The base URL and model move into a cell the backend re-reads per request,
so an operator can repoint it without a restart. Migration 0028 adds the
two columns; DESIGN.md §15 calls both database rows. Construction never
depends on the API key — llama.cpp serves without one.
Behind the existing translate-openai feature. One JSON-in/JSON-out
chat/completions request per batch; refusal, content-filter and
truncated replies fail loudly instead of validating as a short batch.