feat(recall): bilingual keyword search for cross-language recall

[m0Nst3r873]

Problem

Team knowledge recall is purely lexical (BM25 token matching over a
shared tokenize() on both the index and query sides). The knowledge
base mixes Chinese and English — a learning may be titled in English
while a task is described in Chinese, or vice versa. Because Chinese
query tokens and English document tokens are different strings, they
never match: a Chinese-only query cannot recall an English-only
document, and vice versa.

Approach

Rather than maintain a hardcoded synonym dictionary (zero coverage of
long-tail terms, ongoing maintenance burden), this leverages the LLM's
own translation ability at keyword-selection time. The recall agent is
instructed to expand every meaningful domain term into both
languages when building its search keywords, so a single query
carries tokens that can match documents in either language.

• Chinese task → also add the English term (重启推理集群 → restart inference cluster)
• English task → also add the Chinese term (restart inference cluster → 重启 推理 集群)
• Any language-neutral token — anything written the same way in both
  languages or with no meaningful translation (brand/product names,
  code identifiers and symbols, error codes, version strings, file paths,
  CLI commands, API names, e.g. AppID, Kubernetes, GPU,
  RuntimeError, v2.1.3, oauth) — is kept as-is, since translating it
  only hurts matching.

No code-path or index changes, no SEARCH_INDEX_VERSION bump, no
rebuild required — this is purely a prompt/rule refinement.

Changes

• agents/teamai-recall.md — Step 2 (keyword extraction) now instructs bilingual expansion, keeping language-neutral tokens untranslated.
• src/builtin-rules.ts — the built-in teamai-recall rule's Bash-command usage note gets the same bilingual guidance, so the direct teamai recall path benefits too.

Test Plan

• npx tsc --noEmit — passes.
• Behavioral check: fed tasks through the new Step 2 prompt and verified the generated teamai recall command expands domain terms bilingually while keeping language-neutral tokens intact:
  • Chinese 升级到 v2.1.3 后 nginx 报 502 Bad Gateway,如何排查 → v2.1.3 nginx 502 Bad Gateway 排查 troubleshoot 升级 upgrade (version / brand / error code kept as-is)
  • English the oauth token refresh throws InvalidGrantError in production → oauth token refresh 刷新 InvalidGrantError production 生产环境 (protocol / error code kept as-is)
  • Path/command 修改 src/utils/tokenizer.ts 的分词逻辑导致索引失效 → src/utils/tokenizer.ts 分词 tokenizer 索引 index 失效 invalidation (file path kept as-is)
  • Mixed 为什么 kubectl apply 部署推理服务时 pod 一直 CrashLoopBackOff → kubectl apply 推理 inference 服务 service CrashLoopBackOff pod 部署 deploy (CLI command / pod state kept as-is)

🤖 Generated with Claude Code