docs: 修正入门文档里已经失效的运维事实

[oratis]

背景

收到一份 brand-voice embedding 的 bug 报告，说 server/brand-voice-search.js 用错了 llm.embed() 的签名。排查结论：问题真实存在过，但已经在 #20（0c5323b）修掉了 —— 两个调用点现在都是 const { vectors } = await llm.embed([text])，契约测试 server/__tests__/brand-voice-embed-contract.test.js 也已经在仓库里。所以本 PR 不含任何代码修改。

但核对上下文的过程中撞上好几处文档与现状不符，每一处都会让下一个会话走一段冤枉路，所以单独收成这个 docs PR。

改了什么

位置	文档写的	实际
CLAUDE.md ×2 / memory.md ×3	prod DB 密码"在 .env 里"	这台机器上没有 .env，只有 .env.example。改为 gcloud secrets versions access latest --secret=DATABASE_URL --project=gameclaw-492005
CLAUDE.md（两处自相矛盾：377 / 234）、memory.md ×2（234）	服务端测试数	656 个 / 69 个文件，~1 秒
CLAUDE.md「Frontend has no tests」+「4 Vitest component tests」	前端没测试	13 个 vitest 文件 + 5 条 Playwright（3 个 spec），与 memory.md §6 已关闭表一致
memory.md §5.4「没有 Sentry / OTEL（Sprint 1 待加）」	没接入	都已接入（server/sentry.js / server/otel.js），只是要配 SENTRY_DSN / OTLP endpoint 才上报。原文与同文件 §6 和 CLAUDE.md 自相矛盾

顺带查证的生产事实

借这次机会用 Cloud SQL Auth Proxy 只读查了一次 prod：

brand_voices coverage: { total: 0, with_embedding: 0, with_model: 0 }

表是空的 —— embedding 写入路径坏了这么久并没有造成数据损失，不需要 backfill。 下次部署后创建的第一条 brand voice 就会正常写入向量。这条结论记进了 memory.md §6，省得以后再查一遍。

同时新增一条「仍然成立」条目：embedding 只在 POST /api/brand-voices 创建时写（server/index.js:3419 是唯一写点），目前只有 GET/POST/DELETE、没有 update 路由所以暂时无害，但以后加编辑接口必须重新 embed。

验证

• 纯文档改动，未触碰任何代码路径
• npm test → 656/656 通过

提醒

prod 仍停在 00049-w2x，#6–#21 这一批（含 brand-voice 修复）尚未部署。部署前检查项见 docs/MASTER_PLAN_2026-08.md §5 —— 这批改了启动契约（MAILBOX_ENCRYPTION_KEY 现在 fail-fast、webhook 无 secret 时 fail-closed）。

🤖 Generated with Claude Code

[oratis]

Review: approving — every claim checked, but the numbers had themselves gone stale

A docs PR is only worth its accuracy, so I re-verified each claim against main at 7df4183 instead of reading the diff for plausibility.

claim	verdict
no local .env, only .env.example	✅ ls confirms .env: No such file or directory
Sentry + OTEL wired	✅ server/sentry.js, server/otel.js both present
13 Vitest files under client/src/{components,pages,utils}/	✅ exactly 13 — 6 / 5 / 2 across those three dirs
5 Playwright tests across 3 specs	✅ 3 spec files in e2e/tests/
brand_voices empty in prod → no backfill	✅ consistent with 0c5323b; not re-queried
656 tests / 69 files	❌ stale — actually 678 / 71

Fixed in b9dbea5

#23 (+9 tests, +1 file) and #24 (+7, +1) merged while this branch sat open, so 656 / 69 was wrong before it could land. That's precisely the failure mode this PR exists to fix, just with a shorter fuse. Corrected in all five places — CLAUDE.md ×2, memory.md §5.2, §7.3, and the Last reviewed footer, which also still read post #20.

Worth noting for future doc passes: a hardcoded count in five places goes stale on every merge. The prose that doesn't carry a number ("13 Vitest files under …") aged fine here.

Also added

The SQLITE_BUSY flake, because it cost me a diagnostic detour on this very batch. npm test is node --test, which runs files concurrently against the one shared influencex.db at the repo root — writes collide and 2–5 files go red on a rotating cast. It reproduces on main with zero changes applied, and serialized (--test-concurrency=1) is a stable 678/678.

Left undocumented, the next session reads the flake as its own regression — the same wasted-detour cost the rest of this PR is written to prevent, which is why it belongs here rather than in a follow-up.

Verified: 678/678 serialized on main, and on main + #25 + this branch.

Merging.