Telnyx STT provider: design question, barge-in depends on partials

[a692570]

Context

#74 added Telnyx as a TTS provider. Telnyx also runs a streaming STT WebSocket (wss://api.telnyx.com/v2/speech-to-text/transcription) that already fronts Deepgram, Google, and Azure engines, plus a Telnyx in-house engine. I'd like to contribute a Telnyx STT provider as a follow-up, but one design question needs settling first because it touches barge-in.

What the Telnyx engine does (verified against the live API)

• Client sends raw linear16 PCM as binary frames; server sends JSON frames with transcript, is_final, confidence.
• The Telnyx engine emits exactly ONE final frame after audio stops, then closes the socket (1000). No interims, no word timestamps, no speech_started / utterance_end. Confidence is null.
• interim_results=true suppresses ALL output from the Telnyx engine (adding the param returns zero frames). It cannot be requested.
• transcription_engine is case-sensitive: Telnyx works, telnyx closes the socket.
• The Deepgram engine through the same endpoint does emit interims (is_final: false frames with confidence), so the finals-only behavior is specific to the Telnyx engine.

The problem: barge-in is a hard dependency on partials

In internal/pipeline/inbound.go (line numbers at current main):

• hasPartialText is set only by non-final results (~101-108, gate at ~76).
• The barge-in trigger requires bargeInVAD.IsSpeaking() && agentSpeaking && hasPartialText (~196).
• Backchannel classification ("mm-hm" filtering) reads latestPartial (~168-193), also fed only by partials.
• Live transcript captions stream to the client from the partial path (~110-113).

With a finals-only provider, hasPartialText is never set: barge-in never fires (dead, not degraded), backchannel filtering never classifies, and the UI gets no live captions.

Options

1. Provider + docs note (the finals-only OpenAI whisper precedent, docs/providers.md:15): ship the Telnyx-engine provider and document that barge-in and live captions do not work with it.
2. Provider that exposes Telnyx's engine param: transcription_engine = "Telnyx" (in-house, finals-only, no barge-in) or "Deepgram" (proxied through the same endpoint, emits interims, barge-in works). Default engine becomes a product choice rather than a hard limitation.
3. Pipeline capability flag: expose whether the active STT produces partials and degrade gracefully. More pipeline surface, so I don't want to assume that's welcome without asking.

Which way would you like it? Happy to implement any of the three.