A third `InferencePort` beside the connector and the SDK sidecar: one
loopback-only OpenAI-compatible `POST /v1/chat/completions` call per
prepared request, tool calls returned inert for Rust's own evaluator to
drive the next round, exactly as the connector lane already does. No
credential is read or sent; `open_inference` refuses a non-loopback
`LocalBinding` and a model prefix two transports both claim.
`RoutedInferencePort` gains the local slot and routes by longest claimed
prefix, refusing rather than falling back to the connector when a claimed
prefix has no configured port behind it (the same rule the SDK prefix
already had, now made explicit for both). `ControllerModels::each` is
`pub` and `open_inference` takes `models: &[&str]` instead of the whole
struct, so a caller can extend the list with a model the struct does not
carry.
Dungeon's runtime reads `GHOSTLIGHT_LOCAL_ENDPOINT` and
`GHOSTLIGHT_LOCAL_MODEL_PREFIX` and wires the binding through; no other
Dungeon behavior changes in this cut.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
182f8034f7