fix(docker): --prefer-binary + pinned llama-cpp-python==0.3.7 for prebuilt wheel

build: use prebuilt CPU wheel index for llama-cpp-python (avoid source OOM)

fix: bump llama-cpp-python>=0.3.7 + Phi-3.5/Qwen2.5-3B fallback chain

init: Gemma-4-E4B (Q4_K_XL) CPU agent for Trading Floor T12

initial commit