channel for s in sources)
+    return {
+        "seed": seed,
+        "sources": count,
+        "directional": directional,
+        "completed": summary["completed"],
+        "stop_reason": summary["stop_reason"],
+        "all_cleared": sorted(summary["cleared_channels"]) == truth,
+        "measures": summary["measures"],
+        "clears": summary["clears"],
+        "total_requests": summary["total_requests"],
+        "fallback_cells": [entry["cells"] for entry in summary["fallback_clears"]],
+        "fallback_attempts": summary["fallback_clear_attempts"],
+        "virtual_time_s": summary["virtual_time_s"],
+        "virtual_recompute_mismatches": len(report["mismatches"]),
+        "anomaly_count": len(summary["anomalies"]),
+    }
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description=__doc__.splitlines()[0])
+    parser.add_argument("--seeds", default="0-15",
+                        help="种子，如 0-15 / 0,5,9 / 20260910")
+    parser.add_argument("--sources", type=int, default=10)
+    parser.add_argument("--directional", type=int, default=3)
+    parser.add_argument("--bearing-margin-deg", type=float, default=0.005)
+    parser.add_argument("--out", default="output/q4/benchmark.json")
+    args = parser.parse_args()
+
+    rows = [run_one(seed, args.sources, args.directional, args.bearing_margin_deg)
+            for seed in parse_seeds(args.seeds)]
+    completed = sum(1 for row in rows if row["completed"])
+    cleared = sum(1 for row in rows if row["all_cleared"])
+    summary = {
+        "config": {"sources": args.sources, "directional": args.directional,
+                   "bearing_margin_deg": args.bearing_margin_deg},
+        "games": len(rows),
+        "completed": completed,
+        "all_cleared": cleared,
+        "measures_total": sum(row["measures"] for row in rows),
+        "clears_total": sum(row["clears"] for row in rows),
+        "total_requests_total": sum(row["total_requests"] for row in rows),
+        "fallback_attempts_total": sum(row["fallback_attempts"] for row in rows),
+        "fallback_attempts_max": max((row["fallback_attempts"] for row in rows),
+                                     default=0),
+        "virtual_time_s_max": max((row["virtual_time_s"] for row in rows),
+                                  default=0.0),
+        "mismatch_total": sum(row["virtual_recompute_mismatches"] for row in rows),
+        "anomaly_total": sum(row["anomaly_count"] for row in rows),
+    }
+    payload = {"summary": summary, "games": rows}
+    out_path = ROOT / args.out
+    out_path.parent.mkdir(parents=True, exist_ok=True)
+    out_path.write_text(json.dumps(payload, ensure_ascii=False, indent=2) + "\n",
+                        encoding="utf-8")
+    print(json.dumps(summary, ensure_ascii=False, indent=2))
+    header = ("seed", "done", "clear", "measures", "clears", "requests",
+              "fb_attempts", "virtual_s")
+    print("\t".join(header))
+    for row in rows:
+        print("\t".join(str(row[key]) for key in
+                        ("seed", "completed", "all_cleared", "measures", "clears",
+                         "total_requests", "fallback_attempts", "virtual_time_s")))
+    print(f"\n结果已写入 {out_path}")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
@@ -72,7 +72,9 @@ class SearchConfig:
    grid_step_m: float = 700.0        # P4 覆盖网格步长；单元对角 700*sqrt(2)≈989.95<=1000
    grid_half_cells: int = 3          # i,j ∈ [-3,3]，49 点覆盖 [-2100,2100]^2 ⊇ Ω
    fallback_clear_grid_m: float = 20.0  # P5 后备清除网格；10*sqrt(2)<=20
-    bearing_rounding_margin_deg: float = 0.0  # 默认不加舍入裕量；敏感性分析时显式设 0.005
+    # 舍入裕量：服务器对示向度保留两位小数，引入 <=0.005° 偏差 [题面]；
+    # 不加裕量时合法真源可能被角域裁掉，故库默认与 CLI run 默认一致取 0.005 [工程]。
+    bearing_rounding_margin_deg: float = 0.005
    # ---- P3/P4 在线定位与闭环参数 [工程] ----
    polygon_sides: int = 192        # 圆盘外切多边形边数；Ω 近似误差 1800(1-cos(π/192))≈0.24m
    witness_spacing_m: float = 60.0   # 最坏后验评估的边界采样间距；选定后再按 1/4 加密复核
@@ -82,7 +84,8 @@ class SearchConfig:
    max_localize_steps: int = 8     # 单频道定位阶段允许的追加测量上限
    max_anomalies_per_channel: int = 3  # 单频道异常记录上限，防止异常刷屏
    # ---- P5 定向扩展参数 [工程] ----
-    max_blocked_probes: int = 4     # 定位单轮内被朝向遮挡的探点预算（对跖复核+环探）
+    max_blocked_probes: int = 4     # 定位单轮内被朝向遮挡的探点预算
+                                    # （首次选点+对跖复核+2 次环探；达限转后备清除）
    fallback_max_cells: int = 400   # 后备清除格数硬顶；超过则不盲扫（宽候选集继续定位）


@@ -2,6 +2,10 @@

 All calculations are floating point with normalized station coordinates.
 The reference algorithm deliberately favors auditability over asymptotic speed.
+Near-parallel plane pairs are intersected instead of skipped (only exactly
+parallel pairs are discarded), and the feasibility tolerance scales with the
+candidate vertex magnitude, so legitimate far-away vertices of long thin
+regions are never silently dropped.
 """

 from dataclasses import dataclass
@@ -180,6 +184,13 @@ def solve(measurements: list[Measurement], eps: float = EPS) -> dict:
    so they possess a vertex, even if unbounded. Feasibility follows from vertex
    enumeration; unboundedness follows from a nonzero recession direction.
    Empty diameter is null, unbounded diameter is null with diameter_infinite=true.
+
+    Near-parallel plane pairs are intersected rather than skipped: their
+    intersection can be a legitimate vertex of a very long thin region
+    (e.g. stations 1cm apart with bearings 1 and 358.99999999999 deg give a
+    vertex ~5.7e10 m away). Only exactly-parallel pairs (det == 0) have no
+    intersection. The feasibility tolerance scales with the vertex magnitude,
+    since absolute residuals grow with coordinates.
    """
    if not isfinite(eps) or eps <= 0:
        raise ValueError("eps必须为有限正数")
@@ -197,15 +208,16 @@ def solve(measurements: list[Measurement], eps: float = EPS) -> dict:
    planes = [h for m in local for h in halfplanes(m)]

    def inside(p):
-        return all(a * p[0] + b * p[1] <= c + eps for a, b, c in planes)
+        tol = eps * max(1.0, abs(p[0]), abs(p[1]))
+        return all(a * p[0] + b * p[1] <= c + tol for a, b, c in planes)

    feasible_vertices = []
    for (a, b, c), (d, e, f) in combinations(planes, 2):
        det = a * e - b * d
-        if abs(det) <= PARALLEL_EPS:
-            continue
+        if det == 0.0:
+            continue   # 精确平行：无交点（近平行对的交点是合法远端顶点，保留）
        p = ((c * e - b * f) / det, (a * f - c * d) / det)
-        if inside(p):
+        if isfinite(p[0]) and isfinite(p[1]) and inside(p):
            feasible_vertices.append(p)

    if not feasible_vertices:
@@ -14,6 +14,8 @@ P4 全向闭环（验收：10/16 源正确退出；边界朝外不漏；停止
 （清除数达 16，或每频道均有清除成功/不存在证明）→ /exit。
 发现 10 源、连续无信号、候选采样为空都不是停止条件；
 现实预算耗尽退出记为未证明完成（completed=false），不与成功混记。
+断连/协议错误/守恒性破坏中断时同样尽力 /exit 并产出未完成 summary
+（stop_reason 记中断类别，附逐频道状态快照），不静默丢失会话。

 P5 定向扩展（问题4混合源闭环）：
 - 49 点覆盖的发现保证对定向源同样成立（源在单元凸包内，任意发射半平面
@@ -31,13 +33,15 @@ import math
 from dataclasses import dataclass

 from ..belief import (BeliefState, CandidateRegion, ChannelStatus,
-                      clip_halfplane)
+                      InconsistentBelief, clip_halfplane)
 from ..config import ProblemConfig, Vec2
 from ..coverage import CoverageLedger
 from ..evaluation import recompute_virtual_time, summarize
-from ..protocol.client import DeadlineExceeded, SimulatorClient
+from ..protocol.client import (DeadlineExceeded, ProtocolError,
+                               SimulatorClient)
 from ..protocol.rules import CLEAR_SUCCESS, DIRECTION, NEAR, NO_SIGNAL
 from ..protocol.timing import move_time
+from ..protocol.transport import TransportError


 def bearing_deg(a: Vec2, b: Vec2) -> float:
@@ -45,6 +49,11 @@ def bearing_deg(a: Vec2, b: Vec2) -> float:
    return math.degrees(math.atan2(b[1] - a[1], b[0] - a[0])) % 360.0


+def _error_brief(exc: BaseException) -> dict:
+    """异常摘要（summary 留档用）：类型名 + 截断消息。"""
+    return {"type": type(exc).__name__, "message": str(exc)[:500]}
+
+
@dataclass(frozen=True)
 class SecondPointPlan:
    """下一检测点的候选与依据（P3 输出）。"""
@@ -91,11 +100,14 @@ def candidate_points(center: Vec2, safe_radius_m: float, rings, angles: int) ->


 def _worst_posterior_radius(region: CandidateRegion, q: Vec2,
-                            witnesses: list[Vec2], half_angle_deg: float) -> float:
-    """对候选点 q，遍历"目标边界点 × 读数区间端点"的最坏后验包围半径。
+                            witnesses: list[Vec2], half_angle_deg: float,
+                            near_radius_m: float, disk_sides: int) -> float:
+    """对候选点 q，遍历"目标边界点 × 读数分支"的最坏后验包围半径。

-    读数 θ' 与目标 g 一致当且仅当 |θ' - bearing(q→g)| <= δ，因此端点
-    φ(g)±δ 是每个 g 的最不利读数。这是启发式近似，不构成最优性证明 [工程]。
+    读数分支：direction 的区间端点 φ(g)±δ（端点是每个 g 的最不利读数），
+    以及 near——源落在 q 的 near_radius_m 内时后验为 B(q, near)∩候选集。
+    保证接收盘内 no_signal 不可能发生，故不枚举该分支。
+    这是启发式近似，不构成最优性证明 [工程]。
    """
    posterior_radii = []
    for g in witnesses:
@@ -105,6 +117,14 @@ def _worst_posterior_radius(region: CandidateRegion, q: Vec2,
            probe.clip_wedge(q, reading, half_angle_deg)
            if probe.polygon:
                posterior_radii.append(probe.mec()[1])
+    near_probe = CandidateRegion(list(region.polygon), region.slack_m)
+    try:
+        near_probe.clip_disk(q, near_radius_m, disk_sides)
+    except InconsistentBelief:
+        pass   # B(q, near) 与候选集不相交：near 分支不可能发生，跳过
+    else:
+        if near_probe.polygon:
+            posterior_radii.append(near_probe.mec()[1])
    if not posterior_radii:
        return region.mec()[1]   # 无有效采样点时退回当前半径
    return max(posterior_radii)
@@ -144,15 +164,17 @@ def select_measurement_point(region: CandidateRegion, from_position: Vec2,
    coarse = region.witnesses(search.witness_spacing_m)
    samples = candidate_points(center, safe_radius,
                               search.candidate_rings, search.candidate_angles)
+    near_radius, sides = field.near_radius_m, search.polygon_sides
    best_key, best_point, best_worst = None, None, None
    for point in samples:
-        worst = _worst_posterior_radius(region, point, coarse, half_angle)
+        worst = _worst_posterior_radius(region, point, coarse, half_angle,
+                                        near_radius, sides)
        key = (round(worst, 6), round(move_cost(point), 6))
        if best_key is None or key < best_key:
            best_key, best_point, best_worst = key, point, worst
    dense = _worst_posterior_radius(region, best_point,
                                    region.witnesses(search.witness_spacing_m / 4),
-                                    half_angle)
+                                    half_angle, near_radius, sides)
    reason = (f"安全盘 B(centre,{safe_radius:.0f}m) 内 {len(samples)} 候选；"
              f"粗采样最坏后验 {best_worst:.1f}m，1/4 间距加密复核 {dense:.1f}m")
    return SecondPointPlan(tuple(best_point), dense, move_cost(best_point), reason)
@@ -190,19 +212,41 @@ class OmnidirectionalLoop:
        # 证书有效但 /clear 失败后必须先补测再重试，避免在同一证书上
        # 反复消耗 3 秒/次的失败清除。
        self._may_clear: set[int] = set()
+        # 断连/协议错误/守恒性破坏等中断：留档异常，收尾时写入 summary。
+        self._abort_detail: BaseException | None = None

    # ------------------------------------------------------------ 主流程
    def run(self) -> dict:
-        self.client.enter()
+        """主闭环。任何异常都必须尽力 /exit 并产出 summary，不得静默丢会话。"""
        try:
+            self.client.enter()
            self._coverage_phase()
            if self.stop_reason is None:
                self._mark_absent_after_coverage()
                self.stop_reason = self._resolve_detected()
        except DeadlineExceeded:
            self.stop_reason = "real_budget_exhausted"
-        info = self.client.exit()
-        return self._summary(info)
+        except (TransportError, ProtocolError, InconsistentBelief) as exc:
+            self._abort_detail = exc
+            if isinstance(exc, TransportError):
+                self.stop_reason = "transport_failure"
+            elif isinstance(exc, InconsistentBelief):
+                self.stop_reason = "inconsistent_belief"
+            else:
+                self.stop_reason = "protocol_error"
+        except Exception as exc:   # 意外异常同样收尾，不阻断 /exit 与结果汇总
+            self._abort_detail = exc
+            self.stop_reason = "unexpected_error"
+        info = None
+        exit_error = None
+        try:
+            info = self.client.exit()
+        except Exception as exc:   # 接口不可用：记录后仍产出未完成状态
+            exit_error = exc
+        summary = self._summary(info)
+        if exit_error is not None:
+            summary["exit_error"] = _error_brief(exit_error)
+        return summary

    # ------------------------------------------------------------ 覆盖阶段
    def _pending_channels_at(self, index: int) -> list[int]:
@@ -321,6 +365,7 @@ class OmnidirectionalLoop:
        return "all_resolved" if not unresolved else "unresolved_channels"

    def _summary(self, info) -> dict:
+        """结果汇总。info 为 None 表示 /exit 未能完成（断连等），按未完成上报。"""
        events = self.client.log.action_events()
        stats = summarize(events)
        report = recompute_virtual_time(events)
@@ -331,21 +376,37 @@ class OmnidirectionalLoop:
                      if channel not in resolved]
        completed = (self.stop_reason == "sixteen_cleared"
                     or (self.stop_reason == "all_resolved" and not unresolved))
-        return {
+        summary = {
            "completed": completed,
            "stop_reason": self.stop_reason,
-            "exit_reason": info.reason,
-            "virtual_time_s": round(info.virtual_time_s, 3),
+            "exit_reason": "exit_unreachable" if info is None else info.reason,
+            "virtual_time_s": round((info.virtual_time_s if info is not None
+                                     else self.client.state.virtual_time_s), 3),
            "cleared_channels": cleared,
            "absent_channels": absent,
            "unresolved_channels": unresolved,
            "measures": stats["measures"],
            "clears": stats["clears"],
            "clear_success": stats["cleared"],
+            "total_requests": stats["measures"] + stats["clears"],
            "move_distance_m": round(stats["distance_m"], 1),
            "anomalies": self.anomalies,
            "virtual_recompute_mismatches": len(report["mismatches"]),
+            "channel_status": self._channel_status_snapshot(),
        }
+        if self._abort_detail is not None:
+            summary["abort_detail"] = _error_brief(self._abort_detail)
+        return summary
+
+    def _channel_status_snapshot(self) -> dict:
+        """逐频道状态快照：中断收尾时的未完成状态留档（每频道一条）。"""
+        snapshot = {}
+        for channel, belief in sorted(self.belief.channels.items()):
+            entry = {"status": belief.status.value}
+            if belief.region is not None:
+                entry["posterior_radius_m"] = round(belief.region.mec()[1], 3)
+            snapshot[str(channel)] = entry
+        return snapshot


 class DirectionalExtension(OmnidirectionalLoop):
@@ -389,7 +450,7 @@ class DirectionalExtension(OmnidirectionalLoop):
                    return True
                self._may_clear.discard(channel)   # 证书失败：补测后才可重试
                continue
-            if len(self.blocked_probes.get(channel, [])) > search.max_blocked_probes:
+            if len(self.blocked_probes.get(channel, [])) >= search.max_blocked_probes:
                break   # 梯子用尽仍被遮挡：定位停滞，转后备清除
            point = self._next_probe(region, channel)
            outcome = self.client.measure(point, channel)
@@ -424,11 +485,12 @@ class DirectionalExtension(OmnidirectionalLoop):
        """定向遮挡下的探点升级梯子 [工程]。

        - 未被遮挡过：worst-posterior 选点器（保证接收盘内）；
-        - 第 1 次被遮挡：对 MEC 圆心取对跖点，仍在保证接收盘内；
+        - 第 1 次被遮挡：对 MEC 圆心取对跖方向，偏移裁进保证接收盘；
        - 之后：绕 MEC 圆心、半径 = ρ 的环探，方位自首个被遮挡点起
-          每次 +45°。环包围任意候选源（|g-c| <= ρ），发射半平面在环上
-          截出 >=180° 弧，等距采样大概率落入；环点距离不构成接收保证，
-          失败由后备清除兜底。
+          每次 +45°。几何边界必须如实陈述：环上可见弧没有下界——
+          源可位于环上（|g-c| = ρ）且朝向沿外切线，可见集退化为单个
+          切点，固定步进环探可以全部漏过。梯子只提高提前命中概率，
+          不构成接收或命中保证；完备性完全由后备清除兜底。
        """
        blocked = self.blocked_probes.get(channel, [])
        center, radius = region.mec()
@@ -441,10 +503,11 @@ class DirectionalExtension(OmnidirectionalLoop):
            dx, dy = center[0] - first[0], center[1] - first[1]
            norm = math.hypot(dx, dy)
            if norm <= 1e-9:
-                probe = min(safe, 1.0)
-                return (center[0] + probe, center[1])
+                # 首个被遮挡探点恰在圆心：方向信息为零，沿 +x 走满安全盘。
+                return (center[0] + safe, center[1])
            scale = min(1.0, safe / norm)
-            return (center[0] + dx / norm * scale, center[1] + dy / norm * scale)
+            # 对跖方向，偏移模长 = min(norm, safe)：圆心对称点裁进保证接收盘。
+            return (center[0] + dx * scale, center[1] + dy * scale)
        base = math.atan2(first[1] - center[1], first[0] - center[0])
        azimuth = base + math.radians(45.0 * (len(blocked) - 1))
        return (center[0] + radius * math.cos(azimuth),
@@ -496,6 +559,9 @@ class DirectionalExtension(OmnidirectionalLoop):
        report = super()._summary(info)
        report["directional_observed_channels"] = list(self.directional_observed)
        report["fallback_clears"] = self.fallback_log
+        # 后备 /clear 请求计入总请求口径，仅统计检测次数会低估现实耗时。
+        report["fallback_clear_attempts"] = sum(
+            entry.get("attempts", 0) for entry in self.fallback_log)
        return report


@@ -5,15 +5,16 @@

 与官方实现的边界：本类只承诺"协议可观测行为一致"（响应字段、
 耗时推进、幂等与拒绝语义）；示向度误差分布是测试假设，不代表官方实现。
+误差场为逐点固定：同一（源频道，位置）的示向度误差由种子确定性决定，
+重复测量读数不变 [题面/官方演练实测]；不同位置误差独立（白噪声假设）。
 """

 from __future__ import annotations

 import hashlib
 import json
-from dataclasses import dataclass, field
+from dataclasses import dataclass
 from math import atan2, degrees, hypot, isfinite
-from random import Random
 from time import time

 from ..config import ProblemConfig
@@ -58,7 +59,7 @@ class LocalSimulator:
        self.cost = config.cost
        self.error_half_deg = (self.field.bearing_error_half_deg
                               if error_half_deg is None else error_half_deg)
-        self._rng = Random(error_seed)
+        self.error_seed = error_seed
        self._rtt_ms = rtt_ms
        self.entered = False
        self.exited = False
@@ -190,10 +191,21 @@ class LocalSimulator:
        if distance <= self.field.near_radius_m:
            return NEAR, None
        true_bearing = degrees(atan2(source.y - pos[1], source.x - pos[0]))
-        error = self._rng.uniform(-self.error_half_deg, self.error_half_deg)
+        error = self._bearing_error(channel, pos)
        svd = round(true_bearing + error, self.field.bearing_decimals) % 360.0
        return DIRECTION, svd

+    def _bearing_error(self, channel: int, pos) -> float:
+        """逐点固定误差场：同一（频道, 位置）误差恒定 [题面：同一地点误差不变]。
+
+        坐标按 1e-6 m 量化后与种子、频道一起哈希，映射到
+        [-error_half_deg, +error_half_deg)；不同位置近似独立（白噪声假设）。
+        """
+        key = f"{self.error_seed}|{channel}|{round(pos[0], 6)}|{round(pos[1], 6)}"
+        digest = hashlib.sha256(key.encode("utf-8")).digest()
+        unit = int.from_bytes(digest[:8], "big") / 2.0 ** 64
+        return (2.0 * unit - 1.0) * self.error_half_deg
+
    def _source(self, channel: int):
        for source in self.sources:
            if source.channel == channel:
@@ -109,10 +109,26 @@ class BeliefDirectionTests(TestCase):
        with_margin = BeliefState(CONFIG)
        with_margin.update_measure(1, (0.0, 0.0), DIRECTION, 2.01)
        self.assertTrue(with_margin.candidate_set(1).contains(source, tol_m=0.5))
-        no_margin = BeliefState(ProblemConfig())
+        no_margin_config = replace(ProblemConfig(),
+                                   search=replace(ProblemConfig().search,
+                                                  bearing_rounding_margin_deg=0.0))
+        no_margin = BeliefState(no_margin_config)
        no_margin.update_measure(1, (0.0, 0.0), DIRECTION, 2.01)
        self.assertFalse(no_margin.candidate_set(1).contains(source, tol_m=0.5))

+    def test_default_config_carries_rounding_margin(self):
+        """库默认配置必须自带 0.005° 舍入裕量，与 CLI run 默认一致。
+
+        在线 API 若按默认配置消费官方读数，"加误差再舍入"规则下
+        合法真源可能被裁掉；默认口径必须保守。
+        """
+        self.assertEqual(ProblemConfig().search.bearing_rounding_margin_deg, 0.005)
+        default_belief = BeliefState(ProblemConfig())
+        source = (1200 * math.cos(math.radians(1.005)),
+                  1200 * math.sin(math.radians(1.005)))
+        default_belief.update_measure(1, (0.0, 0.0), DIRECTION, 2.01)
+        self.assertTrue(default_belief.candidate_set(1).contains(source, tol_m=0.5))
+
    def test_near_shrinks_to_five_meters(self):
        belief = BeliefState(CONFIG)
        belief.update_measure(2, (100.0, 100.0), NEAR, None)
@@ -58,6 +58,23 @@ class GeometryTests(unittest.TestCase):
        self.assertEqual(r['status'], 'polygon')
        self.assertGreater(r['diameter_m'], 1e9)

+    def test_near_antiparallel_thin_wedge_keeps_far_vertex(self):
+        """近平行细长区域不得丢失远端顶点（审核复现案例）。
+
+        测站相距 1cm、示向度 1 与 358.99999999999：两条近平行边界相交出
+        ~5.7e10 m 处的远端顶点。旧实现按绝对阈值 PARALLEL_EPS 跳过该平面
+        对，漏掉顶点后错误报出 0.143m 直径并给出清除保证。
+        """
+        r = solve([Measurement(0, 0, 1), Measurement(0, 0.01, 358.99999999999)])
+        self.assertEqual(r['status'], 'polygon')
+        self.assertGreater(r['diameter_m'], 5.7e10)
+        self.assertFalse(r['guaranteed_clearable_at_mec_center'],
+                         "细长区域的清除保证必须被拒绝")
+        # 远端方向上的可行源 (1000, 0) 必须仍落在最小包围圆内
+        mec = r['minimum_enclosing_circle']
+        self.assertLessEqual(distance(mec['center'], (1000.0, 0.0)),
+                             mec['radius_m'] + 1e-6)
+
    def test_circle_known_shapes(self):
        c, r = smallest_enclosing_circle([(0,0),(2,0),(2,2),(0,2)])
        self.assertEqual(c, (1,1))
@@ -112,6 +112,38 @@ class LoopAcceptanceTests(unittest.TestCase):
        self.assertEqual(sorted(simulator.cleared), truth)


+class AbortFinalizationTests(unittest.TestCase):
+    """断连超过重试上限：仍须 /exit、产出未完成 summary，不得静默崩溃。"""
+
+    def test_transport_failure_aborts_with_summary_and_exit(self):
+        """首个检测动作三次断连：run() 正常返回，/exit 成功，stop_reason 留档。"""
+        summary, _, _ = _run_scene(
+            random_omni_scene(seed=5, count=2),
+            fault_script=[None, "drop_before", "drop_before", "drop_before"])
+        self.assertFalse(summary["completed"])
+        self.assertEqual(summary["stop_reason"], "transport_failure")
+        self.assertEqual(summary["exit_reason"], "user_exit",
+                         "接口仍可用时必须尝试退出")
+        self.assertEqual(summary["abort_detail"]["type"], "TransportError")
+        self.assertLessEqual(summary["measures"] + summary["clears"],
+                             summary["total_requests"])
+        self.assertEqual(len(summary["channel_status"]), 20,
+                         "未完成状态须有逐频道快照")
+
+    def test_exit_failure_still_produces_summary(self):
+        """检测与 /exit 均三次断连：产出 exit_unreachable 的未完成 summary。"""
+        summary, _, _ = _run_scene(
+            random_omni_scene(seed=7, count=2),
+            fault_script=[None,
+                          "drop_before", "drop_before", "drop_before",
+                          "drop_before", "drop_before", "drop_before"])
+        self.assertFalse(summary["completed"])
+        self.assertEqual(summary["stop_reason"], "transport_failure")
+        self.assertEqual(summary["exit_reason"], "exit_unreachable")
+        self.assertEqual(summary["exit_error"]["type"], "TransportError")
+        self.assertIn("channel_status", summary)
+
+
 class CoverageLedgerIntegrationTests(unittest.TestCase):
    def test_grid_matches_plan_derivation(self):
        points = coverage_points(CONFIG)
@@ -81,6 +81,43 @@ class FallbackClearCellsTests(unittest.TestCase):
        self.assertLessEqual(len(cells), bound)


+class ProbeLadderGeometryTests(unittest.TestCase):
+    """对跖复核探点几何：圆心对称点裁进保证接收盘，不得塌缩到圆心附近。"""
+
+    def _probe_after_block(self, region, blocked_point):
+        simulator = LocalSimulator([], robot_id="T001", config=CONFIG)
+        client = SimulatorClient("T001", config=CONFIG,
+                                 transport=LocalSimTransport(simulator))
+        client.enter()
+        loop = DirectionalExtension(client, CONFIG)
+        loop.blocked_probes[7] = [blocked_point]
+        return loop._next_probe(region, 7)
+
+    def test_antipode_reflects_blocked_probe_through_center(self):
+        """圆心为原点、原探点 (200,0)：对跖复核应返回 (-200,0)。"""
+        region = CandidateRegion.disk((0.0, 0.0), 100.0, 192)   # safe≈900>=200
+        probe = self._probe_after_block(region, (200.0, 0.0))
+        self.assertAlmostEqual(probe[0], -200.0, places=6)
+        self.assertAlmostEqual(probe[1], 0.0, places=6)
+
+    def test_antipode_clipped_into_safe_disk(self):
+        """原探点距离超过安全盘半径时：落在距圆心 safe 处的对跖方向上。"""
+        region = CandidateRegion.disk((0.0, 0.0), 900.0, 192)
+        safe = safe_disk_radius(region, CONFIG)
+        self.assertLess(safe, 200.0)
+        probe = self._probe_after_block(region, (200.0, 0.0))
+        self.assertAlmostEqual(math.hypot(probe[0], probe[1]), safe, places=6)
+        self.assertLess(probe[0], 0.0, "应位于原探点的对跖方向")
+
+    def test_degenerate_center_block_walks_full_safe_radius(self):
+        """被遮挡探点恰在圆心：沿 +x 走满安全盘半径，不再钳到 1m。"""
+        region = CandidateRegion.disk((0.0, 0.0), 100.0, 192)
+        probe = self._probe_after_block(region, (0.0, 0.0))
+        safe = safe_disk_radius(region, CONFIG)
+        self.assertAlmostEqual(probe[0], safe, places=6)
+        self.assertAlmostEqual(probe[1], 0.0, places=6)
+
+
 class MixedLoopAcceptanceTests(unittest.TestCase):
    def test_mixed_ten_sources_all_cleared(self):
        sources = mixed_scene(seed=5, count=10, directional=3)
@@ -9,7 +9,8 @@ from mcm26b.belief import CandidateRegion
 from mcm26b.config import ProblemConfig
 from mcm26b.geometry.localization import Measurement
 from mcm26b.geometry.planning import first_bearing_candidate_disk
-from mcm26b.planner import (bearing_deg, clear_certificate, select_measurement_point,
+from mcm26b.planner import (_worst_posterior_radius, bearing_deg,
+                            clear_certificate, select_measurement_point,
                            select_second_point)
 from mcm26b.protocol.client import SimulatorClient
 from mcm26b.sim import LocalSimTransport, LocalSimulator
@@ -71,6 +72,24 @@ class SelectionGuaranteeTests(TestCase):
                self.assertTrue(probe.polygon and probe.contains(g, tol_m=0.05),
                                f"后验剪掉了生成读数的边界点 {g}")

+    def test_posterior_scores_near_branch(self):
+        """后验评分必须包含 near 分支，否则低报最坏后验半径。
+
+        候选集为线段 [-4,4]×{0}、测点为原点时，任意合法读数必为 near
+        （源距原点 <=4 < 5m），后验仍为整段线段（半径 4m）；只枚举
+        direction 端点会错误报出 2m。
+        """
+        region = CandidateRegion([(-4.0, 0.0), (4.0, 0.0)])
+        half = 1.005
+        witnesses = region.witnesses(60.0)
+        self.assertEqual(witnesses, [(-4.0, 0.0), (4.0, 0.0)])
+        worst = _worst_posterior_radius(region, (0.0, 0.0), witnesses, half,
+                                        CONFIG.field.near_radius_m,
+                                        CONFIG.search.polygon_sides)
+        self.assertGreaterEqual(worst, 4.0 - 1e-6,
+                                "near 分支后验半径应逼近候选集半径 4m")
+        self.assertLessEqual(worst, 4.0 + region.slack_m + 0.1)
+
    def test_move_cost_recorded_and_reason_explains(self):
        region = _k1_region()
        plan = select_measurement_point(region, (500.0, 500.0), CONFIG)
@@ -227,6 +227,32 @@ class LocalSimObservationTests(unittest.TestCase):
        # 误差在 ±1° 内均匀分布，返回值再保留两位小数。
        self.assertLessEqual(abs(outcome.svd_deg - true_bearing), 1.005)

+    def test_same_position_error_is_fixed(self):
+        """同一（源，位置）重复测量读数不变 [题面/官方演练实测]。"""
+        source = Source(channel=1, x=1000, y=300, r_eff_m=1500)
+        simulator = LocalSimulator([source], robot_id="T001")
+        client = SimulatorClient("T001", transport=LocalSimTransport(simulator))
+        client.enter()
+        first = client.measure((0, 0), 1)
+        second = client.measure((0, 0), 1)   # 新 request_id，同位置
+        self.assertEqual(first.result, "direction")
+        self.assertEqual(first.svd_deg, second.svd_deg,
+                         "同位置重复测量的示向度必须相同")
+
+    def test_error_field_deterministic_per_seed(self):
+        """同种子误差场逐点一致；不同种子得到不同误差场（逐点白噪声）。"""
+        source = Source(channel=1, x=1000, y=0, r_eff_m=1500)
+        readings = {}
+        for label, seed in (("a", 1), ("b", 1), ("c", 2)):
+            simulator = LocalSimulator([source], error_seed=seed, robot_id="T001")
+            client = SimulatorClient("T001", transport=LocalSimTransport(simulator))
+            client.enter()
+            readings[label] = [client.measure((float(x), 0.0), 1).svd_deg
+                               for x in (0, 100, 200, 300, 400)]
+        self.assertEqual(readings["a"], readings["b"])
+        self.assertNotEqual(readings["a"], readings["c"],
+                            "不同种子应在至少一个测点产生不同读数")
+

 class SessionLogTests(unittest.TestCase):
    def test_log_records_attempts_and_accepted(self):