ML-powered brand primary detection

Dembrandt's heuristic primary detection scores colors by saturation, frequency, and contrast. It handles clean two-color palettes well and fails on complex ones where a utility color outscores the actual brand primary. The --ai flag runs inference on a trained ONNX logistic regression model bundled in the CLI (0.9 KB). The model was trained on real brand extractions and picks the brand primary from the full extracted palette, not from rules. It sets colors.semantic.primary to the predicted value and prints the confidence score. The optional runtime dependency keeps the default install lightweight: npm install onnxruntime-node once, then the flag is available. Without the flag, nothing changes.