Quasar switches to tail-bucket KL evaluation metric

Quasar migrated from top-128 KL to tail-bucket KL scoring to capture probability mass dumped into model tails, addressing optimization gaming. New temporary gates are KL 6.3 and quality floor 0.06 (old 2.1/0.20 not comparable). Full-vocab KL remains under calibration. GitHub/dashboard updates are rolling out; large eval queue ongoing with 20 models in current round.

• •Switched from top-128 KL to tail-bucket KL; accounts for tail probability mass
• •Temporary gates: KL threshold 6.3, quality floor 0.06 (live, not dashboard text)
• •100% burn/no miner weights when no model clears gates; protects subnet emissions
• •Full-vocab KL deferred due to compute cost (~100GB+ dense logits) and calibration needs