The tokenizer GGUFs were generated with an older policy that quantized
RVQ codebooks and fc / fc2 / project_in / project_out linears. RVQ
nearest-neighbor lookup is sensitive to per-row quantization noise,
even BF16 mantissa drifts codes and degrades voice cloning. New policy
keeps these tensors at F32 in every variant.
Conv kernels with non-alignable rows now fall back to F16 instead of
F32 in Q* variants, restoring BF16 > Q8_0 > Q4_K_M precision order.
Snake activation alphas are stored at F32 (per-channel activation
parameters, not weights).
