Local AI

QVAC-backed inference is optional in the preview. A local model avoids sending prompts to a hosted inference provider, but it still requires compatible hardware, disk space, and an initial model download.

• Keep model initialization in the worker, never the renderer.
• Report download, loading, ready, and error states separately.
• Use mock inference for deterministic UI and protocol tests; label it as mock.
• After a successful model install, test local inference without internet access.
• Do not publish a universal RAM or download-size claim until the chosen model is pinned and measured.