Validator overhaul fixes Goodhart trap, UID 185 crowns

Distil v30.6 released to fix anti-correlated composite scoring that predicted worse real-world performance. Held-out public benchmarks (gsm8k, humaneval, bbh, ifeval, mmlu_pro) now weighted 50% of composite score; skill groups reweighted and aggregation switched to bottom-half mean to prevent saturation gaming. UID 185 (const0312/zeus_03) dethroned UID 213 with composite final of 0.537. Eval GPU utilization bumped to 0.92 after chat migration.

• •Validator now anchors 50% weight on held-out evalscope canaries vs procedural probes
• •Reasoning skill group weight cut from 0.24 to 0.06; other groups rebalanced by correlation
• •Most DQs from one_eval_per_registration policy; reuse requires fresh hotkey per eval
• •UID 33 disqualified for weights/config fraud: ~16B model falsely claimed as 1.13B