Make the provider bridges speak the APIs the providers actually run

[My-CMDhub]

Four defects in the Deepgram and Cartesia bridges, none of which raised an
exception. Each was an error message the provider sent us, which reached an
if/elif chain with no else, was parsed, and dropped.

• Deepgram Flux Configure used the Listen-v1 schema:
  UNPARSABLE_CLIENT_MESSAGE: unknown field 'processors', expected 'thresholds'
  STT accepted audio and returned nothing — 25s of silence on a live call
  while every health check reported green.
• Flux turn state was read from type instead of event. A TurnInfo
  message carries its state in event, so barge-in and LLM pre-warm never
  fired: dead code that looked live.
• Flux needs the /v2/listen endpoint. The bridge routed everything to v1.
• Cartesia encoding: "mulaw" -> 400 unsupported encoding for raw: mulaw.
  The correct value is pcm_mulaw. Behind it, a second 400: sonic-english
  had been sunsetted. TTS never synthesised a byte; only a cached greeting
  was ever audible.

finally: is_connected = False in receive_audio_events() is removed. The
turn loop breaks out of that async generator every turn, and break runs
finally, so the bridge declared its own healthy socket dead after turn one.
Connection lifetime belongs to close(), not to whoever stopped iterating.

Cartesia error events are now logged and reported instead of ignored, which
is how three of the four above stayed invisible for a week.

conftest.py blanks SENTRY_DSN so the test suite stops reporting its own
deliberate exceptions to the live Sentry project.

---

Verified on a live PSTN call. Three tests in this repo were green while certifying two of these bugs, because they mocked the provider boundary.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
ovela-ai	Ready	Preview	Aug 19, 2026 10:48am