The v2 soak never generates a contraction the contraction rule applies to #127

🧭 Claude, the einfold agent

#126 adds a contraction rule to ddx-ad's transpose. It applies to a saved SUM(e_X · e_Y) over one inner equi-join of two inputs, where each factor reads one side, every grouping key is a column of one side, and the other side holds a grouping key. Such a sum's gradient is emitted as the cotangent contracted with the other input, without rebuilding the join.

mutation_test.py gains two mutants for that rule, contraction-wrong-cotangent and contraction-guard-off. Both survived 240 s of soak_v2_query_ad (about 505 cases each), while sum-halved and null-skip-leaks were killed in 0 s and 54 s. So the soak's generator rarely, if ever, writes a loss the rule applies to, and a clean soak says little about the rule. contraction_transpose.rs covers the rule directly and kills both mutants, but the weekly mutation report will list them as survivors until the generator reaches this shape.

Goal. The generator sometimes writes a two-input contraction in the rule's shape: SUM(f(a) * g(w)), joined on equalities, grouped by a column of each side. Include NULL join keys, with both = and IS NOT DISTINCT FROM, so the soak kills both mutants within the mutation workflow's budget.