SN97 eval pipeline hardened, miner configs enforced

Major eval-pipeline cleanup (524 tests passing) landed with strict new miner requirements: DeepseekV3 architecture only, no `auto_map` blocks, bf16/fp16 weights only, no custom Python files beyond teacher copies, and tokenizer/vocab must match teacher exactly. Validators now disqualify configs with undisclosed quantization and models that derail past ~500 tokens. Chat and dashboard bugs fixed; king changed twice in 24 hours as post-Kimi cohort stabilizes.

• •DeepseekV3 architecture + exact teacher tokenizer vocab (163,840) required
• •Delete `auto_map` from config.json; no quantization markers or custom .py files
• •UID 223 now king with composite final 0.298 after UID 151 brief reign
• •Chat latency fixed; KL/RKL now render; cold-start dethronement deadlock resolved