The MLX engine — from protocol stub to Metal-native inference in 48 hours

Built the MLX engine from scratch in two days: foundation, ViewModel migration, VLM image pipeline, multi-file downloads with SHA-256 integrity, and runtime engine switching UI. Day one was the protocol work — InferenceEngine replaced the LiteRT-LM-specific interfaces, and 24 test files migrated to MockInferenceEngine. Day two was the real engine: sampling parameters, native tool calling, streamDetails metrics, and the VLM pipeline that lets Gemma 4 see images through Metal-native compute. The multi-file download system handles MLX's split-weight format (5-10 files per model) with SHA-256 verification per shard. The engine switching dropdown in the toolbar lets you swap backends without restarting. 47 tests for format detection, engine switching, and RuntimeType. The "three engines, one app" log entry from July 6 gets the glory, but this is where the protocol architecture proved it works.