test(evals): expand 17 pairwise behavioral sets from 3 to 25 problems each

Per-skill statistical power for the behavioral A/B sweep. In-domain,
software/agent-biased, diverse scenarios; existing 3 preserved per skill.
Skills: steel-manning, feedback-loops, archetypes, leverage-points, triz,
thought-experiment, lindy-effect, jobs-to-be-done, effectuation, reversibility,
regret-minimization, second-order, opportunity-cost, theory-of-constraints,
pre-mortem, via-negativa, first-principles.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>