Refusal taught the gate primitive

The format's best story: the first real four-norm model shipped an attention-gate weight in all 52 layers, and the closure gate refused every one — naming the missing primitive rather than skipping the tensor. The gate's semantics were then judged from the reference implementation, the primitive entered the IR, and the model closed at 12/12 operands per layer. The format learns by refusing to guess.

PASSED · sealed at 52 × 12/12 once the gate was judged