Module provider 

Unified LLM provider abstraction (ADR-010).

One trait covers both local Candle inference and cloud frontier LLMs. The host picks a backend at session construction time — the rest of the SDK sees only LlmProvider. This is the seam that lets mobile apps swap local ↔ frontier without touching their UI code.

Design notes:

• All types are plain std (no async runtime dep in this crate).
• chat_stream uses a callback so each binding surface wraps it in its own async/stream primitive (FRB → Dart Stream, uniffi → async callback, wasm-bindgen → ReadableStream).
• CredentialRef is a runtime value from the host — never embedded.

Structs

ChatMessage

One turn in the conversation.

ChatRequest

A chat completion request. model is a routing hint:

ChatResponse

A completed chat response.

ChatToken

A single streamed token fragment.

CredentialRef

Runtime API credential. The host resolves this from platform keystore (Android Keystore / iOS Keychain) before calling start_session. The SDK never logs or persists the value.

Enums

ChatRole

Which role a message belongs to.

Traits

LlmProvider

The unified backend trait (ADR-010). Implemented by: