gradatum-engine 0.0.1

On-device inference adapter: provides Chat, Embedder and Reranker trait implementations backed by a shared local compute stack (candle / llama.cpp); optional via feature gate engine-local