antirez/h3.c

This issue exposes a critical precision degradation in the `h3_gqa_causal_bf16` kernel, where scaled BF16 queries are unnecessarily rounded back to BF16 before F32 QK contraction. This intermediate step introduces significant error, as evidenced by drastically higher absolute error, RMSE, and output mismatch rates compared to a correctly implemented F32-scaled path. The problem directly impacts the accuracy of attention scores for the Qwen3-VL prompt encoder, compromising model fidelity without any corresponding benefit in storage or bandwidth. This represents a fundamental flaw in numerical stability for a key component, demanding immediate correction to ensure the reliability and competitive performance of the H3 inference engine on Apple Silicon.