updates to speed up computation of ranked list

Fix JSON coercion: every LLM call returned empty (ported prefill broke)

Downscale screenshots before the vision encoder to cut CPU latency

whatfirst-small: offline Gradio + llama.cpp task prioritizer