test: add replay consistency framework for session and memory

session_memory_summary_diff_report.json

@@ -0,0 +1,30 @@
+{
+  "description": "Sample field-level replay diff report for Session / Memory / Summary consistency tests.",
+  "allowed_diff_policy": [
+    "Event ids and timestamps are normalized before comparison.",
+    "Summary ownership, summary text, event counts, state values, memory text, and tool payloads are compared exactly.",
+    "Backend-specific differences must be listed by exact field path, never by broad wildcard."
+  ],
+  "sample_diffs": [
+    {
+      "case_id": "summary_update",
+      "session_id": "replay-session",
+      "event_index": null,
+      "summary_id": "replay-session",
+      "field_path": "$.summary.session_id",
+      "expected": "replay-session",
+      "actual": "other-session",
+      "allowed_diff": false
+    },
+    {
+      "case_id": "tool_call_and_response",
+      "session_id": "replay-session",
+      "event_index": 2,
+      "summary_id": null,
+      "field_path": "$.events[2].parts[0].response.temperature",
+      "expected": "30C",
+      "actual": "lost",
+      "allowed_diff": false
+    }
+  ]
+}

tests/sessions/replay_consistency_design.md

@@ -0,0 +1,3 @@
+# 回放一致性测试设计
+
+框架以 ReplayCase 描述标准输入，由同一驱动器依次执行会话事件、工具调用与响应、状态覆盖、记忆写读及摘要更新，再对 InMemory 与 SQLite 的快照做递归字段比较。快照会剔除随机事件 ID、时间戳等非确定字段，并固定回放事件顺序；摘要归属、正文、计数、状态覆盖关系及工具载荷保持精确比较。每条差异包含用例、会话、事件索引、摘要标识、字段路径和两侧值，后端特例只能按精确路径声明 allowed_diff。轻量测试使用内存数据库，外部 Redis/SQL 集成可通过环境开关启用并在缺少服务时跳过。