SN97 v30.2 Overhaul Live; Teacher Slowdown Fixed

Distil deployed a major scoring update addressing Goodhart gaming on ranking axes. The new composite.final ranking key (70% worst-3-mean + 30% weighted) replaces single-axis minimization, adds a super_teacher axis rewarding models that beat the baseline, and collapses 13 benches into 4 skill groups. Separately, a vLLM misconfiguration causing 100+ minute eval rounds was fixed; validator now targets 45-minute cycles.

• •Ranking: composite.final replaces worst() floor; super_teacher axis (10% weight) explicitly rewards exceeding Qwen3.6-35B on 16 verifiable benches.
• •Goodhart fix: negative correlation between validator math/reasoning scores and held-out GSM8K (-0.665 to -0.746 r) prompted procedural template rebalance.
• •vLLM bugs: DISTIL_VLLM_CONCURRENCY env var ignored, gpu-memory-utilization too low, no max-num-seqs cap caused OOM fallback to slow HuggingFace.
• •Next: Kimi K2.6 A/B teacher experiment, multi-GPU pod migration (8×H100), batched student forward passes for 2-3× speedup.