perf(chat): skip redundant transcript scans

[Duang777]

Summary

• skip full transcript parsing when append_message generates a fresh UUID
• preserve transcript deduplication for caller-supplied idempotency keys
• add a regression proving generated-ID writes remain append-only while explicit-ID replay remains idempotent

Performance evidence

Seven-run median for one generated-ID append, with fsync stubbed to isolate transcript scan cost:

Existing messages	Before	After
100	4.80 ms	1.27 ms
1,000	5.41 ms	1.25 ms
10,000	40.14 ms	1.00 ms

The common generated-ID path no longer scales with transcript length. No cache, index, schema, or migration was added.

Validation

• 88 Chat tests passed
• Ruff check passed on both changed files
• git diff --check passed
• loopx canary premerge --from-git-diff passed

Scope

A broader transcript index was considered and deferred: deterministic completion messages still need durable replay deduplication, and this change removes the unnecessary scan without adding state or recovery complexity.

[huangruiteng]

Approval conclusion

精确审阅 head：d5279c81ff295e3a7a5be7b1bb9842ff8a43ff21。结论为 APPROVE，未发现阻塞或非阻塞问题。

1. 变更与架构判断

本 PR 将 ChatSessionStore.append_message() 的 transcript 去重扫描限定到调用方显式提供 message_id 的路径。普通消息由 store 生成全新 UUID，可在既有 exclusive file lock 下直接 append；managed/attached completion 等显式稳定 ID 仍走原有扫描和幂等 replay。变更仅涉及 4 行生产代码和一个 46 行回归测试，没有引入索引、缓存、schema 或第二个状态 owner，范围与热路径问题成比例。

2. 语义与调用链复核

我检查了所有 append_message callers、_read_jsonl/_append_jsonl、managed/attached completion writers、Chat runtime consumers 和 PR #3981 引入的 caller-supplied idempotency contract。分支依据是原始可选参数是否存在，而不是生成后的 payload ID，因此显式 ID 始终保留锁内扫描；无 ID 调用才生成 UUID 并跳过无意义的 O(n) 预读。invalid ID 校验、JSONL row shape、fsync、排序和 readback 均未改变。

3. 验证结果

• tests/test_chat_session_active_turn.py：31 passed in 3.59s。
• 全部 tests/test_chat*.py：88 passed in 23.37s。
• Ruff、git diff --check 和 17 项 hosted checks 通过。
• 使用真实 filesystem ChatSessionStore 做 base/head poisoned-read counterfactual：base 会触发 transcript scan，当前 head 在不扫描时成功追加 32 字符 generated ID。
• 同一回归覆盖显式 ID replay，确认重复 completion 返回原 durable row 而不追加第二条。

4. 失败与兼容性判断

最强反例是把优化错误扩大到显式 completion ID，导致重试重复写 terminal message。当前实现由 message_id is not None 精确保护该路径，并在既有 exclusive lock 内保持 dedup；回归同时对 generated path 的“不得读”和 explicit path 的“必须幂等”敏感。无持久化迁移、权限、quota、scheduler 或外部副作用变化，回滚也只是恢复原三行扫描逻辑。

5. Future-facing pass

相邻 owner 内没有更有价值的小型重构：把这一条件抽成 helper 会稀释 caller-intent 语义，新增 transcript index/cache 则明显超出当前问题。保留局部条件和双向语义回归是更可读、可逆的边界。

English verdict: APPROVE exact head d5279c81ff295e3a7a5be7b1bb9842ff8a43ff21. The change safely removes the transcript pre-read only for store-generated message IDs while preserving caller-supplied idempotent replay under the existing lock. Base/head counterfactuals, 88 Chat tests, Ruff, diff hygiene, and 17 hosted checks support the intended behavior; no unresolved finding remains.