fix(v1): explicit knowledge-cutoff disclaimer in system prompt

fix(zerogpu): switch to gr.Blocks + add /run/synth_batch endpoint

fix: minimal gr.Interface (ChatInterface fails on gradio 4.44 force-pin)

fix: 14B + bnb int4 (AWQ build failed; bnb proven, no compile)

feat: Qwen2.5-Coder-32B AWQ (biggest fits A10G 24GB)

fix: revert to 7B INT4 + apply Surrogate-1 v1 LoRA (REAL Surrogate-1)

feat: upgrade to Qwen2.5-Coder-14B + INT4 quant (4x more capable)

fix: switch to Qwen2.5-Coder-3B (faster cold load, fits A10G in <60s)

fix: lazy load (avoid Space init OOM with 7B+LoRA)

fix: use gr.ChatInterface (simpler sig, avoids _json_schema bug)

initial: Qwen2.5-Coder-7B + Surrogate-1 v1 LoRA on ZeroGPU A10G