Card: add honest builder-blind sealed-test result (88.0% acc / 17.7% invented; fails the <5% gate; dev 1.2% was optimistic)

Card: correct training-data sourcing (mixed vendor + community spec tables), dev-vs-sealed honesty

Spec-parser LoRA r5: 91.6% field accuracy, 1.2% invented-field rate (dev set)

initial commit