Add the family-naming model as stage 4; remove the gpu/cpu runtime choice

drop llama-cpp-python (unbuildable on Spaces); explain MiMo-over-Gemma choice

runtime picker (gpu/cpu); shorter GPU grant to stretch the free ZeroGPU quota

ZeroGPU: duration under the 300s cap, weights prefetched outside the grant

dual runtime: @spaces.GPU 4-bit NF4 path (Part B config) + llama.cpp CPU fallback

MiMo-7B PDF injection detector: CPU/GGUF runtime, verbatim Part A+B logic