Function ffn_batch16_enabled 

pub fn ffn_batch16_enabled() -> bool

ATLAS_FFN_BATCH16 opt-in: the value 1 — and only 1 — arms the 5..=32-row tier. Anything else, absence included, leaves those widths on the pre-#927 arms.

VALUE rather than the house PRESENCE convention (ffn_w8a16_only next door) because the polarity is the other way round: for a switch that ARMS an arm, it is ATLAS_FFN_BATCH16=0 meaning “on” that would be the trap. Same shape as moe_grouped_decode_forced in layers/mod.rs, the other lever in this crate that arms rather than disarms.

OnceLock-cached and read ONCE PER LAYER, into DenseFfnLayer’s batch16_enabled: the route must be CONSTANT across CUDA-graph replays, and std::env::var walks the environment block on every call.