Constrain v7 to prevalidated grounded wording choices

Add deterministic untrusted-choice guard and fresh v6 gate

Replace brittle multi-field output with SAMS v5 wording selector

Rebuild SAMS v4 bounded selector and strict 40-case gate

Resume v3 from preserved adapter and apply 90 percent adapter diagnostic gate

Deploy SAMS v3 targeted quantization-robust retraining

Replace rejected v1 with bounded SAMS v2 training pipeline

Add trainer/evaluate.py