MioSub

修复

• 本地 Whisper：修复在未安装 Microsoft Visual C++ 运行库的电脑（常见于新装机）上，本地转录和 VAD 全部以退出码 3221225781 失败的问题。内置的 whisper.cpp 二进制（v1.8.8）改为静态链接运行时，不再依赖 VC++ 运行库；如果二进制仍无法启动，MioSub 会给出带 vc_redist 下载链接的明确提示（而非原始退出码），并在二进制更新检查中主动提供替换下载（修复 MIOSUB-7J、MIOSUB-7K）。
• 本地 Whisper：恢复内置 whisper.cpp 的 --version 参数（v1.8.7 版本中意外丢失），版本检测和更新检查恢复正常。能正常运行但无法报告版本号的二进制现在会收到更新提示，而不再永远停留在「未知版本」。
• 生成管线：在解析边界校验 LLM 返回的 JSON 字段类型——第三方中转返回的畸形术语表条目（嵌套对象）和非字符串说话人名称不会再导致术语表合并、编辑器渲染或字幕导出崩溃；旧版本已持久化的损坏术语表会在加载时自动清理（修复 MIOSUB-76、MIOSUB-6Y、MIOSUB-78、MIOSUB-79）。

Fixes

• Whisper (Local): Fix all local transcription and VAD failing with cryptic exit code 3221225781 on machines without the Microsoft Visual C++ runtime (typically brand-new PCs). The bundled whisper.cpp binaries (v1.8.8) are now built with the static runtime and no longer require the VC++ Redistributable. If a binary still cannot load, MioSub shows an actionable message with the vc_redist download link instead of a raw exit code, and the binary updater now offers a replacement download for this case instead of staying silent (Fixes MIOSUB-7J, MIOSUB-7K).
• Whisper (Local): Restore the --version flag in the bundled whisper.cpp build (lost in v1.8.7), so version detection and update checks work again. Binaries that run fine but cannot report a version are now offered an update instead of being stuck on "unknown" forever.
• Pipeline: Validate LLM JSON field types at the parse boundary — malformed glossary terms (nested objects) and non-string speaker names returned by third-party proxies no longer crash the glossary merger, the editor, or subtitle export; corrupted glossaries persisted by earlier versions are cleaned up on load (Fixes MIOSUB-76, MIOSUB-6Y, MIOSUB-78, MIOSUB-79).

修复

• 编辑器：修复过长的下拉菜单溢出屏幕底部、导致滚动条无法拖动的问题。「说话人」筛选菜单和每行的说话人选择器现在会根据触发按钮上下方的可用空间限制高度，超出部分在菜单内部滚动。
• 导入：导入单语字幕（例如只含译文的导出文件）时，MioSub 会询问这段文本应作为原文还是译文，而不再一律当作原文。重新导入只含译文的文件不会再把译文放进原文栏。双语文件的导入行为保持不变。

Fixes

• Editor: Fix long dropdown menus overflowing past the bottom of the screen, which left their scrollbar unreachable. The Speaker filter and the per-row speaker picker now cap their height to the space available above/below the trigger and scroll internally.
• Import: When importing a single-language subtitle (for example a translation-only export), MioSub now asks whether the text is the original or the translation, instead of always treating it as the original. Re-importing a translation-only file no longer puts the translation in the source column. Bilingual files import as before.

修复

• 编辑器：修复下拉菜单和提示框被容器边缘裁切的问题——每行的说话人选择器（尤其是片段最后一行）、工具栏的「问题」/「说话人」筛选菜单，以及时间编辑校验提示框，现已通过 portal 渲染，不再被遮挡（Fixes LOCAL-L11）。

Fixes

• Editor: Fix dropdown menus and tooltips being clipped at their container edge — the per-row speaker picker (notably on a segment's last row), the Issues/Speaker filter menus in the toolbar, and the time-edit validation tooltip are now rendered in a portal so they are never cut off (Fixes LOCAL-L11).

新功能

• 设置：新增分步骤自定义 Gemini 模型名称——在 设置 > 服务 中新增「自定义模型名称」弹窗，可分别为各处理步骤（精修、翻译、术语提取、说话人分析、批量校对）指定所用的 Gemini 模型。仅限 Gemini 系列（保存时校验）；留空则使用默认值。

修复

• 本地 Whisper：当误将模型文件（如 .bin）选为 Whisper 可执行程序时，在预检阶段给出清晰、可操作的错误提示，而非稍后以晦涩的 spawn 错误失败（Fixes MIOSUB-5X）。
• 本地 Whisper：在冷启动瞬时失败时，短暂预热后重试一次版本/GPU 探测，提升版本与 GPU 支持检测的可靠性。
• 转写（OpenAI）：将网络/连接失败（离线、DNS 被屏蔽、无代理）归类为可操作的「请检查网络/代理」错误，而非笼统的转写失败，避免长时间重试循环。
• 术语表：对所有术语查找做未定义值保护，避免术语缺失时崩溃（Fixes LOCAL-L10）。

Features

• Settings: Add per-step custom Gemini model names — a new "Custom Model Names" dialog under Settings > Services lets you override the Gemini model used for each pipeline step (refinement, translation, glossary extraction, speaker analysis, batch proofread). Gemini-series only (validated on save); blank fields fall back to the defaults.

Fixes

• Whisper (Local): Detect when a model file (e.g. .bin) is mistakenly selected as the Whisper executable and show a clear, actionable error during preflight, instead of failing later with a cryptic spawn error (Fixes MIOSUB-5X).
• Whisper (Local): Retry the binary version/GPU probe once after a brief warm-up on transient cold-start failures, improving the reliability of version and GPU-support detection.
• Transcription (OpenAI): Classify network/connection failures (offline, DNS blocked, no proxy) as an actionable "check your network/proxy" error instead of a generic transcription failure, avoiding long retry loops.
• Glossary: Guard all glossary term lookups against undefined values to prevent a crash when a term is missing (Fixes LOCAL-L10).