SpecPrefill speeds long-context prefill on Apple Silicon

This looks like prompt compression aimed squarely at the prefill bottleneck, and the Apple Silicon angle matters because unified memory can hide more of the draft-model overhead. The biggest gains should come on long contexts, where reducing prefill work compounds with attention’s quadratic cost, and the reported pattern fits a tiny draft model helping most when the target model is much larger. The 20% keep-rate setting reads like the pragmatic middle ground here: aggressive enough to save compute without making structured outputs brittle, which makes this especially interesting for local inference stacks that care about TTFT.