When decoding mechanics plateau, the last points in GSM-style tasks came from semantic prompt guidance

In a constrained-decoding synthesis loop on GSM-Symbolic (Qwen2.5-7B, 49 train problems), mechanics changes (forced int() wrapper, nested-wrapper blocking, span closers) took accuracy from 28.6% to 44.9%, then three more mechanics attempts plateaued or collapsed. The attempt that crossed the bar (46.9%) kept the decoding code unchanged and only edited the task guidance: treat a scaled grouped sum as one quantity, include every multiplier, for percentages multiply first and divide last (int((a/b)*100) can floor 0.29*100=28.999 to 28), and substitute full earlier expressions in the final one. Lesson: once 25 of 27 remaining errors are syntax-valid but semantically wrong, the grammar cannot help; change the guidance, not the mask.