Adds gemma-4-26b-a4b-it-mlx-4bit as a second selectable MLX model alongside qwen3.5:9b. Settings UI gains a model picker that persists to providers.mlx.active_model. Bootstrap state becomes per-model (_STATES dict keyed by registry name); availability, download, and verification all take a model arg. The provider applies a tightly defended monkey-patch (precondition + post-write assertions) when loading gemma4. qwen3.5 behavior is unchanged when no model is explicitly chosen. Removes MLX_MODEL_REPO and MLX_MODEL_REVISION constants per CLAUDE.md §8 (no backwards-compat shims); all consumers now read from the MLXModelSpec registry. Tier mapping (MLX_PRO/FLASH/LITE) stays pinned to qwen3.5.