v3.4.0 - EPUB Context Chat and Reliable Panel Startup | EPUB 上下文对话与面板稳定启动

✨ What's Changed

• 📚 EPUB context chat: AIdea can now use EPUB books as the active document context in Zotero’s reader panel. It supports EPUB 3 navigation documents and EPUB 2 NCX structure, with semantic headings and spine order as fallbacks for less structured books. Thanks @senlinyy for contributing the document-adapter and EPUB parsing work in #69; the integrated commit preserves the original Git authorship.

• 🎯 Deterministic, bounded retrieval: Explicit chapter references, whole-book requests, and ambiguous follow-ups are routed locally. AIdea sends only bounded relevant excerpts to the answering model and does not make an additional model request to plan sections.

• 🌐 Context-aware EPUB selection translation: EPUB selections use bounded, selection-anchored book context directly, without a separate cold-start request. Selection translation remains automatic, with no additional custom-instruction setting.

• 🛡️ Safer EPUB processing: Archive paths, entry counts, individual decompressed entries, total extracted text, and normalized book text are bounded to prevent malformed or oversized EPUB files from exhausting Zotero resources.

• 🐛 Reliable panel startup: Required chat tables are committed before optional historical migrations run. Failed core initialization can be retried, migration failures no longer leave the AIdea panel blank, and Zotero SQLite LIKE patterns use bound parameters. Thanks @senlinyy for reporting #71.

• 🔄 PDF and history compatibility: Existing PDF context behavior and legacy basePdf conversation data remain compatible while the shared document-adapter architecture supports future formats.

• ✅ Validation: The release has passed 438 TypeScript tests, formatting and lint checks, production builds, reproducible EPUB 2/3 fixture tests, GitHub CI, and local Zotero verification covering EPUB reading, panel initialization, and selection translation.

• 🔁 After updating: Restart Zotero, then open a PDF or EPUB and use the same AIdea reader panel. No additional mode switch is required.

• 🔒 Privacy and model requests: EPUB extraction and section routing run locally. Only the bounded context selected for the current request is sent to the configured model provider. No additional planning-model request is made.

• ⚠️ EPUB structure: Retrieval quality benefits from publisher-provided navigation and chapter structure. For less structured books, AIdea falls back to semantic headings and spine reading order.

📝 更新内容

• 📚 EPUB 上下文对话：AIdea 现在可以在 Zotero 阅读器侧边栏中把 EPUB 图书作为当前文档上下文。支持 EPUB 3 Navigation Document 和 EPUB 2 NCX 目录结构；对于结构不完整的图书，会使用语义标题和 spine 阅读顺序作为后备。感谢 @senlinyy 在 #69 中贡献文档适配器与 EPUB 解析工作；整合后的提交保留了原始 Git 作者信息。

• 🎯 本地确定性有界检索：明确章节、全书请求和模糊追问均在本地完成章节路由。AIdea 只把有界的相关原文片段发送给回答模型，不会为了规划章节而额外调用一次模型。

• 🌐 带上下文的 EPUB 划词翻译：EPUB 划词翻译直接使用以选中文本为锚点的有界图书上下文，不再执行单独的冷启动请求。划词翻译仍然自动工作，不增加自定义指令设置。

• 🛡️ 更安全的 EPUB 处理：对压缩包路径、条目数量、单个解压条目、累计提取文本和规范化图书文本设置明确上限，避免异常或超大 EPUB 占用过多 Zotero 资源。

• 🐛 面板启动更加可靠：聊天所需的核心数据表会先提交，再运行可选的历史迁移。核心初始化失败后可以重试，迁移失败不再导致 AIdea 面板空白，同时 Zotero SQLite 的 LIKE 查询改为绑定参数。感谢 @senlinyy 报告 #71。

• 🔄 兼容现有 PDF 和历史记录：原有 PDF 上下文行为保持不变，并继续兼容旧版 basePdf 对话数据；共享文档适配器架构也为后续格式扩展提供统一入口。

• ✅ 验证情况：已通过 438 项 TypeScript 测试、格式与 lint 检查、生产构建、可复现的 EPUB 2/3 fixture 测试、GitHub CI，以及覆盖 EPUB 阅读、面板初始化和划词翻译的本地 Zotero 验证。

• 🔁 更新后操作：请重启 Zotero，然后打开 PDF 或 EPUB，直接使用同一个 AIdea 阅读器侧边栏，无需切换额外模式。

• 🔒 隐私与模型调用：EPUB 内容提取和章节路由均在本地完成，只有针对当前请求选出的有界上下文会发送给已配置的模型服务商，不会额外调用章节规划模型。

• ⚠️ EPUB 结构说明：检索效果会受图书内置目录和章节结构影响。对于结构不完整的图书，AIdea 会使用语义标题和 spine 阅读顺序作为后备。