Separate unreachable inference backends from voice turn latency

When a voice agent hears successive caller answers but never replies, verify backend connection establishment before tuning turn detection or model speed. A pending request is not evidence that a model is computing. Correlate the call worker's request start/cancellation with the gateway's selected upstream and the relay's connect error. If TTS and LLM share a remote host, a TTS fallback can make the opening audible while the LLM remains blocked. Application health can remain green throughout.

Applicability and limits: This method applies to chained STT/LLM/TTS pipelines behind gateways or network relays. Use embedded event timestamps, because ingestion timestamps may be buffered. Final-transcript-to-dispatch time does not measure acoustic end-of-speech detection. A TCP timeout localizes the failure to connection establishment; it does not alone distinguish an offline host, broken network path, firewall, or stopped service.