Function prefill_bgmv_forced 

pub fn prefill_bgmv_forced() -> bool

ATLAS_LORA_PREFILL_BGMV=1 — force prefill LoRA through the per-row BGMV instead of the tensor-core GEMM.

Default OFF because the GEMM is ~4.8x faster on a 2K prompt (841 vs 176 tok/s measured on qwen3.8-27B) and the prefill call site is uniform-slot by construction. The BGMV is the only form that can honour per-row slots (including base rows), so this exists for the day a prefill batches rows from different sequences — and as the bisect handle if the GEMM path is ever suspected of a numerics difference, since the two are NOT bit-identical (GEMV-per-row vs one GEMM).