"""Ground-truth generators for each sub-bench.

Generators are tool-neutral. They read the corpus and produce a JSON
dataset that the bench runners consume — the scoring is then deterministic
and does not depend on any adapter's internal analysis."""