fix: restore llama.cpp with source build, use fine-tuned GGUF model

refactor: remove llama-cpp entirely, use deterministic mock-only mode

Phase 2/3: Gradio Server backend, CRT frontend, engine, agents, mentor, tests, CI/CD