Adds `selfcheck` mode: MUSETALK_REF_SAVE=1 dumps per-stage GPU outputs (encode/unet/decode) as npy; re-running compares same-device same-dtype (PSNR/cosine). Isolates kernel-change numeric drift from the f16-vs-f32 cross-device gap that `verify` conflates. Baseline self-check = 120 dB (exact). Co-authored-by: Hanzo Dev <[email protected]>