Let coverage dominate Household Orders sweep scoring

[axelpey]

Summary

• replace sweep distance-decay utility with the ablation-proven linear coverage utility
• raise uncapped novelty reward so the search spreads through unexplored space
• retain the conservative first-view travel cost and all reliability guards

Why

The independent 40-run ablation improved successful median from 354.6s to 288.2s and successful max from 664.2s to 517.3s, with no observed behavior-valid reliability loss (37/40 control vs 35/37 after excluding three infrastructure failures). This PR applies only that component change.

Validation

• python3 -m py_compile workspace/innate_skills/find_next_person.py
• targeted ROS-image test: 5 passed, 87 deselected
• fresh-source confirmation: 40/40 Batch jobs succeeded and 40/40 behavioral trials passed
• confirmation latency: min 230.7s, median 310.3s, P95 559.4s, max 590.5s
• confirmation infrastructure estimate: $2.60

The confirmation ran commit 95a493880b797e9281afe2af92d7b674c1cc44a0 on one fresh VM per trial, zero retries, with first-person evidence recording.