Use a writable runtime HF cache; drop preload_from_hub

Default reviewer becomes Nemotron 3 Nano 4B, startup-loaded

Bring Nemotron 3 Nano back as the default reviewer

Stop the 1B reviewer echoing its briefing

Render markdown tables and stop truncating reviews

Keep warned checks out of the praise section

Fix dropped review stream and runaway Mellum generation

Rewrite app on gr.Server with a friendly reviewer voice

Disable hybrid-reasoning thinking in the reviewer

Fix generate call for dict-returning apply_chat_template

Default the reviewer to Mellum 2, drop Nemotron

Add ZeroGPU reviewer-model inference