Async Traits

Every trait in this codebase uses #[async_trait]:

#[async_trait]
pub trait LLM: Sync + Send {
    async fn generate(&self, messages: &[Message]) -> Result<GenerateResult, LLMError>;
    // desugars to:
    // fn generate(...) -> Pin<Box<dyn Future<Output=Result<...>> + Send + 'async_trait>>
}

This means every generate() call heap-allocates a Box<dyn Future>. With 124 uses across 60 files, this is pervasive.

For static dispatch (generics), native async fn works without any macro:

// No #[async_trait] needed — works for static dispatch
pub trait LLM: Sync + Send {
    async fn generate(&self, messages: &[Message]) -> Result<GenerateResult, LLMError>;
    async fn stream(&self, messages: &[Message])
        -> Result<Pin<Box<dyn Stream<Item = Result<StreamData, LLMError>> + Send>>, LLMError>;
}

// Generic function — zero-cost, no boxing
async fn run_chain<L: LLM>(llm: &L, messages: &[Message]) -> Result<String, LLMError> {
    Ok(llm.generate(messages).await?.generation)
}

async fn in dyn Trait still requires boxing as of Rust 1.85. At dyn call sites, use a local adapter:

use async_trait::async_trait;

// Thin wrapper that adds boxing only at the dyn boundary
#[async_trait]
trait DynLLM: Send + Sync {
    async fn generate(&self, messages: &[Message]) -> Result<GenerateResult, LLMError>;
}

// Blanket impl bridges native LLM → DynLLM
impl<L: LLM + Send + Sync> DynLLM for L {
    async fn generate(&self, messages: &[Message]) -> Result<GenerateResult, LLMError> {
        LLM::generate(self, messages).await
    }
}

Or keep Arc<dyn LLM> with #[async_trait] on just the trait definition — boxing happens once at the trait boundary, not in every impl.

• DO add #[async_trait] to new trait impls — consistent with existing code
• DO NOT add new async_trait uses to non-trait code (closures, free functions)
• When implementing JOB-252, start with LLM trait (highest call frequency), measure, then proceed to Tool, Chain, Embedder