[test] 补齐 LlmSession / ResponsesAgentToolState 的 durable correctness 回归测试

[YueZh127]

背景

origin/dev 已经把 Responses 相关的核心运行态从 prototype 往正式主链推进，当前至少包括：

• LlmSessionGAgent
• ResponsesAgentToolStateGAgent
• ResponsesCompletionApplicationService
• current-state projection / query reader / registration adapter / command adapter

仓库中已经存在不少单模块测试，但目前更大的风险不是“单个类完全没测”，而是 整条 durable 链路虽然每个零件都有单测，却还缺足够明确的系统级防退化表达。

一旦这条链路退化，外层 API 仍可能表面可用，但 durable truth、tool state 恢复、continuation 一致性会悄悄失真。

目标

为 LlmSession / ResponsesAgentToolState 主链补齐 durable correctness 回归测试，重点锁住：

1. committed fact -> projection -> query 的 authority path
2. tool call emitted / received / resolved / expired 的状态推进
3. live event 缺失时的恢复语义
4. completion / continue / cancel 对同一 authority source 的一致依赖

In Scope

• test/Aevatar.GAgentService.Tests/Application/ResponsesCompletionApplicationServiceTests.cs
• test/Aevatar.GAgentService.Tests/Core/LlmSessionGAgentTests.cs
• test/Aevatar.GAgentService.Tests/Core/ResponsesAgentToolStateGAgentTests.cs
• test/Aevatar.GAgentService.Tests/Infrastructure/LlmSessionRegistrationAdapterTests.cs
• test/Aevatar.GAgentService.Tests/Infrastructure/ResponsesAgentToolStateCommandAdapterTests.cs
• test/Aevatar.GAgentService.Tests/Projection/LlmSessionCurrentStateProjectorTests.cs
• test/Aevatar.GAgentService.Tests/Projection/ResponsesAgentToolStateCurrentStateProjectorTests.cs
• 必要时少量 test/Aevatar.GAgentService.Integration.Tests/*

Out of Scope

• 扩写 Responses feature 范围
• 引入新的 runtime 抽象层
• 将 query path 改成 request-time priming / replay

建议补的测试缺口

• tool emitted -> tool received -> tool resolved 的完整 durable 状态推进测试
• tool_call_expired / session expiry 的 committed + query 可恢复语义
• live observation 缺失时，query/readmodel 仍能恢复 tool state 与 session state
• continue / cancel / completion 对同一 responseId/sessionId authority source 的一致性测试
• adapter / application service 不回退到 request-path 临时拼装或本地事实态缓存
• current-state version / refreshed state 与 committed authority 对齐

完成标准

• 至少新增 5 个 durable correctness 回归测试
• 至少包含 2 个“live miss but durable recover”场景
• 至少包含 1 个 expired / timeout / terminal tool state 场景
• 测试直接表达 authority source 与 recovery contract，而不是只看 DTO 形状

建议验证命令

dotnet test test/Aevatar.GAgentService.Tests/Aevatar.GAgentService.Tests.csproj --nologo --filter "LlmSession|ResponsesAgentToolState|ResponsesCompletion"
dotnet test test/Aevatar.GAgentService.Integration.Tests/Aevatar.GAgentService.Integration.Tests.csproj --nologo --filter "Responses|LlmSession"
bash tools/ci/query_projection_priming_guard.sh
bash tools/ci/projection_state_version_guard.sh
bash tools/ci/projection_state_mirror_current_state_guard.sh
bash tools/ci/test_stability_guards.sh

备注

这张 issue 的重点不是再加一圈普通单测，而是把 Responses 新 durable 主链真正锁成“就算 live 掉了、进程换了、重放恢复了，语义仍然诚实一致”的测试网。