2026-08-19 - Operand layouts, and the limit of the per-family model

Done. Layouts for every family the model can actually describe, plus the two decoder capabilities they needed: an operand may name a later dword, and a field may store a scaled register index.

Encoding::operands became Option<Vec<_>>, so "no layout established" and "checked, takes no register operands" are different states rather than the same empty list.

Verified. Gate green. Seven families of seventeen carry a layout, and all 99 operands they produce match the reference disassembler across the fixture corpus.

Surprises.

Not done. Ten families still have no operand layout, deliberately. Per-opcode operand data - derivable from the reference the way mnemonics already are - is the next mechanism and has not been built. Nothing translates.