When Contact Becomes Support: Auditing and Repairing Contact-Topology Failures in Programmatic Dexterous Grasp Search

Dexterous hands can touch an object, lift it briefly, and still fail the physical task becausecontact has not become load-bearing support. We study this contact-to-support gap inMuJoCo using an open-source Menagerie/Allegro hand, strict no-assist 15-second lift-andholdevaluation, and a coding-agent-maintained programmatic Heuristic Learning (HL)workflow. Recent agentic-robotics work shows that coding agents can improve robot-controlcode, but many such settings emphasize lower-DOF precision manipulation or task-levelcontroller repair. Our focus is narrower: high-DOF multi-finger grasping where the decisivefailure is missing contact topology. The reported systems are best-of-K simulation-guidedsearches, not one-shot perception-to-action policies; at K = 6, success primarily measurescandidate-library coverage under rollout-based selection, while K = 1 is the closest one-shotproxy. On a matched six-family suite, V4.0 achieved 42/120 strict successes at K = 6, whileV4.1 achieved 49/120 on the same cases with 7 V4.1-only successes, 0 V4.0-only regressions,and an exact paired sign/McNemar p = 0.0156. Removing V4.1 extra repair anchors returnedthe result to 42/120, localizing the gain to sphere repair; capsule, bar, and flat-box transferremained weak. Candidate-budget analysis shows V4.1 rising from 20/120 at K = 1 to 45/120at K = 3 and 49/120 at K = 6. Separate cylinder-only fixed-topology tests reached 79/100strict successes at K = 6 and 89/100 at K = 10. These results do not establish generaldexterous grasping, sim-to-real transfer, or an experimentally isolated advantage of codingagents. They provide an auditable case study showing how HL can expose contact/supportfalse positives, bound candidate-search claims, and convert one diagnosed topology failureinto a measurable but narrow repair.