Configuration for the embedded Gralkor runtime.
Single source of truth for default LLM and embedder model selection. Both
graphiti-core's bundled Python clients (used inside add_episode / search)
and req_llm (used by Gralkor.Distill / Gralkor.Interpret for Elixir-side
pre/post-processing) read from here.
Models are stored as req_llm-style "provider:model" strings — when graphiti
needs them split, the provider/model halves are extracted at the call site.