fix#2: UK formal LibriVox rubric, sampling, duration/LUFS, literary reranking bonus

god-v1 fix#1: port v3 best-of-N inference + Hippius-calibrated loudness/silent-frac/brightness scoring (N=34 paired reference clips)

Upgrade weights: epoch-0 -> epoch-5 (KGSS header preserved, format=pt)

Fix: restore format=pt in safetensors __metadata__ (was missing, blocked Qwen3-TTS loader)

Add vocence_config.yaml (miner /health metadata)

Add vocab.json (sourced from tts-epoch-5; required by qwen-tts tokenizer)

Initial upload of God-tts-v1 (Qwen3-TTS 1.7B snapshot with unique safetensors header)

initial commit