fix(ai): add 26B as Gemini fallback before llama-cpp

feat(ai): Gemma 4 31B primary with retry + thinking model extraction

fix: f-string backslash SyntaxError on Python 3.11 — extract _fmt_entity helper

fix(dockerfile): use llama-cpp-python==0.3.20 (0.3.9 does not exist on PyPI)