Realtime Interpreter Breaks Mac Latency Wall

WebRTC VAD likely matters as much as the model swap because it trims dead air before the pipeline spends cycles on it; native whisper-cpp-python and llama-cpp-python bindings should cut overhead and memory churn on Apple Silicon compared with heavier wrappers; Tencent's HY-MT1.5-1.8B-GGUF is a sensible fit here: translation-focused, compact, and explicitly positioned for edge deployment; zero-shot prompting and minimal context are the right latency tradeoff for meetings, but accents, noise, and long clauses will still be the stress test; packaging it as a .dmg is the real validation step, because beta feedback from actual meetings will matter more than a clean demo clip.