Cross-Architecture Sparse Prefill for Large Language Models on Apple Silicon. 3.71–5.45× TTFT reduction on Qwen3.5-122B with a 1.4 GB draft model.