TGI router InferError::Overloaded to HTTP 429 mapping and the max_concurrent_requests semaphore

InferError::Overloaded(#[from] TryAcquireError) carries the display string 'Model is overloaded'. The Infer struct holds limit_concurrent_requests: Arc<Semaphore>, so Overloaded is a tokio Semaphore try_acquire failure rather than a queue-full condition. router/src/server.rs implements From<InferError> for (StatusCode, Json<ErrorResponse>) mapping Overloaded => StatusCode::TOO_MANY_REQUESTS (429), ValidationError => 422 UNPROCESSABLE_ENTITY, GenerationError => 424 FAILED_DEPENDENCY, and IncompleteGeneration / IncompleteGenerationStream / StreamSerializationError => 500. The OpenAPI annotation on /generate documents 429 with the example body {"error": "Model is overloaded", "error_type": "overloaded"}. /health returns 503 {"error":"unhealthy","error_type":"healthcheck"} when the backend health flag is false. The health trait has a start_health() default of false with the comment 'Typically false, or true if the backend includes a warmup phase'.