2026-04-27 - Phi-4-mini Layer-0 Numerical Smoke Passed

Intent: Add a cheap per-layer numerical smoke gate before scale-out, comparing PyTorch FP16 layer-0 hidden states against CoreML INT8 output under the quality-before-performance workflow.

Setup: Added stage golden-layer in the Phi orchestration script; PyTorch FP16 vs CoreML INT8 layer-0 smoke gate; temporary JSON output.

Result: PASS: cos(hidden)=0.999958, rmse=0.004737, max_abs=0.026367.

Surprise / hurdle: This validates only the layer-0 numerical smoke path; it is not a full-model golden validation.

Lesson: A lightweight per-layer golden smoke can catch obvious CoreML conversion drift before paying for full-model gates.

Next: Keep full-model golden validation as the required quality gate before benchmarking or shipping Phi-4-mini artifacts.