v0.5.0: stats, permissions, idle unload, model location, diagnostic logs

[giraphant]

Summary

• 听写统计 pane: total dictation time, characters, average speed, last-30-day and per-app breakdown (counts only, stored in ~/Library/Application Support/Dousha/stats.json).
• 权限管理 pane: Microphone / Speech Recognition / Accessibility status with request / open-settings actions.
• Local Qwen3 idle unload: model is released after a configurable delay (default 1 min, 10 s–20 min), then Memory.clearCache(); RSS drops ~1.1 GB → ~125 MB, reload ~0.3 s.
• Model location: local models now live in ~/Library/Application Support/Dousha/Models (migrated from ~/.cache/huggingface).
• Diagnostic logs: 保存诊断日志 toggle (default on) with reveal / delete, under 历史记录 › 诊断. Logs never contain dictation text.
• Four-character sidebar titles; version bump to 0.5.0 (19).

Test plan

• swift test — 452 tests pass
• make install + manual check of stats, permissions, idle unload, model migration and log toggle

🤖 Generated with Claude Code