//! Inference engines — pluggable decode strategies. //! //! The primary engine is [`MarkovResidualEngine`] which replaces traditional //! KV caching with stored pre-layer residuals. K/V are recomputed from //! stored residuals at decode time (validated KL=0.0 vs full-KV baseline). pub mod markov_residual; pub use markov_residual::{kv_memory_bytes_for_seq, recompute_kv, rs_decode_step, rs_prefill}; pub use markov_residual::{MarkovResidualEngine, RsPrefillResult, RsStore};