v0.49.2 - Accurate Vision and Model Cache Status

This patch makes the primary TUI report multimodal capability separately from loaded backend readiness and clearly reuses cached model artifacts instead of presenting every switch as a new download.

Highlights:

• Qwen and Gemma show vision on-demand before their projectors are loaded.
• Vision ready requires an exact healthy model, context, and verified projector match.
• First image use prepares the projector automatically; exact model and projector caches are reused.
• Model selection separates base-model cache state from lazy projector state.
• Vision fact replies no longer intercept implementation, debugging, analysis, or test requests.
• Production-shaped regressions cover stopped, stale, mismatched, ready, unsupported, unselected, and cache-hit paths.

An exact cache hit does not download the model again, although SHA-256 verification of a multi-gigabyte artifact can still take time.

GitHub Releases remains the only supported binary distribution channel.