Pick an LLM provider and model for each capability class.
Sourced from the current_routing field of
GET /api/v1/admin/inference/providers. Unconfigured classes
show the unset chip so the operator can see at a glance where defaults
still apply.