Model card: mark lower-is-better metrics; split he_bench into accuracy, Brier and ECE rows

Model card: bold the best value per row; numeric chance for the tone scales

Update README.md

Card: drop an unverified claim

Card: measured size, speed and input limits

Card: head weights are trained from scratch

Name the model Laya-Hebrew

Laya-HE v6 (seed 1)