[katstankiewicz]

• Fix bug where a phantom turn-start (turn start event but no audio/transcript) would cause a query cancellation. The change was to use turn start strategy transcription so we only start a turn if we have a transcript
  • Note the config behaviour change. For ink-2 and other self-endpointing STT, the default turn_start_strategy is now transcription but was forced to external before.
• Add stt_latency for ink-2. Since ink-2 does its own vad handling we didn't have the events to measure this. The stt_latency includes some silence time due to cartesia's watchdog waiting ~500ms before emitting user stopped speaking. This makes the stt time seem slow, but is also silence that the user hears so I kept that in. The logs have a No audio received for 509 ms. Sending silence to Cartesia to prevent a dangling turn which we could use to remove that extra silence to measure stt directly.