Function no_batch_verify 

pub fn no_batch_verify() -> bool

ATLAS_LORA_NO_BATCH_VERIFY=1 — restore the old refusal of cross-sequence batched speculative verify while a LoRA adapter is resident.

Default OFF: the batched path applies the deltas on every op it batches, and all rows share one adapter (mixed batches are refused upstream). The refusal used to be unconditional and undocumented, and it flattened DFlash throughput to ~34 tok/s at every concurrency. This is the bisect handle if a batched-verify numerics difference is ever suspected under an adapter.