test: add replay consistency harness

Author: 张志鹏

tests/sessions/REPLAY_CONSISTENCY_DESIGN.md

@@ -0,0 +1,23 @@
+# Replay Consistency Design
+
+The replay harness models each case as a normalized stream of operations:
+session creation, ordinary or tool events, state deltas, memory store/search,
+and summary creation/update. Lightweight CI has an InMemory-only path and an
+InMemory-vs-SQLite path, so contributors do not need Redis or MySQL locally.
+External SQL and Redis backends are explicit opt-in paths through
+`TRPC_AGENT_REPLAY_SQL_URL` and `TRPC_AGENT_REPLAY_REDIS_URL`.
+
+Normalization is intentionally narrow. Timestamps become a placeholder,
+generated summary event ids become a stable marker, JSON objects are key sorted,
+memory hits are sorted after timestamp normalization, and summary text only gets
+whitespace folding. Business data remains strict: event order, authors, content,
+state, memory content, summary ownership, summary version/update, and overwrite
+semantics are all compared.
+
+The diff report uses InMemory as the baseline and recursively compares each
+other backend. Every difference records the case id, expected backend, actual
+backend, field path, and both values. Summary comparison is split into semantic
+text and storage metadata: `session_id`, `summary_id`, version, updated-at
+placeholder, and compressed/original event counts remain independent fields.
+`allowed_diff` only accepts explicit rules; the lightweight mode currently has
+no default allowed differences, so real mismatches cannot disappear silently.