Fix WebSocket reconnection races and Inworld TTS context restore

[hwuiwon]

This PR improves reliability for websocket-based services during transient disconnects.

• WebsocketService: serialize reconnect attempts with an asyncio.Lock and short-circuit if the connection is already
  healthy (reduces “reconnect attempt aborted / unable to reconnect” races).
• InworldTTSService: if a reconnect happens while a context is active, re-send the create message for the existing
  contextId, and only close_context / clear contextId on intentional disconnects (helps prevent “connected but no
  audio / context missing” after reconnects).

[hwuiwon]

@markbackman @cshape would appreciate a review!

[codecov]

Codecov Report

❌ Patch coverage is 0% with 7 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/pipecat/services/inworld/tts.py	0.00%	7 Missing ⚠️

Files with missing lines	Coverage Δ
src/pipecat/services/inworld/tts.py	0.00% <0.00%> (ø)

... and 24 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[markbackman]

@hwuiwon we haven't seen this case. We've also just made changes to the WebsocketService to make it more robust to reconnects. I'm hesitant to make further changes for this release unless we have a documented issue and repro case that occurs on main. If you have that, then we're happy to take a look. I just want to be cautious because changes here affect so many services.

[hwuiwon]

@markbackman I’ve been seeing multiple cases like #3538 (comment) while running hundreds of calls through the Inworld WebSocket service. It doesn’t happen consistently, but it reliably shows up under higher call volume. I would suggest making 1000+ calls that uses inworld websocket service to reproduce this. Not so sure how testing is set up for testing this service in particular from pipecat side.

It is critical because this prevents the speech from getting generated and users won't get any feedbacks.

[markbackman]

@markbackman I’ve been seeing multiple cases like #3538 (comment) while running hundreds of calls through the Inworld WebSocket service. It doesn’t happen consistently, but it reliably shows up under higher call volume. I would suggest making 1000+ calls that uses inworld websocket service to reproduce this. Not so sure how testing is set up for testing this service in particular from pipecat side.

It is critical because this prevents the speech from getting generated and users won't get any feedbacks.

Have you been running from main? Also, is the issue with the Inworld configuration or others? We have customers that do 10K+ calls daily on services like Cartesia and ElevenLabs and they haven't reported issues.

[hwuiwon]

I don't believe it's the configuration issue with Inworld because it works normally with the http service. I can rollback the changes in the src/pipecat/services/websocket_service.py but I think we need fix in the src/pipecat/services/inworld/tts.py

[ianbbqzy]

if a reconnect happens while a context is active, re-send the create message for the existing
contextId

I think this is worth a try. I noticed that Cartesia doesn't have a separate CreateContext message because they expect the configs in every message. On the other hand, 11lab might also not necessitate a createContext message because a lot of configs were already sent as part of query params in the URL.

In the case of Inworld, on Websocket reconnection, it's important to re create a context first

[hwuiwon]

@markbackman rolled back the changes in websocket_service.py. Fix only touches Inworld Websocket TTS now