feat(chat): add optional character RAG with bilingual starter knowledge

[Lucas1479]

Main Chat can miss Kurisu's canonical BBS handle, especially across Chinese/Japanese spellings. This adds optional local character retrieval shared by all eight Chat routes, with a sourced bilingual starter corpus. RAG is off by default and the static persona is unchanged. Based on current public main d5e1d7f, including #55.

• Reuse FAISS + multilingual E5 and JSON source data. Build from a file or directory; include 15 topics / 30 Chinese and Japanese records. Personal data, index binaries and weights are not published.
• Configure enablement, index directory, top-k and threshold through .env or Settings. Show applied configuration, loading/failure state and last retrieval summary. CLI diagnostics show rejected candidates as well as accepted hits.
• Label retrieved passages as fallible current-turn reference data, preserve the original user message/history and existing fallback behavior, and reuse the cache-only model-loading boundary. Accepted excerpts go to the selected model, including remote APIs.
• Add the optional rag extra without changing CPU/cu124/ROCm build selection; add a dedicated CPU retrieval CI job. Update both READMEs, setup/evaluation docs and source-archive selection.

Validation: 107 targeted Python tests passed (1 skipped); all 32 real E5 retrieval development cases passed; 15 Electron tests and TypeScript/Vite build passed. Ruff, architecture views, clean core environment verification and source-release checks passed. Clean model-less core regression: 1,837 passed, 12 skipped. Remote CI is tracked by the checks below.

The evaluation records 100 live Flash requests across successive corpus iterations, not 100 persona passes. Known limitations include irrelevant retrieval for a Chinese greeting, occasional masculine voice, unprompted Christina reactions and invented shared history. Some failures occur without any retrieved reference. These are disclosed; this PR does not claim to solve default-off persona consistency or justify enabling RAG by default.

Related to #56; keep the issue open for contributor retesting and separate persona-baseline follow-up. Setup: optional character knowledge.

---

Main Chat 对红莉栖 BBS 网名及中日文写法的识别存在缺口。这次接入各聊天路径共用的可选本地检索，并提供有来源的中日文基础资料；默认关闭，固定 persona 不变。分支基于包含 #55 的公开主线 d5e1d7f。

复用 FAISS/E5/JSON 基础方式，支持知识目录和文件，附 15 个主题、30 条资料。开关、索引目录、top-k 和阈值可通过 .env 或 Settings 配置；显示实际生效值、加载状态和最近检索摘要，CLI 可查看被过滤的候选。资料作为明确标注的本轮参考传入，保留原始用户输入和历史；命中文字会发给所选模型，包括远程 API。模型仅在显式准备时下载，聊天运行时使用本地缓存。

新增可选 rag 依赖和 CPU 检索 CI，保持现有 CPU/cu124/ROCm 构建选择，补齐中英文 README、配置及评估说明。未上传旧个人库、索引或模型。

验证：107 项相关 Python 测试通过、1 项跳过；32 条真实检索开发用例通过；15 项 Electron 测试、前端构建、Ruff、架构视图、干净 core 环境及源码发布检查通过。干净且未安装模型依赖的 core 完整回归：1,837 项通过、12 项跳过。远程 CI 状态见下方检查。

评估记录了多个资料迭代中的 100 次真实 Flash 调用，不是 100 次角色测试全通过。中文闲聊误召回、偶发男性口吻、无端反应“克里斯蒂娜”和虚构共同经历等仍有反例，部分出现在没有检索资料时。这次先完成 RAG 和资料，人设基线另行处理；#56 保持开启，邀请贡献者重建索引并复测。