Add nine Core ML packages with persistent KV state, context 2048, grouped LUT6 weights, the reproducible toolkit and serialized/CPU evidence. All 205 compressed matrices match the numerical reference; 84 Python tests and three fixed next-token probes passed. Apple SDK compilation, Core ML execution, device parity and ANE placement remain unverified. Preserve existing monolithic model packages and their separate runtime instructions.
Add the stateful runtime contract, pinned tokenizer and config, Python and Swift generation examples, Llama 3.2 license files, exact artifact inventories, and independent compiler validation.
Add export report for Dolphin3.0-Llama3.2-3B-stateful-int4.mlpackage
