v0.5.0

Windows and Linux, tested end to end. Recording, local transcription, frames, the review page and exports were run step by step on Windows and Linux machines, and CI now checks both on every change. Fixes from that run:

• First run on a fresh Windows: the speech-model download failed with CERTIFICATE_VERIFY_FAILED (Windows fetches the root CA behind huggingface.co only on demand). Downloads now verify TLS through the OS trust store.
• Recording over Remote Desktop or on machines without a GPU: capture fell over when ddagrab had no D3D11 device; it now falls back to gdigrab.
• No microphone (or PortAudio missing) records video only instead of crashing.
• items.json saved by PowerShell or Notepad (with a BOM) loads.
• Linux (X11) recording: the Tk recorder and headless capture use x11grab; doctor.py lists the apt packages it needs.
• Review page: confirming no longer races a late autosave that could drop reviewed: true, and failed saves say so (#2, thanks @Hosuke). On a headless Linux box it doesn't try to launch a browser; the agent hands you the URL.

Windows、Linux 都能用了。 录屏、本地转写、截帧、审核页、导出在 Windows 和 Linux 上逐项跑通，CI 每次改动都在两边检查。顺手修了：新装 Windows 首次下载语音模型证书校验失败；远程桌面或无独显机器上录不了屏；没有麦克风时崩溃；PowerShell 或记事本存的 items.json 读不了；Linux（X11）录屏；审核页确认时可能被迟到的自动保存覆盖（#2，感谢 @Hosuke）。

Setup: Windows winget install astral-sh.uv + winget install Gyan.FFmpeg; Ubuntu/Debian sudo apt install ffmpeg libportaudio2 python3-tk + uv. The tray app, the pen and system audio are still macOS-only.