Local inference and in-app updates

Highlights

• The llama.cpp local inference engine ships and updates with the Luxi App; there is no separate engine installer.
• First use requires an explicit model choice and download: Little (Qwen3.5 9B Q4, about 5.63 GB) or Pro (Qwen3.6 35B-A3B Q4, about 21.17 GB).
• GGUF models are not included in the App installer. Downloaded models remain on the Mac.
• Local inference has no cloud fallback. Models that have finished downloading remain available offline.
• Model providers, API keys, base URLs, remote models, and external ACP Agent configuration have been permanently removed. Chat and Luxi's built-in Agent remain available.
• Legacy cloud-model settings are removed during upgrade with a one-time notice. MCP and sandbox settings are unaffected.
• Intel Macs stay on the existing older release and are not offered 0.3.0.
• Adds anonymous product feedback with category, rating, and optional contact email.
• Adds automatic and manual update checks, download progress, install-and-restart, and a website fallback.
• Feedback environment data is limited to the app version, macOS version, CPU architecture, and interface language.
• Logs, task content, prompts, file paths, and user files are never uploaded automatically.

On Apple Silicon Macs, upgrading from 0.2.0 still requires a manual download. Once 0.3.0 is installed, later compatible releases can be installed in the app.

This release performs an irreversible local-only settings migration and does not support downgrading to an older Luxi version.