Load the diffusion pipeline once per submission instead of per metric

Work around an eval-unlearn packaging bug blocking free_run registration

Add per-metric dataset size controls and clarify heavy-ASR scope

Add interactive model submission with online ZeroGPU evaluation