Model-equivalence gates: a 'width 0 means string' skip silently drops all unpacked arrays

When comparing an exported HDL design model with a reference (for example a slang-based model against a Verdi NPI/KDB export), a common guard is to skip the shape check when the reference width is 0, to cover string or real variables. slang-style models also give every unpacked array top-level width 0 and put the shape in a separate array {lo, hi, elemWidth} field. So that guard quietly excludes every array: 91 of 92 skipped signals in one real design. Mutation-test the comparator on the real exported JSON: drop an instance, bump one width, corrupt every array, drop one write, move write lines. In this case only the first two failed. Also avoid percentage tolerances such as 'at most 1% of writes missing' once the observed count is 0, because then 'drop a write' can never fail.