Broken streaming tts retries

[jayeshp19]

Bug Description

it looks like streaming TTS retries are broken. The retry loop reuses the same SynthesizeStream instance, but the _input_ch channel and plugin state are one-shot resources that get consumed during the first attempt.

As a result, retrying the same request sends no text or produces no segments, leading to consistent failures like APIError: no audio frames were pushed or segments mismatch.

related issues: #4135

Expected Behavior

we should be able to replay events during retry

Reproduction Steps

1. 
2.
3.
...
- Sample code snippet, or a GitHub Gist link -

Operating System

Macos

Models Used

No response

Package Versions

livekit-agents=1.3.10

Session/Room/Call IDs

No response

Proposed Solution

Additional Context

No response

Screenshots and Recordings

No response

[longcw]

with the current version, I think the retry only works when the connection establishing failed. it's not changed since the very beginning so it's hard to say it's "broken".

I am not sure if this behavior is by design, but there is one issue I can see if we allow input replaying is that it may generate repeated content after failed in the middle of generation. cc @theomonnom wdyt?

btw, I think it's a different issue if the TTS only fails after the connection created and input sent (maybe 11labs with the multi-context TTS?), we should fix that plugin specifically.

[jayeshp19]

I am not sure if this behavior is by design, but there is one issue I can see if we allow input replaying is that it may generate repeated content after failed in the middle of generation.

added a guard to prevent duplicate speech if a failure happens mid-generation - 6ab2b6c