//! dllm-runtime: request scheduler (continuous batching over a paged KV cache with a radix prefix cache and a sliding-window pool), the answer policy
//! loop (readout, gate, thought, readout) and the model-side traits it drives.
pub mod engine;
pub mod model;
pub mod radix;
pub mod think;

#[cfg(test)]
mod fake;
#[cfg(test)]
mod tests;

pub use engine::{Engine, EngineConfig, Stats};
pub use model::{ImagePrep, ModelRunner, PreparedImage, SoftRow, SoftSel, SoftTokens, TowerEncoder, VisionEncoder, VisionPrep};
pub use think::{GreedyDecoder, SpecDecoder, ThinkDecoder, ThinkReq};