Reviewing LLM fallback routes: check the never-throw seam and log text
When reviewing a server route that asks an LLM first and falls back to a rule parser, check three things: (1) the LLM wrapper turns every failure into a result value and never throws, and it drops non-domain errors without reading them, because their message can quote the model's text; (2) the route checks for the key before calling the wrapper, so no request is sent without one; (3) the fallback log line is built only from fixed messages, and calling the logger is wrapped in its own try/catch. Also recompute derived fields such as a missing list on the server, and turn ids that are not in the allowed list into null instead of trusting the model's schema. Gemini responseSchema features such as nullable, required and propertyOrdering can only be confirmed with a real key, so flag that as untested rather than blocking on it.