Function lora_eager_env 

pub fn lora_eager_env() -> bool

Permanent LoRA debugging hatch: ATLAS_LORA_EAGER=1 (or true) forces eager decode (no CUDA-graph capture) when an adapter is active, so graph-vs-eager output parity can be compared in the field. Read ONCE — the decode graph gate runs per token. Reads the process-wide ModelLevers, which resolve from the environment exactly once. This used to say the value was “resolved at the point of use rather than cached … a getenv is free” — but from_env reads ~30 variables per call, and a sibling caller was invoking it once per layer per prefill.