chore: merge 1.3.10 unified completion verdict (#30/#31)

chore: date 1.3.10 changelog for release
fix: address completion verdict review findings
2026-10-15 00:38:40 -04:00 · 2026-10-15 00:38:19 -04:00 · 2026-10-15 00:28:20 -04:00 · 2026-10-15 00:03:15 -04:00 · 2026-10-14 23:54:11 -04:00 · 2026-10-14 23:39:50 -04:00
40 changed files with 2932 additions and 307 deletions
@@ -41,10 +41,35 @@ for CODE_DIR in "${CODE_DIRS[@]}"; do
        fi
    fi

+    # 复杂度只对"本次改动触及的函数"硬拦: 全量拦截会让存量债务挡住一切提交
+    # (含纯文档提交),久之只剩绕过一条路;存量 ≥C 函数降为告警,随触及随还。
+    # 用 HEAD 对比工作区+暂存区: 钩子在 `git add && git commit` 同一命令里先于 add 运行
    if command -v radon &> /dev/null; then
-        RADON_OUTPUT=$(radon cc "$CODE_DIR" -n C -s 2>&1 || true)
-        if echo "$RADON_OUTPUT" | grep -qE '^\s+[FMC]\s'; then
-            ERRORS+="[radon] 存在圈复杂度 ≥ C 的函数：\n$RADON_OUTPUT\n"
+        mapfile -t CHANGED_PY < <(git diff HEAD --name-only --diff-filter=AM -- "$CODE_DIR" | grep '\.py$' || true)
+        if [[ ${#CHANGED_PY[@]} -gt 0 ]]; then
+            TOUCHED=$(git diff HEAD -U0 -- "${CHANGED_PY[@]}" | python3 -c '
+import json, re, subprocess, sys
+hunks, path = {}, None
+for line in sys.stdin:
+    if line.startswith("+++ "):
+        path = line[6:].strip() if line.startswith("+++ b/") else None
+    m = re.match(r"@@ -\S+ \+(\d+)(?:,(\d+))? @@", line)
+    if m and path:
+        start, count = int(m.group(1)), int(m.group(2) or 1)
+        hunks.setdefault(path, []).append((start, start + max(count, 1) - 1))
+data = json.loads(subprocess.run(["radon", "cc", "-j", "-n", "C", *hunks], capture_output=True, text=True).stdout or "{}")
+for path, blocks in data.items():
+    for b in blocks if isinstance(blocks, list) else []:
+        if any(s <= b["endline"] and e >= b["lineno"] for s, e in hunks.get(path, [])):
+            print("%s:%s %s - %s (%s)" % (path, b["lineno"], b["name"], b["rank"], b["complexity"]))
+' 2>&1 || true)
+            if [[ -n "$TOUCHED" ]]; then
+                ERRORS+="[radon] 本次改动触及的函数圈复杂度 ≥ C：\n$TOUCHED\n"
+            fi
+        fi
+        LEGACY_COUNT=$(radon cc "$CODE_DIR" -n C 2>/dev/null | grep -cE '^\s+[FMC]\s' || true)
+        if [[ "$LEGACY_COUNT" -gt 0 ]]; then
+            WARNINGS+="[radon] $CODE_DIR/ 存量 ${LEGACY_COUNT} 个函数圈复杂度 ≥ C(未被本次改动触及,不阻塞)。\n"
        fi
    fi

@@ -31,7 +31,7 @@ LLM__QWEN__1__TIMEOUT_S=120
 #   模型不支持所配档位时报错并列出它真正支持的档(库带能力表,含出处与实测日期)
 # LLM__QWEN__1__EFFORT_FALLBACK=error    # 档位打空时: error(默认,报错) | nearest(映射到最近的档)
 #   默认报错的理由是钱: 静默的 medium→max 在部分模型上是数倍账单;nearest 等距取弱侧
-# LLM__QWEN__1__MISSING_DONE=retry       # SSE 缺 [DONE]: retry(默认) | salvage
+# LLM__QWEN__1__MISSING_DONE=retry       # SSE 真截断(无 [DONE] 且无 finish+其后实测 usage): retry(默认) | salvage;1.3.10 起 finish+实测 usage+干净关闭的无 DONE 流恒视为完整
 # LLM__QWEN__1__TRUST_ENV=true           # false = 绕过本地代理(LAN 直连)
 # LLM__QWEN__1__EXTRA_BODY={"temperature":0}   # 本源恒定的采样参数(JSON 对象串)
 #   并入请求体,优先级低于 chat(overlay=...);受控实验固定解码用它,免得漏传
@@ -1,6 +1,43 @@
 # Changelog

-## 1.4.0(2026-10-01)
+## 1.3.10(2026-10-15)
+
+#30/#31 同根修复: 1.3.9 以 `[DONE]` 存在、正文非空、字段白名单判定成功,上游的正常结局被当瞬时错误重试到耗尽(magpie/glm 不发 `[DONE]` 时单次调用挂 1216s;只有思考的 length 截断每次写满 max_tokens 后重抽)。本版改以上游 `finish_reason` 与传输完整性为准,流式/非流式/缓存读写共用一份判定。设计 `research-wiki/designs/2026-10-14-issue30-31-completion-verdict-design.md`。
+
+### 修复
+
+- **缺 `[DONE]` 但完整的流**(#31,**两种模式**): finish_reason 已到、其同帧或之后收到实测 usage、连接干净关闭 → 完整交回,用量记 measured,`stream_end="eof_after_finish"`。文本模式默认 `MISSING_DONE=retry` 原本同样会重试到耗尽;该键现在只管真截断(无 finish 或 finish 后无实测 usage)。
+- **只有思考的 length 截断**(#30): 保真模式原样交回(content 空、思考完整、finish_reason=length),不再重抽。
+- **`null` 与未知字段**(#30): 任意层 `null` 视为缺省(`refusal: null`、`tool_calls: null` 不再判失败,**非流式同样修复**);保真模式未知字段按 OpenAI 流式惯例无损合并(字符串拼接、对象递归、带 `index` 数组按 index 合并、`index/type/id/format` 同值幂等),Gemini `extra_content` 一类工具层扩展可原样回传。
+- **`reasoning` 别名**: vLLM/OpenRouter 的 `reasoning` 字段计入活性(此前只有思考的流会被 TTFT 看门狗误杀),文本模式无非空 `reasoning_content` 时投影为 `thinking`。
+- **截断结果不再写缓存**: 文本 salvage 结果此前会被缓存并在 TTL 内反复回放。
+
+### 行为变更
+
+- **文本模式 length/过滤且无正文 → `ResultInvalidError`**(原 `TransientError` 重试到耗尽): 不重试、不换源、熔断记成功不计窗口;**不经 `AllSourcesExhausted` 包装**,直接到调用方。原因码见异常消息(`length_without_content` / `filtered_without_content`)。文本模式"只有思考 + stop"维持瞬时(证据收集中)。
+- **finish_reason 为 `network_error`/`error`**(GLM 等上游自报失败)即使有正文也判瞬时。
+- **保真模式**: 只有思考/拒答/空白 + 明确 finish(非上游错误)→ 原样交回;length 截断的半截工具(缺 id/name)→ `ResultInvalidError` `tool_truncated`;无参函数全程只发 `arguments: null`(其余字段齐全)→ `ResultInvalidError` `tool_arguments_null`(确定性方言,交付值不改写)。
+- 文本模式 `choice.index: null` 视为 0(此前整帧被丢弃);`tool_calls` 中全为 `null` 的元素按 `tool_calls: null` 处理,不产出会被上游拒收的空数组。
+- **缓存**: 只有自然结束(stop/tool_calls/无 finish 的旧口径)且有正文的结果写入;length/过滤/未登记 finish 值、salvaged、空正文均交付但不缓存,读到此类条目按 miss。未登记 finish 值按源告警一次。
+- **活性口径增补**: 未知字段的非空字符串片段(如 `reasoning`)算进展,`index/type/id/format` 等标签键不算;`stall_ms` 计量随之覆盖。
+
+### 新增
+
+- `LLMResponse` / `TransportResult` 末尾追加 `finish_reason: str | None`、`stream_end: StreamEnd | None`(`done` / `eof_after_finish` / `salvaged`;非流式、缓存命中为 None);包根导出 `StreamEnd`。
+- 遥测 `llm_calls` 末尾追加可空列 `stream_end`(INSERT 字段 38 → 39,物理列 39 → 40;旧行 NULL 不回填;auto 档幂等补列,manual 档裁剪写入并点名缺列,补列 SQL 由 `telemetry_schema_sql()` 自取)。
+
+### 破坏性(请先读)
+
+- **`TelemetryRecorder.record_llm_call` 新增必填 keyword `stream_end`**: 自建 recorder 未跟进会在装配期被形状闸拒绝(`**fields` 形态不受影响)。
+
+### 提醒
+
+- 保真模式交付 `finish_reason=length` 的工具调用时 arguments 可能是半截 JSON: **执行工具前先查 `finish_reason`**。
+- 未知字段合并冲突时会**删除该字段**并按(源, 路径)告警一次: 结果可能缺字段,以告警为准。典型形态: 带 `index` 的数组(如 `reasoning_details`)后续帧元素缺 `index` 即判形态改变,**整个字段被删**。
+- 诚实边界: 中转在上游出错后主动正常关闭连接,与真完整在传输层不可区分;防线是"finish 必须存在、不属上游错误、且其后有实测 usage"。
+- 旧版本写入的 salvage/length 文本缓存无法辨认: 受影响集合须在升级后首次读写前更换 namespace/salt(同 1.3.9 提醒)。
+
+## 1.3.9(2026-10-01)

 - 新增 `chat(response_mode="openai")`，返回 `LLMResponse.completion`：assistant 消息、工具调用、finish_reason、上游 response_id、usage 快照；支持流式聚合及非流式。
 - 修复 SSE 顶层 error 后接 DONE 被当成功、同帧 content/reasoning_content 丢失；工具实际进展纳入看门狗、首 token 与中段对冲。
@@ -9,7 +46,6 @@
 - 新模式不绕过 managed/raw 冲突校验；raw-only 仍须源级/请求级均无受管意图。`Transport.complete` 新增必填 `response_mode`，自定义实现需适配。
 - 分发采用 Apache-2.0；README 补充常驻事件循环线程桥接，不新增同步 API。

-
 ## 1.3.8(2026-09-14)

 流式中段沉默的计量与应对(issue #27 / #28):`generation_ms` 口径重定义为**净生成时间**(扣除所有超阈值中段沉默)并配 `stall_ms`/`stall_count` 归因字段与遥测两列;另加一条**默认关闭**的中段沉默对冲(主路首 token 之后相邻增量 chunk 间隔超阈值时,向异源并发再发一次)。两键均不设时行为逐字等于 1.3.7(默认关闭回归门: 既有 unit/contracts 全套件语义断言一行未改)。
@@ -2,10 +2,12 @@

 > [!URGENT]
 > **实验室内部通用基础库(生产级、非 MVP、零业务假设)**
+>
 > 1. 本项目是被多个科研/生产项目依赖的**库**,不是应用:稳定性、并发性、防御性、可观测与测试不可为"简单"让步(YAGNI 仍适用,但不削减健壮性)。库的 bug 会同时击穿所有下游项目。
 > 2. 你的所有思考过程和回复必须使用 **简体中文**。

 ## 1. 项目元数据
+
 - **核心目标**: PolyGateway = 统一的大语言模型(LLM/VLM/OCR,音频预留)调度与中转库。治理单位是**一次模型调用**:请求封装、多源多账号、限流、错误分类与重试、熔断、Redis 响应缓存、流式看门狗、遥测(含成本)、结构化输出策略。全组件端口化可插拔。
 - **架构权威文档**: `research-wiki/ARCHITECTURE.md`(架构单一事实源,含 D1-D14 决策及讨论过程、子系统设计、三项目迁移验收标准;**不受 400 行设计文档限制**,以无歧义传达既有讨论为准绳)。开发顺序见 `research-wiki/ROADMAP.md`;`research-wiki/designs/` 仅存放每次实现具体功能的设计文档。
 - **参考项目**: `reference/` 下三个项目是本库的需求来源与代码蓝本(**只读,勿改**;M4 起"只读"指工作区文件与 main 检出不变——迁移实施经 `git worktree` 在 `~/Projects/m4-worktrees/` 的 feature 分支进行,worktree 的 git 操作会写 `reference/*/.git` 元数据,属预期);库必须能按 ARCHITECTURE.md §11 被它们迁移接入,否则即边界缺口。
@@ -35,11 +37,13 @@ make ci        # 只读验证(check + test)
 > **同一理由适用于长跑命令**: 需要后台跑时(如全套件测试),命令末尾**不得接管道**,否则退出码失真;要判完成用 `wait`/轮询 PID,不要用会匹配到自身的 `pgrep -f "<完整命令串>"`。

 ### Phase 1: 规划与设计
+
 1. 涉及**公共 API、端口签名、架构边界、新子系统**的变更**必须**调用 `brainstorming`(产出 2-3 备选方案+权衡)并经**人类确认**后实施;其余任务自判(判据: 是否改变库对下游的承诺)。动手前查阅 `research-wiki/`(单一事实源)。
 2. 功能产生运行时数据时**必须**调用 `structured-logging`。
 3. **里程碑级/跨多文件**功能编码前**必须**调用 `writing-plans`;小改动自判。审核门控: design 走 Claude 自审 → Codex 审 → **人类审**;plan 走 Claude 自审 → Codex 审 → 直接执行。

 ### Phase 2: 执行与验证
+
 1. **测试结果门**: 合并前每个行为变更必须有"先失败后通过"的测试证据(`test-driven-development`);bug 修复必带回归测试;不规定中间怎么走。
 2. **独立验证**: 里程碑级/跨多文件/合并前**必须**派全新上下文的 verifier subagent(`verification-before-completion`);任何规模的完成声明都必须逐条对应本会话内的工具输出(证据化声明,禁止虚报)。
 3. **反 gold-plating**: 不做任务外的重构、抽象与"顺手清理"。
@@ -47,6 +51,7 @@ make ci        # 只读验证(check + test)
 ## 4. 核心规则

 ### 4.1 核心原则(按优先级)
+
 - **P1 YAGNI**: 不写当前用不到的代码;但并发控制、防御校验、可观测埋点、错误隔离与测试是"当前需要",不在削减之列。
 - **P2 高可读性**: 领域术语命名;注释解释"为什么"。
 - **P3 单一职责**: 一句话说不清职责(需要"和")= 拆分。
@@ -58,7 +63,7 @@ make ci        # 只读验证(check + test)
 ### 4.2 库铁律(本项目特有,违反即 bug)

 | 铁律 | 内容 |
-|---|---|
+| --- | --- |
 | 零业务假设 | 库内禁止出现任何下游业务领域词汇(视频/文书/超声等)与业务 fixtures;扩展点一律 Protocol |
 | 纯 asyncio 中立 | 无全局状态、无框架假设、无模块级单例;同一 `GatewayClient` 在 arq worker 与裸脚本中行为一致 |
 | 取消可穿透 | `asyncio.CancelledError` 永不捕获吞没;重试循环、限流等待、流式读取全部可被取消;in-flight 资源在 finally 释放 |
@@ -69,6 +74,7 @@ make ci        # 只读验证(check + test)
 | 无缓存毒化 | 缓存 key 必含 model + messages 摘要 + 命名空间/租户 + salt;多模态 content 先摘要再 hash |

 ### 4.3 代码开发规范
+
 - 模块/类/方法必须有**中文 Docstring**;复杂逻辑用 `# Phase N` 注释组织。
 - 校验分层: 外部输入校验用显式异常(Python `-O` 移除 assert,禁止 assert 承担生产校验);assert 仅用于内部不变量。
 - 类名 `PascalCase`,私有前缀 `_`;导入顺序标准库→第三方→项目内。
@@ -76,6 +82,7 @@ make ci        # 只读验证(check + test)
 - 不考虑向后兼容,直接修改原文件。**例外**: `LLMResponse` 等已被三项目消费的公共类型,字段只增不删不改名(迁移兼容约束,见 ARCHITECTURE.md §5.1)。

 ### 4.4 Git 工作流
+
 - 一切开发在 feature 分支,严禁直改 main;频繁语义化提交;提交**必须**调用 `commit` skill;大改动前先提交回滚点。

 ### 4.4.1 发布流程(每步都是历史欠账换来的,不得跳步)
@@ -87,7 +94,7 @@ make ci        # 只读验证(check + test)
 按顺序执行,**构建之前**必须先改完所有文档:

 | # | 动作 | 要点 |
-|---|---|---|
+| --- | --- | --- |
 | 1 | **更新 README** | 打包会把当时的 README 固化进 sdist,**发布后再改就来不及了**(包里那份永远是旧的)。逐项核对: 安装命令的版本约束(`==1.1.*` 这类**极易漏改**,漏了下游就被锁在旧版)、能力表是否覆盖新行为、数字型断言是否仍成立(如遥测字段数,须用 `inspect.signature` 实测而非凭记忆) |
 | 2 | CHANGELOG 定版 | "未发布" → `## X.Y.Z(日期)` |
 | 3 | 版本号 | `pyproject.toml` + `src/polygateway/__init__.py` 两处必须一致 |
@@ -106,11 +113,13 @@ Gitea 包 registry 是 **owner 级**(`/iomgaa/-/packages/`)不是仓库级;PyPI
 发布验证补充（2026-09-09 实测）：私有索引不镜像 setuptools，下载 sdist 时 pip 即使带 `--no-deps` 仍可能尝试安装构建依赖；已具备构建工具的 conda 环境可加 `--no-build-isolation`，仍须核对下载来源与散列。包关联 POST 本次返回400，但匿名包页面和认证 GET 包 API 均证实已关联目标仓库；**400本身不算成功，也不能直接推断已关联**，必须读取实际 `repository.full_name` 与页面链接。包 API 可能要求认证（匿名401），页面仍应匿名亲查；不要为重试关联先解绑现有正确关系。

 ### 4.5 配置管理
+
 - 工程配置走 `pydantic-settings` + `.env`(模板 `.env.example`,敏感项不提交);严禁硬编码默认值;缺失关键配置直接报错。
 - 多源命名约定 `{SCOPE}__{PROVIDER}__{N}__{FIELD}`;韧性参数键名沿用三项目习惯(`LLM_TIMEOUT` 等),降低迁移成本。
 - 装配只有两条路: `GatewayClient.from_env()`/`from_settings()`(工厂)或构造函数全量注入(测试/高级);库内部任何组件不得自读环境变量。

 ### 4.6 测试组织
+
 - `tests/{unit,integration,e2e}`;真实场景优先(录制的真实网关响应二次构造优于凭空 mock)。
 - 覆盖率目标 80%;并发/韧性行为是一等测试对象: 重试穿透取消、熔断开路半开、限流结算退款、Redis 掉线降级方向、缓存 key 隔离。
 - Redis 相关测试用真实 Redis(integration),不 mock Lua 行为;限流契约测试随实现一起交付(参考 CHSAnalyzer `tests/contracts_limiter.py`)。
@@ -140,7 +149,7 @@ project_root/
 ## 6. 上下文导航

 | 需求 | 路径 |
-|---|---|
+| --- | --- |
 | 架构全貌: 决策 D1-D14 及讨论过程、端口清单、错误分类、子系统设计、迁移验收 | `research-wiki/ARCHITECTURE.md`(单一事实源) |
 | 开发顺序与里程碑状态 | `research-wiki/ROADMAP.md`(活文档,随进度更新) |
 | 三项目迁移文档(ARCHITECTURE §11 的展开,库设计的常驻约束) | `research-wiki/migrations/`(govdoc-saas / video-tree-trm5 / chsanalyzer) |
@@ -154,6 +163,7 @@ project_root/
 | 第一次抽库尝试(教训与蓝本) | `reference/GovDoc-SaaS/packages/docagent-core/` |

 ## 7. 输出规范
+
 - 所有输出**中文**;文档优先表格/伪代码/Mermaid;禁止超 15 行代码块入文档、禁止连续超 5 条碎片列点;设计 ≤400 行、计划 ≤1000 行、报告 ≤300 行。
 - **例外**: `research-wiki/ARCHITECTURE.md` 不受行数限制——它的准绳是"后来的 AI/人类无需还原原始讨论即可准确理解全部决策及理由",宁详勿略(详细 ≠ 琐碎:记录论证与取舍,不堆砌实现细节)。

@@ -163,7 +173,7 @@ project_root/
 > **Skill 的 description 即触发边界。** 下表标注 **MANDATORY** 的情形是硬门(设计人类门、合并前测试证据门、合并前独立验证门、commit 格式),不得跳过;其余情形按 description 边界自判——自判看任务实质(规模/风险/是否触及公共承诺),不是省事。任何 skill 流程不得引入任务外的重构或抽象。优先级: 用户显式指令 > Skill 详细流程 > 本文件宏观规则。

 | Skill | 触发边界 |
-|-------|---------|
+| ------- | --------- |
 | `brainstorming` | **MANDATORY**: 公共 API/端口签名/架构边界/新子系统变更;其余自判 |
 | `writing-plans` | **MANDATORY**: 里程碑级/跨多文件功能;小改动自判 |
 | `test-driven-development`(测试结果门) | **MANDATORY**: 合并前——行为变更须有先失败后通过的测试证据 |
@@ -19,12 +19,13 @@
 | 调用期限 | 一次逻辑调用可选一条**墙钟硬边界**(`{SCOPE}__CALL_DEADLINE_S` 或 `chat(call_deadline_s=...)`,缺省不启用):治理的是**等待**——重试退避、配额轮询、熔断冷却、结构化重问与 embedding 分批共享同一份期限。三条须知:①**返回时刻 = 期限 + 清理耗时**(遥测/结算/缓存收尾在 `finally` 里跑完,允许超期;实测构造达期限的 5–7 倍),库只承诺切断等待、不给返回上界;②**到期 ≠ 未产出、≠ 未计费**,上游可能已算完并计费;③**不配置即逐字保持 1.3.5 语义**(纯 429 序列仍可能长等、有限大 `Retry-After` 仍照睡)。到期抛 `CallDeadlineExceeded`,**不属四分类、不属 `GatewayUnavailableError` 族** |
 | 长尾对冲 | **默认关闭**的可选加速(`{SCOPE}__HEDGE__AFTER_S`,仅 chat):一次尝试挂起超阈值时向**异源**并发再发一次,先回者赢、输家取消。流式以「首 token 未至」为触发判据(不误杀慢生成),非流式只有纯总时长阈值;对冲走完整限流/熔断准入,拿不到配额静默放弃。成本须知:开启即用配额换延迟(触发窗口内 in-flight 翻倍),输家取消按 est 保留预扣且**可能已被上游计费**(取消止不住上游);输家不喂熔断,遥测 attempt 行标 `error="hedge_cancelled"` 可按 `logical_call_id` join 对账;两路皆败只计一次重试预算;与 `call_deadline_s` 正交组合(装配守卫强制对冲阈值 < 期限)
 | 中段沉默对冲(可选,超业界现状能力) | **默认关闭**(`{SCOPE}__HEDGE__STALL_AFTER_S`,仅流式 chat):主路**首 token 之后**相邻增量 chunk 沉默超阈值时向异源并发再发一次,与首 token 对冲 `HEDGE__AFTER_S` **独立生效**、可独设可组合;仅设它且请求非流式 → 逐字走单路(中段沉默无非流式概念)。**诚实标注:中段沉默触发并发副本在主流网关/库中无公开先例**(业界收敛为看门狗掐流重试)。成本须知:下游实测沉默 >20s 占 19.0%,触发率远高于首 token 对冲,开启即用配额换延迟,输家可能已被上游计费;看门狗掐死**喂**熔断而对冲输家取消**不喂**(不对称声明),详见[1.3.8 生成时间口径与中段沉默对冲](#138-生成时间口径与中段沉默对冲) |
-| OpenAI 消息保真(1.4.0) | `chat(response_mode="openai")` 返回完整 assistant 消息/工具调用、finish_reason、上游 ID 与 usage；单 choice，复用治理，不执行工具；默认仍为文本模式 |
+| OpenAI 消息保真(1.3.9) | `chat(response_mode="openai")` 返回完整 assistant 消息/工具调用、finish_reason、上游 ID 与 usage；单 choice，复用治理，不执行工具；默认仍为文本模式 |
+| 完成判定(1.3.10) | 以上游 `finish_reason` 为准判定"生成结束、结果可交付":缺 `[DONE]` 但 finish 与其后的实测 usage 齐全即完整(两模式);length/过滤等确定性结局不再被当瞬时错误重试;未知/`null` 字段不再判失败;`LLMResponse.finish_reason` / `stream_end` 给出结束原因与方式,详见[1.3.10 统一完成判定](#1310-统一完成判定) |
 | 响应缓存 | Redis/内存;key 含 model + messages 摘要 + namespace(缓存隔离单位)+ salt + 采样参数 + 请求级推理档位(同 messages 跑 low 与 max 不互相命中),多模态 content 先摘要再 hash(防毒化);可 per-call 绕过(科研重采样) |
-| 流式看门狗 | TTFT / inter-token / 总超时三层活性;thinking token 刷活性不计结果;截断流(缺 `[DONE]`)判瞬时不入缓存 |
+| 流式看门狗 | TTFT / inter-token / 总超时三层活性;thinking token(含 `reasoning` 别名)刷活性不计结果;截断流(无 `[DONE]` 且无 finish+其后实测 usage)默认判瞬时(`MISSING_DONE=salvage` 的文本打捞结果交付但不入缓存) |
 | 推理可观测性 | "这次到底推理没推理"由多信号裁定(推理正文压倒 usage 明细),三态落在 `LLMResponse.thinking_observation`:`observed` / `absent` / `unknown`——**`unknown` 是"本次判不出",不是"没推理"**;本次实发档位与实测观测矛盾时按 `(源, 模型, 生效档位)` 各告警一次(能力表过期、开启未生效、注入了却观测不到;同一模型的 low 与 max 是两个独立的矛盾,不共用节流键);裁定结果随遥测落库 |
 | 推理档位 | 推理是**八档**(`none`/`auto`/`minimal`/`low`/`medium`/`high`/`xhigh`/`max`)而非开关:源级 `REASONING_EFFORT` + 请求级 `chat(reasoning_effort=...)`,`ENABLE_THINKING` 保留为语法糖;库带 24 条能力表(逐条 evidence 自报实测/文档推定),档位打空**默认报错并给出该模型最省的可用档与该配的键**,要静默映射需显式配 `EFFORT_FALLBACK=nearest`;实发档随 `LLMResponse.applied_effort` 与遥测落库 |
-| 遥测与成本 | 每次调用(含缓存命中与失败)必录 38 字段;三类行(`event_kind` = `attempt` / `cache_hit` / `terminal_failure`)加逐源诊断列(`http_status_code` / `error_type` / `cause_type` / `error_body`);SQLite / Postgres 后端(表已存在时**不需要** schema 建表权限,最小权限账号可直接用);按价格表折算成本落库(注意 `LLMResponse.cost` 本身恒为 `None`,成本只进遥测);多模态内容摘要落库不存原图 |
+| 遥测与成本 | 每次调用(含缓存命中与失败)必录 39 字段;三类行(`event_kind` = `attempt` / `cache_hit` / `terminal_failure`)加逐源诊断列(`http_status_code` / `error_type` / `cause_type` / `error_body`);SQLite / Postgres 后端(表已存在时**不需要** schema 建表权限,最小权限账号可直接用);按价格表折算成本落库(注意 `LLMResponse.cost` 本身恒为 `None`,成本只进遥测);多模态内容摘要落库不存原图 |
 | 逻辑调用统计 | 治理单位是**一次逻辑调用**而非一次尝试:四种响应(chat / embedding / OCR 两种)带 `call_stats`(`logical_call_id` / `attempts` / `total_latency_ms`,及生成时间 `generation_ms`——**1.3.8 起为净口径**:赢家那次 transport 调用墙钟 − Σ超阈值中段沉默,仍含 TTFT 与尾部,排除准入排队/退避/触发前等待;1.3.7 的墙钟值由 `generation_ms + stall_ms` 还原——与沉默归因 `stall_ms` / `stall_count`(成对报告,None = 无沉默观测)、对冲计数 `hedges` / `hedge_won`),重试、换源、结构化重问、embedding 分批共享同一逻辑 ID;每次**领域失败**另落一条 `terminal_failure` 行,失败调用数从此是一条 `WHERE event_kind = 'terminal_failure'`,详见[1.3.5 逻辑调用统计与失败诊断](#135-逻辑调用统计与失败诊断)与[1.3.8 生成时间口径与中段沉默对冲](#138-生成时间口径与中段沉默对冲) |
 | 遥测的资源与降级 | Postgres 池**闲时占 0 条连接**、忙时上限可配(`PGW_TELEMETRY_PG_POOL_MAX`,缺省 4),每次写入有硬预算(`PGW_TELEMETRY_PG_WRITE_TIMEOUT_S`,缺省 5s);后端不可用是**可恢复的降级**(冷却 60s 后自动重试,DBA 建完表/放开权限即自愈),永久失能只留给 DSN 本身写错;降级状态可编程查询——`client.telemetry_status` 给出 `degraded`/`fatal`/`reason`/`dropped_rows` 等只读快照,不必再靠人工对账。**对账要同时看 `degraded` 与 `dropped_rows`**: 池饱和超预算丢的行走行级丢弃,`degraded` 保持 `False`(后端没挂,是本进程并发超了),只按 `degraded` 告警会看不见这一类丢行——而它恰是 `pool_max` 配小了的唯一信号 |
 | 调用方维度 | 每次调用可带 `tenant_id`(遥测表的真实列,可挂 RLS、可建复合索引)与 `meta`(≤16 个自定义 KV);四个公共方法全覆盖,校验超限即报错;**库只交付列,不启用 RLS、不建索引** |
@@ -125,7 +126,7 @@ stats.total_latency_ms             # 含缓存 IO、退避、准入等待、重
 | 定义 | **净生成时间** = 赢家那次 transport 调用的墙钟 − Σ(所有 > `STALL_THRESHOLD_MS`(缺省 1000ms)的相邻增量 chunk 沉默);仍**排除**准入排队/退避/对冲触发前等待(同 1.3.7) |
 | 边界 | 仍含 **TTFT 与尾部**(末 chunk → 流结束归活跃侧,墙钟基底口径的有意选择) |
 | 还原公式 | **墙钟 = `generation_ms + stall_ms`**(`stall_ms is None` 时墙钟 = `generation_ms`) |
-| `stall_ms` / `stall_count` | 沉默总额(毫秒)与超阈值相邻间隔个数,**成对报告**(只报净数无法行动);chunk = 任何增量产出(`content` 或 `reasoning_content` 均计),空 chunk / usage 帧 / SSE keepalive 不计时 |
+| `stall_ms` / `stall_count` | 沉默总额(毫秒)与超阈值相邻间隔个数,**成对报告**(只报净数无法行动);chunk = 任何增量产出(`content`、`reasoning_content`、工具实际增量,1.3.10 起另含未知字段的非空字符串片段如 `reasoning`),空 chunk / usage 帧 / SSE keepalive / 仅标签键(`index`/`type`/`id`/`format`)不计时 |
 | `stall_ms is None` 分支条件 | 非流式 / 增量 chunk <2 → 两字段 None 且 `generation_ms` **不退减**;None = 「无沉默观测」而非 0,下游预算公式以此为分支条件 |
 | 与基准口径的区隔 | NVIDIA TTFT/ITL/E2EL 与 vLLM 的基准口径**均不扣除**沉默——扣沉默是本库自造口径,与基准数字比较时不得混用 |

@@ -146,7 +147,7 @@ stats.total_latency_ms             # 含缓存 IO、退避、准入等待、重

 ```bash
 pip install --extra-index-url https://gitea.iomgaa.online/api/packages/iomgaa/pypi/simple/ \
-    "polygateway[redis,postgres,structured]>=1.4.0,<2"
+    "polygateway[redis,postgres,structured]>=1.3.10,<2"
 ```

 核心仅依赖 `httpx` + `pydantic`;按需选 extras:
@@ -612,7 +613,7 @@ make ci        # 只读全量验证

 `LLMResponse` 等被下游消费的公共类型,字段**只增不删不改名**且新增字段必带默认值;`{SCOPE}__{PROVIDER}__{N}__{FIELD}` 与平铺韧性键名(`LLM_TIMEOUT` 等)沿用三项目既有习惯,不做破坏性改名。实验室内部库,随实验室项目需求演进。

-## 1.4.0 OpenAI 工具消息保真（单 choice）
+## 1.3.9 OpenAI 工具消息保真（单 choice）

 `chat(..., response_mode="openai")` 仍返回 `LLMResponse`，完整协议结果在
 `response.completion`：`message`、`finish_reason`、`response_id`、`usage`。
@@ -632,16 +633,17 @@ messages.append(completion.message)

 保真是 JSON 字段和值保真，不是 HTTP 字节透明代理。仅支持单 choice（`n=1`），
 可含多个并行 function tool_calls；arguments 逐字拼接，不解析、修复或执行。
-SSE 支持 role/content/reasoning_content/tool_calls，未知增量字段会明确失败；
+SSE 的 role/content/reasoning_content/tool_calls 按原语义聚合；1.3.10 起 `null` 视为缺省，
+未知增量字段按流式惯例无损合并（冲突删键并告警，见下节）；
 非流式保留整个 message 扩展字段，不返回顶层 logprobs/created 等其他字段。
 上游实际模型仍读 `model_reported`，配置别名是 `model`。

 | 边界 | 行为 |
-|---|---|
-| 空文本且有完整工具 | 正常成功；不完整工具或无任何有效产物仍失败 |
+| --- | --- |
+| 空文本且有完整工具 | 正常成功；不完整工具失败（length 截断的半截工具归 ResultInvalid），其余见 1.3.10 结局表 |
 | error 帧 | 所有模式立即判失败并走已有重试；后续 DONE 不挽救 |
-| 缺 DONE | 保真模式不受 salvage 放宽，防截断工具被执行 |
-| finish_reason=length | 原样返回，不修补参数；调用方自行判断是否执行 |
+| 缺 DONE | 1.3.10 起 finish + 其后实测 usage + 干净 EOF 即完整；真截断仍不受 salvage 放宽，防截断工具被执行 |
+| finish_reason=length | 原样返回（含只有思考的回答），不修补参数；**执行工具前先查 finish_reason**，arguments 可能是半截 JSON |
 | structured | 不可与保真并用；自行传 overlay.response_format 不启用库的修复/重问阶梯 |
 | raw 推理 | 保真不是校验逃生口；raw-only 要求源 ENABLE_THINKING/REASONING_EFFORT 均未设置，请求档为 None，再通过 overlay.reasoning_effort 传原值 |

@@ -676,6 +678,43 @@ except TimeoutError:
 应用退出时先停止接单、取消/等待在途调用，在同一个 loop 中 await client.aclose()，
 再停 loop 并 join 线程；注入组件仍由所有者关闭。不要从 loop 本身阻塞调用 future.result。

+## 1.3.10 统一完成判定
+
+#30/#31 同根: 1.3.9 以 `[DONE]` 存在、正文非空、字段白名单作为成功判据,上游的正常结局
+(不发 `[DONE]`、只有思考的 length 截断、`refusal: null`)被当瞬时错误重试到耗尽。1.3.10
+改以上游 `finish_reason` 与传输完整性为准,流式/非流式/缓存四处共用一份判定。
+
+| 结束方式(`stream_end`) | 条件 |
+| --- | --- |
+| `done` | 收到 `[DONE]` |
+| `eof_after_finish` | 未发 `[DONE]`,但 finish 已到、其同帧或之后收到的 usage 为实测,连接干净关闭(用量记 measured) |
+| `salvaged` | 文本模式 `MISSING_DONE=salvage` 打捞的截断结果(用量降 estimated,**不写缓存**) |
+| `None` | 非流式 / 缓存命中 / 失败行 |
+
+| 结局(首个匹配行生效) | `openai` 模式 | `text` 模式 |
+| --- | --- | --- |
+| finish 为 `network_error`/`error`(上游自报失败) | 瞬时错误 | 瞬时错误 |
+| 工具快照不成立 | length → `ResultInvalidError`;其余瞬时 | — |
+| 无任何产物 + length/过滤 | 原样交回 | `ResultInvalidError`(不重试、不计熔断) |
+| 无任何产物 + 其余(零增量抖动) | 瞬时错误 `empty_completion` | 同左 |
+| 有正文 | 交付 | 交付 |
+| 只有思考/拒答/空白 + length/过滤 | 原样交回 | `ResultInvalidError` |
+| 只有思考/拒答/空白 + stop 等 | 原样交回 | 瞬时错误(证据收集中) |
+| 只有思考/拒答/空白 + 无 finish | 瞬时错误 | 同左 |
+
+`ResultInvalidError` 不经 `AllSourcesExhausted` 包装,直接到调用方;原因码在异常消息里
+(`length_without_content` / `filtered_without_content` / `tool_truncated` 等,遥测 `error`
+列可检索)。只有自然结束(stop/tool_calls/function_call,或未上报 finish 的旧口径)且投影正文非空
+(或工具完整)的结果才写缓存;未登记的 finish 值按正常交付、不缓存并告警一次。无参函数只发
+`arguments: null` 的工具判 `ResultInvalidError`(`tool_arguments_null`,确定性方言,不改写不重试)。文本模式 `thinking` 在无非空 `reasoning_content` 时取
+`reasoning`(vLLM/OpenRouter 新名)。保真模式未知字段:字符串拼接、对象递归、带 `index` 的
+数组按 index 合并、`index/type/id/format` 同值幂等;冲突时**删除该字段**并按(源, 路径)
+告警一次——结果可能缺字段,以告警为准。诚实边界:中转在上游出错后主动正常关闭连接与真
+完整在传输层不可区分,防线是"finish 必须存在、不属上游错误、且其后有实测 usage"。
+
+遥测 `llm_calls` 追加可空列 `stream_end`;`TelemetryRecorder.record_llm_call` 新增必填
+keyword `stream_end`(**自建 recorder 须跟进,否则装配期被形状闸拒绝**)。
+
 ## 许可证

 PolyGateway 使用 [Apache License 2.0](LICENSE)。内部移植代码的原创 owner 已明确授权；
@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"

 [project]
 name = "polygateway"
-version = "1.4.0"
+version = "1.3.10"
 description = "PolyGateway：实验室统一的大语言模型（LLM/VLM/OCR）调度与中转库——多源、限流、重试、熔断、缓存、遥测"
 # registry 包页面的正文只认这一项:缺了页面就是一片空白(1.1.2 的教训,twine 会警告
 # long_description missing 但不阻塞上传)。README 在打包时被固化进产物,发布后再改无效。
@@ -413,7 +413,7 @@ flowchart TB
 | 值 | 含义 | 生产者 | cost |
 | --- | --- | --- | --- |
 | `measured` | usage 帧完整可信 | 正常路径;OCR 成功行(0 token 是**事实**而非未知) | 按 token 换算 |
-| `estimated` | 有实测数字但可信度降级 | 打捞路径(收到 usage 帧但流被截断,§7.1) | 按 token 换算 |
+| `estimated` | 有实测数字但可信度降级 | 打捞路径(收到 usage 帧但流被截断,§7.1;1.3.10 起即 `stream_end="salvaged"`) | 按 token 换算 |
 | `unavailable` | 用量信息不可得 | usage 帧缺失、失败尝试、终态失败 | **NULL** |

 值域在 `types.py` 以模块级 frozenset 常量 `USAGE_SOURCES` 落地,**仅约束库内生产侧**(所有写入点从该常量取值),不在 `LLMResponse`/`Usage`/`TransportResult` 上加 `__post_init__` 值域校验——它们是运行时构造点,裸 `ValueError` 不属 §6 四分类、`RetryMW` 不捕会逃出 `chat()`;且 `LLMResponse` 是三项目已消费的公共类型,新增运行时校验属下游可见行为变更。历史库里既有的 `estimated` 行在新值域中依然合法可读。
@@ -470,7 +470,9 @@ flowchart TB
 | HTTP 429(body 无 insufficient_quota)、500/502/503/504 | `TransientError`(携 `Retry-After` 解析值,仅支持秒数形态) |
 | HTTP 429 + body 含 insufficient_quota、401、403 | `SourceDeadError` |
 | HTTP 400 | `RequestRejectedError` |
-| **空补全**: 200 且流程完整([DONE]/usage 正常)但 content 为空(2026-07-20 M1 验证发现,人类裁决) | `TransientError`(服务抖动,重试/换源;绝不缓存空响应) |
+| **空补全**: 200 且流程完整([DONE]/usage 正常)但 content 为空(2026-07-20 M1 验证发现,人类裁决) | `TransientError`(服务抖动,重试/换源;绝不缓存空响应)。1.3.10 起仅限**零产物**或无法区分抖动与结局的情形;上游明确的 length/过滤结局见 §15 |
+| finish_reason 为 `network_error`/`error`(上游自报失败,1.3.10) | `TransientError` |
+| 文本模式 length/过滤且无正文;保真模式 length 截断的半截工具(1.3.10) | `ResultInvalidError`(确定性结局,不重试、不计熔断) |
 | 解析层失败(结构化输出/OCR ZIP) | `ResultInvalidError` |

 **响应体留存(2026-08-16,Gitea issue #10;设计 `designs/2026-08-16-issue10-error-body-retention-design.md`)**: 上表每一条 HTTP 翻译**都必须携带响应体摘要**——摘要同时进入异常 message 与 `PolyGatewayError.body_text`(1.2.0 新增基类字段)。两者都要,因为逐次遥测写的是 `str(exc)`,只加字段进不了遥测表,而"事后可查"正是这条要求的目的。摘要口径由 `transports/_http_errors.summarize_body` 单点实现(折叠空白 → 限长 2048 字符 → 超长保留头 1400 + 尾 600 并记省略字数),两个 transport 共用,**不得各写一份**——issue #10 的成因正是"只有 429 那一支用了响应体"。`body_text` 是旁路数据,不参与任何治理判定;`_translate_429` 的类型细分仍解析未截断原文(摘要会破坏 JSON,改用它会让超长 body 的 `insufficient_quota` 退化成普通限速)。
@@ -566,7 +568,7 @@ flowchart TB

 ### 7.8 遥测与成本

-**必录字段**(继承三项目 15 字段规范;当前 38 个 INSERT 字段,物理表列 39 = 38 + 数据库自填的 `created_at`,两套口径的区分见 `telemetry/schema.py` 模块 docstring): call_id、parent_call_id、session_id、model、provider、source_name、messages(JSON)、response、thinking、prompt_tokens、completion_tokens、usage_source、latency_ms、ttft_ms、max_inter_token_ms、cache_hit、error、**cost**、**cached_prompt_tokens**、**model_reported**、**sampling**、**reasoning_tokens**、**tenant_id**、**meta**、**thinking_observation**、**reasoning_effort**、**scope**、**operation**、**logical_call_id**、**event_kind**、**http_status_code**、**error_type**、**cause_type**、**error_body**、**attempts**、**total_latency_ms**、**stall_ms**、**stall_count**。
+**必录字段**(继承三项目 15 字段规范;当前 39 个 INSERT 字段,物理表列 40 = 39 + 数据库自填的 `created_at`,两套口径的区分见 `telemetry/schema.py` 模块 docstring): call_id、parent_call_id、session_id、model、provider、source_name、messages(JSON)、response、thinking、prompt_tokens、completion_tokens、usage_source、latency_ms、ttft_ms、max_inter_token_ms、cache_hit、error、**cost**、**cached_prompt_tokens**、**model_reported**、**sampling**、**reasoning_tokens**、**tenant_id**、**meta**、**thinking_observation**、**reasoning_effort**、**scope**、**operation**、**logical_call_id**、**event_kind**、**http_status_code**、**error_type**、**cause_type**、**error_body**、**attempts**、**total_latency_ms**、**stall_ms**、**stall_count**、**stream_end**(1.3.10,§15)。

 **`stall_ms`/`stall_count` 两列(2026-09-11,1.3.8,issue #27,端口 36 → 38)**: 沉默归因进库,口径见 §7.2 净口径段。取 INTEGER 而非 REAL——落库值是 retry 层已取整的 int,与换算唯一处纪律(N8)一致。追加在 `total_latency_ms` 之后,可空无默认、旧行 NULL 不回填;`emit_cache_hit`/`emit_terminal_failure` 与失败/取消 attempt 行恒 NULL(缓存命中与终态行无 transport 观测;被取消的对冲输家无 TransportResult,其 `hedge_cancelled` 行 stall 恒 NULL,存在性由该行表达)。`generation_ms` **仍不进遥测列**(YAGNI)——墙钟由 `generation_ms + stall_ms` 还原。

@@ -806,3 +808,19 @@ src/polygateway/
 | 迁移 | 旧 text 污染无法自动判断，受影响集合必须迁移 namespace/salt；新模式天然隔离；无新遥测列/表，统一 helper 与取消结算不变 |

 本增量不会执行工具；length 结束原因与参数原样交回，由下游判断可执行性。嵌套 JSON 是独立可变快照，不承诺深冻结。许可已由原创 owner 明确批准 Apache-2.0。
+
+## 15. 统一完成判定（2026-10-14，#30/#31 已授权增量）
+
+本节修订 §6.2 空补全行、§7 `usage_source` 映射与 §14 协议/失败/缓存三行，详细边界见 [设计](designs/2026-10-14-issue30-31-completion-verdict-design.md)。
+
+| 接缝 | 增量契约 |
+|---|---|
+| 根因 | 1.3.9 以 `[DONE]` 存在、正文非空、字段白名单判定成功，上游正常结局落 Transient 被重试放大；本版以上游 `finish_reason` 与传输完整性为准 |
+| 内核 | `types.py` 纯函数 `_stream_is_complete` / `_classify_finish` / `_judge_outcome` / `_is_cacheable`，流式、非流式、缓存读写四处共用；只返回判定值，错误由调用点翻译 |
+| 流完整性 | `[DONE]` → `done`；无 DONE 但 finish 已到且其同帧或之后的 usage 为实测、干净 EOF → `eof_after_finish`；其余为截断：保真不打捞，文本按 `missing_done` |
+| 结局 | 首匹配表：上游错误 finish → Transient；坏工具（length → ResultInvalid）；零产物 + length/过滤 → 保真交付、文本 ResultInvalid；零产物其余 → Transient（保 2026-07-20 裁决）；有正文交付；有产物无正文按 finish 分模式 |
+| 字段 | 五层 null 视缺省；保真未知字段无损合并（字符串拼接、index 数组按 index、标签键幂等），冲突删键 + 有界告警；未知字符串片段计活性 |
+| 缓存 | 可缓存 ⊂ 可交付：salvaged、无正文、length/过滤/未知 finish 不写，读到即 miss；判定先于命中覆盖 |
+| 公共面 | `LLMResponse`/`TransportResult` 追加 `finish_reason`、`stream_end`；遥测加 `stream_end` 列，`record_llm_call` 加必填参数 |
+
+`usage_source` 与 `stream_end` 分工：前者答"账单可信吗"，后者答"内容完整吗"。`eof_after_finish ⇒ measured`；`salvaged ⇒ estimated|unavailable`；`estimated ⇒ salvaged`，逆命题不成立。诚实边界：中转在上游出错后主动正常关闭连接，与真完整在传输层不可区分。
@@ -13,7 +13,7 @@
 ## 1. 总览与状态

 | 阶段 | 内容 | 接入方 | 状态 |
-|---|---|---|---|
+| --- | --- | --- | --- |
 | P0 奠基 | 调研、ARCHITECTURE.md、CLAUDE.md、skill Fable 5 改造、hooks 硬边界、脚手架(git/pyproject/Makefile/conda 环境/CI 绿) | — | ✅ 完成(2026-07-20) |
 | M1 核心 | 内核类型 + httpx transport + 治理中间件(内存后端)+ 多源多账号 + 缓存 + SQLite 遥测 + 结构化输出 + from_env | GovDoc、Video-Tree | ✅ 完成(2026-07-20;239 测试全绿、覆盖 92%、真实网关+真实 Redis 验收通过、独立 verifier 问题清零;见 designs/2026-07-20-m1-core-design.md 状态行) |
 | M2 分布式 | Redis 限流/熔断后端、多源 × Redis 联合验证、背压、Postgres 遥测、成本、Embedding(Q3)、压测 harness | CHSAnalyzer(治理) | ✅ 完成(2026-07-21;契约双后端全绿 + 时间语义真实等待变体 10/10、跨连接联合验证、CHS 对标十项打钩、P3 真实验收 300 次双进程七不变量 PASS、独立 verifier 0 Critical 且 Important 清零;见 designs/2026-07-20-m2-distributed-design.md 与 findings/m2-verifier-fixes.md。压测收官: P5 故障混编 40 次八不变量 PASS,P6 首跑 58.1% 暴露三大治理盲区(findings/2026-07-21-p6-soak-baseline.md)→ **M2.5 治理韧性**六轮迭代(双通道熔断/健康选源/AIMD/429 免预算/健康门槛降权/连败抑制)后同场景 **98.96%**、八不变量全 PASS,验收见 findings/2026-07-21-m25-acceptance.md 与 designs/2026-07-21-m25-resilience-design.md) |
@@ -25,7 +25,7 @@
 **内部开发顺序**(编号即依赖顺序,同号可并行):

 | 步 | 内容 | 顺序理由 |
-|---|---|---|
+| --- | --- | --- |
 | 1 | `types.py` + `errors.py` + `ports.py` 全量设计与冻结 | 原则 2:公共承诺先行;这是 M1 设计文档(人类门)的主体 |
 | 2a | `streaming.py` 看门狗移植 | 原则 4:纯函数,零依赖,直接移植+补测 |
 | 2b | `providers.py` 注册表 | 叶子模块;transport 的前置(思考流字段声明;thinking 注入的**决策**已于 issue #16/#17 搬到 `thinking.py`,这里只留形态声明) |
@@ -72,9 +72,11 @@

 **已交付增补(2026-09-14)**: 流式中段沉默计量与对冲(issue #27/#28,1.3.8)——`generation_ms` 口径重定义为**净生成时间**(墙钟 − Σ超阈值中段沉默,仍含 TTFT 与尾部;1.3.7 墙钟口径由 `generation_ms + stall_ms` 还原,披露见 CHANGELOG 未发布段),`CallStats` 增 `stall_ms`/`stall_count` 归因字段,遥测表加同两列(INSERT 字段 36→38);`{SCOPE}__HEDGE__STALL_AFTER_S` 中段沉默对冲(独立于 `HEDGE__AFTER_S` 生效、默认关闭;无业界公开先例,超业界现状能力),`{SCOPE}__STREAM__STALL_THRESHOLD_MS` 计量阈值(默认 1000,三层可调)。设计 designs/2026-09-11-streaming-stall-design.md(人类已批准 + 独立审查补遗 B1-B4/N1-N9),实施 plans/2026-09-11-streaming-stall.md,验证 findings/2026-09-11-streaming-stall-validation.md。后续储备:自适应分位阈值(防反馈环,现不做)、流式透传下的 commit 点重审、embedding/OCR 对冲(现不做)。

+**已交付增补(2026-10-15,1.3.10 已发布)**: 统一完成判定(issue #30/#31,1.3.10)——以上游 `finish_reason` 与传输完整性判定流完成与结局,流式/非流式/缓存读写共用内核纯函数;缺 `[DONE]` 但 finish+实测 usage 完整即成功(两模式),文本 length/过滤无正文改 `ResultInvalidError`,保真模式 null/未知字段无损合并;`LLMResponse` 增 `finish_reason`/`stream_end`,遥测加 `stream_end` 列。设计 designs/2026-10-14-issue30-31-completion-verdict-design.md(人类授权 Claude 与独立 subagent 两轮审查后直接执行),实施 plans/2026-10-14-issue30-31-completion-verdict.md。后续储备:文本 n>1 混流修复、文本"只有思考+stop"证据收集。
+
 ## 7. 开放决策依赖

 | 决策 | 阻塞点 | 需拍板时间 |
-|---|---|---|
+| --- | --- | --- |
 | Q1 打包分发(git+ssh vs 内网 index) | 不阻塞开发,阻塞 M4 发布 | M4 前 |
 | Q3 Embedding 是否纳入 | ~~影响 M2 范围~~ **已拍板纳入 M2**(2026-07-20) | ✅ 已决 |
@@ -7,7 +7,9 @@ date: 2026-10-01

 # #29 OpenAI 消息保真设计

-> 状态：方向已获人类批准；计划独立审查 P1/P2 已修订并实施，首轮独立代码审查发现的缓存空产物及测试双向依赖已修复，等待父会话最终复核；1.4.0 发布准备，尚未发布。
+> **修订指针(2026-10-14)**: §4.1"未知 delta 拒绝"、§4.2"reasoning-only 不算成功"与"缺 DONE 一律失败"、§5 活性口径、§6 缓存判据已由 [统一完成判定设计](2026-10-14-issue30-31-completion-verdict-design.md)(1.3.10)修订,以后者为准。
+
+> 状态：方向已获人类批准；计划独立审查 P1/P2 已修订并实施，首轮独立代码审查发现的缓存空产物及测试双向依赖已修复，等待父会话最终复核；按人类最终指定准备 1.3.9，取代未发布的 1.4.0；尚未发布。
 > 基线：`0e29deac1a72bfee348ed2f8061ea2c33be217e8` / 1.3.8；日期：2026-10-01。
 > 需求：[issue #29](https://gitea.iomgaa.online/iomgaa/PolyGateway/issues/29)。
 > 权威约束：ARCHITECTURE §2.3、§5、§6、§7.1–7.8；实现前按本文修订对应架构段落。
@@ -23,7 +25,7 @@ Codex 不可用，人类已授权原生 subagent 替代独立审查。具体 API
 ## 2. 源码核验与方案比较

 | 已核实接缝 | 现状与修复目标 |
-|---|---|
+| --- | --- |
 | `transports/openai_compat.py::_sse_delta` | 只读首 choice；content 早返回丢同帧 reasoning；忽略 tool_calls/error |
 | `_complete_stream` / `_complete_once` | 纯文本判空；raw 只存 usage；需新增协议快照与 tool-only 成功语义 |
 | `client.py::chat` / `_build_payload` | messages 已直传，overlay 已合并，无需白名单重写 |
@@ -32,7 +34,7 @@ Codex 不可用，人类已授权原生 subagent 替代独立审查。具体 API
 | `middleware/cache.py` | 已有完整消息/overlay 身份，但嵌套新类型需重建、模式需隔离 |

 | 方案 | 优势 | 代价/结论 |
-|---|---|---|
+| --- | --- | --- |
 | A：chat 显式模式 + LLMResponse 内嵌协议结果 | 生命周期、取消、缓存、遥测和对冲单一实现；兼容旧投影 | 需定义组合约束和缓存身份；**选定** |
 | B：新增 chat_completion 入口，复用私有核心 | 接口像 OpenAI | 重复公共签名、输入校验和终态封装，收益不足，否决 |
 | C：独立 SDK 客户端或代理自聚合 | 局部改动直观 | 无法自动解决响应上浮/缓存/线路错误，复制治理，否决 |
@@ -64,7 +66,7 @@ completion 是独立于上游容器的深拷贝 JSON 快照；frozen 只保证
 ### 3.2 请求与组合校验

 | 条件 | 规则 |
-|---|---|
+| --- | --- |
 | 未知模式 | 公共入口 ValueError，尚未创建逻辑调用上下文 |
 | openai + structured 非 None | 公共入口 ValueError；不注入 schema、不自动重问；overlay 的 response_format 仍可直传 |
 | openai + n | 仅允许缺省或严格整数 1（不接受 bool）；调用 overlay 在入口校验，选中源 extra_body 及最终 payload 在 HTTP 前校验，不接受被 overlay 遮蔽的非法源级 n |
@@ -93,7 +95,7 @@ response_id 是上游 id，允许未上报为 None，绝不拿 call_id 补齐；
 每次 complete 独立创建聚合器，不放 transport 实例共享状态。逐 JSON 帧先校验对象并查 error，再读取 choices/delta、元信息和 usage；坏类型统一转 TransientError，不让 AttributeError/TypeError 逃出治理。

 | 字段 | 合并规则 |
-|---|---|
+| --- | --- |
 | content / reasoning_content | 非 null 必为 str，逐片拼接；同帧同时收集；不 trim；从未出现则不造键，仅 null 则保留 null |
 | role | openai 模式若出现必须 assistant；只出现 role 不算活性；最终成功必须含 assistant role；缺失失败，不伪造字段 |
 | tool_calls | 必为列表；每项 index 必为非负严格整数；按 index 分桶，末尾按 index 排序，索引不输出到最终工具对象 |
@@ -110,7 +112,7 @@ SSE、非流式及缓存最终快照均要求 role="assistant"；缺失或其他
 ### 4.2 成功、截断与错误

 | 场景 | 行为 |
-|---|---|
+| --- | --- |
 | openai：空正文 + 有完整工具 | 成功，不触发空补全重试 |
 | openai：非空投影正文、无工具 | 成功；存在任何不完整工具则仍拒绝，不用正文掩盖坏工具 |
 | openai：无正文且无有效工具 | 维持 empty_completion 瞬时失败；reasoning-only 不算成功 |
@@ -146,7 +148,7 @@ CacheMW 显式序列化和重建 OpenAIChatCompletion，读时验证容器、单
 ## 7. 旧行为审计与非功能约束

 | 既有行为 | 处置 |
-|---|---|
+| --- | --- |
 | 旧字段/默认文本/structured 阶梯/受管推理 | 保留；默认文本同帧漏 reasoning 属修复 |
 | HTTP 分类、retry/backoff、两类对冲、TPM/熔断 | 保留，不改算法与预算 |
 | 空补全、missing_done salvage | 文本保留；openai tool-only 放行、缺 DONE 强制失败 |
@@ -0,0 +1,251 @@
+---
+type: design
+node_id: design:2026-10-14-issue30-31-completion-verdict-design
+title: 统一完成判定设计(#30/#31)
+date: 2026-10-14
+---
+
+# #30/#31 统一完成判定设计
+
+> 状态: 方向已获人类批准(2026-10-14,四项取舍与五个拍板点按推荐);人类授权"Claude 与独立 subagent 确认无误后直接执行",不再走人类设计审阅。两轮独立审查已闭合(R1 3 阻塞 + 9 应修;R2 通过,2 应修已并入),见 §11。
+> 基线: `22609c9`(1.3.9);目标版本 **1.3.10**;分支 `fix/issue30-31-completion-verdict`。
+> 需求: [#30](https://gitea.iomgaa.online/iomgaa/PolyGateway/issues/30)、[#31](https://gitea.iomgaa.online/iomgaa/PolyGateway/issues/31)。
+> 修订对象: [1.3.9 保真设计](2026-10-01-issue29-openai-fidelity-design.md) §4.1/§4.2/§5/§6、ARCHITECTURE §6.2/§7/§14。
+
+## 1. 问题与根因
+
+两条 issue 是**同一根因**: 库判定"生成结束了、结果可交付"时,依赖库自己的推断(`[DONE]` 存在、正文非空、字段在白名单内、快照字段类型),而不是上游的权威声明(`finish_reason`)与传输完整性(干净 EOF + 结尾 usage)。推断失败一律落 `TransientError` → 重试/换源/计熔断,把确定性结局放大成 `max_attempts` 次付费重抽并推高健康源的熔断计数。
+
+| 表现 | 位置(1.3.9) | 后果 |
+| --- | --- | --- |
+| #31 缺 `[DONE]` 但 finish+usage 完整 | `openai_compat.py:668-671`(保真无条件拒);`_check_done:714`(文本 `missing_done=retry` 默认同样拒) | magpie/glm 两模式均重试到耗尽;实测 1216s 挂起、dissect3 38 次终败 |
+| #30-1 reasoning-only + `length` | `types.py:478-490`、`_reject_empty_completion` | 每次写满 max_tokens 后丢弃重抽 |
+| #30-2 未知键 / `tool_calls:null` | `_openai_completion.py:97-98,112-114,139-140,146-147`;**非流式也中**: `OpenAIChatCompletion` 拒 `tool_calls: None`(`types.py:451-453`,已实测) | 渠道一换 100% 失败,且以"瞬时"形态出现 |
+| 隐患: 只有 `reasoning` 字段的流 | 活性只认 content/reasoning_content | 两模式被看门狗当无进展 → TTFT 超时 |
+| 隐患: 文本 salvage 的截断结果被缓存 | `cache.py:197-204` 写路径无判定 | 半截结果 TTL 内反复回放 |
+
+补充事实: GLM 的 finish_reason 含 `network_error`/`sensitive`;2026-07-20 空补全裁决针对的是 stop/无 finish 的"零增量"抖动(`findings/2026-07-21-m25-acceptance.md:48`);vLLM 等实现每帧带累计 usage,finish 之前的 usage 不证明走完。
+
+## 2. 方案比较
+
+| 方案 | 内容 | 结论 |
+| --- | --- | --- |
+| A 局部放行 | 按 issue 字面在三处各加例外 | 否决: 流式/非流式/缓存判据继续分叉;文本 #31 残留;`network_error` 被当成功 |
+| B 统一完成判定(选定) | 内核纯函数分两步(流完整性 → 结局分类),流式/非流式/缓存读写四处共用;"可交付"与"可缓存"拆两判据;未知字段无损合并 | 一次消除根因;代价是公共字段追加与遥测端口加参 |
+| C 透明透传 | 保真模式不做结局判定,原样交回 | 否决: 零增量抖动回归(违背 2026-07-20 裁决),截断流也被交出 |
+
+## 3. 公共契约变更(只增不改)
+
+| 位置 | 变更 |
+| --- | --- |
+| `types.py` | 新增 `StreamEnd = Literal["done", "eof_after_finish", "salvaged"]`,包根导出 |
+| `TransportResult` / `LLMResponse` 末尾 | 追加 `finish_reason: str \| None = None`、`stream_end: StreamEnd \| None = None` |
+| `OpenAIChatCompletion` 校验 | `tool_calls: null` 合法(视同无工具,原样保留键) |
+| `TelemetryRecorder.record_llm_call` | 追加**无默认值** keyword-only `stream_end: str \| None`(**破坏性**: 自建 recorder 装配期被形状闸拒绝,CHANGELOG 单列) |
+| `llm_calls` 表 | 末尾追加可空列 `stream_end TEXT`(落点见 §7) |
+| `Transport.complete` 签名 | **不变** |
+
+`finish_reason` 取值: 流式取 choice 0(保真已强制单 choice;文本取 `index` 缺省或为 0 的 choice)首个非 null 值;非流式取 `choices[0].finish_reason`。保真模式非字符串维持 1.3.9 拒绝(Transient);文本模式非字符串记 None。`finish_reason` 不进遥测列(未获批)。
+
+| 场景 | `stream_end` |
+| --- | --- |
+| 流式收到 `[DONE]` | `"done"` |
+| 干净 EOF + finish + 结尾 measured usage | `"eof_after_finish"` |
+| 文本 `missing_done=salvage` 打捞 | `"salvaged"` |
+| 非流式 / 缓存命中 / 失败行 | `None`(与 `ttft_ms`/`stall_ms` 同"无从观测"口径) |
+
+与 `usage_source` 分工: 后者答"账单可信吗",前者答"内容完整吗"。`eof_after_finish ⇒ measured`(由 §4.2 条件保证);`salvaged ⇒ estimated|unavailable`;`estimated ⇒ salvaged`,逆命题不成立。
+
+## 4. 统一完成判定
+
+### 4.1 位置与形态
+
+全部放 `types.py` 内部纯函数(下划线前缀,不导出): CacheMW 要复用,import-linter 禁止 middleware 依赖 transports,而内核层 `ports : types : errors : streaming` 允许。函数只返回判定值,**不抛领域错误**(错误需 `source_name` 等上下文,由调用点翻译),不捕获任何异常,取消天然穿透。
+
+| 函数 | 职责 |
+| --- | --- |
+| `_tool_snapshot_ok(message)` | 工具快照是否成立(id/type/function.name/arguments 规则 + 唯一 id);`OpenAIChatCompletion.__post_init__` 复用 |
+| `_classify_finish(value)` | 映射到 finish 类别(下表) |
+| `_stream_is_complete(...)` | 第一步,返回 `StreamEnd \| None`(None = 截断) |
+| `_judge_outcome(...)` | 第二步,返回 `deliver` 或 `(reject_transient \| reject_result, 原因码)` |
+| `_is_cacheable(...)` | §6 判据 |
+
+| finish 类别 | 值(大小写敏感) |
+| --- | --- |
+| natural | `stop`、`tool_calls`、`function_call` |
+| truncated | `length` |
+| filtered | `content_filter`、`sensitive` |
+| upstream_error | `network_error`、`error` |
+| unknown | 其余非空字符串: 按 natural 交付、**不可缓存**、按源+值告警一次(文案写明"该值不缓存",便于下游定位命中率下降);`end_turn`/`stop_sequence` 等无证据不预先归类 |
+| absent | None |
+
+### 4.2 第一步: 流完整性(仅流式)
+
+| 条件 | 结果 |
+| --- | --- |
+| 读取异常 / 活性超时 / error 帧 / 畸形帧 | 照旧 Transient(不动) |
+| 收到 `[DONE]` | `done` |
+| 干净 EOF、无 DONE、finish 已见、finish 帧**同帧或之后**收到的 usage 被 `_resolve_usage` 判为 measured | `eof_after_finish`,usage 记 measured(不降级) |
+| 干净 EOF、其余情形 | 截断: openai → Transient `missing_done`(仍不打捞);text → 原 `_check_done` 语义(零产物 early_eof、`retry` 抛、`salvage` 记 `salvaged`) |
+
+聚合器每尝试记录 `finish_seen` 与 `usage_after_finish`: 收到非 null usage 时,若 `finish_seen` 已真或同帧带 finish,则置后者为真;是否 measured 以**最后一个**非 null usage(与计费所用 `sink["usage"]` 同一对象)在流结束时判定;usage-only 帧(`choices: []`)在 finish 之后到达是常规形态,正确置位。状态随重试/对冲每尝试新建。
+
+诚实边界: 中转站在上游出错后主动正常结束响应,与真完整在传输层不可区分;防线是"finish 必须存在、不属 upstream_error、且其后有 measured usage",文档如实写明。
+
+### 4.3 第二步: 结局分类
+
+产物定义——**保真模式**: 任何产物 = 消息中任一非空字符串片段(content、reasoning_content、reasoning、refusal、合并保留的未知字符串,空白也算)或至少一个非 null 工具条目;**文本模式**: 任何产物 = 投影 content 或 thinking(含 §4.4 别名)非空串(含空白),未知字段与工具增量只计活性不计产物,`_check_done` 零产物判据同口径。两模式**正文产物** = 投影 content strip 后非空,或(仅保真)工具快照成立且非空。
+
+表格**自上而下取首个匹配行**:
+
+| # | 条件 | openai 模式 | text 模式 |
+| --- | --- | --- | --- |
+| 1 | finish 为 upstream_error | Transient `upstream_finish_error` | 同左 |
+| 2 | (仅保真)存在工具条目但 `_tool_snapshot_ok` 不成立 | truncated → ResultInvalid `tool_truncated`;其余 → Transient `tool_incomplete` | 不适用 |
+| 3 | 无任何产物 + truncated/filtered | 交付(#30;隐藏推理模型写满预算的形态) | ResultInvalid `length_without_content` / `filtered_without_content` |
+| 4 | 无任何产物 + natural/unknown/absent | Transient `empty_completion`(保 2026-07-20 裁决) | 同左 |
+| 5 | 有正文产物 | 交付 | 交付 |
+| 6 | 有产物无正文 + truncated | 交付(#30) | ResultInvalid `length_without_content` |
+| 7 | 有产物无正文 + filtered | 交付 | ResultInvalid `filtered_without_content` |
+| 8 | 有产物无正文 + natural/unknown | 交付 | Transient `empty_completion`(待证据;消息带 finish 与 thinking 长度) |
+| 9 | 有产物无正文 + absent(流式仅 DONE 收尾,或非流式 body finish 为 null) | Transient `empty_completion` | 同左 |
+
+ResultInvalid 理由: `retry.py:409-414` 对其"不重试、不换源、熔断记成功不计窗口";length/过滤是预算或内容策略的确定性结局,不是源坏了。对冲下原路 ResultInvalid 等对冲路裁定,两败原样上抛(`retry.py:548-560`);StructuredMW 只捕 parse 抛出的 ResultInvalid,不会误重问(`structured.py:79`)。已知代价: 该分支 TPM 全额退款、用量不落成本(`actual=0`),与 Transient 丢成本同级,登记为已知边界。文本 length 改抛 ResultInvalid 后不再经 `AllSourcesExhausted` 包装,§10 要求核对下游 catch 点。
+
+原因码以固定格式 `"{源} {code}: …"` 写进异常 message,从而进遥测 `error` 列(LIKE 可检索);不新增列、不改 `_failure_reason`(Transient 一族熔断原因记 `network_error` 属已知口径)。
+
+保真模式行 2 之外的工具快照成立 + `length`: 交付,arguments 可能是半截 JSON——docstring 与 CHANGELOG 提醒调用方先查 `finish_reason` 再执行工具。非流式跳过第一步,直接以 body 的 finish_reason 走第二步。
+
+### 4.4 文本模式 reasoning 别名
+
+投影 `thinking`: 本次尝试出现过**非空** `reasoning_content` 则只用它,否则用 `reasoning`。保真 `completion.message` 保留上游原键不改名。两模式活性、产物摘要、`thinking_observation` 均基于该投影。
+
+## 5. 字段合并与活性
+
+### 5.1 null 与已知字段
+
+null 规则**优先于**已知字段类型校验: chunk、choice、delta、工具条目、function 五层任意键值为 null 视为缺省,数组中的 null 元素(如 `tool_calls: [null]`)跳过;键从未出现时保留 `key: null`(含 `tool_calls`、`refusal`、`role`、`function.arguments`)。保真成功仍须最终 role=assistant(全程只有 `role: null` 即缺 role,照旧失败)。其余已知字段(content/reasoning_content/role/tool_calls 及工具 id/type/function.name/arguments)非 null 时语义逐字沿用 1.3.9。
+
+### 5.2 未知字段(delta、`tool_calls[]` 条目、`function` 三层统一)
+
+| 值 | 规则 |
+| --- | --- |
+| 标签键 `index`/`type`/`id`/`format`(任意未知层级) | 同值幂等,不拼接;不同值 = 冲突 |
+| 其他字符串 | 按到达顺序拼接(OpenAI SDK 流式累加惯例) |
+| 数字/布尔 | 同值幂等;不同值 = 冲突(不学 SDK 相加) |
+| 对象 | 递归套用本表 |
+| 数组且元素均为带整数 `index` 的对象 | 按 index 合并(`reasoning_details` 一类);合并策略由该键首次出现形态决定,之后形态改变 = 冲突 |
+| 其他数组 | 追加 |
+| 冲突 | 删除该键路径,本尝试内不再接收;按 (源, 归一路径) 告警一次 |
+
+冲突删键是有损的: docstring 与 CHANGELOG 写明"可能缺字段,以告警为准"。上游方言确定,重拉同样结果,不视为缓存毒化。
+
+### 5.3 活性
+
+已知字段活性逐字沿用 1.3.9(非空 content/reasoning_content、工具 name/arguments 非空增量、首次建立工具身份)。未知字段仅当本帧在非标签键上并入非空字符串片段(含追加型数组中的字符串元素)时算进展(`reasoning` 由此计入);同帧只记一次。残余风险如实登记: 上游每帧重复下发的非标签字符串(如 `status`)会被拼接并持续刷新看门狗。文本模式同口径。README 中"增量 chunk = content 或 reasoning_content"的 stall/TTFT 口径表述同步更新为本节口径。
+
+### 5.4 告警节流
+
+节流集合放 transport 实例(与 `_warned_mismatches` 分立,纯 asyncio 中立),键为 (源, 归一路径) 与 (源, 未知 finish 值):路径中数组下标归一为 `[]`,finish 值截断至 64 字符;集合上限 1024,饱和后记一次"告警已饱和"不再增长。日志只写源名与路径,不回显值(可能含工具参数)。聚合器以只读集合交出冲突路径,transport 负责告警。
+
+## 6. 可交付 vs 可缓存
+
+可缓存 ⊂ 可交付。两模式同一判据 `_is_cacheable`,读写共用:
+
+| 条件(任一成立即不可缓存) | 理由 |
+| --- | --- |
+| `stream_end == "salvaged"` | 截断结果,修现存毒化 |
+| 无正文产物 | 不把空回答固化进 TTL |
+| finish 类别为 truncated/filtered/unknown | 截断有随机性,同 key 换时可成功;未知值不可证明是自然结束 |
+| openai 模式 completion 原文 content 为空且无工具(沿 1.3.9) | 不信任旧投影 |
+
+| 路径 | 规则 |
+| --- | --- |
+| 写 | 不可缓存 → 跳过,debug 日志 |
+| 读 | `_is_cacheable` 在命中覆盖(`cache.py:224` `fields.update`)**之前**、对存储原值判定;不可缓存 → miss + warning |
+| 读: 非法 `stream_end` | 不在 Literal 域内 → miss + warning(关乎内容完整性,故与 `_coerce_observation` 的宽容方向相反) |
+| 读: 顶层无 finish_reason | 回退 `completion.finish_reason`(1.3.9 保真 length 条目因此可识别);文本旧条目两字段缺省 → None → 仍可命中 |
+| 命中返回 | `stream_end=None`;`finish_reason` 取判定所用值(顶层缺省时为回退值),非字符串视 miss |
+| 序列化 | `_RESPONSE_FIELDS` 由 `dataclasses.fields` 自动派生,无需改动 |
+
+旧 salvage/length 文本污染无法辨认: CHANGELOG 继续要求受影响集合迁移 namespace/salt。
+
+## 7. 遥测落点
+
+| 位置 | 变更 |
+| --- | --- |
+| `middleware/telemetry.py` `_AttemptUsage.of(response)` | 读 `response.stream_end`(本次 LLMResponse 有字段,不像 1.3.8 走 `emit_attempt` kw,避免约 50 处调用点) |
+| `_record` | 新增参数,单一 helper 下沉 recorder |
+| `emit_cache_hit` / `emit_terminal_failure` / 失败与取消 attempt 行 | 显式 None |
+| `telemetry/schema.py` | SQLite/PG 两份 CREATE 末尾、`SQLITE_BACKFILL`、`_PG_BACKFILL_DECLS`(列序对齐)、`COLUMNS`;`telemetry_schema_sql()` 随之输出 |
+| sqlite/postgres recorder | 按 `COLUMNS` 写入;auto 档幂等补列、manual 档裁剪写入并点名缺列(沿 1.3.8) |
+| tests | 约 20 处测试替身 `record_llm_call` 补参数 |
+| 文档 | README "必录字段数"按 `inspect.signature` 实测更新;`schemas/llm-calls.md` 加列并修正陈旧计数 |
+
+## 8. 旧行为审计
+
+| 既有行为 | 处置 |
+| --- | --- |
+| 保真缺 DONE 一律失败 | **替换**: §4.2 判完整时成功;其余仍失败、仍不打捞 |
+| 文本 `missing_done=retry` 遇缺 DONE 即抛 | **替换**: 同上;`missing_done` 只管真截断 |
+| 文本 salvage + estimated | 保留,另标 `salvaged`,不再写缓存 |
+| 零增量 + DONE → Transient | **保留**(`test_stream_zero_content_with_done_is_transient` 红线回归) |
+| 保真 reasoning-only/空产物失败 | **替换**: §4.3 行 3/6/7/8 |
+| 未知 delta/工具/function 字段拒绝;null 已知字段拒绝 | **替换**: §5 |
+| 非流式 `tool_calls: null` 拒绝 | **替换**: 合法 |
+| 多 choice/index≠0 拒绝、最终 role 必须 assistant、error 帧立即失败 | 保留 |
+| 工具快照强校验 | 保留;仅 length 截断时改归 ResultInvalid |
+| finish_reason 冲突拒绝(Transient) | 保留(登记: 确定性方言,将来可改 ResultInvalid) |
+| 已知字段活性口径 | 保留;未知字段按 §5.3 增补 |
+| 缓存写路径无判定 | **替换**: 按 `_is_cacheable` |
+| 文本 n>1 混流 | 不在范围,另开 issue |
+
+## 9. 非功能维度
+
+| 维度 | 回答 |
+| --- | --- |
+| 并发与取消 | 新状态在每尝试聚合器/sink 内;判定为纯函数,不新增 try/except,`CancelledError` 原路穿透;finally 结算与 permit 释放不变 |
+| 降级方向 | 缓存跳写/视 miss 只 debug/warning;遥测新列缺失按 manual 档裁剪 + 告警,不冒泡;限流/熔断不涉及 |
+| 幂等与重复 | 减少重复付费重抽;标签键与身份字段幂等;告警节流有界 |
+| 持久化与原子性 | 仅新增一可空列,补列沿既有探测后 ALTER 幂等流程;缓存仍成功后单次原子 set |
+| 错误分类 | 仅落 Transient 与 ResultInvalid 两类,不新增类;原因码见 §4.3 |
+| 可观测 | `stream_end` 字段 + 遥测列;冲突路径与未知 finish 有界告警;遵 structured-logging 更新 `schemas/llm-calls.md` |
+
+## 10. 测试、文档与交付
+
+全部用假 SSE / MockTransport 复现,每条先红后绿;真实中转冒烟标 slow 按发布清单 diff 交集跑。
+
+| 组 | 用例 |
+| --- | --- |
+| #31 验收 | 1 stop+usage+EOF 成功、`eof_after_finish`、measured、attempts=1;2 tool_calls 同上工具完整;3 无 finish 即关 → 失败;4 finish 无 usage → 失败;5 中途 `RemoteProtocolError` → 失败;6 带 DONE → `done` 不变 |
+| #31 补充 | usage 早于 finish(累计 usage)+EOF → 不完整;finish 后 usage 为 `{}` → 不完整;`network_error`+usage+EOF → Transient;文本模式 1/3/4(`retry`、`salvage` 两档) |
+| #30 验收 | reasoning-only+length → 交付、content 空、思考完整、attempts=1;流式 `refusal:null`+`tool_calls:null` 成功;非流式 message 带 `tool_calls:null`+`refusal:null` 成功 |
+| 结局表 | §4.3 每行至少一例,两模式、流式与非流式;文本 length → ResultInvalid 且熔断未计失败;零增量+DONE(带 stop/不带 finish)仍 Transient;content 非空 + 坏工具 + stop → Transient |
+| 合并/活性 | 只有 `reasoning` 刷新活性不超时;二次构造 OpenRouter `reasoning_details` 往返(`format`/`id` 不拼接);工具层嵌套对象原样保留可回传;数值冲突删键、告警一次、调用成功;节流饱和;`role:null`/`arguments:null` 视缺省 |
+| 缓存 | 四类不可缓存均不写、注入后读视 miss;非法 `stream_end` 视 miss;1.3.9 旧文本条目仍命中;1.3.9 保真 length 条目经回退识别为不可缓存;命中 `stream_end=None` |
+| 遥测 | sqlite/pg(integration)新列写入、旧库补列、manual 档裁剪;形状闸拒绝缺参 recorder |
+| 改写 | `test_openai_never_salvages`;`test_invalid_delta_is_domain_error` 中 `{"audio": {}}`、`{"index":0,"extra":1}`、null 已知字段参数;提交说明列出 |
+
+同步: 源码 docstring、README(能力表、`1.3.*` 安装约束核对、活性口径、字段数)、CHANGELOG 1.3.10(行为变更段、破坏性段: recorder 端口、namespace/salt 提醒、半截工具参数提醒、冲突删键提醒)、ARCHITECTURE §6.2/§7/§14、1.3.9 设计加修订指针、`schemas/llm-calls.md`、`research-wiki/migrations/` 三份核对下游对 ResultInvalid 的 catch 点、`docs-convention.md` §2 清单。另开 issue: 文本 n>1 混流;文本"有产物无正文 + stop"证据收集。发布按 CLAUDE.md §4.4.1 九步,#30/#31 在 registry 可安装且验收证据齐备后关闭。
+
+证据: `retry.py:404-450,548-560`;`cache.py:197-236`;`structured.py:79`;[openai-python `_streaming.py`](https://github.com/openai/openai-python/blob/main/src/openai/_streaming.py) 与 `lib/streaming/_deltas.py`;[vLLM reasoning outputs](https://docs.vllm.ai/en/latest/features/reasoning_outputs/);GLM finish 枚举来自第三方镜像,实现前以真实中转探针再核;Gemini `extra_content` 路径未核实原文,测试按通用嵌套对象构造。
+
+## 11. 审查修订记录
+
+| 轮次 | 意见 | 处置 |
+| --- | --- | --- |
+| R1 | B1 `tool_calls:null` 被快照拒(非流式现存) | 采纳: §3/§5.1/§10 |
+| R1 | B2 结局表行重叠 | 采纳: 首匹配行序 + `_tool_snapshot_ok` 前置 |
+| R1 | B3 文本"任何产物"不可计算 | 采纳: §4.3 分模式定义 |
+| R1 | S1 eof_after_finish 但 usage 非 measured | 采纳: 条件加 measured |
+| R1 | S2/S3 活性冲突与标签拼接 | 采纳: §5.2/§5.3 |
+| R1 | S4 告警集合无界 | 采纳: §5.4 |
+| R1 | S5 无产物 + length 与 #30 原文冲突 | 采纳(人类授权代决): 保真交付、文本 ResultInvalid、stop/absent 仍 Transient |
+| R1 | S6-S9 缓存顺序、遥测落点、原因码、null 优先 | 采纳: §6/§7/§4.3/§5.1 |
+| R1 | 建议 1-7 | 采纳 1/2/3/5/6/7;建议 4 不标不可缓存(上游确定性,非毒化) |
+| R2 | 结论通过;N1 measured 以最后 usage 判定、N2 非空 reasoning_content | 采纳: §4.2/§4.4 |
+| R2 | 建议 1-4: 命中回退值、未知 finish 告警文案、数组字符串活性与残余风险、null 五层与数组 null 元素 | 全部采纳: §6/§4.1/§5.3/§5.1 |
+| 代码审查 | I1 保真可缓存须同时要求投影正文非空(think 标签剥净后为空不缓存) | 采纳: `_is_cacheable` |
+| 代码审查 | I2 无参函数 `arguments` 全程 null → 快照不成立被判 Transient 重试 | 人类授权代决: 归 ResultInvalid `tool_arguments_null`(确定性方言);不改写交付值(改写破坏保真,重试救不回) |
+| 代码审查 | I3 文本 `index:null` 整帧丢弃、`index:False` 冒充 0 | 采纳: null 视 0,bool 拒 |
+| 代码审查 | Minor: `[null]` 产出空 `tool_calls`、非流式缺 role 先于判定、告警路径长度有界、README 表述 | 采纳;index 数组缺 index 整键删除登记于 CHANGELOG 提醒 |
@@ -240,6 +240,16 @@
      "id": "review:2026-10-01-issue29-cache-review",
      "label": "保真缓存审查修复与1.4.0准备",
      "type": "review"
+    },
+    {
+      "id": "plan:2026-09-10-24-hedged-requests",
+      "label": "issue #24 长尾对冲请求与裸生成时间实施计划",
+      "type": "plan"
+    },
+    {
+      "id": "design:2026-09-10-24-hedged-requests-design",
+      "label": "长尾对冲请求(issue #24)设计",
+      "type": "design"
    }
  ],
  "links": [
@@ -508,6 +518,13 @@
      "relation": "informs",
      "evidence": "统一缓存与transport的有效产物判据，保持内核依赖边界",
      "added": "2026-10-01T18:27:16.434742+00:00"
+    },
+    {
+      "source": "plan:2026-09-10-24-hedged-requests",
+      "target": "design:2026-09-10-24-hedged-requests-design",
+      "relation": "implements",
+      "evidence": "计划头部与设计 §9 批准状态(2026-09-10 H1-H8 全数获批)",
+      "added": "2026-09-10T16:48:37.849653+00:00"
    }
  ]
 }
@@ -1,8 +1,9 @@
 # Research Wiki 索引

-> 自动生成，更新时间：2026-10-01 18:27 UTC
+> 自动生成，更新时间：2026-10-02 02:23 UTC

 ## design (47)
+
 - [1.3.4 推理意图与测试证据设计](designs/2026-09-09-134-thinking-contracts-design.md) `design:2026-09-09-134-thinking-contracts-design`
 - [2026-07-20-m1-core-design](designs/2026-07-20-m1-core-design.md) `design:2026-07-20-m1-core-design`
 - [2026-07-20-m2-distributed-design](designs/2026-07-20-m2-distributed-design.md) `design:2026-07-20-m2-distributed-design`
@@ -52,6 +53,7 @@
 - [采样参数透传设计(issue #4)](designs/sampling-params.md) `design:sampling-params`

 ## finding (22)
+
 - [1.3.4 推理契约验证与发布准备](findings/2026-09-09-134-thinking-contracts-validation.md) `finding:2026-09-09-134-thinking-contracts-validation`
 - [1.3.5 T2/T3/T4 验收证据：36 列遥测、失败终态、PG 存储兼容与变异矩阵](findings/2026-09-09-135-call-observability-validation.md) `finding:2026-09-09-135-call-observability-validation`
 - [1.3.5 发布完成与外部验收：合并后门、registry 产物与页面亲查](findings/2026-09-09-135-release-completion.md) `finding:2026-09-09-135-release-completion`
@@ -76,6 +78,7 @@
 - [推理开关与 reasoning_tokens: 供应商实测与业界做法](findings/2026-08-02-thinking-switch-and-reasoning-tokens.md) `finding:2026-08-02-thinking-switch-and-reasoning-tokens`

 ## plan (42)
+
 - [1.3.4 推理契约实施计划](plans/2026-09-09-134-thinking-contracts.md) `plan:2026-09-09-134-thinking-contracts`
 - [1.3.5 逻辑调用统计与结构化失败诊断实施计划](plans/2026-09-09-135-call-observability.md) `plan:2026-09-09-135-call-observability`
 - [1.3.6 可选调用期限与取消结算修复实施计划](plans/2026-09-10-136-call-deadline.md) `plan:2026-09-10-136-call-deadline`
@@ -120,6 +123,7 @@
 - [采样参数透传实现计划(issue #4)](plans/sampling-params-plan.md) `plan:sampling-params-plan`

 ## review (6)
+
 - [2026-09-11-138-branch-review](reviews/2026-09-11-138-branch-review.md) `review:2026-09-11-138-branch-review`
 - [2026-09-11-138-verifier](reviews/2026-09-11-138-verifier.md) `review:2026-09-11-138-verifier`
 - [2026-09-11-streaming-stall-design-review](reviews/2026-09-11-streaming-stall-design-review.md) `review:2026-09-11-streaming-stall-design-review`
@@ -128,8 +132,10 @@
 - [整分支审查: issue #14 熔断等待档](reviews/issue14-branch-review.md) `review:issue14-branch-review`

 ## schema (1)
+
 - [表结构: llm_calls(遥测 36 字段)](schemas/llm-calls.md) `schema:llm-calls`

 ## metric (2)
+
 - [OCR 治理调用成功率与错误分类分布](metrics/ocr-call-success.md) `metric:ocr-call-success`
 - [每次调用必录覆盖率(含缓存命中/失败/取消)](metrics/call-telemetry-coverage.md) `metric:call-telemetry-coverage`
@@ -160,6 +160,8 @@
 - [2026-09-09 13:42 UTC] 重建索引: 100 篇页面
 - [2026-09-09 16:49 UTC] 重建索引: 101 篇页面
 - [2026-09-09 18:01 UTC] 重建索引: 102 篇页面
+- [2026-09-10 16:48 UTC] 新增边: plan:2026-09-10-24-hedged-requests --implements--> design:2026-09-10-24-hedged-requests-design
+- [2026-09-10 16:48 UTC] 重建索引: 107 篇页面
 - [2026-10-01 17:52 UTC] 新增 design: OpenAI消息保真设计 (design:2026-10-01-issue29-openai-fidelity-design)
 - [2026-10-01 17:52 UTC] 新增 plan: OpenAI消息保真实现计划 (plan:2026-10-01-issue29-openai-fidelity)
 - [2026-10-01 17:52 UTC] 新增边: plan:2026-10-01-issue29-openai-fidelity --implements--> design:2026-10-01-issue29-openai-fidelity-design
@@ -168,3 +170,4 @@
 - [2026-10-01 18:24 UTC] 新增 review: 保真缓存审查修复与1.4.0准备 (review:2026-10-01-issue29-cache-review)
 - [2026-10-01 18:27 UTC] 新增边: review:2026-10-01-issue29-cache-review --informs--> design:2026-10-01-issue29-openai-fidelity-design
 - [2026-10-01 18:27 UTC] 重建索引: 120 篇页面
+- [2026-10-02 02:23 UTC] 重建索引: 120 篇页面
@@ -202,3 +202,7 @@ stack = ExtractionProviderStack(
 | **AIMD 自适应并发**(库常量: 每源初始 8、429 削减 ×0.5、成功 +1/limit、下限 1、ceiling=max(64, 源级 max_concurrency)) | 冷启动每源并发从"无上限"变为 8 起步爬升——高吞吐下游首分钟吞吐低于旧行为;无 env 开关(韧性常量,同 jitter 系数先例),需要禁用者构造函数注入自定义 pacer |
 | 连败通道受健康证据抑制(窗口样本 ≥ min_calls 且失败率 < fail_rate 时 5 连败不开路) | CHS "连败必开"在高流量健康源上不再成立(防随机噪声误熔唯一好源);冷启动/低流量语义不变 |
 | 结构化重问缺省 1→2 | 解析失败时最多多一跳成本;`PGW_STRUCTURED_MAX_RETRIES=1` 可显式还原 |
+
+## 1.3.10 影响核对(2026-10-14,#30/#31)
+
+CHS 已把 `ResultInvalidError` 作终态(坏结果≠坏服务,熔断记成功),1.3.10 文本 length/过滤且无正文落该分类与既有语义同向,worker `_TERMINAL` 分支无需改;缺 `[DONE]` 而完整的流从此直接成功。
@@ -170,3 +170,7 @@ loop = AgentLoop(client, max_steps=...,
 | R7 | 内存限流在多进程(arq)下的口径须在文档/配置中显式警示(B7 过渡期语义),避免误配全局限额 | M1 |

 以上 G1/G2 若按本文方向解决,GovDoc 全量迁移中业务代码(agent/workflow/govdoc api)对治理层的消费**零改动**,改动收敛于: 装配净新增、types.py/protocols.py/test_imports.py 三处改写、`.env` 改键。
+
+## 1.3.10 影响核对(2026-10-14,#30/#31)
+
+装配处 `retryable_exceptions=(TransientError, AllSourcesExhausted)`(§4)**不含** `ResultInvalidError`:文本 length/过滤且无正文改抛它后,AgentLoop 步级重试**不再**重抽该步——这正是本次修复的意图(确定性结局重抽只花钱)。业务侧若想对 length 截断做"加大 max_tokens 再试",须显式 except `ResultInvalidError` 并按异常消息原因码 `length_without_content` 分支;其余缺 `[DONE]` 而完整的流(#31)从此直接成功。
@@ -200,3 +200,7 @@ ocr = FrameOcrAdapter(OcrTextPort_from_env()) # 业务适配器包装库 OCR 端
 ---

 *不确定项声明*: 库侧 `from_env(role=...)`、OCR 装配工厂的具体 API 名称以 M1/M3 设计文档定稿为准，本文 §5.2/§6 中相应伪代码届时同步修订；`telemetry.db` 是否存在项目外的离线分析脚本消费方未能实测（项目内 grep 无读取方）。
+
+## 1.3.10 影响核对(2026-10-14,#30/#31)
+
+AgentLoop 注入的 `(TransientError, AllSourcesExhausted)`(§4 点 3)**不含** `ResultInvalidError`:文本 length/过滤且无正文改抛它后,步级重试不再重抽该步(意图如此)。需要"加大 max_tokens 再试"的业务须显式 except 并按原因码 `length_without_content` 分支。§保留项 3"缺 `[DONE]` 判瞬时、不写缓存"在 1.3.10 收窄为"无 finish 或 finish 后无实测 usage 才算截断";finish+实测 usage+干净 EOF 视完整。
@@ -228,7 +228,7 @@ generation_sink.append((wall_ms, stall_ms_i, stall_count_i))
 | --- | --- |
 | loader | `_load_hedge`(:745-769)扩读 `STALL_AFTER_S`:`_first` + `_cast` + `ensure_call_deadline`(origin 传实际命中键名),返回字典加 `"hedge_stall_after_s"` 键;未设 = None(关闭) |
 | 字段 | `GatewaySettings` 加 `hedge_stall_after_s: float \| None = None`;`_validate_hedge`(:382-392)扩传 |
-| 守卫(全部落进 `check_hedge_assembly` :772-849,签名加 keyword-only `stall_after_s` 参;**返回契约钉死(审查 B2)**:返回 `tuple[float | None, float | None]` = (归一化 after_s, 归一化 stall_after_s),`config.py:384` 与 `client.py:252` 两处调用点随之解构,`test_config.py:1174/:1184` 列入 T4 同步清单(`is None` 断言按新返回形态改写);**两阈值皆 None 才早退**——"同 `after is None` 早退形态"字面执行会让仅 stall 配置绕过守卫②③④,禁止) | ①值域 = `ensure_call_deadline` 同款(有限正数);②`stall_after_s ≥ min over sources of (inter_token_timeout_s or timeout_s)`(有效最小值,与 openai_compat.py:573 回落同款)→ **ValueError**(对冲窗永不开,装配期直接拒绝;stall 场景无非流式残留收益,warning 论证不足,N5);③`stall_after_s ≥ call_deadline_s`(两者皆设)→ ValueError(与 `after_s` 同款,config.py:825-828 邻域);④单源 scope 设阈值 → warning 允许(1.3.7 §5 同款);⑤**B4**:Embedding/OCR scope 装配 warning 落 `EmbeddingSettings.__post_init__`(:865-869 邻域)与 `OcrSettings`(新增 `__post_init__`):内嵌 gateway 的 `hedge_after_s` **或** `hedge_stall_after_s` 非 None → warning"对冲键仅对 chat scope 生效,本 scope 非流式静默退化"(1.3.7 `AFTER_S` 同病顺带补齐,缺陷修复不扩公共面);chat 侧装配 warning 明确不做(per-call stream,装配期无信号,B4 裁决) |
+| 守卫(全部落进 `check_hedge_assembly` :772-849,签名加 keyword-only `stall_after_s` 参;**返回契约钉死(审查 B2)**:返回 `tuple[float | None, float | None]`= (归一化 after_s, 归一化 stall_after_s),`config.py:384` 与 `client.py:252`两处调用点随之解构,`test_config.py:1174/:1184`列入 T4 同步清单(`is None` 断言按新返回形态改写);**两阈值皆 None 才早退**——"同 `after is None` 早退形态"字面执行会让仅 stall 配置绕过守卫②③④,禁止) | ①值域 = `ensure_call_deadline` 同款(有限正数);②`stall_after_s ≥ min over sources of (inter_token_timeout_s or timeout_s)`(有效最小值,与 openai_compat.py:573 回落同款)→ **ValueError**(对冲窗永不开,装配期直接拒绝;stall 场景无非流式残留收益,warning 论证不足,N5);③`stall_after_s ≥ call_deadline_s`(两者皆设)→ ValueError(与 `after_s` 同款,config.py:825-828 邻域);④单源 scope 设阈值 → warning 允许(1.3.7 §5 同款);⑤**B4**:Embedding/OCR scope 装配 warning 落 `EmbeddingSettings.__post_init__`(:865-869 邻域)与 `OcrSettings`(新增 `__post_init__`):内嵌 gateway 的 `hedge_after_s` **或** `hedge_stall_after_s` 非 None → warning"对冲键仅对 chat scope 生效,本 scope 非流式静默退化"(1.3.7 `AFTER_S` 同病顺带补齐,缺陷修复不扩公共面);chat 侧装配 warning 明确不做(per-call stream,装配期无信号,B4 裁决) |
 | client | `GatewayClient.__init__` 加 `hedge_stall_after_s: float \| None = None` keyword-only,:252-257 守卫调用扩传;RetryMW 构造(:264-283)传 `stall_after_s=` 归一化值(取守卫返回二元组第二元,B2);`from_settings`(:525 邻域)透传 |

 ### 3.6 T4:retry 编排改动(P2 唯一实现形态)
@@ -7,7 +7,7 @@ date: 2026-10-01

 # #29 OpenAI 消息保真实现计划

-> 状态：计划及首轮代码独审已完成，P1/P2 均采纳；缓存空产物和测试双向依赖已修复，1.4.0 发布准备由本轮落实，最终独审与 T8 外部发布仍由父会话执行。
+> 状态：计划及首轮代码独审已完成，P1/P2 均采纳；缓存空产物和测试双向依赖已修复，发布目标按人类最终指定改为 1.3.9（取代未发布的 1.4.0），最终独审与 T8 外部发布仍由父会话执行。
 > 设计：[批准方向的设计落实稿](../designs/2026-10-01-issue29-openai-fidelity-design.md)。
 > 目标：同一 chat 治理链返回可供下游继续工具对话的完整单 choice 协议结果，拒绝截断和 error 帧伪成功。
 > 方案：增加 response_mode 与嵌套 completion；协议解析与聚合内聚在 transport 内部模块，治理层仅传模式/搬结果；旧文本默认保持兼容。
@@ -16,7 +16,7 @@ date: 2026-10-01
 ## 0. 工作区、审批与执行纪律

 | 项目 | 已核验值 |
-|---|---|
+| --- | --- |
 | 工作目录 | `/home/iomgaa/.config/superpowers/worktrees/PolyGateway/issue29-openai-fidelity` |
 | 分支 | `feat/issue29-openai-fidelity` |
 | 基线 | `0e29deac1a72bfee348ed2f8061ea2c33be217e8` / 1.3.8 |
@@ -33,7 +33,7 @@ date: 2026-10-01
 ## 1. 文件结构与职责锁定

 | 路径（相对工作目录） | 动作与职责 |
-|---|---|
+| --- | --- |
 | `src/polygateway/types.py` | 新 ResponseMode/OpenAIChatCompletion；末尾追加 ChatRequest.response_mode、两个响应 completion |
 | `src/polygateway/__init__.py` | 导出新公共类型；最终发布时同步版本 |
 | `src/polygateway/ports.py` | complete 增必填模式参数；说明工具进展与单 choice 契约 |
@@ -147,12 +147,12 @@ conda run -n PolyGateway pytest tests/unit/test_types.py tests/unit/test_client.
 补测 SSE/非流式/缓存缺 role、非 assistant role；合法 role-only 起始帧后工具成功。先红测试至少含下表各组；复用现有 SSE/MockTransport，不引入第二套 HTTP fake 框架。

 | 组 | 输入与精确预期 |
-|---|---|
+| --- | --- |
 | 并行工具 | 2 工具交错 index；同帧两工具；name/arguments 分片；空白/Unicode/引号逐字相等；最终按 index 排序 |
 | 身份 | id/type 同值重复只保留一次，冲突失败；缺 id/type/name/arguments、重复 id、bool/负数 index、非 function 类型失败 |
 | 文本 | 同帧正文+推理均保留；缺省/null/空串区别；原消息 think 标签不剥，旧投影仍剥 |
 | envelope | 顶层数组/null、choices/delta/tool_calls 非容器、坏字段类型、多 choice/非零 choice index 均受控失败 |
-| 非流式 |完整消息、扩展字段、usage/finish/id 全保留；工具参数不做业务 JSON 修复；原响应对象改动不影响快照 |
+| 非流式 | 完整消息、扩展字段、usage/finish/id 全保留；工具参数不做业务 JSON 修复；原响应对象改动不影响快照 |
 | error | 首帧 error、正文后 error+usage+DONE、error=null、error={}、salvage 配置下 error 都立即失败；text/openai 两模式参数化 |

 复用现有 summarize_body/compose_message，断言异常与遥测诊断有界，不打印工具参数正文。默认文本模式不因无用工具字段的严格 finish 校验而新增失败，但不吞顶层 error。无新 HTTP 状态映射。
@@ -198,7 +198,7 @@ conda run -n PolyGateway lint-imports
 先红→绿证据须覆盖真实 OpenAICompatTransport 接入 GatewayClient，而不是只用返回预制 completion 的 fake 证明解析成功：

 | 场景 | 断言 |
-|---|---|
+| --- | --- |
 | 第一次部分正文/工具→error→DONE，第二次正常 | 只返回第二次 completion，attempts=2；第一次 error 的字段不会污染结果或缓存 |
 | 工具持续进展 | 首 token 对冲及时止窗，中段 event 复位，未错误触发 |
 | 真沉默触发对冲 | 副路完整工具结果胜出，返回自己的 id/usage，输家不泄露半成品 |
@@ -220,7 +220,7 @@ conda run -n PolyGateway pytest tests/integration/test_governance_stack.py tests
 文件：LICENSE、pyproject.toml、test_package.py、新 license-audit finding、README、CHANGELOG、.env.example、ARCHITECTURE；设计/计划 Wiki 注册关系。

 | 工作 | 验收 |
-|---|---|
+| --- | --- |
 | 来源审计复核（初审前置 T0） | 查 tracked 分发文件的移植/参考/版权注释及具体片段；列内部原创授权记录，外部只读 reference 的路径/许可/实际使用范围；未发现不等于法律认证 |
 | 第三方义务 | 对实际复制片段核实许可、保留通知；发现 AGPL 不兼容复制须上报，不借用户授权覆盖第三方 |
 | LICENSE | 使用标准 Apache-2.0 全文，不自行重写；版权归属沿项目元数据，不猜用户真实姓名；NOTICE 仅在实际需要时新增 |
@@ -258,10 +258,10 @@ make ci

 ### T8 — 发布与 issue 外部验收（父会话持有，不由本设计任务执行）

-发布版本由父会话按新增公共能力确定，不能在规划中冒称已发布。遵守 CLAUDE.md §4.4.1 全步骤：README/CHANGELOG/两处版本先完成 → 合并 main/push → 合并后 lint/全套件/相交 slow → tag push → 构建/twine check → registry 上传 → 实际下载核验 → Release/挂库/包页核验。
+发布版本已由人类最终指定为 1.3.9（取代未发布的 1.4.0），不能在规划中冒称已发布。遵守 CLAUDE.md §4.4.1 全步骤：README/CHANGELOG/两处版本先完成 → 合并 main/push → 合并后 lint/全套件/相交 slow → tag push → 构建/twine check → registry 上传 → 实际下载核验 → Release/挂库/包页核验。

 | 发布门 | 可接受证据 |
-|---|---|
+| --- | --- |
 | 构建 | `conda run -n PolyGateway python -m build` 和 twine check 成功；解包核 LICENSE、许可元数据、completion 类型/参数/解析器均在产物中 |
 | 安装 | 在独立临时环境从私有索引下载/安装已发布版本；必要时 --no-build-isolation 避免索引无 setuptools；核来源/hash，不误用本地 editable |
 | 外部页面 | registry 包正文与仓库链接、Gitea Release/tag 实际可见；不以 POST 返回码代替最终状态 |
@@ -271,7 +271,7 @@ make ci
 ## 4. 自审清单与残余边界

 | 需求 | 覆盖任务 |
-|---|---|
+| --- | --- |
 | 工具双并行、消息往返、raw=max、非流式 | T1–T3、T7 |
 | error+DONE 所有模式失败、attempts=2 | T2、T5 |
 | 缺 DONE 不 salvage 工具、length 不修复 | T3 |
@@ -288,7 +288,7 @@ P1 最终 assistant role、P2 来源核查前置均已采纳，证据见来源
 T1–T6 代码/文档/许可已落实；T7 工具真实冒烟、Redis、质量门已执行，独立代码审查由父会话接续；T8 禁止本 worker 执行。

 | 验证 | 实际结果/证据 |
-|---|---|
+| --- | --- |
 | 基线红 | tests/outputs/issue29/baseline-red.log：error+DONE 未抛、双文本丢 reasoning、模式 API 缺失，3 failed |
 | 扩展红绿 | expanded-red.log 19 failed（git archive 0e29dea 的 src，不改主树）；expanded-green.log 同19 passed |
 | 防御红绿 | defense-red.log 3 failed：usage 裸错误/空 tool_calls 丢失；defense-green.log 同3 passed |
@@ -305,13 +305,12 @@ Wiki add_entity 会覆写无frontmatter同名文件，现场检查发现后已

 补充最终证据：unit-final.log 1678 passed（23.84s），smoke.log + smoke.exit=0 为既有真实 chat 冒烟4 passed（97.18s）；未跑昂贵全模型矩阵。源代码最后变更早于 ci-final，后续只增测试/文档；最后单元覆盖全部新增测试。

-
 ## 6. 审查修复与1.4.0准备（2026-10-01，未发布）

 已读取首轮独立报告并核验采纳 P2 空产物缓存、Minor 测试双向依赖，详见 [审查处置](../reviews/2026-10-01-issue29-cache-review.md)。本轮只在隔离 feature 树写入，主树既有脏文档未改；不合并、不 push/tag/上传、不关issue。

 | 本轮命令/证据（日志均在 tests/outputs/issue29，不提交） | 实际结果 |
-|---|---|
+| --- | --- |
 | 新缓存回归修复前：review-cache-red.log | 10 failed、1 passed：五种无效产物×两种缓存投影误命中；合法tool-only通过 |
 | 四模块定向：review-cache-green.log | 268 passed，17.71s；包含11项新增缓存回归 |
 | 真实Redis+包测试：review-redis-package.log | 13 passed，2.66s（6 Redis、7包） |
@@ -326,3 +325,21 @@ wheel SHA256 `3186d01d7710c3b5ef727b59f95e8c2db0821277a79491ffb913dd2ce34a888d`
 跳过项仍为17个独立slow Redis时间变体及6个缺私有图像语料OCR；110 slow未选中。本轮不重复付费网关：复用前轮流式/非流式工具双轮2通过和chat冒烟4通过，当前生产改动只统一空产物判据，工具/文本正常行为已有定向与全套件证明；合并后slow发布门由父执行。构建有setuptools许可表格弃用提醒（兼容setuptools>=68设计），并有基础conda RequestsDependencyWarning；均不等于当前构建或测试失败。

 按commit skill提交feature全部实现、修复、许可与发布准备，原main脏文件及运行产物不纳入。提交后交父最终独立复核；T8外部发布尚未完成，不能提前关闭#29。
+
+## 7. 版本定案与历史文档归档（2026-10-01 EDT / 2026-10-02 UTC）
+
+用户最终指定本次发布 __1.3.9__，取代尚未打 tag/上传的 1.4.0。上节 1.4.0 构建散列、测试结果与历史提交标题均是当时实证，保留原样，不冒充 1.3.9 产物。README 安装约束、能力标题、CHANGELOG 与两处版本同步；自定义 `Transport.complete(response_mode=...)` 适配要求及旧缓存更换 namespace/salt 提醒保持不变。CHANGELOG 使用执行环境本地日期 2026-10-01（EDT），知识库时间戳明确使用 UTC。
+
+历史六文档独立归档提交 `ce10881`；采纳独审发现，仅补 #24 implements 边缺少的 plan/design 两节点。验证旧新节点/边/日志并集完整、120 个索引链接与计数一致、三文件仅空白/空引用格式、无冲突标记。剩余 15 个历史未注册端点以及 wiki lint 的 15 条孤立/稀疏提示均未扩大处理。原始备份 `/tmp/pgw140-release` 及 stash 保留。
+
+本轮只准备与本地构建，不合并、不 push/tag/上传、不关闭 issue；后续独审及发布外部验收仍是硬门。相关功能代码未变，可复用 `/tmp/pgw140-release/main-ci.log` 的 1828 项通过、95% 覆盖及 `main-slow.log` 的 6 项真实网关通过（205.81s）；本轮另执行包/协议定向、质量检查和 1.3.9 构建解包验证。
+
+| 本轮证据（`/tmp/pgw139-prepare/`） | 实际结果 |
+| --- | --- |
+| source.log / tests.log | 显式 `PYTHONPATH=/home/iomgaa/Projects/PolyGateway/src` 确认源码位置及1.3.9；包+缓存+协议四模块276 passed，12.43s，无新增行为测试 |
+| check.log | `make check` 格式、ruff、依赖契约通过，1 kept/0 broken |
+| docs-validation.log / wiki-lint.log | #24两端点补齐；原新记录保全、120链接计数通过；存量15个端点缺口及15条lint提示未清理 |
+| remote-tags.txt / registry-simple.html | 2026-10-02 UTC读取远端tag与匿名simple均无1.3.9/1.4.0，无版本冲突；发布前仍须复查 |
+| build.log / twine.log / artifacts.log / unpacked-import.log | 1.3.9 wheel/sdist构建、twine及全Python源码/README/LICENSE逐字核验通过；解包路径独立导入成功，38遥测字段实测不变 |
+
+本地 wheel SHA256 `45306557152fd44f8135928041e47924834c3b4039c586f72d521f51b4590dd3`；sdist SHA256 `8f455973e7fef71811e2b86f5a14cb03469d893ac9720b902f522d2b1a8dcd40`。产物位于 `/tmp/pgw139-prepare/dist`，不是已发布包。构建仅有已知许可证表格弃用提示、conda基础环境依赖警告；不为清警告增加依赖。
@@ -0,0 +1,230 @@
+---
+type: plan
+node_id: plan:2026-10-14-issue30-31-completion-verdict
+title: 统一完成判定实现计划(#30/#31)
+date: 2026-10-14
+---
+
+# #30/#31 统一完成判定实现计划
+
+**目标**: 按[设计](../designs/2026-10-14-issue30-31-completion-verdict-design.md)(下称"设计")实现统一完成判定,修复 #30/#31 并发布 1.3.10。
+
+**方案概述**: 在 `types.py` 内核放一组纯判定函数(流完整性 → 结局分类 → 可缓存),transport 两条路径、CacheMW 读写共用;聚合器改为 null 优先 + 未知字段无损合并;`LLMResponse`/`TransportResult` 追加 `finish_reason`/`stream_end`,遥测加 `stream_end` 列。
+
+**技术**: Python 3.12、httpx、pydantic、pytest(asyncio)、SQLite/PG 遥测、Redis 缓存。环境 `conda run -n PolyGateway`。
+
+本计划不涉及参考实现迁移,保真校验不适用。
+
+## 0. 执行纪律
+
+- 分支 `fix/issue30-31-completion-verdict`。每个任务一次提交,经 `commit` skill;pre-commit 钩子会跑全套件 + ruff + "改动触及函数复杂度 < C",故**每个提交点必须全绿**,改到的既有高复杂度函数(`add_chunk` F42、`_complete_stream`/`_complete_once` C、`_rehydrate` C、`OpenAIChatCompletion.__post_init__` D)须在同一提交内拆到 B 及以下;钩子(2026-10-14 经人类同意改为只拦"本次改动触及"的块,存量 ≥C 降为告警,提交 5d2ad40)按 radon 块判定,**类块**(`ChatCompletionAccumulator` C16、`OpenAIChatCompletion` D24)同样须降到 B 及以下。不改 `OpenAICompatTransport.complete()`(C14),新增逻辑一律放在拆出的 helper 中。
+- 每个任务的统一验证(与钩子一致): `conda run -n PolyGateway pytest tests/ -q`(含 integration,本地有 Redis/PG 时真跑)+ `conda run -n PolyGateway ruff check src` + `conda run -n PolyGateway radon cc -n C <本任务改动文件>`(输出必须为空,或仅含本任务未触及的存量块)。各任务列出的子集命令只作开发期快速回路。
+- 测试结果门: 每个任务的新增/改写测试先在未改实现的代码上运行并记录**失败**输出(粘贴进提交说明正文的"Red:"一行,写明失败用例名与断言),再实现至通过。
+- 后台长跑命令末尾禁止接管道(CLAUDE.md §3)。
+
+## 1. 文件结构
+
+| 文件 | 职责变化 |
+| --- | --- |
+| `src/polygateway/types.py` | 新增 `StreamEnd`、`FinishClass`、`_OutputSummary`、`_Verdict` 与七个纯函数(§2 列出;另 `_effective_finish`);`TransportResult`/`LLMResponse` 末尾追加两字段;`OpenAIChatCompletion` 接受 `tool_calls: None` 并复用 `_tool_snapshot_ok` |
+| `src/polygateway/__init__.py` | 导出 `StreamEnd`;`__version__` → 1.3.10(T6) |
+| `src/polygateway/transports/_openai_completion.py` | 聚合器: null 优先、未知字段合并、标签键、冲突路径、finish/usage 位置记录、reasoning 别名;`parse_chat_completion` 容忍 null |
+| `src/polygateway/transports/openai_compat.py` | 两条路径改走判定;有界告警节流;`finish_reason`/`stream_end` 填充;投影别名 |
+| `src/polygateway/middleware/retry.py` | `_build_response` 搬运两字段 |
+| `src/polygateway/middleware/cache.py` | 写/读按 `_is_cacheable`;`stream_end` 校验;finish 回退 |
+| `src/polygateway/ports.py` | `record_llm_call` 追加 `stream_end` |
+| `src/polygateway/middleware/telemetry.py` | `_AttemptUsage.stream_end`、`_record` 搬运、两 emit 显式 None |
+| `src/polygateway/telemetry/schema.py`、`sqlite.py`、`postgres.py` | 新列 + 补列 |
+| tests | 新增 `tests/unit/test_completion_verdict.py`;改写/扩充 `test_openai_completion.py`、`test_openai_fidelity.py`、`test_openai_compat.py`、`test_cache.py`、`test_telemetry.py`、`test_types.py`、`test_ports.py`、`tests/integration/test_postgres_telemetry.py` 及各 `record_llm_call` 替身 |
+| 文档 | README、CHANGELOG、ARCHITECTURE §6.2/§7/§14、1.3.9 设计修订指针、`schemas/llm-calls.md`、`migrations/` 三份、pyproject 版本 |
+
+## 2. 跨任务接口(T1 定义,后续任务只消费)
+
+```python
+StreamEnd = Literal["done", "eof_after_finish", "salvaged"]
+FinishClass = Literal["natural", "truncated", "filtered", "upstream_error", "unknown", "absent"]
+
+def _classify_finish(value: str | None) -> FinishClass: ...
+def _tool_snapshot_ok(message: Mapping[str, Any]) -> bool: ...
+def _stream_is_complete(
+    *, done: bool, finish_reason: str | None, usage_after_finish: bool, usage_measured: bool
+) -> StreamEnd | None: ...
+```
+
+```python
+@dataclass(frozen=True)
+class _OutputSummary:
+    any_output: bool      # 设计 §4.3 "任何产物"
+    body_output: bool     # 设计 §4.3 "正文产物"
+    tools_present: bool   # 至少一个非 null 工具条目(仅保真有意义)
+    tools_ok: bool        # tools_present 时 _tool_snapshot_ok 结果;无工具恒 True
+
+@dataclass(frozen=True)
+class _Verdict:
+    deliver: bool
+    error_kind: Literal["transient", "result_invalid"] | None  # deliver 时为 None
+    code: str | None      # 设计 §4.3 原因码
+```
+
+```python
+def _summarize_text(content: str, thinking: str) -> _OutputSummary: ...
+def _summarize_message(message: Mapping[str, Any], projected_content: str) -> _OutputSummary: ...
+def _judge_outcome(*, mode: ResponseMode, summary: _OutputSummary, finish: FinishClass) -> _Verdict: ...
+def _is_cacheable(
+    *, mode: ResponseMode, content: str, finish_reason: str | None,
+    stream_end: str | None, completion: OpenAIChatCompletion | None,
+) -> bool: ...
+```
+
+缓存读写共用(T1 定义,T4 消费),顶层优先、回退 `completion.finish_reason`,避免读写口径分叉:
+
+```python
+def _effective_finish(finish_reason: Any, completion: OpenAIChatCompletion | None) -> str | None: ...
+```
+
+`types.py` 无 `from __future__ import annotations`: `StreamEnd`/`FinishClass` 必须定义在 `TransportResult`/`LLMResponse` 之前。
+
+聚合器对外只读属性(T2 定义,T3 消费): `finish_reason: str | None`、`finish_seen: bool`、`usage_after_finish: bool`、`conflicts: frozenset[str]`(归一路径)、`raw_message() -> dict[str, Any]`(未校验的合并消息深拷贝,含按 index 排序组装的 `tool_calls`,与 `finish()` 同一组装函数,供 `_tool_snapshot_ok` 判定)、`finish() -> OpenAIChatCompletion`(保真快照,校验失败抛 Transient,与 1.3.9 同)。`ChatDelta` 追加字段 `reasoning: str`(别名片段,默认 `""`)。
+
+## 3. 任务
+
+### T1 内核判定函数与公共字段
+
+- 文件: `src/polygateway/types.py`、`src/polygateway/__init__.py`;测试 `tests/unit/test_completion_verdict.py`(新)、`tests/unit/test_types.py`、`tests/unit/test_package.py`(导出)。
+- 行为:
+  - `_classify_finish` 按设计 §4.1 表,大小写敏感;None → `absent`。
+  - `_tool_snapshot_ok`: `tool_calls` 缺省或 None → True;否则必须是 list,每个非 None 元素满足 1.3.9 规则(非空唯一 id、`type == "function"`、非空 `function.name`、字符串 `arguments`);list 中 None 元素跳过。`OpenAIChatCompletion.__post_init__` 改为调用它(并拆分以降复杂度),`tool_calls: None` 合法且原样保留键。
+  - `_stream_is_complete`: `done` → `"done"`;否则 `finish_reason is not None and usage_after_finish and usage_measured` → `"eof_after_finish"`;否则 None。
+  - `_summarize_text`: any = `content != "" or thinking != ""`;body = `content.strip() != ""`;tools_present False、tools_ok True。
+  - `_summarize_message`: any = 递归遍历 message 中除 `role`、`tool_calls` 外所有值存在非空字符串,或 tools_present;tools_present = `tool_calls` 为含非 None 元素的 list;tools_ok = `_tool_snapshot_ok(message)`;body = `projected_content.strip() != ""` 或 (tools_present 且 tools_ok)。
+  - `_judge_outcome`: 严格按设计 §4.3 九行首匹配;原因码字面量: `upstream_finish_error`、`tool_truncated`、`tool_incomplete`、`length_without_content`、`filtered_without_content`、`empty_completion`。行 2 仅 `mode == "openai"` 时生效。
+  - `_is_cacheable`: 设计 §6 四条,任一成立 False;`stream_end` 不在 `StreamEnd` 域且非 None → False;openai 模式且 `completion is None` → False。
+  - `TransportResult`、`LLMResponse` 末尾追加 `finish_reason: str | None = None`、`stream_end: StreamEnd | None = None`(放在现有最后一个字段之后,不扰动位置构造);docstring 写清 None 语义与 `usage_source` 分工。
+  - 包根导出 `StreamEnd`;`_has_completion_output` 本任务**保留不动**(T3/T4 迁移后删除)。
+- 测试(新文件): `_classify_finish` 全值表参数化;`_stream_is_complete` 真值表(含 usage 早于 finish、usage 非 measured);`_judge_outcome` 九行 × 两模式参数化(每行至少一例,含"正文非空+坏工具+stop → transient tool_incomplete");`_summarize_message` 对 refusal-only、reasoning-only、whitespace content、未知嵌套字符串、`tool_calls: [None]`;`_is_cacheable` 四类否例 + 正例 + 非法 stream_end。`test_types.py`: `OpenAIChatCompletion` 接受 `tool_calls=None`、拒非 list、`[None]` 合法;两新字段默认 None。
+- 验证: `conda run -n PolyGateway pytest tests/unit/test_completion_verdict.py tests/unit/test_types.py tests/unit/test_package.py -q` → 全过;`make lint` 无新增问题。
+- [ ] 提交 `feat: add kernel completion verdict functions and response fields`
+
+### T2 聚合器: null 优先、未知字段合并、finish/usage 记录
+
+- 文件: `src/polygateway/transports/_openai_completion.py`;测试 `tests/unit/test_openai_completion.py`。
+- 保留 `message`、`tools` 实例属性名与语义(`test_openai_completion.py:70,93` 直接依赖)。
+- 行为(设计 §5、§4.2):
+  - 每帧: 校验 chunk 对象;`usage` 非 null 时记录为 `last_usage`,若 `finish_seen` 或同帧 choice 带非 null finish → `usage_after_finish = True`。
+  - choices: 保真维持单 choice/index 0 校验;文本模式取 `index` 缺省或为 0 的第一个 choice,其余忽略。
+  - finish_reason: 保真走 `_identity`(非字符串/冲突 → Transient,同 1.3.9);文本模式首个字符串值生效,非字符串忽略。任一模式见到非 null 即 `finish_seen = True`。
+  - null 五层(设计 §5.1): chunk 层 `choices: null`、choice 层 `delta: null` 视为缺省;例外: 顶层 `error` 键(含 `error: null`)仍由 `_check_chat_error` 按 error 帧拒绝(不动,保 `test_error_before_done_is_failure`)。
+  - delta 与工具条目、function 三层: 值为 None 的键 → 若目标无此键则 `setdefault(key, None)`(保真),不做类型校验;`tool_calls` list 中 None 元素跳过。
+  - 已知字段非 None 时逐字沿用 1.3.9 校验与合并(content/reasoning_content 拼接、role 必须 assistant、工具 id/type `_identity`、name/arguments 拼接)。
+  - 未知字段(delta 层除 role/content/reasoning_content/tool_calls;工具层除 index/id/type/function;function 层除 name/arguments): 交给 `_merge_unknown(target, key, value, path)`,规则见设计 §5.2;标签键集合 `{"index", "type", "id", "format"}`;冲突 → 从 target 删除该键、把归一路径(数组下标替换为 `[]`,如 `delta.reasoning_details[].format`)加入 `conflicts` 与本尝试"封禁路径"集合,后续同路径值丢弃。`_merge_unknown` 返回"是否并入非空字符串片段"。
+  - `reasoning`: 在保真模式作为普通未知字符串合并进 message;两模式都把其片段放入 `ChatDelta.reasoning`。
+  - progress: 1.3.9 已知字段口径 **或** 任一 `_merge_unknown` 返回 True(文本模式也调用合并函数计算活性,但合并结果不对外)。
+  - 新增只读属性见 §2;`finish()` 改为: 组装 message(含排序后工具)后构造 `OpenAIChatCompletion`(校验失败仍抛 Transient)。
+  - `parse_chat_completion` 拆为两步: `check_completion_body(body) -> tuple[dict, str | None]`(只校验 body 为对象、恰好一个 choice、index 缺省或为 0、message 为对象、保真下 finish 为字符串或 null,失败抛 Transient;返回原始 message 与 finish)与 `build_completion(message, body) -> OpenAIChatCompletion`(构造快照,失败抛 Transient)。快照构造前剔除 `tool_calls` 中的 None 元素(与流式跳过一致,避免回传 `[null]` 被上游 400);`tool_calls: null` 原样保留。
+  - `_merge_unknown` 按值类型拆分私有函数(标签键、字符串、标量、对象、index 数组、其他数组),每个 ≤ B。
+  - 拆分 `add_chunk` 为 `_take_meta`、`_take_choice`、`_take_delta`、`_take_tools`、`_take_tool` 等私有方法,每个 ≤ B。
+- 测试: 改写 `test_invalid_delta_is_domain_error` 参数: 删除 `{"audio": {}}`、`{"index":0,"extra":1}` 与 null 已知字段参数(改为正例),保留 role 非 assistant、content 非字符串、tool index 非法等仍应拒绝的参数。新增: `refusal:null`+`tool_calls:null` 帧成功且快照含 `refusal: None`;只有 `reasoning` 的帧 progress=True 且 `ChatDelta.reasoning` 有值;二次构造的 OpenRouter `reasoning_details` 分片(每帧重复 `format`/`id`,`text` 分片)合并后 `format` 不重复、`text` 拼接完整;工具层嵌套对象 `extra_content={"google":{"thought_signature":"s"}}` 原样保留;数值冲突 → 键被删、`conflicts` 含归一路径、调用成功;同路径冲突后再到达的值被丢弃;形态改变(字符串→对象、index 数组→普通数组)即冲突;布尔值冲突;其他数组追加;同帧多个字段/多个未知字符串只产生一次 progress;chunk `choices:null`、choice `delta:null` 视缺省;`role:null`、`arguments:null` 视缺省;全程只有 `role:null` → `finish()` 失败;非流式 `tool_calls:[None, 合法工具]` 快照只含合法工具;`usage_after_finish` 四种顺序(finish 前 usage、finish 同帧、finish 后 usage-only 帧、无 usage);文本模式 index=1 的 choice 被忽略;重复 tool id 不拼接(既有用例保持)。
+- 验证: §0 统一验证;快速回路 `conda run -n PolyGateway pytest tests/unit/test_openai_completion.py tests/unit/test_openai_fidelity.py tests/unit/test_openai_compat.py -q`。T2 期间 transport 仍调用 `parse_chat_completion` 的组合(保留为两步的薄包装),行为不变。
+- [ ] 提交 `fix: merge unknown and null stream fields losslessly in completion accumulator`
+
+### T3 transport 两条路径接入判定
+
+- 文件: `src/polygateway/transports/openai_compat.py`;测试 `tests/unit/test_openai_fidelity.py`、`tests/unit/test_openai_compat.py`。
+- 行为:
+  - `_complete_stream`: 流结束后先算 `usage_measured = _resolve_usage(sink.get("usage") or {})[2] == "measured"`,再 `stream_end = _stream_is_complete(done=..., finish_reason=acc.finish_reason, usage_after_finish=acc.usage_after_finish, usage_measured=...)`。
+    - None 且 openai → Transient `"{源} missing_done: …"`;None 且 text → 沿 `_check_done`(零产物按 `_summarize_text` 的 any 判定 early_eof;`retry` 抛;`salvage` → `stream_end="salvaged"`)。
+    - `salvaged` 时 usage 沿 `_resolve_stream_usage(..., salvaged=True)`;`done`/`eof_after_finish` 时 salvaged=False。
+  - 投影: `thinking` = 非空 reasoning_content 片段拼接,否则 `ChatDelta.reasoning` 片段拼接;再走既有 `_finalize_text`(think 标签)。
+  - 判定: openai → `summary = _summarize_message(acc.raw_message(), content)`;text → `_summarize_text(content, thinking)`;`verdict = _judge_outcome(mode, summary, _classify_finish(acc.finish_reason))`;非 deliver → 按 `error_kind` 抛 `TransientError` 或 `ResultInvalidError`,message 固定格式 `f"{source.name} {code}: finish_reason={fr!r}, thinking_len={len(thinking)}"`,带 `source_name`、`operation="chat"`;`ResultInvalidError` 显式 `raw_text=""`(不回显推理内容)。deliver 且 openai → `acc.finish()` 构造快照。
+  - 删除 `_reject_empty_completion` 与 1.3.9 的"保真缺 DONE"分支;`_check_done` 只保留文本截断语义。
+  - `_complete_once`: 跳过第一步。保真: 先 `check_completion_body` 取原始 message 与 finish,用原始 message(剔除 None 工具后)算 `_summarize_message` 与 `_judge_outcome`,**仅 deliver 时**调 `build_completion` 构造快照——否则坏工具 + length 会在构造时抛 Transient,到不了 `tool_truncated`。文本: finish 取 `choices[0].get("finish_reason")`,非字符串记 None;同一判定;`stream_end=None`;投影别名同上(`message.get("reasoning_content") or message.get("reasoning")`,仅取字符串)。
+  - 两路径 `TransportResult` 填 `finish_reason`、`stream_end`。
+  - 告警节流: 实例属性 `_warned_fields: set[tuple[str, str]]`,上限 1024,饱和时记一次 `"告警集合已饱和,后续同类告警省略"`;对 `acc.conflicts` 中每个路径、以及 `_classify_finish == "unknown"` 的 finish 值(截断 64 字符,文案写明"该 finish_reason 不缓存")各告警一次;日志只含源名与路径/值,不回显字段内容。
+  - 拆分 `_complete_stream`/`_complete_once` 至 ≤ B(建议抽 `_consume_stream`、`_verdict_or_raise`、`_project_text`)。
+- 测试(假 SSE,复用文件内既有 helper):
+  - #31 验收 1-6(设计 §10),断言 `stream_end`、`usage_source`、attempts=1(经 GatewayClient 或 RetryMW 的既有测试设施)。
+  - usage 早于 finish + EOF → 失败;finish 后 usage `{}` → 失败;`network_error`+usage+EOF → Transient。
+  - 文本模式: stop+usage+EOF 无 DONE(`missing_done="retry"`)→ 成功 `eof_after_finish`;无 finish + EOF → retry 抛、salvage 记 `salvaged`。
+  - #30 验收: 保真 reasoning-only+length → 交付,`completion.message["content"]` 缺省或空、`reasoning_content` 完整、`finish_reason=="length"`;文本同输入 → `ResultInvalidError` 且 message 含 `length_without_content`;保真 content 空+length(无任何产物)→ 交付;文本同 → ResultInvalid;保真/文本 零增量+DONE(带 stop、不带 finish)→ Transient(`test_stream_zero_content_with_done_is_transient` 保持不改);非流式 message 带 `tool_calls:null`+`refusal:null` → 成功;工具缺 name + length → ResultInvalid `tool_truncated`;content 非空 + 坏工具 + stop → Transient。
+  - 只有 `reasoning` 字段的流不触发 TTFT 超时(用既有看门狗测试设施设置小 ttft)且文本 `thinking` 等于拼接值;同时有非空 reasoning_content 时以其为准。
+  - 冲突/未知 finish 告警各只一次(loguru sink: `logger.add(list.append, level="WARNING")`,先例 `test_openai_compat.py:877`;caplog 抓不到 loguru);饱和用例预先往 `_warned_fields` 灌 1024 条再触发一次,断言饱和告警一次。
+  - 结局表传输层覆盖: 设计 §4.3 九行 × 两模式,每行至少流式一例 + 非流式一例(行 2 只保真);非流式"工具缺 name + length → ResultInvalid `tool_truncated`";文本 length 用例经 GatewayClient 断言 attempts=1 且熔断窗口失败计数为 0;#30 reasoning-only+length 经 GatewayClient 断言 attempts=1。
+  - 改写 `test_openai_never_salvages`: 断言"无 finish 的 EOF 在 `missing_done=salvage` 下仍失败",并新增"finish+usage+EOF 成功"作为对照。
+- 验证: `conda run -n PolyGateway pytest tests/unit -q` → 全过。
+- [ ] 提交 `fix: judge stream completion by finish reason and trailing usage`
+
+### T4 retry 搬运与缓存读写判据
+
+- 文件: `src/polygateway/middleware/retry.py`、`src/polygateway/middleware/cache.py`、`src/polygateway/types.py`(删除 `_has_completion_output`);测试 `tests/unit/test_cache.py`、`tests/unit/test_retry.py`、`tests/integration/test_redis_cache.py`。
+- 行为:
+  - `_build_response` 搬运 `finish_reason`、`stream_end`。
+  - 写路径: 序列化前 `_is_cacheable(mode=request.response_mode, content=response.content, finish_reason=response.finish_reason, stream_end=response.stream_end, completion=response.completion)` 为 False → `logger.debug` 跳写。
+  - 读路径(`_rehydrate`): 重建 completion 后、`fields.update` 之前: `stored_end = fields.get("stream_end")` 非 None 且不在域 → miss+warning;`finish = fields.get("finish_reason")`,非 None 且非 str → miss;为 None 且 completion 存在 → 回退 `completion.finish_reason`;`_is_cacheable(...)` False → miss+warning;命中覆盖加 `stream_end=None`、`finish_reason=finish`。拆分 `_rehydrate` 至 ≤ B。
+  - 删除 `_has_completion_output` 及其 import(全仓 grep 为零)。
+- 读写判定的 finish 统一取 `_effective_finish(response.finish_reason, response.completion)`(读路径对存储值同样调用)。
+- 测试: 不可缓存六类(salvaged、reasoning-only length 保真、文本 length 带正文、unknown finish、保真 reasoning-only + stop(交付但无正文产物)、`content_filter` 带正文)均不写入(fake cache backend 断言 set 未调用)且第二次调用 attempts=1;手工注入上述条目 → 读视 miss;非法 `stream_end` 注入 → miss;1.3.9 旧文本条目(无新键)→ 命中;1.3.9 保真条目(无顶层 finish、completion.finish_reason="length")→ miss;命中返回 `stream_end is None`、`finish_reason` 为存储/回退值;retry 测试断言 LLMResponse 两字段来自 TransportResult。Redis integration 一例往返两新字段。
+- 验证: `conda run -n PolyGateway pytest tests/unit -q` → 全过;`conda run -n PolyGateway pytest tests/integration/test_redis_cache.py -q`(需本地 Redis)→ 全过。
+- [ ] 提交 `fix: gate cache writes and reads on a shared cacheability verdict`
+
+### T5 遥测 `stream_end` 列与 recorder 端口
+
+- 文件: `src/polygateway/ports.py`、`src/polygateway/middleware/telemetry.py`、`src/polygateway/telemetry/schema.py`、`sqlite.py`、`postgres.py`;测试 `tests/unit/test_telemetry.py`、`tests/unit/test_ports.py`、`tests/integration/test_postgres_telemetry.py`,以及全仓 `def record_llm_call` 测试替身(`grep -rn "def record_llm_call" tests`)。
+- 行为:
+  - `TelemetryRecorder.record_llm_call` 在 `stall_count` 之后追加 keyword-only `stream_end: str | None`(无默认),docstring 注明 1.3.10。
+  - `_AttemptUsage` 加 `stream_end: str | None = None`,`of()` 读 `response.stream_end`;`emit_attempt` 把它传给 `_record`;`_record` 加同名参数并传给 recorder;`emit_cache_hit`、`emit_terminal_failure` 显式 `stream_end=None`。
+  - `schema.py`: SQLite 与 PG CREATE 末尾 `stream_end TEXT`;`SQLITE_BACKFILL`、`_PG_BACKFILL_DECLS` 追加 `("stream_end", "TEXT")`;`COLUMNS` 追加 `"stream_end"`。sqlite/postgres 若有硬编码列数或参数位次随之更新(以 `COLUMNS` 为准)。
+- 替身与计数(按审查实测清单): ① 固定字段 fixture `tests/unit/test_telemetry.py:142` `_record_minimal`、`tests/integration/test_postgres_telemetry.py:121` `_minimal_fields` 补 `stream_end: None`(sqlite 按 `fields[col]` 取值,缺键 KeyError);② 硬编码计数 38→39 / 列清单追加: `test_ports.py:325`、`test_telemetry.py:53`(`_EXPECTED_COLUMNS`)`:271`、`:301-303`、`:779`、`:829`,`test_postgres_telemetry.py:33`、`:778`、`:1151`、`:1352`(执行时以 grep 复核行号);③ `test_ports.py:112` `_DummyRecorder` 补参数,`stream_end` 加入 `:332` 无默认值参数化列表;④ `**fields` 替身(14 处)无需改动。源码 docstring "38 字段"同步: `ports.py:293`、`middleware/telemetry.py:373`、`sqlite.py:146`、`schema.py:8`。形状闸 `middleware/telemetry.py:323` `_assert_recorder_shape` 从协议签名派生,不改闸本身。
+- 测试: 成功流式行写 `done`/`eof_after_finish`;非流式、缓存命中、失败、取消、终态行为 NULL;旧库(无列)auto 档补列后写入;manual 档缺列 → 裁剪写入且告警点名 `stream_end`;形状闸拒绝缺 `stream_end` 参数的 recorder;PG integration 同上(需本地 PG,沿既有 fixture)。所有替身补参数。
+- 验证: `conda run -n PolyGateway pytest tests/unit -q` → 全过;`conda run -n PolyGateway pytest tests/integration/test_postgres_telemetry.py -q` → 全过;`conda run -n PolyGateway python -c "import inspect, polygateway.ports as p; print(len(inspect.signature(p.TelemetryRecorder.record_llm_call).parameters) - 1)"` 记录字段数供 T6 使用。
+- [ ] 提交 `feat: record stream end mode in telemetry`
+
+### T6 文档、版本与另开 issue
+
+- 文件: `README.md`、`CHANGELOG.md`、`research-wiki/ARCHITECTURE.md`、`research-wiki/designs/2026-10-01-issue29-openai-fidelity-design.md`(顶部加"§4.1/§4.2/§5/§6 部分规则已由 2026-10-14 设计修订"指针)、`research-wiki/schemas/llm-calls.md`、`research-wiki/migrations/*.md`、`pyproject.toml`、`src/polygateway/__init__.py`。
+- 内容:
+  - CHANGELOG `## 1.3.10(未发布)`: 修复段(#31 两模式、#30 两部分、非流式 `tool_calls:null`、reasoning 别名、salvage 不再缓存);**行为变更**段(文本 length/过滤无正文 → ResultInvalid 不再经 `AllSourcesExhausted`;未知 finish 不缓存;活性口径增补);**破坏性**段(`record_llm_call` 新增必填 `stream_end`);提醒(工具快照成立+length 时 arguments 可能半截;冲突删键以告警为准;旧缓存污染迁移 namespace/salt)。
+  - README: 能力表补完成判定/`stream_end`;安装约束 `1.3.*` 核对;遥测字段数用 T5 实测值;活性口径一句。
+  - ARCHITECTURE: §6.2 空补全/missing_done 行、§7 `usage_source` 与 `stream_end` 映射、§14 协议/失败/缓存行按设计改写。
+  - `schemas/llm-calls.md`: 加 `stream_end` 列与取值表,修正陈旧计数。
+  - `migrations/` 三份: 核对下游对 `ResultInvalidError` 的 catch 点,写一行结论(有/无受影响路径)。
+  - 版本 1.3.10 两处一致。
+  - `docs-convention.md` §2 清单逐项过: Gitea Wiki 站点下线,由 README/CHANGELOG/ARCHITECTURE/schemas 承接;`.env.example` 无配置键变更(写明);`research-wiki/ROADMAP.md` 状态同步。
+  - 另开 Gitea issue(tea CLI): "文本模式 n>1 时多 choice 增量混流";"文本模式有产物无正文 + stop 的证据收集(决定是否改 ResultInvalid)"。
+- 验证: `grep -n "1.3.10" pyproject.toml src/polygateway/__init__.py`;`conda run -n PolyGateway pytest tests/unit/test_package.py -q` → 过。
+- [ ] 提交 `docs: document completion verdict and bump version to 1.3.10`
+
+### T7 合并前验证、审查与发布
+
+- [ ] `make lint`、`make test` 全绿(覆盖率不降);`tests/integration` 中 Redis/PG 相关全绿。
+- [ ] 派全新上下文 verifier subagent(`verification-before-completion`),逐条对照设计 §4-§7、§10 与本计划验收,输出证据表。
+- [ ] `requesting-code-review` 整分支审查;按 `receiving-code-review` 核验并修复。
+- [ ] slow 子集(按 diff 交集): 动了公开入口与 transport → `pytest -m slow tests/e2e/test_smoke_gateway.py tests/e2e/test_openai_tools_live.py`;不跑 `test_thinking_live.py`(未动 thinking 能力表)。另做一次 magpie/glm 真实探针: 记录是否出现 `eof_after_finish`,作为 #31 外部证据。
+- [ ] 按 CLAUDE.md §4.4.1 九步发布 1.3.10(CHANGELOG 定版日期、合并 main `--no-ff`、main 上重跑、tag、构建、上传、下载验证、Release + 包关联 + 页面核对)。
+- [ ] 在 #30/#31 评论验收证据与版本,关闭。
+
+## 4. 设计覆盖自查
+
+| 设计条目 | 任务 |
+| --- | --- |
+| §3 公共字段、`StreamEnd` 导出、`tool_calls:null` | T1 |
+| §3 recorder 端口、列 | T5 |
+| §4.1 纯函数与 finish 分类 | T1 |
+| §4.2 流完整性、usage 位置与 measured | T2(记录)+T3(判定) |
+| §4.3 结局表、原因码、错误格式 | T1(判定)+T3(翻译) |
+| §4.4 reasoning 别名 | T2+T3 |
+| §5.1-5.3 null、合并、标签键、活性 | T2 |
+| §5.4 告警节流 | T3 |
+| §6 缓存读写 | T4 |
+| §7 遥测落点 | T5 |
+| §8 旧行为审计中的保留项(零增量+DONE、error 帧、多 choice) | T3 回归测试 |
+| §10 文档、另开 issue、发布 | T6+T7 |
+
+## 5. 审查修订记录
+
+| 轮次 | 意见 | 处置 |
+| --- | --- | --- |
+| P1 | B1 非流式先构造快照导致 `tool_truncated` 不可达 | 采纳: T2 拆 `check_completion_body`/`build_completion`,T3 先判定后构造 |
+| P1 | S1 T5 替身/计数清单 | 采纳: T5 逐项列出 |
+| P1 | S2 类块复杂度、S3 验证与钩子一致 | 采纳: §0 |
+| P1 | S4 保留内部属性、S5 五层 null、S6 缓存漏两类、S7 结局表与熔断断言、S8 合并规则测试 | 采纳: T2/T3/T4 |
+| P1 | 建议 1-8 | 全部采纳;建议 4 选"剔除 None 工具元素" |
@@ -7,11 +7,10 @@ date: 2026-10-01

 # 保真缓存审查修复与1.4.0准备

-
 ## 审查采纳

 | 级别/问题 | 核验与处置 |
-|---|---|
+| --- | --- |
 | Important/P2：空协议缓存绕过成功判据 | 已实证：新测试在修复前 10 failed、1 passed；role-only、null/空/空白正文+空工具、reasoning-only，分别带空投影/伪造旧投影均误命中。采纳并修复。 |
 | Minor：协议/治理测试互相导入 | 已核验双向依赖；将帧构造放基础协议测试中，上层 fidelity 单向复用，没有新增 helpers 抽象。 |

@@ -24,3 +23,7 @@ date: 2026-10-01
 ## 发布准备边界

 1.4.0 对应新增公共协议支持及 transport 端口参数；2026-10-01 远端 tags 和匿名 registry simple 最大版本均1.3.8，无1.4.0冲突。README安装约束、能力表、CHANGELOG定版、两处版本同步。新增 `.env.example` 仅四行说明，无凭据。源码/许可来源核查复用已独审 finding；新产物仍须父会话最终复核后发布。本轮不合并、不 push/tag/上传、不关issue；原main脏文件未改。
+
+## 后续版本定案（2026-10-01 EDT / 2026-10-02 UTC）
+
+用户最终指定发布 **1.3.9**，取代本文当时准备但从未发布的 1.4.0。本文标题、历史构建及审查证据保留原值；当前发布目标与新构建核验见[计划 §7](../plans/2026-10-01-issue29-openai-fidelity.md)。版本号变更不撤销自定义 Transport 的 `response_mode` 适配要求。
@@ -184,3 +184,19 @@ OpenAI 保真没有新增遥测表/列；仍经 RetryMW._emit → TelemetryEmitt
 error 帧是失败 attempt，正文与思考仍为兼容投影，不新增工具全文列。
 本页原标题/表格停留 36 字段；实际 `telemetry/schema.py` 已含 stall_ms/stall_count
 共 38 个 INSERT 字段（§7.8 已登记）。本次不改 DDL，真实定义以源码为准。
+
+## 1.3.10 `stream_end` 列（#31，2026-10-14）
+
+`llm_calls` 末尾追加可空 `stream_end TEXT`：INSERT 字段 38 → 39，物理列（含 `created_at`）39 → 40；
+上文"36 字段"标题为历史口径，以 `telemetry/schema.py` 的 `COLUMNS` 为准。
+
+| 值 | 含义 | 行 |
+| --- | --- | --- |
+| `done` | 收到 `[DONE]` | 流式成功 attempt |
+| `eof_after_finish` | 无 DONE，finish + 其后实测 usage + 干净 EOF | 流式成功 attempt |
+| `salvaged` | 文本 `missing_done=salvage` 打捞 | 流式成功 attempt |
+| NULL | 非流式 / 缓存命中 / 失败 / 取消 / 终态行，或补列前旧行 | — |
+
+来源: emitter 经 `_AttemptUsage.of(response)` 读 `LLMResponse.stream_end`（不走 `emit_attempt` kw）；
+命中与终态入口显式 None。评估复用 `metric:call-telemetry-coverage`；新增观察口径
+`SELECT stream_end, count(*) FROM llm_calls WHERE event_kind='attempt' GROUP BY 1` 统计中转不发 DONE 的比例。
@@ -50,11 +50,12 @@ from polygateway.types import (
    OpenAIChatCompletion,
    ResponseMode,
    SourceConfig,
+    StreamEnd,
    TelemetryStatus,
    ThinkingObservation,
 )

-__version__ = "1.4.0"
+__version__ = "1.3.10"

 __all__ = [
    "DEFAULT_PROFILES",
@@ -73,6 +74,7 @@ __all__ = [
    "GovernanceBackendError",
    "LLMResponse",
    "ResponseMode",
+    "StreamEnd",
    "OpenAIChatCompletion",
    "ModelPrice",
    "OcrClient",
@@ -24,7 +24,8 @@ from polygateway.types import (
    OpenAIChatCompletion,
    ResponseMode,
    ThinkingObservation,
-    _has_completion_output,
+    _effective_finish,
+    _is_cacheable,
 )

 if TYPE_CHECKING:
@@ -148,6 +149,54 @@ def build_cache_key(
    return _KEY_PREFIX + hashlib.sha256(payload.encode("utf-8")).hexdigest()


+def _response_cacheable(mode: ResponseMode, response: LLMResponse) -> bool:
+    """写路径判据;与读路径共用 `_is_cacheable` 与 `_effective_finish`,口径不分叉。"""
+    return _is_cacheable(
+        mode=mode,
+        content=response.content,
+        finish_reason=_effective_finish(response.finish_reason, response.completion),
+        stream_end=response.stream_end,
+        completion=response.completion,
+    )
+
+
+def _restore_fields(raw: str, mode: ResponseMode) -> dict[str, Any]:
+    """缓存原文 → LLMResponse 字段;不可缓存或形态坏抛 ValueError(调用方按 miss)。
+
+    可缓存判定对**存储原值**进行、先于命中覆盖: 覆盖后 `stream_end` 恒 None,
+    在那之后判会把 salvaged 条目放行(设计 §6)。
+    """
+    data = json.loads(raw)
+    fields = {k: v for k, v in data.items() if k in _RESPONSE_FIELDS}
+    completion = fields.get("completion")
+    if completion is not None:
+        if not isinstance(completion, dict):
+            raise ValueError("缓存 completion 必须为对象")
+        fields["completion"] = OpenAIChatCompletion(**completion)
+    finish = _effective_finish(fields.get("finish_reason"), fields.get("completion"))
+    if finish is not None and not isinstance(finish, str):
+        raise ValueError("缓存 finish_reason 必须为字符串或 null")
+    if not _is_cacheable(
+        mode=mode,
+        content=fields.get("content", ""),
+        finish_reason=finish,
+        stream_end=fields.get("stream_end"),
+        completion=fields.get("completion"),
+    ):
+        # 非法 stream_end 也落这里: 它关乎内容完整性,不像 thinking_observation
+        # 那样可以宽容降级
+        raise ValueError("缓存条目不可缓存(截断/空正文/非自然结束/非法 stream_end)")
+    fields["finish_reason"] = finish
+    # JSON 里存的是 StrEnum 的字符串值,不转就复活成裸 str,与字段注解分叉
+    # (下游 `is ThinkingObservation.OBSERVED` 会在命中路径上静默为 False);
+    # 键缺失即升级前写入的旧条目,交给 dataclass 默认值
+    if "thinking_observation" in fields:
+        fields["thinking_observation"] = _coerce_observation(fields["thinking_observation"])
+    if "applied_effort" in fields:
+        fields["applied_effort"] = _coerce_applied_effort(fields["applied_effort"])
+    return fields
+
+
 class CacheMW:
    """洋葱第二层(遥测内、结构化外)。

@@ -195,6 +244,11 @@ class CacheMW:
            if hit is not None:
                return hit
        response = await call_next(request)
+        if not _response_cacheable(request.response_mode, response):
+            # 交付照常,只是不进 TTL: 截断/空正文/非自然结束固化后会把一次偶然
+            # 变成必然(设计 §6)。debug 级: 这是常规分流,不是故障
+            logger.debug("结果不可缓存(截断/空正文/非自然结束),跳过写入")
+            return response
        try:
            serialized = self._serialize(response)
        except Exception as exc:
@@ -208,31 +262,15 @@ class CacheMW:
    def _rehydrate(self, raw: str, request: ChatRequest) -> LLMResponse | None:
        """反序列化 + 按本次调用的 structured 档零网络重建;失败按未命中。"""
        try:
-            data = json.loads(raw)
-            fields = {k: v for k, v in data.items() if k in _RESPONSE_FIELDS}
-            completion = fields.get("completion")
-            if completion is not None:
-                if not isinstance(completion, dict):
-                    raise ValueError("缓存 completion 必须为对象")
-                fields["completion"] = OpenAIChatCompletion(**completion)
-            if request.response_mode == "openai":
-                if fields.get("completion") is None:
-                    raise ValueError("保真缓存缺 completion")
-                if not _has_completion_output(fields.get("content", ""), fields["completion"]):
-                    raise ValueError("保真缓存无有效正文或工具产物")
+            fields = _restore_fields(raw, request.response_mode)
            structured_data = self._rebuild_structured(fields.get("content", ""), request)
-            # JSON 里存的是 StrEnum 的字符串值,不转就复活成裸 str,与字段注解分叉
-            # (下游 `is ThinkingObservation.OBSERVED` 会在命中路径上静默为 False);
-            # 键缺失即升级前写入的旧条目,交给 dataclass 默认值
-            if "thinking_observation" in fields:
-                fields["thinking_observation"] = _coerce_observation(fields["thinking_observation"])
-            if "applied_effort" in fields:
-                fields["applied_effort"] = _coerce_applied_effort(fields["applied_effort"])
            fields.update(
                cache_hit=True,
                latency_ms=0,
                ttft_ms=None,
                max_inter_token_ms=None,
+                # 结束方式描述的是"那次流",命中没有流可言(与 ttft_ms 同口径)
+                stream_end=None,
                call_id=str(uuid.uuid4()),
                structured_data=structured_data,
                # 显式覆盖: 历史条目里的 `call_stats` 是个 dict,而 `_RESPONSE_FIELDS`
@@ -741,6 +741,8 @@ class RetryMW:
        return LLMResponse(
            content=result.content,
            completion=result.completion,
+            finish_reason=result.finish_reason,
+            stream_end=result.stream_end,
            thinking=result.thinking,
            model=source.model,
            provider=source.provider,
@@ -221,6 +221,9 @@ class _AttemptUsage:
    # 两字段留在此处只为让"尝试视图"形状完整,默认 None 即"无沉默观测"
    stall_ms: int | None = None
    stall_count: int | None = None
+    # 1.3.10(#31): 与 stall 相反,结束方式**从 response 取**——LLMResponse 本次有
+    # 该字段,免去约 50 处 emit_attempt 调用点加参;失败尝试 None = 无流可言
+    stream_end: str | None = None

    @classmethod
    def of(cls, response: LLMResponse | None) -> _AttemptUsage:
@@ -239,6 +242,7 @@ class _AttemptUsage:
            model_reported=response.model_reported,
            reasoning_tokens=response.reasoning_tokens,
            thinking_observation=response.thinking_observation,
+            stream_end=response.stream_end,
        )


@@ -370,7 +374,7 @@ def _assert_recorder_shape(recorder: TelemetryRecorder) -> None:


 class TelemetryEmitter:
-    """从请求与结果组装 38 字段并写入 recorder;一切写失败降级 warning。"""
+    """从请求与结果组装 39 字段并写入 recorder;一切写失败降级 warning。"""

    def __init__(
        self,
@@ -464,6 +468,7 @@ class TelemetryEmitter:
            total_latency_ms=None,
            stall_ms=stall_ms,
            stall_count=stall_count,
+            stream_end=usage.stream_end,
        )

    async def emit_cache_hit(
@@ -514,6 +519,8 @@ class TelemetryEmitter:
            # 缓存命中行未产生 transport 观测: 无沉默可言,显式 None
            stall_ms=None,
            stall_count=None,
+            # 命中/终态行无流可言(缓存命中已在 CacheMW 置 None,此处显式不信任)
+            stream_end=None,
        )

    async def emit_terminal_failure(
@@ -569,6 +576,8 @@ class TelemetryEmitter:
            # 终态行无具体源与 transport 观测: 显式 None
            stall_ms=None,
            stall_count=None,
+            # 命中/终态行无流可言(缓存命中已在 CacheMW 置 None,此处显式不信任)
+            stream_end=None,
        )

    async def _record(
@@ -615,6 +624,8 @@ class TelemetryEmitter:
        # 本方法只搬运(同 reasoning_effort 先例)——
        stall_ms: int | None,
        stall_count: int | None,
+        # —— 1.3.10: 结束方式(#31);入口已定型,本方法只搬运 ——
+        stream_end: str | None,
    ) -> None:
        try:
            # 成本换算(M2 §6): 成功行按单价换算;缓存命中 0.0(未产生新调用);
@@ -685,6 +696,7 @@ class TelemetryEmitter:
                total_latency_ms=total_latency_ms,
                stall_ms=stall_ms,
                stall_count=stall_count,
+                stream_end=stream_end,
            )
        except asyncio.CancelledError:
            raise
@@ -290,7 +290,7 @@ class TelemetryStatusProvider(Protocol):

@runtime_checkable
 class TelemetryRecorder(Protocol):
-    """遥测后端;38 字段冻结(M1 设计 §4.4 + issue #3/#4/#11/#16/#20 + 1.3.5/1.3.8),唯一调用点是 TelemetryEmitter。
+    """遥测后端;39 字段冻结(M1 设计 §4.4 + issue #3/#4/#11/#16/#20 + 1.3.5/1.3.8/1.3.10),唯一调用点是 TelemetryEmitter。

    新增参数不设默认值: 库外无第三方实现者(三项目迁移时删除了各自的同名
    Protocol),完整签名的成本为零,而少写一列会被 emitter 的降级吞成 warning。
@@ -357,4 +357,5 @@ class TelemetryRecorder(Protocol):
        total_latency_ms: int | None,
        stall_ms: int | None,
        stall_count: int | None,
+        stream_end: str | None,
    ) -> None: ...
@@ -5,8 +5,8 @@
 多处各存一份必然漂移,而漂移的表现是"下游照打印的 SQL 建完表,库仍报缺列"。

 **`COLUMNS` 是 INSERT 字段序,不是物理列序**: 数据库自填的 `created_at` 不在其中(它带
-`DEFAULT now()` / `datetime('now')`,库从不显式写它)。物理表列 = 38 个 INSERT 字段 +
-`created_at` = 39;列数断言一律按物理列数写,两套口径混用是最易错处。
+`DEFAULT now()` / `datetime('now')`,库从不显式写它)。物理表列 = 39 个 INSERT 字段 +
+`created_at` = 40;列数断言一律按物理列数写,两套口径混用是最易错处。

 本模块只依赖标准库: `telemetry/` 与 `backends/`、`transports/`、`structured/` 同层且
 互不依赖(import-linter 契约执法)。
@@ -64,7 +64,8 @@ CREATE TABLE IF NOT EXISTS llm_calls (
    attempts             INTEGER,
    total_latency_ms     INTEGER,
    stall_ms             INTEGER,
-    stall_count          INTEGER
+    stall_count          INTEGER,
+    stream_end           TEXT
 );
 """

@@ -108,7 +109,8 @@ CREATE TABLE IF NOT EXISTS llm_calls (
    attempts             INTEGER,
    total_latency_ms     INTEGER,
    stall_ms             INTEGER,
-    stall_count          INTEGER
+    stall_count          INTEGER,
+    stream_end           TEXT
 );
 """

@@ -145,6 +147,8 @@ SQLITE_BACKFILL = (
    # 没记过这件事",与"观测过且零沉默"(0)是两回事
    ("stall_ms", "INTEGER"),
    ("stall_count", "INTEGER"),
+    # 1.3.10 结束方式(#31): 可空无默认,旧行 NULL = 补列之前没记过;值域见 StreamEnd
+    ("stream_end", "TEXT"),
 )

 # PG 补列的列定义。语句由此派生成两份文本(见下),使"库内执行的那份"与"打印给
@@ -174,6 +178,8 @@ _PG_BACKFILL_DECLS = (
    # 1.3.8 两列,列序与 SQLITE_BACKFILL 逐项对齐(理由同上)
    ("stall_ms", "INTEGER"),
    ("stall_count", "INTEGER"),
+    # 1.3.10 结束方式(#31): 可空无默认,旧行 NULL = 补列之前没记过;值域见 StreamEnd
+    ("stream_end", "TEXT"),
 )

 # 新列排在 created_at 之后: 与旧表 ALTER 追加的位置一致(见 SQLITE_BACKFILL 同款注释)。
@@ -226,6 +232,8 @@ COLUMNS = (
    # —— 1.3.8 沉默归因(issue #27)——
    "stall_ms",
    "stall_count",
+    # —— 1.3.10 结束方式(#31)——
+    "stream_end",
 )

 _COLUMN_SET = frozenset(COLUMNS)
@@ -143,7 +143,7 @@ class SQLiteRecorder:
                    logger.warning("SQLite 遥测补列失败(写入将逐行降级): {}", exc)

    async def record_llm_call(self, **fields: object) -> None:
-        """写一行遥测;字段集合即 38 字段冻结签名(ports.TelemetryRecorder)。
+        """写一行遥测;字段集合即 39 字段冻结签名(ports.TelemetryRecorder)。

        取值按 `self._columns`(manual 档可能已被裁剪),与 `self._insert` 的
        占位符同序——两者必须一起改,分开改就是把值写进错位的列。
@@ -1,11 +1,23 @@
-"""OpenAI 单 choice 消息聚合：仅协议值处理，不做 I/O 或治理。"""
+"""OpenAI 单 choice 消息聚合：仅协议值处理，不做 I/O 或治理。

+1.3.10(#30): null 一律视为缺省;已知字段沿用 1.3.9 语义;未知字段按 OpenAI 流式
+累加惯例无损合并,冲突删键并登记路径——方言是确定性的,拒绝只会让重试重演同一失败。
+"""
+
+import copy
 from dataclasses import dataclass
 from typing import Any

 from polygateway.errors import TransientError
 from polygateway.types import OpenAIChatCompletion, ResponseMode

+_DELTA_KNOWN = frozenset({"role", "content", "reasoning_content", "tool_calls"})
+_TOOL_KNOWN = frozenset({"index", "id", "type", "function"})
+_FUNCTION_KNOWN = frozenset({"name", "arguments"})
+# 标签键: 每帧可能重复下发的身份/格式标记,同值幂等,绝不拼接(拼成 "x-v1x-v1"
+# 回传上游会被拒);也不算生成进展
+_LABEL_KEYS = frozenset({"index", "type", "id", "format"})
+

 def _invalid(reason: str) -> TransientError:
    """协议形态失败，不回显可能敏感的工具参数。"""
@@ -43,6 +55,14 @@ def _append(target: dict[str, Any], key: str, value: Any) -> str:
    return value


+def _assemble(message: dict[str, Any], tools: dict[int, dict[str, Any]]) -> dict[str, Any]:
+    """按 index 排序组装工具;`finish()` 与 `raw_message()` 共用,判定与快照同源。"""
+    assembled = copy.deepcopy(message)
+    if tools:
+        assembled["tool_calls"] = [copy.deepcopy(tools[index]) for index in sorted(tools)]
+    return assembled
+
+
 def _completion(message: dict[str, Any], metadata: dict[str, Any]) -> OpenAIChatCompletion:
    """复用内核最终快照校验并翻译领域错误。"""
    try:
@@ -53,13 +73,129 @@ def _completion(message: dict[str, Any], metadata: dict[str, Any]) -> OpenAIChat
        raise _invalid(str(exc)) from exc


+# —— 未知字段合并(设计 §5.2) ——
+
+
+class _MergeConflictError(Exception):
+    """合并冲突(内部控制流): 由 `_ExtensionMerger.merge_key` 就地收口为删键。"""
+
+
+def _is_indexed(items: list[Any]) -> bool:
+    """数组元素(忽略 null)均为带整数 index 的对象。"""
+    present = [item for item in items if item is not None]
+    return bool(present) and all(
+        isinstance(item, dict) and type(item.get("index")) is int for item in present
+    )
+
+
+def _fragment_progress(key: str, value: Any) -> bool:
+    """不落盘的活性判据(文本模式): 非标签键上的非空字符串片段。"""
+    if isinstance(value, str):
+        return key not in _LABEL_KEYS and value != ""
+    if isinstance(value, dict):
+        return any(_fragment_progress(k, v) for k, v in value.items())
+    if isinstance(value, list):
+        return any(_fragment_progress(key, item) for item in value)
+    return False
+
+
+class _ExtensionMerger:
+    """每尝试私有的未知字段合并器;冲突路径即封禁,后续同路径值丢弃。"""
+
+    def __init__(self) -> None:
+        self.conflicts: set[str] = set()
+        self._array_modes: dict[str, bool] = {}
+
+    def merge_key(self, target: dict[str, Any], key: str, value: Any, path: str) -> bool:
+        """合并一个键;返回是否并入了非空字符串片段(活性)。"""
+        if path in self.conflicts:
+            return False
+        if value is None:
+            target.setdefault(key, None)
+            return False
+        try:
+            merged, grew = self._merge(target.get(key), value, key, path)
+        except _MergeConflictError:
+            target.pop(key, None)
+            self.conflicts.add(path)
+            return False
+        target[key] = merged
+        return grew
+
+    def _merge(self, old: Any, new: Any, key: str, path: str) -> tuple[Any, bool]:
+        if isinstance(new, dict):
+            return self._merge_dict(old, new, path)
+        if isinstance(new, list):
+            return self._merge_list(old, new, key, path)
+        if key in _LABEL_KEYS or not isinstance(new, str):
+            return self._merge_scalar(old, new)
+        if old is not None and not isinstance(old, str):
+            raise _MergeConflictError
+        return (old or "") + new, new != ""
+
+    @staticmethod
+    def _merge_scalar(old: Any, new: Any) -> tuple[Any, bool]:
+        """标签键与数字/布尔: 同值幂等,不同值冲突(不学 SDK 相加,计数相加会造假)。"""
+        if old is None:
+            return new, False
+        if type(old) is type(new) and old == new:
+            return old, False
+        raise _MergeConflictError
+
+    def _merge_dict(self, old: Any, new: dict[str, Any], path: str) -> tuple[Any, bool]:
+        if old is not None and not isinstance(old, dict):
+            raise _MergeConflictError
+        target = old if old is not None else {}
+        grew = False
+        for sub_key, sub_value in new.items():
+            grew |= self.merge_key(target, sub_key, sub_value, f"{path}.{sub_key}")
+        return target, grew
+
+    def _merge_list(self, old: Any, new: list[Any], key: str, path: str) -> tuple[Any, bool]:
+        if old is not None and not isinstance(old, list):
+            raise _MergeConflictError
+        target = old if old is not None else []
+        items = [item for item in new if item is not None]
+        if not items:
+            return target, False
+        indexed = _is_indexed(items)
+        if self._array_modes.setdefault(path, indexed) != indexed:
+            raise _MergeConflictError  # 形态改变: 首次出现决定合并策略
+        if indexed:
+            return target, self._merge_indexed(target, items, f"{path}[]")
+        return target, self._merge_appended(target, items, key, f"{path}[]")
+
+    def _merge_appended(self, target: list[Any], items: list[Any], key: str, path: str) -> bool:
+        grew = False
+        for item in items:
+            merged, item_grew = self._merge(None, item, key, path)
+            target.append(merged)
+            grew |= item_grew
+        return grew
+
+    def _merge_indexed(self, target: list[Any], items: list[Any], path: str) -> bool:
+        grew = False
+        for item in items:
+            slot = next((e for e in target if e.get("index") == item["index"]), None)
+            if slot is None:
+                slot = {}
+                target.append(slot)
+            grew |= self._merge_dict(slot, item, path)[1]
+        return grew
+
+
@dataclass(frozen=True)
 class ChatDelta:
-    """一帧提取结果；progress 表示真实生成进展。"""
+    """一帧提取结果；progress 表示真实生成进展。
+
+    `reasoning` 是 `reasoning` 字段(vLLM/OpenRouter 新名)的片段,投影 thinking 时
+    作为 `reasoning_content` 缺席时的别名(设计 §4.4)。
+    """

    content: str
    reasoning_content: str
    progress: bool
+    reasoning: str = ""


 class ChatCompletionAccumulator:
@@ -71,104 +207,249 @@ class ChatCompletionAccumulator:
        self.message: dict[str, Any] = {}
        self.metadata: dict[str, Any] = {}
        self.tools: dict[int, dict[str, Any]] = {}
+        self._merger = _ExtensionMerger()
+        self._finish_seen = False
+        self._usage_after_finish = False
+
+    # —— 只读视图(transport 判定消费) ——
+
+    @property
+    def finish_reason(self) -> str | None:
+        """choice 0 首个非 null finish_reason;未见为 None。"""
+        return self.metadata.get("finish_reason")
+
+    @property
+    def finish_seen(self) -> bool:
+        """是否已收到 finish_reason。"""
+        return self._finish_seen
+
+    @property
+    def usage_after_finish(self) -> bool:
+        """finish 同帧或之后是否收到非 null usage(设计 §4.2)。"""
+        return self._usage_after_finish
+
+    @property
+    def conflicts(self) -> frozenset[str]:
+        """被删除的冲突键路径(数组下标已归一为 `[]`)。"""
+        return frozenset(self._merger.conflicts)
+
+    def raw_message(self) -> dict[str, Any]:
+        """未校验的合并消息深拷贝(含排序工具),供构造快照前的结局判定。"""
+        return _assemble(self.message, self.tools)
+
+    # —— 聚合 ——

    def add_chunk(self, chunk: dict[str, Any]) -> ChatDelta:
        """聚合一个已解码帧，一帧至多一个活性事件。"""
        chunk = _object(chunk, "chunk")
-        strict = self.mode == "openai"
-        if strict:
+        if self.mode == "openai":
            _identity(self.metadata, "id", chunk.get("id"))
-        if chunk.get("usage") is not None:
-            _object(chunk["usage"], "usage")
-            self.metadata["usage"] = chunk["usage"]
-        choices = chunk.get("choices", [])
+        choice = self._take_choice(chunk.get("choices"))
+        delta = self._take_delta(choice) if choice is not None else ChatDelta("", "", False)
+        self._take_usage(chunk.get("usage"))
+        return delta
+
+    def _take_usage(self, usage: Any) -> None:
+        """usage 在 finish 处理之后登记,使同帧 finish+usage 算作"之后"。"""
+        if usage is None:
+            return
+        self.metadata["usage"] = _object(usage, "usage")
+        if self._finish_seen:
+            self._usage_after_finish = True
+
+    def _take_choice(self, choices: Any) -> dict[str, Any] | None:
+        """选出本帧 choice 并登记 finish;usage-only 帧返回 None。"""
+        if choices is None:
+            return None
        if not isinstance(choices, list):
            raise _invalid("choices 必须为数组")
        if not choices:
-            return ChatDelta("", "", False)
-        if strict and len(choices) != 1:
-            raise _invalid("仅支持单 choice")
-        choice = _object(choices[0], "choice")
-        if strict:
+            return None
+        if self.mode == "openai":
+            if len(choices) != 1:
+                raise _invalid("仅支持单 choice")
+            choice = _object(choices[0], "choice")
            if type(choice.get("index")) is not int or choice["index"] != 0:
                raise _invalid("choice index 必须为 0")
            _identity(self.metadata, "finish_reason", choice.get("finish_reason"))
-        delta = _object(choice.get("delta", {}), "delta")
-        if strict and delta.keys() - {"role", "content", "reasoning_content", "tool_calls"}:
-            raise _invalid("未知 delta 字段")
-        if "role" in delta and strict:
-            if delta["role"] != "assistant":
-                raise _invalid("role 必须为 assistant")
-            self.message["role"] = delta["role"]
-        content = _append(self.message, "content", delta["content"]) if "content" in delta else ""
-        reasoning = (
-            _append(self.message, "reasoning_content", delta["reasoning_content"])
-            if "reasoning_content" in delta
-            else ""
-        )
-        progress = bool(content or reasoning)
+        else:
+            choice = self._text_choice(choices)
+            if choice is None:
+                return None
+        self._finish_seen = self._finish_seen or self.finish_reason is not None
+        return choice
+
+    def _text_choice(self, choices: list[Any]) -> dict[str, Any] | None:
+        """文本模式取 index 缺省或为 0 的首个 choice,不混入其他 choice(n>1)。"""
+        for raw in choices:
+            choice = _object(raw, "choice")
+            index = choice.get("index")
+            # null 视缺省(=0);bool 不得冒充 0(`False == 0` 为真)
+            if index is None or (type(index) is int and index == 0):
+                finish = choice.get("finish_reason")
+                if isinstance(finish, str) and "finish_reason" not in self.metadata:
+                    self.metadata["finish_reason"] = finish
+                return choice
+        return None
+
+    def _take_delta(self, choice: dict[str, Any]) -> ChatDelta:
+        """拆分已知与未知字段;已知字段语义逐字沿用 1.3.9。"""
+        raw_delta = choice.get("delta")
+        delta = _object(raw_delta if raw_delta is not None else {}, "delta")
+        if self.mode == "openai" and "role" in delta:
+            self._take_role(delta["role"])
+        content = self._take_text(delta, "content")
+        reasoning_content = self._take_text(delta, "reasoning_content")
+        alias = delta.get("reasoning") if isinstance(delta.get("reasoning"), str) else ""
+        progress = bool(content or reasoning_content)
        if "tool_calls" in delta:
-            tools = delta["tool_calls"]
-            if not isinstance(tools, list):
-                if strict:
-                    raise _invalid("tool_calls 必须为数组")
-                return ChatDelta(content, reasoning, progress)
-            if strict:
-                self.message.setdefault("tool_calls", [])
-            for raw in tools:
-                if not strict:
-                    # 文本只消费可识别的工具进展，不承诺工具完整性。
-                    if isinstance(raw, dict) and type(raw.get("index")) is int:
-                        index = raw["index"]
-                        tool = self.tools.setdefault(index, {})
-                        for key in ("id", "type"):
-                            if isinstance(raw.get(key), str) and raw[key] and key not in tool:
-                                tool[key] = raw[key]
-                                progress = True
-                        fn = raw.get("function")
-                        if isinstance(fn, dict):
-                            progress |= any(
-                                isinstance(fn.get(k), str) and bool(fn[k])
-                                for k in ("name", "arguments")
-                            )
-                    continue
-                raw = _object(raw, "tool")
-                index = raw.get("index")
-                if type(index) is not int or index < 0:
-                    raise _invalid("tool index 必须为非负整数")
-                if raw.keys() - {"index", "id", "type", "function"}:
-                    raise _invalid("未知工具增量字段")
-                tool = self.tools.setdefault(index, {})
-                for key in ("id", "type"):
-                    progress |= _identity(tool, key, raw.get(key))
-                if "function" in raw:
-                    fn = _object(raw["function"], "function")
-                    if fn.keys() - {"name", "arguments"}:
-                        raise _invalid("未知 function 增量字段")
-                    target = tool.setdefault("function", {})
-                    for key in ("name", "arguments"):
-                        if key in fn:
-                            if not isinstance(fn[key], str):
-                                raise _invalid(f"function.{key} 必须为字符串")
-                            progress |= bool(_append(target, key, fn[key]))
-        return ChatDelta(content, reasoning, progress)
+            progress |= self._take_tools(delta["tool_calls"])
+        progress |= self._take_unknown(delta)
+        return ChatDelta(content, reasoning_content, progress, alias)
+
+    def _take_role(self, role: Any) -> None:
+        if role is None:
+            self.message.setdefault("role", None)
+            return
+        if role != "assistant":
+            raise _invalid("role 必须为 assistant")
+        self.message["role"] = role
+
+    def _take_text(self, delta: dict[str, Any], key: str) -> str:
+        if key not in delta:
+            return ""
+        if self.mode == "openai" or delta[key] is not None:
+            return _append(self.message, key, delta[key])
+        return ""
+
+    def _take_unknown(self, delta: dict[str, Any]) -> bool:
+        """未知 delta 键: 保真合并进 message;文本只算活性。"""
+        grew = False
+        for key, value in delta.items():
+            if key in _DELTA_KNOWN:
+                continue
+            if self.mode == "openai":
+                grew |= self._merger.merge_key(self.message, key, value, f"delta.{key}")
+            else:
+                grew |= _fragment_progress(key, value)
+        return grew
+
+    def _take_tools(self, tools: Any) -> bool:
+        if self.mode != "openai":
+            return isinstance(tools, list) and self._each_tool(tools, self._text_tool_progress)
+        if tools is None:
+            self.message.setdefault("tool_calls", None)
+            return False
+        if not isinstance(tools, list):
+            raise _invalid("tool_calls 必须为数组")
+        present = [tool for tool in tools if tool is not None]
+        if self.message.get("tool_calls") is None and (present or not tools):
+            # 显式 [] 原样保留;[null] 不造空数组(回传空 tool_calls 会被上游拒)
+            self.message["tool_calls"] = []
+        return self._each_tool(present, lambda raw: self._take_tool(_object(raw, "tool")))
+
+    @staticmethod
+    def _each_tool(tools: list[Any], take: Any) -> bool:
+        """逐个非 null 工具增量聚合,返回本帧是否有进展。"""
+        progress = False
+        for raw in tools:
+            if raw is not None:
+                progress |= take(raw)
+        return progress
+
+    def _text_tool_progress(self, raw: Any) -> bool:
+        """文本只消费可识别的工具进展，不承诺工具完整性。"""
+        if not isinstance(raw, dict) or type(raw.get("index")) is not int:
+            return False
+        tool = self.tools.setdefault(raw["index"], {})
+        progress = False
+        for key in ("id", "type"):
+            if isinstance(raw.get(key), str) and raw[key] and key not in tool:
+                tool[key] = raw[key]
+                progress = True
+        fn = raw.get("function")
+        if isinstance(fn, dict):
+            progress |= any(
+                isinstance(fn.get(k), str) and bool(fn[k]) for k in ("name", "arguments")
+            )
+        return progress
+
+    def _take_tool(self, raw: dict[str, Any]) -> bool:
+        index = raw.get("index")
+        if type(index) is not int or index < 0:
+            raise _invalid("tool index 必须为非负整数")
+        tool = self.tools.setdefault(index, {})
+        progress = _identity(tool, "id", raw.get("id"))
+        progress |= _identity(tool, "type", raw.get("type"))
+        if raw.get("function") is not None:
+            progress |= self._take_function(tool, _object(raw["function"], "function"))
+        for key, value in raw.items():
+            if key not in _TOOL_KNOWN:
+                progress |= self._merger.merge_key(tool, key, value, f"tool_calls[].{key}")
+        return progress
+
+    def _take_function(self, tool: dict[str, Any], fn: dict[str, Any]) -> bool:
+        target = tool.setdefault("function", {})
+        progress = False
+        for key, value in fn.items():
+            if key in _FUNCTION_KNOWN:
+                progress |= bool(_append(target, key, value))
+            else:
+                path = f"tool_calls[].function.{key}"
+                progress |= self._merger.merge_key(target, key, value, path)
+        return progress

    def finish(self) -> OpenAIChatCompletion:
        """线路完整性由调用者保证；此处只校验最终消息。"""
-        message = dict(self.message)
-        if self.tools:
-            message["tool_calls"] = [self.tools[index] for index in sorted(self.tools)]
-        return _completion(message, self.metadata)
+        return _completion(self.raw_message(), self.metadata)


-def parse_chat_completion(body: dict[str, Any]) -> OpenAIChatCompletion:
-    """非流式完整消息保留扩展字段，拒绝静默截取多 choice。"""
-    body = _object(body, "body")
+def _strip_null_tools(message: dict[str, Any]) -> dict[str, Any]:
+    """null 工具元素剔除: 与流式跳过同口径,原样回传 [null] 会被上游拒。"""
+    tools = message.get("tool_calls")
+    if isinstance(tools, list) and None in tools:
+        kept = [tool for tool in tools if tool is not None]
+        # 全为 null 视同 tool_calls:null,不造会被上游拒收的空数组
+        return {**message, "tool_calls": kept or None}
+    return message
+
+
+def _single_choice(body: dict[str, Any]) -> dict[str, Any]:
+    """恰好一个 choice,index 缺省或为 0;拒绝静默截取多 choice。"""
    choices = body.get("choices")
    if not isinstance(choices, list) or len(choices) != 1:
        raise _invalid("仅支持单 choice")
    choice = _object(choices[0], "choice")
    if "index" in choice and (type(choice["index"]) is not int or choice["index"] != 0):
        raise _invalid("choice index 必须为 0")
-    metadata = {**body, "finish_reason": choice.get("finish_reason")}
-    return _completion(_object(choice.get("message"), "message"), metadata)
+    return choice
+
+
+def check_completion_body(body: Any) -> tuple[dict[str, Any], str | None]:
+    """非流式保真形态校验(不构造快照): 单 choice、index 为 0、finish 为字符串或 null。
+
+    与快照构造拆开: 结局判定要先于构造(截断的半截工具须判 ResultInvalid,
+    构造时抛 Transient 就到不了那一行,设计 §4.3 行 2)。
+    """
+    choice = _single_choice(_object(body, "body"))
+    finish = choice.get("finish_reason")
+    if finish is not None and not isinstance(finish, str):
+        raise _invalid("finish_reason 必须为字符串或 null")
+    message = _object(choice.get("message"), "message")
+    if message.get("role") != "assistant":
+        # 形态失败先于结局判定,否则缺 role 的截断工具会被归成 ResultInvalid
+        raise _invalid("role 必须为 assistant")
+    return _strip_null_tools(message), finish
+
+
+def build_completion(
+    message: dict[str, Any], body: dict[str, Any], finish: str | None
+) -> OpenAIChatCompletion:
+    """非流式快照构造;保留消息扩展字段。"""
+    return _completion(message, {**body, "finish_reason": finish})
+
+
+def parse_chat_completion(body: dict[str, Any]) -> OpenAIChatCompletion:
+    """非流式完整消息保留扩展字段，拒绝静默截取多 choice。"""
+    message, finish = check_completion_body(body)
+    return build_completion(message, body, finish)
@@ -12,7 +12,7 @@ import json
 import math
 import re
 import time
-from dataclasses import replace
+from dataclasses import dataclass, field, replace
 from typing import TYPE_CHECKING, Any

 import httpx
@@ -41,16 +41,21 @@ from polygateway.transports._http_errors import compose_message, summarize_body
 from polygateway.transports._openai_completion import (
    ChatCompletionAccumulator,
    ChatDelta,
-    parse_chat_completion,
+    build_completion,
+    check_completion_body,
 )
 from polygateway.types import (
    Effort,
    EmbeddingTransportResult,
-    OpenAIChatCompletion,
    ResponseMode,
    SourceConfig,
+    StreamEnd,
    TransportResult,
-    _has_completion_output,
+    _classify_finish,
+    _judge_outcome,
+    _stream_is_complete,
+    _summarize_message,
+    _summarize_text,
 )

 if TYPE_CHECKING:
@@ -58,6 +63,10 @@ if TYPE_CHECKING:
    from collections.abc import AsyncIterator, Callable, Mapping

 _THINK_PATTERN = re.compile(r"<think>(.*?)</think>", re.DOTALL)
+# 方言告警节流上限与 finish 值截断(设计 §5.4): 键由上游决定,必须有界
+_WARN_CAP = 1024
+_FINISH_WARN_CHARS = 64
+_PATH_WARN_CHARS = 128

 # —— SSE 纯函数(VT llm.py 同款)——

@@ -135,6 +144,125 @@ async def _iter_sse_deltas(
            yield delta


+@dataclass
+class _StreamTiming:
+    """一次流式尝试的计时与文本片段(每尝试新建,不跨重试/对冲共享)。"""
+
+    content_parts: list[str] = field(default_factory=list)
+    reasoning_parts: list[str] = field(default_factory=list)
+    alias_parts: list[str] = field(default_factory=list)
+    chunk_times: list[float] = field(default_factory=list)
+    ttft_ms: float | None = None
+    max_gap: float = 0.0
+
+    @property
+    def has_reasoning(self) -> bool:
+        """出现过非空 reasoning_content: 别名 `reasoning` 让位(设计 §4.4)。"""
+        return bool(self.reasoning_parts)
+
+
+async def _consume_deltas(
+    guarded: AsyncIterator[ChatDelta],
+    started: float,
+    first_token_event: asyncio.Event | None,
+    content_event: asyncio.Event | None,
+) -> _StreamTiming:
+    """消费进展事件: 计时、置位对冲事件、收集文本片段。"""
+    timing = _StreamTiming()
+    last = started
+    async for delta in guarded:
+        now = time.monotonic()
+        # 增量 chunk 时间戳(1.3.8 沉默计量): 循环体只会被真实增量触发——
+        # 空 chunk/usage 帧/SSE keepalive 已在聚合器被过滤
+        timing.chunk_times.append(now)
+        if content_event is not None:
+            # 每个增量 chunk 置位一次: 中段沉默对冲的活性复位信号
+            content_event.set()
+        if timing.ttft_ms is None:
+            timing.ttft_ms = (now - started) * 1000
+            # 首个增量即对冲语义上的"首 token"(思考增量同样是存活证据,
+            # 与看门狗活性口径一致);None = 调用方未启用对冲,零分支成本
+            if first_token_event is not None:
+                first_token_event.set()
+        else:
+            timing.max_gap = max(timing.max_gap, (now - last) * 1000)
+        last = now
+        for parts, text in (
+            (timing.content_parts, delta.content),
+            (timing.reasoning_parts, delta.reasoning_content),
+            (timing.alias_parts, delta.reasoning),
+        ):
+            if text:
+                parts.append(text)
+    return timing
+
+
+def _raise_unless_deliverable(
+    mode: ResponseMode,
+    message: dict[str, Any] | None,
+    content: str,
+    thinking: str,
+    finish_reason: str | None,
+    source: SourceConfig,
+) -> None:
+    """结局判定(设计 §4.3)的唯一翻译点: 判定值 → 领域错误。
+
+    原因码以固定格式进 message(遥测 error 列可检索);不回显推理/工具内容。
+    """
+    if mode == "openai" and message is not None:
+        summary = _summarize_message(message, content)
+    else:
+        summary = _summarize_text(content, thinking)
+    verdict = _judge_outcome(mode=mode, summary=summary, finish=_classify_finish(finish_reason))
+    if verdict.deliver:
+        return
+    text = (
+        f"{source.name} {verdict.code}: finish_reason={finish_reason!r}, "
+        f"thinking_len={len(thinking)}"
+    )
+    if verdict.error_kind == "result_invalid":
+        raise ResultInvalidError(text, raw_text="", source_name=source.name, operation="chat")
+    raise TransientError(text, source_name=source.name, operation="chat")
+
+
+def _text_choice(body: dict[str, Any], source: SourceConfig) -> tuple[Any, str | None]:
+    """文本模式取首个 choice 的 message 与 finish(非字符串 finish 记 None)。"""
+    choices = body.get("choices") or []
+    if not choices:
+        raise TransientError(
+            f"{source.name} 响应缺 choices", source_name=source.name, operation="chat"
+        )
+    if not isinstance(choices, list) or not isinstance(choices[0], dict):
+        raise TransientError("chat choices 形态非法", operation="chat")
+    finish = choices[0].get("finish_reason")
+    return choices[0].get("message", {}), finish if isinstance(finish, str) else None
+
+
+def _nonstream_message(
+    body: dict[str, Any], mode: ResponseMode, source: SourceConfig
+) -> tuple[dict[str, Any], str | None]:
+    """非流式取单 choice 的原始 message 与 finish;只校验形态,不构造快照。"""
+    if mode == "openai":
+        message, finish = check_completion_body(body)
+    else:
+        message, finish = _text_choice(body, source)
+    if not isinstance(message, dict) or any(
+        message.get(key) is not None and not isinstance(message[key], str)
+        for key in ("content", "reasoning_content")
+    ):
+        raise TransientError("chat message 形态非法", operation="chat")
+    return message, finish
+
+
+def _reasoning_of(message: dict[str, Any]) -> str:
+    """非流式思考投影: 非空 reasoning_content 优先,否则 `reasoning` 别名。"""
+    primary = message.get("reasoning_content")
+    if isinstance(primary, str) and primary:
+        return primary
+    alias = message.get("reasoning")
+    return alias if isinstance(alias, str) else ""
+
+
 # —— 错误翻译(CHS invokers.py 同款)——


@@ -392,6 +520,9 @@ class OpenAICompatTransport:
        # 都会连带改掉另一侧的行为。(键空间恰好不相交,故当下**不会**互相压制;
        # 分开维护的理由是语义,不是碰撞)
        self._warned_mismatches: set[tuple[str, str, Effort | None]] = set()
+        # 方言告警(1.3.10)另立集合: 冲突删键/未知 finish 与推理对账语义无关
+        self._warned_fields: set[tuple[str, str]] = set()
+        self._warn_saturated = False
        self._client_factory = client_factory or _default_client_factory
        self._clients: dict[str, httpx.AsyncClient] = {}

@@ -636,54 +767,48 @@ class OpenAICompatTransport:
                inter_token_s=source.inter_token_timeout_s or source.timeout_s,
                total_s=source.timeout_s,
            )
-            content_parts: list[str] = []
-            thinking_parts: list[str] = []
-            ttft_ms: float | None = None
-            last = started
-            max_gap = 0.0
-            chunk_times: list[float] = []
-            async for delta in guarded:
-                now = time.monotonic()
-                # 增量 chunk 时间戳(1.3.8 沉默计量): 循环体只会被真实增量触发——
-                # 空 chunk/usage 帧/SSE keepalive 已在 _sse_data_payload/_sse_delta
-                # 被过滤,这里逐字追加即可;阈值判定不在热路径,派生收敛在
-                # _stall_from_chunk_times(设计 §3 派生原则)
-                chunk_times.append(now)
-                if content_event is not None:
-                    # 每个增量 chunk 置位一次: 中段沉默对冲的活性复位信号
-                    content_event.set()
-                if ttft_ms is None:
-                    ttft_ms = (now - started) * 1000
-                    # 首个增量即对冲语义上的"首 token"(思考增量同样是存活证据,
-                    # 与看门狗活性口径一致);None = 调用方未启用对冲,零分支成本
-                    if first_token_event is not None:
-                        first_token_event.set()
-                else:
-                    max_gap = max(max_gap, (now - last) * 1000)
-                last = now
-                if delta.content:
-                    content_parts.append(delta.content)
-                if delta.reasoning_content:
-                    thinking_parts.append(delta.reasoning_content)
-        if response_mode == "openai" and not sink.get("done"):
-            raise TransientError(
-                "保真响应缺 [DONE]，禁止打捞", source_name=source.name, operation="chat"
-            )
+            timing = await _consume_deltas(guarded, started, first_token_event, content_event)
+        return self._stream_result(sink, accumulator, timing, source, profile, response_mode)
+
+    def _stream_result(
+        self,
+        sink: dict[str, Any],
+        accumulator: ChatCompletionAccumulator,
+        timing: _StreamTiming,
+        source: SourceConfig,
+        profile: ProviderProfile,
+        response_mode: ResponseMode,
+    ) -> TransportResult:
+        """流结束后的统一判定(设计 §4.2-§4.3): 完整性 → 结局 → 组装。"""
+        self._warn_dialect(source, accumulator.conflicts, accumulator.finish_reason)
+        thinking_parts = timing.reasoning_parts if timing.has_reasoning else timing.alias_parts
+        content, thinking = self._finalize_text(timing.content_parts, thinking_parts, profile)
+        stream_end = _stream_is_complete(
+            done=bool(sink.get("done")),
+            finish_reason=accumulator.finish_reason,
+            usage_after_finish=accumulator.usage_after_finish,
+            usage_measured=_resolve_usage(sink.get("usage") or {})[2] == "measured",
+        )
+        if stream_end is None:
+            stream_end = self._truncated(timing, thinking_parts, source, response_mode)
+        message = accumulator.raw_message() if response_mode == "openai" else None
+        _raise_unless_deliverable(
+            response_mode, message, content, thinking, accumulator.finish_reason, source
+        )
        completion_result = accumulator.finish() if response_mode == "openai" else None
-        salvaged = self._check_done(sink, content_parts, thinking_parts, source)
-        content, thinking = self._finalize_text(content_parts, thinking_parts, profile)
-        self._reject_empty_completion(content, completion_result, source)
-        prompt, completion, usage_source = _resolve_stream_usage(sink, salvaged)
+        prompt, completion, usage_source = _resolve_stream_usage(sink, stream_end == "salvaged")
        reasoning_tokens = _coerce_reasoning_tokens(sink.get("usage"))
-        stall_ms, stall_count = _stall_from_chunk_times(chunk_times, source.stall_threshold_ms)
+        stall_ms, stall_count = _stall_from_chunk_times(
+            timing.chunk_times, source.stall_threshold_ms
+        )
        return TransportResult(
            content=content,
            thinking=thinking,
            prompt_tokens=prompt,
            completion_tokens=completion,
            usage_source=usage_source,
-            ttft_ms=ttft_ms,
-            max_inter_token_ms=(max_gap if ttft_ms is not None else None),
+            ttft_ms=timing.ttft_ms,
+            max_inter_token_ms=(timing.max_gap if timing.ttft_ms is not None else None),
            stall_ms=stall_ms,
            stall_count=stall_count,
            raw={"usage": sink.get("usage")},
@@ -696,40 +821,56 @@ class OpenAICompatTransport:
            thinking_observation=observe_thinking(
                thinking=thinking, reasoning_tokens=reasoning_tokens
            ),
+            finish_reason=accumulator.finish_reason,
+            stream_end=stream_end,
        )

-    def _check_done(
+    def _truncated(
        self,
-        sink: dict[str, Any],
-        content_parts: list[str],
+        timing: _StreamTiming,
        thinking_parts: list[str],
        source: SourceConfig,
-    ) -> bool:
-        """缺 [DONE] 语义(设计 §6): 零内容恒 retry;有内容按 missing_done 策略。"""
-        if sink.get("done"):
-            return False
+        response_mode: ResponseMode,
+    ) -> StreamEnd:
+        """截断流(设计 §4.2): 保真不打捞;文本按 missing_done,零产物恒 retry。"""
        ctx: dict[str, Any] = {"source_name": source.name, "operation": "chat"}
-        if not content_parts and not thinking_parts:
+        if response_mode == "openai":
+            raise TransientError(f"{source.name} missing_done: 保真响应截断,禁止打捞", **ctx)
+        if not any(timing.content_parts) and not any(thinking_parts):
            raise TransientError(f"{source.name} SSE early_eof: 零内容断流", **ctx)
        if source.missing_done == "retry":
            raise TransientError(f"{source.name} SSE missing_done: 截断且无 [DONE]", **ctx)
-        return True
+        return "salvaged"

-    def _reject_empty_completion(
-        self, content: str, completion: OpenAIChatCompletion | None, source: SourceConfig
+    def _warn_dialect(
+        self, source: SourceConfig, conflicts: frozenset[str], finish_reason: str | None
    ) -> None:
-        """空补全 → 瞬时错误(2026-07-20 人类裁决,M1 验证发现)。
-
-        服务 200 且流程完整([DONE]/usage 正常)但 content 为空——MiniMax 等
-        网关的间歇异常形态。视为服务抖动: 退避重试/换源,**绝不缓存空响应**;
-        承 CHS "VLM 零 content"归瞬时的先例(invokers.py:309)。
-        """
-        if not _has_completion_output(content, completion):
-            raise TransientError(
-                f"{source.name} 空补全(empty_completion): 流程完整但零内容",
-                source_name=source.name,
-                operation="chat",
+        """方言告警: 冲突删键与未登记 finish 各按 (源, 路径/值) 只喊一次。"""
+        for raw_path in sorted(conflicts):
+            path = raw_path[:_PATH_WARN_CHARS]  # 键名由上游决定,长度同样须有界
+            self._warn_once(
+                source.name, f"field:{path}", f"字段 {path} 合并冲突已删除,保真结果缺该字段"
            )
+        if finish_reason is not None and _classify_finish(finish_reason) == "unknown":
+            value = finish_reason[:_FINISH_WARN_CHARS]
+            self._warn_once(
+                source.name,
+                f"finish:{value}",
+                f"未登记的 finish_reason {value!r},按正常结束交付,该值的结果不缓存",
+            )
+
+    def _warn_once(self, source_name: str, key: str, text: str) -> None:
+        """有界节流: 键由上游决定,长跑 worker 不能让集合无限增长(设计 §5.4)。"""
+        if (source_name, key) in self._warned_fields:
+            return
+        if len(self._warned_fields) >= _WARN_CAP:
+            if not self._warn_saturated:
+                self._warn_saturated = True
+                logger.warning("方言告警节流集合已饱和({}),后续同类告警省略", _WARN_CAP)
+            return
+        self._warned_fields.add((source_name, key))
+        # 单参数传入 loguru: 文案可能含上游给的花括号(同 `_warn_on_thinking_mismatch`)
+        logger.warning("源 {} —— {}", source_name, text)

    def _finalize_text(
        self, content_parts: list[str], thinking_parts: list[str], profile: ProviderProfile
@@ -759,6 +900,9 @@ class OpenAICompatTransport:
        为纯时间阈值(对冲只能靠 `hedge_after_s` 触发)。`content_event`(1.3.8)
        同款: 接收但**永不置位**——非流式无"中段"概念,stall 两值走
        TransportResult 默认 None(排除条款)。
+
+        无流完整性可言,直接以 body 的 finish_reason 走结局判定;保真模式**先判定
+        后构造快照**,否则半截工具在构造时就抛 Transient,到不了 `tool_truncated`。
        """
        resp = await client.post(url, json=payload)
        if resp.status_code != 200:
@@ -772,24 +916,16 @@ class OpenAICompatTransport:
                f"{source.name} 非流式响应非法 JSON", source_name=source.name, operation="chat"
            ) from exc
        _check_chat_error(body)
-        completion_result = parse_chat_completion(body) if response_mode == "openai" else None
-        choices = body.get("choices") or []
-        if not choices:
-            raise TransientError(
-                f"{source.name} 响应缺 choices", source_name=source.name, operation="chat"
-            )
-        if not isinstance(choices, list) or not isinstance(choices[0], dict):
-            raise TransientError("chat choices 形态非法", operation="chat")
-        message = choices[0].get("message", {})
-        if not isinstance(message, dict) or any(
-            message.get(key) is not None and not isinstance(message[key], str)
-            for key in ("content", "reasoning_content")
-        ):
-            raise TransientError("chat message 形态非法", operation="chat")
+        message, finish = _nonstream_message(body, response_mode, source)
        content, thinking = self._finalize_text(
-            [message.get("content") or ""], [message.get("reasoning_content") or ""], profile
+            [message.get("content") or ""], [_reasoning_of(message)], profile
+        )
+        self._warn_dialect(source, frozenset(), finish)
+        raw_message = message if response_mode == "openai" else None
+        _raise_unless_deliverable(response_mode, raw_message, content, thinking, finish, source)
+        completion_result = (
+            build_completion(message, body, finish) if response_mode == "openai" else None
        )
-        self._reject_empty_completion(content, completion_result, source)
        prompt, completion, usage_source = _resolve_usage(body.get("usage") or {})
        reasoning_tokens = _coerce_reasoning_tokens(body.get("usage"))
        return TransportResult(
@@ -810,6 +946,7 @@ class OpenAICompatTransport:
            thinking_observation=observe_thinking(
                thinking=thinking, reasoning_tokens=reasoning_tokens
            ),
+            finish_reason=finish,
        )

    async def aclose(self) -> None:
@@ -425,6 +425,13 @@ class _CallContext:

 ResponseMode = Literal["text", "openai"]

+StreamEnd = Literal["done", "eof_after_finish", "salvaged"]
+"""流式结束方式(1.3.10, #31): `done` = 收到 `[DONE]`;`eof_after_finish` = 未发
+`[DONE]` 但 finish_reason 与其后的实测 usage 齐全、连接干净关闭;`salvaged` = 文本
+模式按 `missing_done=salvage` 打捞的截断结果。非流式/缓存命中/失败行为 `None`。"""
+
+FinishClass = Literal["natural", "truncated", "filtered", "upstream_error", "unknown", "absent"]
+

@dataclass(frozen=True)
 class OpenAIChatCompletion:
@@ -438,56 +445,256 @@ class OpenAIChatCompletion:
    def __post_init__(self) -> None:
        """校验协议快照；transport 翻译错误，缓存坏条目按 miss。"""
        message = self.message
-        if not isinstance(message, dict) or message.get("role") != "assistant":
-            raise ValueError("completion message 必须含 assistant role")
-        for key in ("content", "reasoning_content"):
-            if message.get(key) is not None and not isinstance(message[key], str):
-                raise ValueError(f"completion {key} 必须为字符串或 null")
-        for value in (self.finish_reason, self.response_id):
-            if value is not None and not isinstance(value, str):
-                raise ValueError("completion 身份/结束原因必须为字符串或 null")
-        if self.usage is not None and not isinstance(self.usage, dict):
-            raise ValueError("completion usage 必须为对象或 null")
-        if "tool_calls" in message:
-            tools = message["tool_calls"]
-            if not isinstance(tools, list):
-                raise ValueError("completion tool_calls 必须为数组")
-            seen: set[str] = set()
-            for tool in tools:
-                if not isinstance(tool, dict):
-                    raise ValueError("completion tool 必须为对象")
-                identity = tool.get("id")
-                fn = tool.get("function")
-                if (
-                    not isinstance(identity, str)
-                    or not identity
-                    or identity in seen
-                    or tool.get("type") != "function"
-                    or not isinstance(fn, dict)
-                    or not isinstance(fn.get("name"), str)
-                    or not fn["name"]
-                    or not isinstance(fn.get("arguments"), str)
-                ):
-                    raise ValueError("completion tool 身份/函数字段缺失、冲突或非法")
-                seen.add(identity)
+        _check_completion_shape(message, self.finish_reason, self.response_id, self.usage)
        json.dumps({"message": message, "usage": self.usage}, allow_nan=False)
        object.__setattr__(self, "message", copy.deepcopy(message))
        object.__setattr__(self, "usage", copy.deepcopy(self.usage))


-def _has_completion_output(content: str, completion: OpenAIChatCompletion | None) -> bool:
-    """成功须有投影正文或完整工具；协议原文为空时不信任缓存旧投影。
+def _check_completion_shape(message: Any, finish_reason: Any, response_id: Any, usage: Any) -> None:
+    """快照字段形态校验;违约抛 ValueError(由 transport/缓存各自翻译)。"""
+    if not isinstance(message, dict) or message.get("role") != "assistant":
+        raise ValueError("completion message 必须含 assistant role")
+    if not all(_str_or_none(message.get(key)) for key in ("content", "reasoning_content")):
+        raise ValueError("completion content/reasoning_content 必须为字符串或 null")
+    if not (_str_or_none(finish_reason) and _str_or_none(response_id)):
+        raise ValueError("completion 身份/结束原因必须为字符串或 null")
+    if not isinstance(usage, dict | None):
+        raise ValueError("completion usage 必须为对象或 null")
+    if not isinstance(message.get("tool_calls", []), list | None):
+        raise ValueError("completion tool_calls 必须为数组或 null")
+    if not _tool_snapshot_ok(message):
+        raise ValueError("completion tool 身份/函数字段缺失、冲突或非法")

-    只判定产物存在性，不替代协议结构校验，也不收紧快照类型的构造用途。
-    transport 与缓存共用此规则，避免空响应从缓存旁路成功。
+
+def _str_or_none(value: Any) -> bool:
+    """协议字符串字段的形态判据: 字符串或 null。"""
+    return value is None or isinstance(value, str)
+
+
+def _tool_ok(tool: Any, seen: set[str]) -> bool:
+    """单个工具条目是否满足快照不变式;成立时登记其 id 供查重。"""
+    if not isinstance(tool, dict):
+        return False
+    identity, fn = tool.get("id"), tool.get("function")
+    ok = (
+        isinstance(identity, str)
+        and bool(identity)
+        and identity not in seen
+        and tool.get("type") == "function"
+        and isinstance(fn, dict)
+        and isinstance(fn.get("name"), str)
+        and bool(fn["name"])
+        and isinstance(fn.get("arguments"), str)
+    )
+    if ok:
+        seen.add(identity)
+    return ok
+
+
+def _tool_snapshot_ok(message: Mapping[str, Any]) -> bool:
+    """工具快照不变式(1.3.9 规则): 缺省/null 视无工具;null 元素跳过。
+
+    抽成纯判据而非留在快照构造里: 结局判定(#30)要在构造**之前**知道工具是否
+    成立——截断的半截工具归 ResultInvalid,构造时抛 Transient 就到不了那一行。
    """
-    if completion is not None:
-        if completion.message.get("tool_calls"):
-            return True
-        raw_content = completion.message.get("content")
-        if not isinstance(raw_content, str) or not raw_content.strip():
+    tools = message.get("tool_calls")
+    if tools is None:
+        return True
+    if not isinstance(tools, list):
+        return False
+    seen: set[str] = set()
+    return all(tool is None or _tool_ok(tool, seen) for tool in tools)
+
+
+_FINISH_CLASSES: Mapping[str, FinishClass] = MappingProxyType(
+    {
+        "stop": "natural",
+        "tool_calls": "natural",
+        "function_call": "natural",
+        "length": "truncated",
+        "content_filter": "filtered",
+        "sensitive": "filtered",  # GLM: 安全审核拦截
+        "network_error": "upstream_error",  # GLM: 模型推理异常
+        "error": "upstream_error",
+    }
+)
+
+
+def _classify_finish(value: str | None) -> FinishClass:
+    """finish_reason → 类别;大小写敏感(上游原值),未登记非空值为 unknown。"""
+    if value is None:
+        return "absent"
+    return _FINISH_CLASSES.get(value, "unknown")
+
+
+def _stream_is_complete(
+    *, done: bool, finish_reason: str | None, usage_after_finish: bool, usage_measured: bool
+) -> StreamEnd | None:
+    """流完整性(设计 §4.2): None = 截断。
+
+    缺 `[DONE]` 时只认"上游声明结束 + 其后到达的实测 usage": 前者证明生成结束,
+    后者证明流走到了 include_usage 的自然末尾(finish 之前的累计 usage 不算)。
+    """
+    if done:
+        return "done"
+    if finish_reason is not None and usage_after_finish and usage_measured:
+        return "eof_after_finish"
+    return None
+
+
+@dataclass(frozen=True)
+class _OutputSummary:
+    """一次结果的产物摘要(设计 §4.3 定义),判定的唯一输入形态。"""
+
+    any_output: bool
+    body_output: bool
+    tools_present: bool
+    tools_ok: bool
+    null_arguments: bool = False
+    """工具快照仅因 `function.arguments` 全程 null 而不成立(无参函数方言,审查 I2)。"""
+
+
+def _summarize_text(content: str, thinking: str) -> _OutputSummary:
+    """文本模式: 产物只看投影正文与思考;空白也算产物,正文须非空白。"""
+    return _OutputSummary(
+        any_output=content != "" or thinking != "",
+        body_output=content.strip() != "",
+        tools_present=False,
+        tools_ok=True,
+    )
+
+
+def _has_string(value: Any) -> bool:
+    """递归查找任一非空字符串片段。"""
+    if isinstance(value, str):
+        return value != ""
+    if isinstance(value, dict):
+        return any(_has_string(item) for item in value.values())
+    if isinstance(value, list):
+        return any(_has_string(item) for item in value)
+    return False
+
+
+def _summarize_message(message: Mapping[str, Any], projected_content: str) -> _OutputSummary:
+    """保真模式: 任一非空字符串字段(role 除外)或工具条目即产物。"""
+    tools = message.get("tool_calls")
+    tools_present = isinstance(tools, list) and any(tool is not None for tool in tools)
+    tools_ok = _tool_snapshot_ok(message)
+    fields = {k: v for k, v in message.items() if k not in ("role", "tool_calls")}
+    return _OutputSummary(
+        any_output=tools_present or _has_string(fields),
+        body_output=projected_content.strip() != "" or (tools_present and tools_ok),
+        tools_present=tools_present,
+        tools_ok=tools_ok,
+        null_arguments=not tools_ok and _tool_snapshot_ok(_fill_null_arguments(message)),
+    )
+
+
+def _fill_null_arguments(message: Mapping[str, Any]) -> dict[str, Any]:
+    """仅供判定: 把 null arguments 视作空串后再验快照,区分"只缺参数"与其他缺陷。
+
+    不改写交付值——保真结果原样;此函数的产物只用于分类。
+    """
+    tools = message.get("tool_calls")
+    if not isinstance(tools, list):
+        return dict(message)
+    filled = []
+    for tool in tools:
+        fn = tool.get("function") if isinstance(tool, dict) else None
+        if isinstance(fn, dict) and "arguments" in fn and fn["arguments"] is None:
+            tool = {**tool, "function": {**fn, "arguments": ""}}
+        filled.append(tool)
+    return {**message, "tool_calls": filled}
+
+
+@dataclass(frozen=True)
+class _Verdict:
+    """结局判定;`deliver` 为真时另两字段为 None。"""
+
+    deliver: bool
+    error_kind: Literal["transient", "result_invalid"] | None
+    code: str | None
+
+
+_DELIVER = _Verdict(True, None, None)
+_EMPTY = _Verdict(False, "transient", "empty_completion")
+_WITHOUT_CONTENT: Mapping[str, str] = MappingProxyType(
+    {"truncated": "length_without_content", "filtered": "filtered_without_content"}
+)
+
+
+def _judge_broken(
+    *, mode: ResponseMode, summary: _OutputSummary, finish: FinishClass
+) -> _Verdict | None:
+    """行 1-2: 上游自报失败、坏工具——先于产物判断,正文不得掩盖。"""
+    if finish == "upstream_error":
+        return _Verdict(False, "transient", "upstream_finish_error")
+    if mode == "openai" and summary.tools_present and not summary.tools_ok:
+        if finish == "truncated":
+            return _Verdict(False, "result_invalid", "tool_truncated")
+        if summary.null_arguments:
+            # 无参函数只发 arguments:null 是确定性方言: 重拉同样结果,不重试也不改写
+            return _Verdict(False, "result_invalid", "tool_arguments_null")
+        return _Verdict(False, "transient", "tool_incomplete")
+    return None
+
+
+def _judge_outcome(*, mode: ResponseMode, summary: _OutputSummary, finish: FinishClass) -> _Verdict:
+    """结局分类(设计 §4.3),自上而下首个匹配行生效。
+
+    确定性结局(length/过滤)在文本模式归 ResultInvalid 而非 Transient: 重抽只会
+    再花一份钱,还会把健康源推向熔断;保真模式原样交回,由下游按 finish 决定。
+    """
+    broken = _judge_broken(mode=mode, summary=summary, finish=finish)
+    if broken is not None:
+        return broken
+    if summary.body_output:
+        return _DELIVER
+    if finish in _WITHOUT_CONTENT:  # 行 3/6/7: 确定性结局
+        if mode == "openai":
+            return _DELIVER
+        return _Verdict(False, "result_invalid", _WITHOUT_CONTENT[finish])
+    # 行 4/8/9: 零产物恒为抖动;有产物无正文时,只有保真且上游明确声明结束才交回
+    if summary.any_output and mode == "openai" and finish != "absent":
+        return _DELIVER
+    return _EMPTY
+
+
+def _effective_finish(finish_reason: Any, completion: OpenAIChatCompletion | None) -> str | None:
+    """缓存读写共用的 finish 口径: 顶层优先,缺省回退协议快照(1.3.9 条目无顶层)。"""
+    if finish_reason is None and completion is not None:
+        return completion.finish_reason
+    return finish_reason
+
+
+def _is_cacheable(
+    *,
+    mode: ResponseMode,
+    content: str,
+    finish_reason: str | None,
+    stream_end: str | None,
+    completion: OpenAIChatCompletion | None,
+) -> bool:
+    """可缓存 ⊂ 可交付(设计 §6): 截断、空正文、非自然结束一律不进 TTL。
+
+    缓存回放的是"同一请求的确定答案";length/过滤/未知结束带随机性或不可证明
+    是自然结束,固化进 TTL 会把一次偶然变成必然(无缓存毒化铁律)。
+    """
+    if stream_end not in (None, "done", "eof_after_finish"):
+        return False
+    if _classify_finish(finish_reason) not in ("natural", "absent"):
+        return False
+    if mode == "openai":
+        # 判原文而非投影: 协议原文为空时不信任缓存里的旧投影(1.3.9 同口径)
+        if completion is None:
            return False
-    return isinstance(content, str) and bool(content.strip())
+        raw = completion.message.get("content")
+        summary = _summarize_message(completion.message, raw if isinstance(raw, str) else "")
+        # 原文与投影都须有正文: 原文空不信任旧投影;投影空(think 标签剥净)不固化空回答
+        has_tools = summary.tools_present and summary.tools_ok
+        return summary.body_output and (content.strip() != "" or has_tools)
+    return content.strip() != ""


@dataclass(frozen=True)
@@ -557,6 +764,13 @@ class LLMResponse:

    completion: OpenAIChatCompletion | None = None

+    finish_reason: str | None = None
+    """上游结束原因原值(1.3.10, #30/#31),两模式同口径;`None` = 未上报或不可观测。"""
+
+    stream_end: StreamEnd | None = None
+    """流式结束方式(1.3.10, #31),见 `StreamEnd`。与 `usage_source` 分工: 那个答
+    "账单可信吗",这个答"内容完整吗";`eof_after_finish` 必为 measured。"""
+

@dataclass(frozen=True)
 class ChatRequest:
@@ -701,6 +915,13 @@ class TransportResult:

    completion: OpenAIChatCompletion | None = None

+    finish_reason: str | None = None
+    """上游结束原因原值(1.3.10, #30/#31),两模式同口径;`None` = 未上报或不可观测。"""
+
+    stream_end: StreamEnd | None = None
+    """流式结束方式(1.3.10, #31),见 `StreamEnd`。与 `usage_source` 分工: 那个答
+    "账单可信吗",这个答"内容完整吗";`eof_after_finish` 必为 measured。"""
+

@dataclass(frozen=True)
 class SourceConfig:
@@ -72,6 +72,8 @@ _EXPECTED_COLUMNS = [
    # —— 1.3.8 沉默归因两列(issue #27)——
    "stall_ms",
    "stall_count",
+    # —— 1.3.10 结束方式(#31)——
+    "stream_end",
 ]

 # 1.3.5 之前那张表的 27 个物理列(26 个 INSERT 字段 + created_at)。写成固定切片
@@ -173,6 +175,8 @@ def _minimal_fields(call_id: str | None = None, **overrides) -> dict[str, object
        # —— 1.3.8 两列: 沉默归因只属流式 chat 成功 attempt 行(issue #27),普通行 NULL ——
        "stall_ms": None,
        "stall_count": None,
+        # —— 1.3.10: 结束方式只属流式成功 attempt 行,普通行 NULL ——
+        "stream_end": None,
    }
    fields.update(overrides)
    return fields
@@ -775,7 +779,7 @@ class TestCallerDimensionsAcceptance:
                "WHERE table_schema = $1 AND table_name = 'llm_calls' ORDER BY ordinal_position",
                schema,
            )
-            # 22 → 38 个 recorder 字段(加 created_at 共 39 个物理列),且新列追加在末尾
+            # 22 → 39 个 recorder 字段(加 created_at 共 40 个物理列),且新列追加在末尾
            assert [r["column_name"] for r in cols] == _EXPECTED_COLUMNS
            rows = await _fetch(
                schema_dsn,
@@ -1106,7 +1110,7 @@ class TestCallObservabilityColumnsAcceptance:
    """

    async def test_pre_135_table_gains_the_ten_columns_and_old_rows_stay_null(self, pre_135_schema):
-        """27 列旧表 auto 补齐到 39 列(1.3.5 十列 + 1.3.8 两列),新行两类取值读得回,**历史行新增列一律 NULL**。
+        """27 列旧表 auto 补齐到 40 列(1.3.5 十列 + 1.3.8 两列 + 1.3.10 一列),新行两类取值读得回,**历史行新增列一律 NULL**。

        旧行不回填是本版的明示决策(设计 §5): NULL 表达的是"补列之前根本没记过
        这件事",与任何哨兵值都不是一回事。若哪天有人给这十列加了 DEFAULT,历史行
@@ -1148,7 +1152,7 @@ class TestCallObservabilityColumnsAcceptance:
                "WHERE table_schema = $1 AND table_name = 'llm_calls' ORDER BY ordinal_position",
                schema,
            )
-            # 26 → 38 个 recorder 字段(加 created_at 共 39 个物理列),新列追加在末尾:
+            # 26 → 39 个 recorder 字段(加 created_at 共 40 个物理列),新列追加在末尾:
            # 列序与新建库一致才不会让两条升级路径分叉
            assert [r["column_name"] for r in cols] == _EXPECTED_COLUMNS

@@ -1176,6 +1180,8 @@ class TestCallObservabilityColumnsAcceptance:
                # 1.3.8 两列: 普通成功 attempt 行无沉默观测 → NULL
                "stall_ms": None,
                "stall_count": None,
+                # 1.3.10: 该行未传结束方式 → NULL
+                "stream_end": None,
            }
            assert dict(by_id["term"]) == {
                "call_id": "term",
@@ -1192,6 +1198,8 @@ class TestCallObservabilityColumnsAcceptance:
                # 1.3.8 两列: 终态行无 transport 观测 → NULL
                "stall_ms": None,
                "stall_count": None,
+                # 1.3.10: 终态行无流可言 → NULL
+                "stream_end": None,
            }
            # 历史行: 十列逐列 NULL(整体比对,漏掉任一列都红)
            assert dict(by_id["pre135"]) == {
@@ -1349,7 +1357,7 @@ class TestPublishedSchemaScript:

        await _execute_script(fresh_dsn, script)
        actual = [r["column_name"] for r in await _fetch(fresh_dsn, _PHYSICAL_COLUMNS_SQL, schema)]
-        # 物理列 = 38 个 INSERT 字段 + 库从不显式写的 created_at;对着库常量比,不另抄一份
+        # 物理列 = 39 个 INSERT 字段 + 库从不显式写的 created_at;对着库常量比,不另抄一份
        assert set(actual) == set(COLUMNS) | {"created_at"}
        # 列序也不许漂: 新列必须排在 created_at 之后,否则新建库与 ALTER 升级的列序分叉
        assert actual == _EXPECTED_COLUMNS
@@ -30,6 +30,13 @@ def _resp(content="cached"):
    return LLMResponse(content, "", "m", "p", 1, 2, 30, None, None, False, "orig")


+def _const(response):
+    async def terminal(request):
+        return response
+
+    return terminal
+
+
@pytest.fixture
 async def redis_cache():
    import redis.asyncio as aioredis
@@ -82,6 +89,23 @@ class TestRealRedisRoundtrip:
        assert first.cache_hit is False and second.cache_hit is True
        assert calls["n"] == 1

+    async def test_completion_verdict_fields_roundtrip(self, redis_cache, namespace):
+        """1.3.10: finish_reason/stream_end 经真实 Redis 往返;不可缓存结果不落库。"""
+        mw = CacheMW(
+            backend=redis_cache, model_fingerprint="m", default_namespace=namespace, ttl_s=60
+        )
+        from dataclasses import replace
+
+        ok = replace(_resp(), finish_reason="stop", stream_end="eof_after_finish")
+        first = await mw(ChatRequest(messages=_MSGS), _const(ok))
+        second = await mw(ChatRequest(messages=_MSGS), _const(ok))
+        assert first.cache_hit is False and second.cache_hit is True
+        assert second.finish_reason == "stop" and second.stream_end is None
+        truncated = replace(_resp(), stream_end="salvaged")
+        other = [{"role": "user", "content": "integration-salvaged"}]
+        await mw(ChatRequest(messages=other), _const(truncated))
+        assert await redis_cache.get(build_cache_key("m", other, namespace, None)) is None
+
    async def test_namespace_isolation_on_shared_instance(self, redis_cache, namespace):
        """多项目共用同一 Redis 时,namespace 不同绝不互相命中(防毒化铁律)。"""
        other = f"pgw:test:{uuid.uuid4().hex}"
@@ -1002,3 +1002,128 @@ async def test_openai_tool_only_cache_hits_without_origin():
    hit = await mw(request, terminal)
    assert hit.cache_hit and terminal.calls == 0
    assert hit.completion == completion
+
+
+# —— 1.3.10(#30/#31): 可缓存 ⊂ 可交付,读写共用同一判据(设计 §6)——
+
+_OK_COMPLETION = OpenAIChatCompletion({"role": "assistant", "content": "ok"}, "stop", "up", None)
+_THINK_LENGTH = OpenAIChatCompletion(
+    {"role": "assistant", "reasoning_content": "t"}, "length", "up", None
+)
+_THINK_STOP = OpenAIChatCompletion(
+    {"role": "assistant", "reasoning_content": "t"}, "stop", "up", None
+)
+
+
+@pytest.mark.parametrize(
+    ("mode", "response"),
+    [
+        ("text", _resp(stream_end="salvaged")),
+        ("text", _resp(finish_reason="length")),
+        ("text", _resp(finish_reason="content_filter")),
+        ("text", _resp(finish_reason="end_turn")),
+        ("openai", _resp(content="", finish_reason="length", completion=_THINK_LENGTH)),
+        ("openai", _resp(content="", finish_reason="stop", completion=_THINK_STOP)),
+    ],
+)
+async def test_uncacheable_results_are_delivered_but_not_written(mode, response):
+    """截断/空正文/非自然结束: 照常交回,但不写缓存,下次回源。"""
+    backend = InMemoryCache()
+    mw = _mw(backend)
+    terminal = _Terminal(response)
+    request = ChatRequest(messages=_MSGS, response_mode=mode)
+    assert (await mw(request, terminal)) is response
+    assert (await mw(request, terminal)).cache_hit is False
+    assert terminal.calls == 2 and not backend._store
+
+
+def _stored(mode, **fields):
+    """直接注入一条缓存原文(模拟旧版本或别的进程写入)。"""
+    backend = InMemoryCache()
+    data = dataclasses.asdict(_resp(**fields))
+    data.pop("structured_data", None)
+    data.pop("call_stats", None)
+    key = build_cache_key("m", _MSGS, "proj", None, response_mode=mode)
+    backend._store[key] = (float("inf"), json.dumps(data))
+    return backend, key
+
+
+async def _hit(backend, mode):
+    terminal = _Terminal(_resp(content="fresh", completion=_OK_COMPLETION, finish_reason="stop"))
+    response = await _mw(backend)(ChatRequest(messages=_MSGS, response_mode=mode), terminal)
+    return response, terminal.calls
+
+
+@pytest.mark.parametrize(
+    ("mode", "fields"),
+    [
+        ("text", {"stream_end": "salvaged"}),
+        ("text", {"stream_end": "bogus"}),
+        ("text", {"finish_reason": "length"}),
+        ("text", {"finish_reason": 5}),
+        ("openai", {"content": "", "completion": _THINK_LENGTH, "finish_reason": "length"}),
+    ],
+)
+async def test_injected_uncacheable_entries_read_as_miss(mode, fields):
+    """读路径对存储原值判定(覆盖之前): 不可缓存即 miss 回源。"""
+    backend, _ = _stored(mode, **fields)
+    response, calls = await _hit(backend, mode)
+    assert response.cache_hit is False and calls == 1
+
+
+async def test_legacy_fidelity_length_entry_falls_back_to_completion_finish():
+    """1.3.9 保真条目无顶层 finish: 回退 completion.finish_reason 识别为不可缓存。"""
+    completion = OpenAIChatCompletion({"role": "assistant", "content": "a"}, "length", None, None)
+    backend, key = _stored("openai", completion=completion)
+    data = json.loads(backend._store[key][1])
+    for name in ("finish_reason", "stream_end"):
+        data.pop(name)
+    backend._store[key] = (float("inf"), json.dumps(data))
+    response, calls = await _hit(backend, "openai")
+    assert response.cache_hit is False and calls == 1
+
+
+async def test_legacy_text_entry_without_new_fields_still_hits():
+    """1.3.9 文本条目无新键: 仍命中;命中 stream_end=None。"""
+    backend, key = _stored("text")
+    data = json.loads(backend._store[key][1])
+    for name in ("finish_reason", "stream_end"):
+        data.pop(name)
+    backend._store[key] = (float("inf"), json.dumps(data))
+    response, calls = await _hit(backend, "text")
+    assert response.cache_hit is True and calls == 0
+    assert response.stream_end is None and response.finish_reason is None
+
+
+async def test_hit_resets_stream_end_and_keeps_effective_finish():
+    """命中: stream_end 置 None;finish_reason 取判定所用值(顶层缺省时回退)。"""
+    backend, key = _stored("openai", completion=_OK_COMPLETION, stream_end="eof_after_finish")
+    data = json.loads(backend._store[key][1])
+    data.pop("finish_reason")
+    backend._store[key] = (float("inf"), json.dumps(data))
+    response, calls = await _hit(backend, "openai")
+    assert response.cache_hit is True and calls == 0
+    assert response.stream_end is None and response.finish_reason == "stop"
+
+
+async def test_new_fields_roundtrip_through_cache():
+    """两新字段随序列化往返;命中只重置 stream_end。"""
+    backend = InMemoryCache()
+    mw = _mw(backend)
+    terminal = _Terminal(_resp(finish_reason="stop", stream_end="eof_after_finish"))
+    await mw(ChatRequest(messages=_MSGS), terminal)
+    stored = json.loads(next(iter(backend._store.values()))[1])
+    assert stored["finish_reason"] == "stop" and stored["stream_end"] == "eof_after_finish"
+    hit = await mw(ChatRequest(messages=_MSGS), terminal)
+    assert hit.cache_hit is True and hit.finish_reason == "stop" and hit.stream_end is None
+
+
+async def test_fidelity_think_only_projection_not_cached():
+    """审查 I1: qwen think 标签剥离后投影为空的保真结果不写缓存。"""
+    completion = OpenAIChatCompletion(
+        {"role": "assistant", "content": "<think>x</think>"}, "stop", "up", None
+    )
+    response = _resp(content="", thinking="x", finish_reason="stop", completion=completion)
+    backend = InMemoryCache()
+    await _mw(backend)(ChatRequest(messages=_MSGS, response_mode="openai"), _Terminal(response))
+    assert not backend._store
@@ -0,0 +1,453 @@
+"""统一完成判定在 transport 两条路径上的落点(#30/#31,设计 §4.2-§4.4、§5.4)。
+
+假 SSE 按 magpie/glm 实测形态二次构造: 末块 finish_reason + usage 后干净关闭、
+不发 [DONE];以及只有思考的 length 截断、null/未知字段等方言。
+"""
+
+import asyncio
+import json
+
+import httpx
+import pytest
+from loguru import logger
+
+from polygateway.errors import AllSourcesExhausted, ResultInvalidError, TransientError
+from polygateway.transports.openai_compat import OpenAICompatTransport
+from tests.unit.test_openai_compat import _complete, _source, _sse_stream, _transport_for
+
+_USAGE = {"prompt_tokens": 11, "completion_tokens": 7}
+_TOOL = {
+    "index": 0,
+    "id": "call_1",
+    "type": "function",
+    "function": {"name": "lookup", "arguments": '{"q": 1}'},
+}
+
+
+def _f(delta=None, finish=None, usage=None, choices=True):
+    """单 choice 帧;`choices=False` 构造 usage-only 帧。"""
+    body = {"id": "up-1"}
+    if choices:
+        choice = {"index": 0, "delta": delta or {}}
+        if finish is not None:
+            choice["finish_reason"] = finish
+        body["choices"] = [choice]
+    else:
+        body["choices"] = []
+    if usage is not None:
+        body["usage"] = usage
+    return f"data: {json.dumps(body)}\n\n"
+
+
+def _assistant(**delta):
+    return _f({"role": "assistant", **delta})
+
+
+async def _run(frames, *, done, mode="text", **source_kw):
+    transport = _transport_for(lambda _: _sse_stream(*frames, done=done))
+    try:
+        return await _complete(transport, _source(**source_kw), response_mode=mode)
+    finally:
+        await transport.aclose()
+
+
+async def _run_once(body, *, mode="text"):
+    transport = _transport_for(lambda _: httpx.Response(200, json=body))
+    try:
+        return await _complete(transport, _source(), stream=False, response_mode=mode)
+    finally:
+        await transport.aclose()
+
+
+# —— #31: 缺 [DONE] 但 finish + 其后实测 usage 完整 ——
+
+
+@pytest.mark.parametrize("mode", ["text", "openai"])
+@pytest.mark.parametrize("finish", ["stop", "length"])
+async def test_eof_after_finish_with_usage_succeeds(mode, finish):
+    """验收 1: 末块 finish + usage 后干净关闭 → 成功、measured、eof_after_finish。"""
+    frames = [_assistant(content="hel"), _f({"content": "lo"}, finish=finish, usage=_USAGE)]
+    result = await _run(frames, done=False, mode=mode)
+    assert result.content == "hello" and result.stream_end == "eof_after_finish"
+    assert result.usage_source == "measured" and result.finish_reason == finish
+    assert (result.prompt_tokens, result.completion_tokens) == (11, 7)
+    if mode == "openai":
+        assert result.completion.finish_reason == finish
+
+
+async def test_eof_after_finish_usage_only_frame_and_tools():
+    """验收 2: tool_calls 结束、usage 在其后的 usage-only 帧 → 工具完整交回。"""
+    frames = [
+        _assistant(content=None, tool_calls=[_TOOL]),
+        _f({}, finish="tool_calls"),
+        _f(choices=False, usage=_USAGE),
+    ]
+    result = await _run(frames, done=False, mode="openai")
+    tool = result.completion.message["tool_calls"][0]
+    assert tool["function"] == {"name": "lookup", "arguments": '{"q": 1}'}
+    assert result.stream_end == "eof_after_finish" and result.usage_source == "measured"
+
+
+@pytest.mark.parametrize(
+    "frames",
+    [
+        [_assistant(content="partial"), _f(choices=False, usage=_USAGE)],  # 验收 3: 无 finish
+        [_assistant(content="partial"), _f({}, finish="stop")],  # 验收 4: 无 usage
+        [_assistant(content="a", **{}), _f({}, usage=_USAGE), _f({}, finish="stop")],  # usage 早
+        [_assistant(content="a"), _f({}, finish="stop", usage={})],  # usage 不可用
+    ],
+)
+async def test_incomplete_eof_still_fails(frames):
+    """截断不放宽: 保真仍不打捞,文本 retry 档仍抛。"""
+    with pytest.raises(TransientError, match="missing_done"):
+        await _run(frames, done=False, mode="openai", missing_done="salvage")
+    with pytest.raises(TransientError, match="missing_done"):
+        await _run(frames, done=False)
+
+
+async def test_text_salvage_marks_stream_end():
+    """文本 salvage 打捞 → stream_end=salvaged,usage 降为 estimated。"""
+    frames = [_assistant(content="partial"), _f(choices=False, usage=_USAGE)]
+    result = await _run(frames, done=False, missing_done="salvage")
+    assert result.stream_end == "salvaged" and result.usage_source == "estimated"
+
+
+async def test_read_error_mid_stream_fails():
+    """验收 5: 读取中途 RemoteProtocolError → 失败,不当干净 EOF。"""
+
+    class Broken(httpx.AsyncByteStream):
+        async def __aiter__(self):
+            yield _f({"role": "assistant", "content": "a"}, finish="stop", usage=_USAGE).encode()
+            raise httpx.RemoteProtocolError("peer closed")
+
+        async def aclose(self):
+            return None
+
+    transport = _transport_for(lambda _: httpx.Response(200, stream=Broken()))
+    try:
+        with pytest.raises(TransientError, match="网络错误"):
+            await _complete(transport, _source(), response_mode="openai")
+    finally:
+        await transport.aclose()
+
+
+@pytest.mark.parametrize("mode", ["text", "openai"])
+async def test_done_stream_marks_done(mode):
+    """验收 6: 带 [DONE] 的正常流行为不变,stream_end=done。"""
+    result = await _run([_assistant(content="ok"), _f({}, "stop", _USAGE)], done=True, mode=mode)
+    assert result.stream_end == "done" and result.finish_reason == "stop"
+
+
+@pytest.mark.parametrize("done", [True, False])
+@pytest.mark.parametrize("mode", ["text", "openai"])
+async def test_upstream_error_finish_is_transient(done, mode):
+    """GLM network_error: 上游自报失败,有正文也不交付。"""
+    frames = [_assistant(content="half"), _f({}, finish="network_error", usage=_USAGE)]
+    with pytest.raises(TransientError, match="upstream_finish_error"):
+        await _run(frames, done=done, mode=mode)
+
+
+# —— #30 与结局表(设计 §4.3)——
+
+_THINK = [_assistant(reasoning_content="long thought")]
+
+
+@pytest.mark.parametrize("done", [True, False])
+async def test_openai_reasoning_only_length_delivered(done):
+    """#30 验收: 只有思考 + length → 交回,content 空,思考完整。"""
+    result = await _run([*_THINK, _f({}, "length", _USAGE)], done=done, mode="openai")
+    message = result.completion.message
+    assert message.get("content") in (None, "") and message["reasoning_content"] == "long thought"
+    assert result.finish_reason == "length" and result.thinking == "long thought"
+
+
+@pytest.mark.parametrize(
+    ("frames", "finish", "openai_ok", "text_error"),
+    [
+        (_THINK, "length", True, (ResultInvalidError, "length_without_content")),
+        ([_assistant()], "length", True, (ResultInvalidError, "length_without_content")),
+        (_THINK, "content_filter", True, (ResultInvalidError, "filtered_without_content")),
+        ([_assistant()], "sensitive", True, (ResultInvalidError, "filtered_without_content")),
+        (_THINK, "stop", True, (TransientError, "empty_completion")),
+        ([_assistant(refusal="no")], "stop", True, (TransientError, "empty_completion")),
+        ([_assistant()], "stop", False, (TransientError, "empty_completion")),
+    ],
+)
+async def test_outcome_rows_stream(frames, finish, openai_ok, text_error):
+    """行 3/4/6/7/8: 两模式的分叉。"""
+    full = [*frames, _f({}, finish, _USAGE)]
+    if openai_ok:
+        assert (await _run(full, done=True, mode="openai")).finish_reason == finish
+    else:
+        with pytest.raises(TransientError, match="empty_completion"):
+            await _run(full, done=True, mode="openai")
+    error, code = text_error
+    with pytest.raises(error, match=code) as caught:
+        await _run(full, done=True)
+    assert caught.type is error  # ResultInvalid 不得被 Transient 子类掩盖
+
+
+@pytest.mark.parametrize("mode", ["text", "openai"])
+async def test_reasoning_only_without_finish_is_transient(mode):
+    """行 9: 只靠 DONE 收尾、无 finish → 无法区分抖动与结局,维持瞬时。"""
+    with pytest.raises(TransientError, match="empty_completion"):
+        await _run([*_THINK, _f(choices=False, usage=_USAGE)], done=True, mode=mode)
+
+
+async def test_openai_whitespace_content_stop_delivered():
+    """正文只有空白 + stop: 保真交回(#30 一-2)。"""
+    result = await _run(
+        [_assistant(content="  "), _f({}, "stop", _USAGE)], done=True, mode="openai"
+    )
+    assert result.completion.message["content"] == "  "
+
+
+async def test_truncated_tool_is_result_invalid():
+    """行 2: length 截断的半截工具(缺 name)→ ResultInvalid tool_truncated。"""
+    broken = {"index": 0, "id": "c", "type": "function", "function": {"arguments": '{"q'}}
+    frames = [_assistant(tool_calls=[broken]), _f({}, "length", _USAGE)]
+    with pytest.raises(ResultInvalidError, match="tool_truncated"):
+        await _run(frames, done=True, mode="openai")
+
+
+async def test_content_cannot_mask_broken_tool():
+    """行 2: 正文非空 + 坏工具 + stop → Transient tool_incomplete。"""
+    broken = {"index": 0, "id": "c", "type": "function", "function": {"arguments": "{}"}}
+    frames = [_assistant(content="text", tool_calls=[broken]), _f({}, "stop", _USAGE)]
+    with pytest.raises(TransientError, match="tool_incomplete"):
+        await _run(frames, done=True, mode="openai")
+
+
+async def test_null_refusal_and_tool_calls_stream():
+    """#30 验收二: delta 带 refusal:null 与 tool_calls:null → 正常交回。"""
+    frames = [_assistant(content="ok", refusal=None, tool_calls=None), _f({}, "stop", _USAGE)]
+    result = await _run(frames, done=True, mode="openai")
+    assert result.completion.message["content"] == "ok"
+
+
+# —— 非流式镜像 ——
+
+
+def _body(message, finish):
+    choice = {"index": 0, "message": {"role": "assistant", **message}}
+    if finish is not None:
+        choice["finish_reason"] = finish
+    return {"id": "up", "choices": [choice], "usage": _USAGE}
+
+
+async def test_nonstream_null_tool_calls_and_refusal():
+    """非流式 message 带 tool_calls:null + refusal:null → 成功。"""
+    body = _body({"content": "hi", "tool_calls": None, "refusal": None}, "stop")
+    result = await _run_once(body, mode="openai")
+    assert result.completion.message["tool_calls"] is None and result.stream_end is None
+
+
+@pytest.mark.parametrize(
+    ("message", "finish", "openai_expect", "text_expect"),
+    [
+        (
+            {"reasoning_content": "t"},
+            "length",
+            None,
+            (ResultInvalidError, "length_without_content"),
+        ),
+        ({"content": ""}, "content_filter", None, (ResultInvalidError, "filtered_without_content")),
+        ({"reasoning_content": "t"}, "stop", None, (TransientError, "empty_completion")),
+        ({"reasoning_content": "t"}, None, (TransientError, "empty_completion"), None),
+        ({"content": "a"}, "network_error", (TransientError, "upstream_finish_error"), None),
+        ({"content": "a"}, "length", None, None),
+    ],
+)
+async def test_outcome_rows_nonstream(message, finish, openai_expect, text_expect):
+    """非流式跳过流完整性,直接以 body finish 走结局表。"""
+    body = _body(message, finish)
+    for mode, expect in (("openai", openai_expect), ("text", text_expect or openai_expect)):
+        if expect is None:
+            result = await _run_once(body, mode=mode)
+            assert result.finish_reason == finish and result.stream_end is None
+        else:
+            with pytest.raises(expect[0], match=expect[1]) as caught:
+                await _run_once(body, mode=mode)
+            assert caught.type is expect[0]
+
+
+async def test_nonstream_truncated_tool_is_result_invalid():
+    """非流式先判定后构造快照: 缺 name + length → tool_truncated 而非构造期 Transient。"""
+    tool = {"id": "c", "type": "function", "function": {"arguments": "{"}}
+    with pytest.raises(ResultInvalidError, match="tool_truncated"):
+        await _run_once(_body({"tool_calls": [tool]}, "length"), mode="openai")
+
+
+# —— reasoning 别名(设计 §4.4)——
+
+
+async def test_reasoning_alias_projects_thinking():
+    """只有 reasoning 字段: 文本 thinking 取别名;非空 reasoning_content 优先。"""
+    alias = await _run(
+        [_assistant(reasoning="r1"), _f({"content": "a"}, "stop", _USAGE)], done=True
+    )
+    assert alias.thinking == "r1"
+    both = [
+        _assistant(reasoning="r1", reasoning_content="rc"),
+        _f({"content": "a"}, "stop", _USAGE),
+    ]
+    assert (await _run(both, done=True)).thinking == "rc"
+    empty_rc = [_assistant(reasoning="r1", reasoning_content=""), _f({"content": "a"}, "stop")]
+    assert (await _run(empty_rc, done=True)).thinking == "r1"
+    nonstream = await _run_once(_body({"content": "a", "reasoning": "nr"}, "stop"))
+    assert nonstream.thinking == "nr"
+
+
+async def test_reasoning_alias_keeps_ttft_alive():
+    """只有 reasoning 的思考期刷新活性,不被 TTFT 看门狗误杀。"""
+
+    class Slow(httpx.AsyncByteStream):
+        async def __aiter__(self):
+            yield _assistant().encode()
+            for _ in range(5):  # 0.5s 只有 reasoning,远超 ttft,但每帧间隔 < inter_token
+                yield _f({"reasoning": "t"}).encode()
+                await asyncio.sleep(0.1)
+            yield _f({"content": "answer"}, "stop", _USAGE).encode()
+            yield b"data: [DONE]\n\n"
+
+        async def aclose(self):
+            return None
+
+    transport = _transport_for(lambda _: httpx.Response(200, stream=Slow()))
+    try:
+        source = _source(ttft_timeout_s=0.3, inter_token_timeout_s=0.25)
+        result = await _complete(transport, source)
+        assert result.content == "answer" and result.thinking == "ttttt"
+    finally:
+        await transport.aclose()
+
+
+# —— 有界告警节流(设计 §5.4)——
+
+
+def _sink():
+    records = []
+    handle = logger.add(lambda m: records.append(m.record["message"]), level="WARNING")
+    return records, handle
+
+
+async def test_conflict_and_unknown_finish_warn_once():
+    """同源同路径冲突、同一未知 finish 值各只告警一次,不回显字段值。"""
+    frames = [
+        _assistant(content="a", ext=1),
+        _f({"ext": 2}),
+        _f({}, "end_turn", _USAGE),
+    ]
+    transport = _transport_for(lambda _: _sse_stream(*frames))
+    records, handle = _sink()
+    try:
+        for _ in range(2):
+            result = await _complete(transport, _source(), response_mode="openai")
+            assert "ext" not in result.completion.message
+    finally:
+        logger.remove(handle)
+        await transport.aclose()
+    conflict = [r for r in records if "delta.ext" in r]
+    unknown = [r for r in records if "end_turn" in r]
+    assert len(conflict) == 1 and len(unknown) == 1 and "不缓存" in unknown[0]
+
+
+async def test_warning_set_is_bounded():
+    """节流集合饱和后只记一次饱和告警,不再增长。"""
+    transport = OpenAICompatTransport()
+    transport._warned_fields.update((f"s{i}", "k") for i in range(1024))
+    records, handle = _sink()
+    try:
+        transport._warn_once("qwen_1", "field:x", "a")
+        transport._warn_once("qwen_1", "field:y", "b")
+    finally:
+        logger.remove(handle)
+        await transport.aclose()
+    assert len(transport._warned_fields) == 1024
+    assert sum("饱和" in r for r in records) == 1
+
+
+# —— 治理链: 确定性结局不重试、不计熔断 ——
+
+
+async def test_text_length_without_content_not_retried_nor_breaker_failure():
+    """文本 length 无正文 → ResultInvalid 直抛,attempts=1,熔断不记失败。"""
+    from tests.unit.test_client import _client
+
+    calls = []
+
+    def handler(_):
+        calls.append(1)
+        return _sse_stream(*_THINK, _f({}, "length", _USAGE))
+
+    client = _client(handler=handler)
+    failures = []
+    original = client._breaker_backend.record_failure
+
+    async def spy(*args, **kwargs):
+        failures.append(args)
+        return await original(*args, **kwargs)
+
+    client._breaker_backend.record_failure = spy
+    try:
+        with pytest.raises(ResultInvalidError, match="length_without_content"):
+            await client.chat([{"role": "user", "content": "hi"}])
+    finally:
+        await client._transport.aclose()
+    assert len(calls) == 1 and failures == []
+
+
+async def test_openai_reasoning_length_single_attempt_via_client():
+    """#30 验收经治理链: attempts=1;#31 形态经治理链不再重试到耗尽。"""
+    from tests.unit.test_client import _client
+
+    calls = []
+
+    def handler(_):
+        calls.append(1)
+        return _sse_stream(*_THINK, _f({}, "length", _USAGE), done=False)
+
+    client = _client(handler=handler)
+    try:
+        response = await client.chat([{"role": "user", "content": "hi"}], response_mode="openai")
+    except AllSourcesExhausted:  # pragma: no cover - 失败即回归
+        pytest.fail("reasoning-only length 不应重试到耗尽")
+    finally:
+        await client._transport.aclose()
+    assert len(calls) == 1 and response.call_stats.attempts == 1
+    assert response.stream_end == "eof_after_finish" and response.finish_reason == "length"
+
+
+async def test_eof_after_finish_stop_single_attempt_via_client():
+    """#31 验收 1 经治理链: stop + usage 无 DONE,attempts=1、不重试到耗尽。"""
+    from tests.unit.test_client import _client
+
+    calls = []
+
+    def handler(_):
+        calls.append(1)
+        return _sse_stream(_assistant(content="ok"), _f({}, "stop", _USAGE), done=False)
+
+    client = _client(handler=handler)
+    try:
+        response = await client.chat([{"role": "user", "content": "hi"}])
+    finally:
+        await client._transport.aclose()
+    assert len(calls) == 1 and response.call_stats.attempts == 1
+    assert response.stream_end == "eof_after_finish" and response.usage_source == "measured"
+
+
+async def test_text_salvage_with_length_partial_content_delivered():
+    """文本 salvage + length + 部分正文(finish 后无 usage): 交付、salvaged。"""
+    frames = [_assistant(content="part"), _f({}, finish="length")]
+    result = await _run(frames, done=False, missing_done="salvage")
+    assert result.content == "part" and result.stream_end == "salvaged"
+    assert result.finish_reason == "length"
+
+
+async def test_null_arguments_tool_is_result_invalid_stream():
+    """审查 I2: 无参函数 arguments 全程 null → ResultInvalid,不重试。"""
+    tool = {"index": 0, "id": "c", "type": "function", "function": {"name": "f", "arguments": None}}
+    frames = [_assistant(tool_calls=[tool]), _f({}, "tool_calls", _USAGE)]
+    with pytest.raises(ResultInvalidError, match="tool_arguments_null"):
+        await _run(frames, done=True, mode="openai")
@@ -0,0 +1,265 @@
+"""统一完成判定(#30/#31): 内核纯函数真值表。"""
+
+import pytest
+
+from polygateway.types import (
+    LLMResponse,
+    OpenAIChatCompletion,
+    TransportResult,
+    _classify_finish,
+    _effective_finish,
+    _is_cacheable,
+    _judge_outcome,
+    _OutputSummary,
+    _stream_is_complete,
+    _summarize_message,
+    _summarize_text,
+    _tool_snapshot_ok,
+)
+
+_TOOL = {"id": "t1", "type": "function", "function": {"name": "f", "arguments": "{}"}}
+
+
+@pytest.mark.parametrize(
+    ("value", "expected"),
+    [
+        ("stop", "natural"),
+        ("tool_calls", "natural"),
+        ("function_call", "natural"),
+        ("length", "truncated"),
+        ("content_filter", "filtered"),
+        ("sensitive", "filtered"),
+        ("network_error", "upstream_error"),
+        ("error", "upstream_error"),
+        ("end_turn", "unknown"),
+        ("STOP", "unknown"),
+        (None, "absent"),
+    ],
+)
+def test_classify_finish(value, expected):
+    assert _classify_finish(value) == expected
+
+
+@pytest.mark.parametrize(
+    ("done", "finish", "after", "measured", "expected"),
+    [
+        (True, None, False, False, "done"),
+        (True, "stop", True, True, "done"),
+        (False, "stop", True, True, "eof_after_finish"),
+        (False, "length", True, True, "eof_after_finish"),
+        (False, None, True, True, None),  # 无 finish 即截断
+        (False, "stop", False, True, None),  # usage 早于 finish(累计 usage)
+        (False, "stop", True, False, None),  # finish 后 usage 不可用
+    ],
+)
+def test_stream_is_complete(done, finish, after, measured, expected):
+    got = _stream_is_complete(
+        done=done, finish_reason=finish, usage_after_finish=after, usage_measured=measured
+    )
+    assert got == expected
+
+
+def test_tool_snapshot_ok():
+    assert _tool_snapshot_ok({"role": "assistant"})
+    assert _tool_snapshot_ok({"tool_calls": None})
+    assert _tool_snapshot_ok({"tool_calls": [None, _TOOL]})
+    assert not _tool_snapshot_ok({"tool_calls": {}})
+    assert not _tool_snapshot_ok({"tool_calls": [{"id": "t1", "type": "function"}]})
+    assert not _tool_snapshot_ok({"tool_calls": [_TOOL, _TOOL]})  # 重复 id
+
+
+def test_summarize_text():
+    assert _summarize_text("", "") == _OutputSummary(False, False, False, True)
+    assert _summarize_text("  ", "") == _OutputSummary(True, False, False, True)
+    assert _summarize_text("", "think") == _OutputSummary(True, False, False, True)
+    assert _summarize_text("hi", "") == _OutputSummary(True, True, False, True)
+
+
+@pytest.mark.parametrize(
+    ("message", "projected", "expected"),
+    [
+        ({"role": "assistant"}, "", _OutputSummary(False, False, False, True)),
+        ({"role": "assistant", "content": None}, "", _OutputSummary(False, False, False, True)),
+        ({"role": "assistant", "refusal": "no"}, "", _OutputSummary(True, False, False, True)),
+        (
+            {"role": "assistant", "reasoning_content": "r"},
+            "",
+            _OutputSummary(True, False, False, True),
+        ),
+        ({"role": "assistant", "content": " "}, " ", _OutputSummary(True, False, False, True)),
+        (
+            {"role": "assistant", "x": {"y": ["", "z"]}},
+            "",
+            _OutputSummary(True, False, False, True),
+        ),
+        (
+            {"role": "assistant", "tool_calls": [None]},
+            "",
+            _OutputSummary(False, False, False, True),
+        ),
+        (
+            {"role": "assistant", "tool_calls": [_TOOL]},
+            "",
+            _OutputSummary(True, True, True, True),
+        ),
+        (
+            {"role": "assistant", "tool_calls": [{"id": "t"}]},
+            "",
+            _OutputSummary(True, False, True, False),
+        ),
+        ({"role": "assistant", "content": "a"}, "a", _OutputSummary(True, True, False, True)),
+    ],
+)
+def test_summarize_message(message, projected, expected):
+    assert _summarize_message(message, projected) == expected
+
+
+_NONE = _OutputSummary(False, False, False, True)
+_THINK = _OutputSummary(True, False, False, True)
+_BODY = _OutputSummary(True, True, False, True)
+_BAD_TOOL = _OutputSummary(True, True, True, False)
+
+
+@pytest.mark.parametrize(
+    ("mode", "summary", "finish", "expected"),
+    [
+        # 行 1: upstream_error 优先于一切
+        ("openai", _BODY, "upstream_error", ("transient", "upstream_finish_error")),
+        ("text", _BODY, "upstream_error", ("transient", "upstream_finish_error")),
+        # 行 2: 坏工具(仅保真),正文不得掩盖
+        ("openai", _BAD_TOOL, "truncated", ("result_invalid", "tool_truncated")),
+        ("openai", _BAD_TOOL, "natural", ("transient", "tool_incomplete")),
+        ("text", _BAD_TOOL, "natural", None),
+        # 行 3: 无产物 + truncated/filtered
+        ("openai", _NONE, "truncated", None),
+        ("openai", _NONE, "filtered", None),
+        ("text", _NONE, "truncated", ("result_invalid", "length_without_content")),
+        ("text", _NONE, "filtered", ("result_invalid", "filtered_without_content")),
+        # 行 4: 零增量抖动(2026-07-20 裁决)
+        ("openai", _NONE, "natural", ("transient", "empty_completion")),
+        ("openai", _NONE, "absent", ("transient", "empty_completion")),
+        ("openai", _NONE, "unknown", ("transient", "empty_completion")),
+        ("text", _NONE, "absent", ("transient", "empty_completion")),
+        # 行 5: 有正文
+        ("openai", _BODY, "truncated", None),
+        ("text", _BODY, "filtered", None),
+        ("text", _BODY, "absent", None),
+        # 行 6-9: 有产物无正文
+        ("openai", _THINK, "truncated", None),
+        ("text", _THINK, "truncated", ("result_invalid", "length_without_content")),
+        ("openai", _THINK, "filtered", None),
+        ("text", _THINK, "filtered", ("result_invalid", "filtered_without_content")),
+        ("openai", _THINK, "natural", None),
+        ("openai", _THINK, "unknown", None),
+        ("text", _THINK, "natural", ("transient", "empty_completion")),
+        ("openai", _THINK, "absent", ("transient", "empty_completion")),
+        ("text", _THINK, "absent", ("transient", "empty_completion")),
+    ],
+)
+def test_judge_outcome_rows(mode, summary, finish, expected):
+    verdict = _judge_outcome(mode=mode, summary=summary, finish=finish)
+    if expected is None:
+        assert verdict.deliver and verdict.error_kind is None and verdict.code is None
+    else:
+        assert not verdict.deliver
+        assert (verdict.error_kind, verdict.code) == expected
+
+
+def _completion(content="a", finish="stop", tools=None):
+    message = {"role": "assistant", "content": content}
+    if tools is not None:
+        message["tool_calls"] = tools
+    return OpenAIChatCompletion(message, finish, None, None)
+
+
+@pytest.mark.parametrize(
+    ("kwargs", "expected"),
+    [
+        ({"mode": "text", "content": "a", "finish_reason": "stop", "stream_end": "done"}, True),
+        ({"mode": "text", "content": "a", "finish_reason": None, "stream_end": None}, True),
+        (
+            {
+                "mode": "text",
+                "content": "a",
+                "finish_reason": "stop",
+                "stream_end": "eof_after_finish",
+            },
+            True,
+        ),
+        ({"mode": "text", "content": "a", "finish_reason": None, "stream_end": "salvaged"}, False),
+        ({"mode": "text", "content": " ", "finish_reason": "stop", "stream_end": "done"}, False),
+        ({"mode": "text", "content": "a", "finish_reason": "length", "stream_end": None}, False),
+        ({"mode": "text", "content": "a", "finish_reason": "sensitive", "stream_end": None}, False),
+        ({"mode": "text", "content": "a", "finish_reason": "end_turn", "stream_end": None}, False),
+        ({"mode": "text", "content": "a", "finish_reason": "stop", "stream_end": "bogus"}, False),
+    ],
+)
+def test_is_cacheable_text(kwargs, expected):
+    assert _is_cacheable(completion=None, **kwargs) is expected
+
+
+def test_is_cacheable_openai():
+    ok = _completion()
+    assert _is_cacheable(
+        mode="openai", content="a", finish_reason="stop", stream_end="done", completion=ok
+    )
+    assert not _is_cacheable(
+        mode="openai", content="a", finish_reason="stop", stream_end="done", completion=None
+    )
+    # 原文为空且无工具: 不信任旧投影
+    assert not _is_cacheable(
+        mode="openai",
+        content="a",
+        finish_reason="stop",
+        stream_end="done",
+        completion=_completion(content=""),
+    )
+    tool_only = _completion(content=None, finish="tool_calls", tools=[_TOOL])
+    assert _is_cacheable(
+        mode="openai", content="", finish_reason="tool_calls", stream_end=None, completion=tool_only
+    )
+
+
+def test_effective_finish():
+    assert _effective_finish("stop", _completion(finish="length")) == "stop"
+    assert _effective_finish(None, _completion(finish="length")) == "length"
+    assert _effective_finish(None, None) is None
+
+
+def test_completion_accepts_null_tool_calls():
+    snap = OpenAIChatCompletion({"role": "assistant", "tool_calls": None}, "stop", None, None)
+    assert "tool_calls" in snap.message and snap.message["tool_calls"] is None
+    with pytest.raises(ValueError):
+        OpenAIChatCompletion({"role": "assistant", "tool_calls": {}}, "stop", None, None)
+
+
+def test_response_fields_default_none():
+    result = TransportResult("a", "", 0, 0, "unavailable", None, None, {})
+    assert result.finish_reason is None and result.stream_end is None
+    response = LLMResponse("a", "", "m", "p", 0, 0, 0, None, None, False, "c")
+    assert response.finish_reason is None and response.stream_end is None
+
+
+def test_stream_end_exported():
+    import polygateway
+
+    assert "StreamEnd" in polygateway.__all__
+
+
+def test_openai_cacheable_requires_projected_body():
+    """审查 I1: think 标签剥离后投影为空,即使原文非空也不可缓存。"""
+    think_only = _completion(content="<think>x</think>")
+    assert not _is_cacheable(
+        mode="openai", content="", finish_reason="stop", stream_end="done", completion=think_only
+    )
+
+
+def test_null_arguments_only_tool_is_result_invalid():
+    """审查 I2: 无参函数 arguments 全程 null(其余齐全)→ 确定性方言,ResultInvalid。"""
+    tool = {"id": "t1", "type": "function", "function": {"name": "f", "arguments": None}}
+    summary = _summarize_message({"role": "assistant", "tool_calls": [tool]}, "")
+    assert summary.null_arguments and not summary.tools_ok
+    verdict = _judge_outcome(mode="openai", summary=summary, finish="natural")
+    assert (verdict.error_kind, verdict.code) == ("result_invalid", "tool_arguments_null")
+    broken = {"id": "t1", "type": "function", "function": {"arguments": None}}
+    assert not _summarize_message({"role": "assistant", "tool_calls": [broken]}, "").null_arguments
@@ -99,13 +99,13 @@ def test_incomplete_tool_is_not_success(change):
    "delta",
    [
        [],
-        {"audio": {}},
        {"tool_calls": {}},
        {"content": []},
+        {"role": "user"},
        {"tool_calls": [{"index": True}]},
        {"tool_calls": [{"index": -1}]},
-        {"tool_calls": [{"index": 0, "extra": 1}]},
-        {"tool_calls": [{"index": 0, "function": {"arguments": None}}]},
+        {"tool_calls": [{"index": None}]},
+        {"tool_calls": [{"index": 0, "function": {"arguments": 12}}]},
    ],
 )
 def test_invalid_delta_is_domain_error(delta):
@@ -225,3 +225,244 @@ def test_fragmented_name_and_arguments_with_usage():
    assert result.message["tool_calls"][0]["function"] == {"name": "echo", "arguments": " {} "}
    assert result.usage["extra"]["nested"] == 9
    assert result.finish_reason == "length"
+
+
+# —— 1.3.10(#30): null 优先、未知字段无损合并、finish/usage 位置 ——
+
+
+def _feed(state, *chunks):
+    """依次喂入已解码帧,返回每帧的 ChatDelta。"""
+    return [state.add_chunk(c) for c in chunks]
+
+
+def _frame(delta, finish=None, **top):
+    """单 choice 已解码帧。"""
+    choice = {"index": 0, "delta": delta}
+    if finish is not None:
+        choice["finish_reason"] = finish
+    return {"choices": [choice], **top}
+
+
+def test_null_known_and_unknown_fields_are_absent():
+    """refusal/tool_calls/role/arguments 为 null 不判失败,键按缺省保留。"""
+    state = ChatCompletionAccumulator(response_mode="openai")
+    _feed(
+        state,
+        _frame({"role": None, "refusal": None, "tool_calls": None, "content": None}),
+        _frame({"role": "assistant", "content": "hi"}, finish="stop"),
+    )
+    message = state.finish().message
+    assert message["refusal"] is None and message["content"] == "hi"
+    assert message["role"] == "assistant"
+
+
+def test_role_only_null_fails_at_finish():
+    """全程只有 role:null 即缺 role,照旧失败。"""
+    state = ChatCompletionAccumulator(response_mode="openai")
+    _feed(state, _frame({"role": None, "content": "x"}, finish="stop"))
+    with pytest.raises(TransientError, match="role"):
+        state.finish()
+
+
+def test_null_arguments_and_null_tool_elements():
+    """arguments:null 视缺省;tool_calls 中 null 元素跳过。"""
+    state = ChatCompletionAccumulator(response_mode="openai")
+    tool = {"index": 0, "id": "a", "type": "function", "function": {"name": "f"}}
+    _feed(
+        state,
+        _frame({"role": "assistant", "tool_calls": [None, tool]}),
+        _frame({"tool_calls": [{"index": 0, "function": {"arguments": None}}]}),
+        _frame({"tool_calls": [{"index": 0, "function": {"arguments": "{}"}}]}, "tool_calls"),
+    )
+    assert state.finish().message["tool_calls"][0]["function"] == {"name": "f", "arguments": "{}"}
+
+
+def test_null_choices_and_delta_are_absent():
+    """chunk 层 choices:null 与 choice 层 delta:null 视缺省。"""
+    state = ChatCompletionAccumulator(response_mode="openai")
+    assert not state.add_chunk({"choices": None}).progress
+    assert not state.add_chunk({"choices": [{"index": 0, "delta": None}]}).progress
+
+
+def test_reasoning_alias_is_progress_and_preserved():
+    """只有 reasoning 字段: 计活性、交别名片段、保真原键保留。"""
+    for mode in ("openai", "text"):
+        state = ChatCompletionAccumulator(response_mode=mode)
+        delta = state.add_chunk(_frame({"role": "assistant", "reasoning": "think"}))
+        assert delta.progress and delta.reasoning == "think" and delta.reasoning_content == ""
+    state.add_chunk(_frame({"reasoning": " more"}))
+    openai = ChatCompletionAccumulator(response_mode="openai")
+    _feed(openai, _frame({"role": "assistant", "reasoning": "a"}), _frame({"reasoning": "b"}))
+    assert openai.raw_message()["reasoning"] == "ab"
+
+
+def _detail(text, **extra):
+    return {"type": "reasoning.text", "index": 0, "format": "x-v1", "text": text, **extra}
+
+
+def test_reasoning_details_merge_by_index_without_label_concat():
+    """OpenRouter reasoning_details 分片: index 合并,format/id/type 不拼接。"""
+    state = ChatCompletionAccumulator(response_mode="openai")
+    _feed(
+        state,
+        _frame({"role": "assistant", "reasoning_details": [_detail("he", id="r1")]}),
+        _frame({"reasoning_details": [_detail("llo", id="r1")]}),
+        _frame({"content": "ok"}, finish="stop"),
+    )
+    details = state.finish().message["reasoning_details"]
+    assert details == [
+        {"type": "reasoning.text", "index": 0, "format": "x-v1", "text": "hello", "id": "r1"}
+    ]
+    assert not state.conflicts
+
+
+def test_tool_level_nested_extension_preserved():
+    """工具层未知嵌套对象(如 thought signature)原样保留以便回传。"""
+    state = ChatCompletionAccumulator(response_mode="openai")
+    extra = {"google": {"thought_signature": "sig"}}
+    tool = {"index": 0, "id": "a", "type": "function", "extra_content": extra}
+    _feed(
+        state,
+        _frame({"role": "assistant", "tool_calls": [{**tool, "function": {"name": "f"}}]}),
+        _frame({"tool_calls": [{"index": 0, "function": {"arguments": "{}", "strict": True}}]}),
+    )
+    snapshot = state.finish().message["tool_calls"][0]
+    assert snapshot["extra_content"] == extra and snapshot["function"]["strict"] is True
+
+
+@pytest.mark.parametrize(
+    ("first", "second"),
+    [(1, 2), (True, False), ("s", {"o": 1}), ([{"index": 0, "t": "a"}], ["plain"])],
+)
+def test_conflict_drops_key_and_bans_path(first, second):
+    """数值/布尔/形态冲突: 删键、登记路径、后续同路径丢弃、调用成功。"""
+    state = ChatCompletionAccumulator(response_mode="openai")
+    _feed(
+        state,
+        _frame({"role": "assistant", "content": "a", "ext": first}),
+        _frame({"ext": second}),
+        _frame({"ext": first}, finish="stop"),
+    )
+    assert "ext" not in state.finish().message
+    assert state.conflicts == frozenset({"delta.ext"})
+
+
+def test_plain_arrays_append_and_same_scalar_idempotent():
+    """普通数组追加;同值标量幂等;追加数组中的字符串算活性。"""
+    state = ChatCompletionAccumulator(response_mode="openai")
+    deltas = _feed(
+        state,
+        _frame({"role": "assistant", "notes": ["a"], "n": 1}),
+        _frame({"notes": ["b", None], "n": 1}),
+    )
+    message = state.raw_message()
+    assert message["notes"] == ["a", "b"] and message["n"] == 1
+    assert all(d.progress for d in deltas[1:])
+
+
+def test_one_progress_per_frame_with_many_fragments():
+    """同帧多个字段只产生一次进展事件(ChatDelta 单值)。"""
+    state = ChatCompletionAccumulator(response_mode="text")
+    delta = state.add_chunk(_frame({"content": "a", "reasoning_content": "b", "x": "c"}))
+    assert delta.progress is True
+
+
+def test_label_only_unknown_is_not_progress():
+    """未知层仅标签键(type/format/id/index)不刷新活性。"""
+    state = ChatCompletionAccumulator(response_mode="openai")
+    assert not state.add_chunk(_frame({"meta": {"type": "x", "format": "y"}})).progress
+
+
+@pytest.mark.parametrize(
+    ("frames", "after"),
+    [
+        ([_frame({"content": "a"}, usage={"x": 1}), _frame({}, finish="stop")], False),
+        ([_frame({"content": "a"}, finish="stop", usage={"x": 1})], True),
+        ([_frame({"content": "a"}, finish="stop"), {"choices": [], "usage": {"x": 1}}], True),
+        ([_frame({"content": "a"}, finish="stop")], False),
+    ],
+)
+def test_usage_after_finish_tracking(frames, after):
+    """usage 必须在 finish 同帧或之后到达才算流走到自然末尾。"""
+    for mode in ("openai", "text"):
+        state = ChatCompletionAccumulator(response_mode=mode)
+        _feed(state, *frames)
+        assert state.finish_seen and state.finish_reason == "stop"
+        assert state.usage_after_finish is after
+
+
+def test_text_mode_ignores_other_choice_index():
+    """文本模式只取 index 缺省或为 0 的 choice,不混入 index=1。"""
+    state = ChatCompletionAccumulator(response_mode="text")
+    delta = state.add_chunk(
+        {
+            "choices": [
+                {"index": 1, "delta": {"content": "B"}, "finish_reason": "length"},
+                {"index": 0, "delta": {"content": "A"}},
+            ]
+        }
+    )
+    assert delta.content == "A" and state.finish_reason is None
+
+
+def test_text_mode_finish_non_string_ignored():
+    """文本模式 finish 非字符串记 None,不判失败。"""
+    state = ChatCompletionAccumulator(response_mode="text")
+    state.add_chunk(_frame({"content": "a"}, finish=5))
+    assert state.finish_reason is None and not state.finish_seen
+
+
+def test_nonstream_null_tool_elements_stripped():
+    """非流式 tool_calls 中 null 元素剔除(回传 [null] 会被上游拒),null 列表原样保留。"""
+    tool = {"id": "a", "type": "function", "function": {"name": "f", "arguments": "{}"}}
+    body = {
+        "choices": [{"message": {"role": "assistant", "tool_calls": [None, tool], "refusal": None}}]
+    }
+    assert parse_chat_completion(body).message["tool_calls"] == [tool]
+    body["choices"][0]["message"]["tool_calls"] = None
+    assert parse_chat_completion(body).message["tool_calls"] is None
+
+
+@pytest.mark.parametrize(("index", "taken"), [(None, True), (0, True), (False, False), (1, False)])
+def test_text_mode_choice_index_null_is_zero(index, taken):
+    """审查 I3: 文本模式 index:null 视缺省(=0);bool 不冒充 0。"""
+    state = ChatCompletionAccumulator(response_mode="text")
+    delta = state.add_chunk(
+        {"choices": [{"index": index, "delta": {"content": "hi"}, "finish_reason": "stop"}]}
+    )
+    assert (delta.content == "hi") is taken and (state.finish_reason == "stop") is taken
+
+
+def test_null_only_tool_elements_do_not_create_empty_list():
+    """审查 Minor1: tool_calls null 后 [null] 不得产出会被上游拒收的空数组。"""
+    state = ChatCompletionAccumulator(response_mode="openai")
+    _feed(
+        state,
+        _frame({"role": "assistant", "content": "a", "tool_calls": None}),
+        _frame({"tool_calls": [None]}, finish="stop"),
+    )
+    assert state.finish().message["tool_calls"] is None
+    body = {"choices": [{"message": {"role": "assistant", "content": "a", "tool_calls": [None]}}]}
+    assert parse_chat_completion(body).message["tool_calls"] is None
+
+
+def test_tool_calls_null_then_list_merges():
+    """tool_calls 先 null 后列表: 正常聚合工具。"""
+    state = ChatCompletionAccumulator(response_mode="openai")
+    tool = {"index": 0, "id": "a", "type": "function", "function": {"name": "f", "arguments": "{}"}}
+    _feed(
+        state,
+        _frame({"role": "assistant", "tool_calls": None}),
+        _frame({"tool_calls": [tool]}, finish="tool_calls"),
+    )
+    assert state.finish().message["tool_calls"][0]["id"] == "a"
+
+
+def test_nonstream_role_checked_before_verdict():
+    """审查 Minor3: 非流式缺 role 是形态失败,先于结局判定。"""
+    from polygateway.transports._openai_completion import check_completion_body
+
+    with pytest.raises(TransientError, match="role"):
+        check_completion_body(
+            {"choices": [{"message": {"content": "a"}, "finish_reason": "length"}]}
+        )
@@ -146,10 +146,10 @@ async def test_invalid_mode_combinations_fail_before_http(kwargs):


 async def test_openai_never_salvages():
-    """已收完整工具也必须有 DONE。"""
+    """无 finish 后 usage 的截断流即使收全工具也不打捞(1.3.10: 判据由 DONE 改为完成判定)。"""
    transport = _transport_for(lambda _: _sse_stream(*tool_frames(), done=False))
    try:
-        with pytest.raises(TransientError, match="DONE"):
+        with pytest.raises(TransientError, match="missing_done"):
            await transport.complete(
                messages=[],
                source=_source(missing_done="salvage"),
@@ -165,6 +165,28 @@ async def test_openai_never_salvages():
        await transport.aclose()


+async def test_openai_finish_with_trailing_usage_is_complete_without_done():
+    """对照: 同一工具流补上 finish 之后的 usage 帧,即使无 DONE 也完整交回(#31)。"""
+    usage = 'data: {"choices": [], "usage": {"prompt_tokens": 3, "completion_tokens": 4}}\n\n'
+    transport = _transport_for(lambda _: _sse_stream(*tool_frames(), usage, done=False))
+    try:
+        result = await transport.complete(
+            messages=[],
+            source=_source(),
+            stream=True,
+            overlay={},
+            call_id="c",
+            reasoning_effort=None,
+            first_token_event=None,
+            content_event=None,
+            response_mode="openai",
+        )
+    finally:
+        await transport.aclose()
+    assert result.stream_end == "eof_after_finish"
+    assert [t["id"] for t in result.completion.message["tool_calls"]] == ["a", "b"]
+
+
 async def test_tool_progress_sets_events():
    """工具首进展驱动两个事件与时间观测。"""
    import asyncio
@@ -150,6 +150,7 @@ class _DummyRecorder:
        total_latency_ms,
        stall_ms,
        stall_count,
+        stream_end,
    ) -> None: ...


@@ -322,7 +323,7 @@ class TestTelemetryRecorderSignature:
            "attempts",
            "total_latency_ms",
        } <= set(params)
-        assert len(params) - 1 == 38  # 减掉 self
+        assert len(params) - 1 == 39  # 减掉 self

    @pytest.mark.parametrize(
        "name",
@@ -341,6 +342,7 @@ class TestTelemetryRecorderSignature:
            "error_body",
            "attempts",
            "total_latency_ms",
+            "stream_end",
        ],
    )
    def test_caller_dimensions_have_no_default(self, name):
@@ -92,6 +92,8 @@ _EXPECTED_COLUMNS = [
    # —— 1.3.8 沉默归因两列(issue #27)——
    "stall_ms",
    "stall_count",
+    # —— 1.3.10 结束方式(#31)——
+    "stream_end",
 ]

 # 1.3.5 之前的 26 个 INSERT 字段(旧表形态),供冻结 INSERT 语句与旧表补列用例复用
@@ -187,6 +189,8 @@ async def _record_minimal(recorder, call_id="c1", **overrides):
        # —— 1.3.8 沉默归因两列: 默认 None = 非流式/无沉默观测(合法事实)——
        "stall_ms": None,
        "stall_count": None,
+        # —— 1.3.10 结束方式: 默认 None = 非流式/缓存命中/失败行 ——
+        "stream_end": None,
    }
    fields.update(overrides)
    await recorder.record_llm_call(**fields)
@@ -268,7 +272,7 @@ class TestSchemaModule:

        # COLUMNS 是 INSERT 字段序,不含数据库自填的 created_at
        assert list(COLUMNS) == [c for c in _EXPECTED_COLUMNS if c != "created_at"]
-        assert len(COLUMNS) == 38
+        assert len(COLUMNS) == 39
        # 两端 DDL 的列出现顺序 == 物理列序(created_at 在第 19 位)
        for ddl in (SQLITE_DDL, PG_DDL):
            assert _first_occurrence_order(ddl, _EXPECTED_COLUMNS) == _EXPECTED_COLUMNS
@@ -282,8 +286,8 @@ class TestSchemaModule:
            "ALTER TABLE llm_calls ADD COLUMN cached_prompt_tokens INTEGER",
        )
        assert PG_BACKFILL[-1] == (
-            "stall_count",
-            "ALTER TABLE llm_calls ADD COLUMN stall_count INTEGER",
+            "stream_end",
+            "ALTER TABLE llm_calls ADD COLUMN stream_end TEXT",
        )
        assert all("IF NOT EXISTS" not in stmt for _, stmt in PG_BACKFILL)

@@ -297,10 +301,10 @@ class TestSchemaModule:

        assert insert_sql("sqlite", _PRE_135_COLUMNS) == _FROZEN_SQLITE_INSERT
        assert insert_sql("postgres", _PRE_135_COLUMNS) == _FROZEN_PG_INSERT
-        # 全量 38 列: 占位符随列数增长,且不留空洞
-        assert insert_sql("postgres", COLUMNS).count("$") == 38
-        assert "$38)" in insert_sql("postgres", COLUMNS)
-        assert insert_sql("sqlite", COLUMNS).count("?") == 38
+        # 全量 39 列: 占位符随列数增长,且不留空洞
+        assert insert_sql("postgres", COLUMNS).count("$") == 39
+        assert "$39)" in insert_sql("postgres", COLUMNS)
+        assert insert_sql("sqlite", COLUMNS).count("?") == 39
        # 裁剪列表按位置占位符重新编号,不留空洞
        assert insert_sql("postgres", ["call_id", "model"]) == (
            "INSERT INTO llm_calls (call_id, model) VALUES ($1, $2) ON CONFLICT DO NOTHING"
@@ -377,13 +381,14 @@ class TestBackendColumnParity:
        """新列只能追加在末尾: 旧表经 ALTER 补列必落末尾,插在中间会让两条路径分叉。"""
        from polygateway.telemetry.schema import COLUMNS

-        assert COLUMNS[-6:] == (
+        assert COLUMNS[-7:] == (
            "cause_type",
            "error_body",
            "attempts",
            "total_latency_ms",
            "stall_ms",
            "stall_count",
+            "stream_end",
        )


@@ -776,7 +781,7 @@ class TestSQLiteCallerDimensionsAcceptance:

        conn = sqlite3.connect(db)
        cols = [r[1] for r in conn.execute("PRAGMA table_info(llm_calls)")]
-        assert cols == _EXPECTED_COLUMNS  # 22 → 38 个 recorder 字段(+ created_at 共 39 物理列)
+        assert cols == _EXPECTED_COLUMNS  # 22 → 39 个 recorder 字段(+ created_at 共 40 物理列)
        rows = dict(conn.execute("SELECT call_id, tenant_id FROM llm_calls").fetchall())
        assert rows["new-row"] == "tenant-a"
        assert rows["old-row"] == ""  # 不是 None: NULL 会被 RLS 静默吞掉
@@ -826,7 +831,7 @@ class TestSQLiteSchemaMode:
    """issue #13: `auto_migrate` 两档——auto 保持自动补列,manual 只裁剪写入不发 DDL。

    列数断言一律按**物理列数**写: 旧表 22 个 INSERT 字段 + `created_at` = 23,
-    补齐后 38 + `created_at` = 39。混用 INSERT 字段数与物理列数是本处最易错的地方。
+    补齐后 39 + `created_at` = 40。混用 INSERT 字段数与物理列数是本处最易错的地方。
    """

    def _physical_columns(self, db: Path) -> list[str]:
@@ -865,7 +870,7 @@ class TestSQLiteSchemaMode:
        assert "ALTER TABLE" in message  # 给出可直接执行的补列 SQL

    async def test_auto_mode_still_upgrades_the_legacy_table(self, tmp_path):
-        """auto + 同款旧表: 现状回归,补列后物理列数 23 → 39。"""
+        """auto + 同款旧表: 现状回归,补列后物理列数 23 → 40。"""
        db = tmp_path / "auto_legacy.db"
        _make_pre_tenant_db(db)

@@ -874,10 +879,10 @@ class TestSQLiteSchemaMode:
        recorder.close()

        assert self._physical_columns(db) == _EXPECTED_COLUMNS
-        assert len(self._physical_columns(db)) == 39
+        assert len(self._physical_columns(db)) == 40

    async def test_manual_mode_still_creates_a_fresh_table(self, tmp_path):
-        """manual 只管 ALTER,不管 CREATE: 全新库照建,39 个物理列齐全(设计 §4.2)。"""
+        """manual 只管 ALTER,不管 CREATE: 全新库照建,40 个物理列齐全(设计 §4.2)。"""
        db = tmp_path / "manual_fresh.db"
        recorder = SQLiteRecorder(db, auto_migrate=False)
        await _record_minimal(recorder, call_id="c-fresh", tenant_id="tenant-a")
@@ -1117,6 +1122,8 @@ class TestPostgresBackfillDiscipline:
        # 1.3.8 两列: 稳态定义随补列前移,否则每进程首写都去抢锁
        "stall_ms",
        "stall_count",
+        # 1.3.10 一列,理由同上
+        "stream_end",
    ]

    def _recorder(self, conn):
@@ -3586,3 +3593,92 @@ class TestTerminalRowSqlSemantics:
            ).fetchone()[0]
            == 0
        )
+
+
+# 1.3.8 形态(1.3.10 补列前)的表: 38 个 INSERT 字段 + created_at = 39 物理列。
+# 冻结字面量(由 _PRE_138_DDL 末行补两列),不从常量切片派生,理由同 _PRE_138_DDL。
+_PRE_1310_DDL = _PRE_138_DDL.replace(
+    "    total_latency_ms     INTEGER\n);",
+    "    total_latency_ms     INTEGER,\n    stall_ms             INTEGER,\n"
+    "    stall_count          INTEGER\n);",
+)
+
+
+class TestStreamEndColumn:
+    """1.3.10 结束方式列(#31): 落行、缺省 NULL、补列幂等、manual 告警、各入口取值。"""
+
+    async def test_value_round_trips_and_default_is_null(self, tmp_path):
+        db = tmp_path / "stream_end.db"
+        recorder = SQLiteRecorder(db, auto_migrate=True)
+        await _record_minimal(recorder, call_id="c-eof", stream_end="eof_after_finish")
+        await _record_minimal(recorder, call_id="c-none")
+        recorder.close()
+        conn = sqlite3.connect(db)
+        rows = dict(conn.execute("SELECT call_id, stream_end FROM llm_calls").fetchall())
+        conn.close()
+        assert rows == {"c-eof": "eof_after_finish", "c-none": None}
+
+    async def test_auto_migrate_backfills_from_138_table(self, tmp_path):
+        db = tmp_path / "pre1310.db"
+        conn = sqlite3.connect(db)
+        conn.execute(_PRE_1310_DDL)
+        conn.commit()
+        conn.close()
+        for call_id in ("c1", "c2"):  # 二次初始化幂等
+            recorder = SQLiteRecorder(db, auto_migrate=True)
+            await _record_minimal(recorder, call_id=call_id, stream_end="done")
+            recorder.close()
+        conn = sqlite3.connect(db)
+        assert [r[1] for r in conn.execute("PRAGMA table_info(llm_calls)")] == _EXPECTED_COLUMNS
+        conn.close()
+
+    async def test_manual_mode_names_the_new_alter(self, tmp_path, captured_warnings):
+        db = tmp_path / "manual1310.db"
+        conn = sqlite3.connect(db)
+        conn.execute(_PRE_1310_DDL)
+        conn.commit()
+        conn.close()
+        recorder = SQLiteRecorder(db, auto_migrate=False)
+        await _record_minimal(recorder, call_id="c-manual", stream_end="done")
+        recorder.close()
+        conn = sqlite3.connect(db)
+        assert len([r[1] for r in conn.execute("PRAGMA table_info(llm_calls)")]) == 39
+        assert conn.execute("SELECT call_id FROM llm_calls").fetchone() == ("c-manual",)
+        conn.close()
+        assert "ALTER TABLE llm_calls ADD COLUMN stream_end TEXT" in captured_warnings[0]
+
+    async def test_attempt_takes_stream_end_from_response(self):
+        rec = _MemoryRecorder()
+        await TelemetryEmitter(rec, text_cap=None, scope="LLM").emit_attempt(
+            request=_REQ,
+            source=_source(),
+            call_id="c",
+            latency_ms=1,
+            response=_resp(stream_end="eof_after_finish"),
+            error=None,
+            reasoning_applies=True,
+            operation="chat",
+        )
+        assert rec.rows[0]["stream_end"] == "eof_after_finish"
+
+    async def test_failed_hit_and_terminal_rows_are_null(self):
+        rec = _MemoryRecorder()
+        emitter = TelemetryEmitter(rec, text_cap=None, scope="LLM")
+        await emitter.emit_attempt(
+            request=_REQ,
+            source=_source(),
+            call_id="c",
+            latency_ms=1,
+            response=None,
+            error="boom",
+            reasoning_applies=True,
+            operation="chat",
+        )
+        # 命中行显式不信任 response 上的值(即使上游误带)
+        await emitter.emit_cache_hit(
+            request=_REQ, response=_resp(cache_hit=True, stream_end="done"), operation="chat"
+        )
+        await emitter.emit_terminal_failure(
+            request=_REQ, call_id="c", error="dead", operation="chat", stats=_MIGRATED_STATS
+        )
+        assert [row["stream_end"] for row in rec.rows] == [None, None, None]
Author	SHA1	Message	Date
iomgaa	25a90ebd05	chore: merge 1.3.10 unified completion verdict (#30/#31)	2026-10-15 00:38:40 -04:00
iomgaa	1201a04449	chore: date 1.3.10 changelog for release	2026-10-15 00:38:19 -04:00
iomgaa	abcdf6930e	fix: address completion verdict review findings - fidelity cacheability also requires a non-empty projected body, so think-tag-only answers are no longer cached - a no-argument tool sent with arguments null only is a deterministic dialect: ResultInvalid tool_arguments_null instead of retried - text mode treats choice index null as 0 and no longer lets False pose as 0 - null-only tool_calls elements no longer produce an empty tool_calls list; non-stream fidelity checks role before the verdict; warning paths are length-bounded - docs: ARCHITECTURE 7.8 field counts, README cache/watchdog wording, CHANGELOG notes, review record in the design, hook note in the plan Red: 8 new tests failed before this change (e.g. test_openai_cacheable_requires_projected_body, test_null_arguments_tool_is_result_invalid_stream, test_text_mode_choice_index_null_is_zero[None-True]).	2026-10-15 00:28:20 -04:00
iomgaa	31942f55f1	docs: document completion verdict and bump version to 1.3.10 CHANGELOG 1.3.10 (unreleased) with behaviour-change and breaking sections, README capability rows and a 1.3.10 section, ARCHITECTURE section 15 plus 6.2/7 updates, schema doc for the stream_end column, migration impact notes, ROADMAP entry, .env.example MISSING_DONE semantics, and a revision pointer on the 1.3.9 design. Follow-ups filed as #32 and #33.	2026-10-15 00:03:15 -04:00
iomgaa	779a9b29a3	feat: record stream end mode in telemetry llm_calls gains a nullable stream_end TEXT column (auto backfill, manual-mode trimming as in 1.3.8). TelemetryRecorder.record_llm_call takes a new required keyword stream_end; the emitter reads it from the response on attempt rows and writes NULL for failed, cache-hit and terminal rows. Red: 18 tests in tests/unit/test_telemetry.py and test_ports.py failed against the pre-change source (missing column, 38-field signature, stream_end absent from emitted rows).	2026-10-14 23:54:11 -04:00
iomgaa	48ba36dd94	fix: gate cache writes and reads on a shared cacheability verdict Cacheable is now a strict subset of deliverable: salvaged streams, content-less results and length/filter/unknown finishes are delivered but never written, and the read path applies the same verdict to the stored values before hit overrides (invalid stream_end or non-string finish read as a miss; 1.3.9 fidelity entries fall back to the completion's finish). Hits reset stream_end. The legacy _has_completion_output helper is folded into _is_cacheable. Red: 13 tests in tests/unit/test_cache.py failed before this change (uncacheable results were written and replayed; new fields not reset).	2026-10-14 23:39:50 -04:00
iomgaa	39a319e34e	fix: judge stream completion by finish reason and trailing usage Both transport paths now share one verdict: a stream is complete on [DONE] or on finish_reason followed by measured usage and a clean EOF (both modes; missing_done only governs real truncation). Outcomes follow the design's first-match table: upstream_error finishes stay transient, fidelity mode delivers reasoning-only/length results, text mode raises ResultInvalid for content-less length/filter endings, and non-stream fidelity judges before building the snapshot so truncated tools reach tool_truncated. Projections honour the reasoning alias, dialect warnings are bounded, and responses carry finish_reason and stream_end. Red: 39 of 46 tests in tests/unit/test_completion_transport.py failed before this change (e.g. eof_after_finish raised 保真响应缺 [DONE]; reasoning-only length raised empty_completion; reasoning-only stream hit the TTFT watchdog, re-confirmed by disabling alias liveness).	2026-10-14 23:31:54 -04:00
iomgaa	16743ca11d	fix: merge unknown and null stream fields losslessly in completion accumulator Null is treated as absent at every layer; unknown delta, tool and function fields merge by OpenAI streaming conventions (strings concat, label keys idempotent, indexed arrays by index) and conflicts drop the key and record its path instead of failing the call. The accumulator now tracks finish and usage-after-finish in both modes, exposes the reasoning alias, and non-stream parsing splits shape checks from snapshot construction. Red: 20 of 47 tests in tests/unit/test_openai_completion.py failed before this change (e.g. test_null_known_and_unknown_fields_are_absent raised TransientError 未知 delta 字段; finish_seen/raw_message missing).	2026-10-14 23:21:05 -04:00
iomgaa	e5ad67408a	feat: add kernel completion verdict functions and response fields Pure functions in types.py classify finish reasons, judge stream completeness and outcome, and decide cacheability; TransportResult and LLMResponse gain finish_reason and stream_end; snapshots accept tool_calls null. Red: tests/unit/test_completion_verdict.py failed at collection (ImportError: _classify_finish et al. undefined) before this change.	2026-10-14 23:12:00 -04:00
iomgaa	ab0e924849	docs: add implementation plan for unified completion verdict	2026-10-14 23:04:05 -04:00
iomgaa	607523ef5c	docs: fold round-2 review fixes into completion verdict design	2026-10-14 22:46:56 -04:00
iomgaa	95104af2a5	docs: add unified completion verdict design for issues 30 and 31	2026-10-14 22:39:47 -04:00
iomgaa	5d2ad40336	chore: gate complexity only on functions touched by the change Whole-tree radon blocking let 16 legacy C+ functions block every commit, including docs-only ones. Touched functions still block; legacy debt is reported as a warning.	2026-10-14 22:32:52 -04:00
iomgaa	22609c90d6	chore: merge approved 1.3.9 release preparation	2026-10-01 23:16:18 -04:00
iomgaa	a0cf8546dc	chore: prepare approved 1.3.9 release version	2026-10-01 23:08:37 -04:00
iomgaa	ce10881eb0	docs: archive preserved wiki records and register hedge nodes	2026-10-01 23:03:51 -04:00