Update source scripts and card for a local embedding-only int8 trial. FP16 remains the published artifact; no compressed weights are uploaded. Three selected native/Core ML fixtures passed the 0.02 probability gate with All compute units, but latency did not improve.
Fix unlabelled Core ML serving and Hub snapshot loading
Use upstream Kev typed-request rendering; retain the published FP16 weights. Stage symlinked Hugging Face snapshot packages for Core ML compilation. Verified one unlabelled Choice request against the pinned published package.
