research(ai): choose a local inference runtime and baseline model #75

Before shipping AI features, choose a runtime and model setup that preserves the offline-first promise and remains supportable across desktop platforms.

Questions to answer

• Which embeddable runtimes fit the Rust/Tauri stack and their licenses?
• What CPU, RAM, disk, and startup costs should users expect?
• What works on Apple Silicon, Windows x64, and Linux without requiring a separate server?
• Can models be imported locally as well as downloaded explicitly?
• How will artifacts be versioned, checksummed, removed, and upgraded?
• How do we prove there is no silent cloud fallback or background upload?

Deliverable

Add a short architecture decision record with a small benchmark table, a throwaway prototype, and a recommended baseline. This issue does not add a production AI feature.