AWS Transcribe errors are silently swallowed in voice module

[tigy32]

Summary

When AWS Transcribe streaming fails to start or encounters an error during streaming, the error is only logged via tracing::error! and then the tokio task exits silently. Callers have no way to receive or handle these errors - they only observe symptoms like channel closures.

Location

tycode-core/src/voice/stt/aws_transcribe.rs

Problem

In the start() method, when AWS Transcribe fails:

let response = client
    .start_stream_transcription()
    // ...
    .send()
    .await;

let output = match response {
    Ok(output) => output,
    Err(e) => {
        tracing::error!("Failed to start AWS Transcribe stream: {e:?}");
        return;  // <-- Task exits silently, dropping audio_tx
    }
};

What happens:

1. AWS Transcribe fails (auth, region, quota, invalid audio format, etc.)
2. The tokio task logs via tracing::error! and returns
3. This drops audio_tx, closing the audio channel
4. The caller's audio_sink.send() starts returning errors
5. Caller sees "send failed" but has NO visibility into the actual AWS error

The same pattern occurs for streaming errors:

while let Ok(Some(event)) = transcript_stream.recv().await {
    // ...
}
// When loop exits (stream closed or error), task exits silently
// No indication of whether this was normal completion vs error

Impact

When integrating tycode-core's voice capabilities:

• AWS configuration errors are invisible to the application
• Authentication failures appear as generic "channel closed" errors
• Debugging requires adding logging inside tycode-core source
• Applications cannot display meaningful error messages to users

Suggested Fix

1. Propagate errors through the TranscriptionStream:

pub struct TranscriptionStream {
    rx: mpsc::Receiver<Result<TranscriptionChunk, TranscriptionError>>,
}

2. Or add an error channel alongside the transcription channel:

async fn start(&self) -> Result<(AudioSink, TranscriptionStream, ErrorReceiver)>

3. At minimum, surface the error when the stream ends:

pub enum TranscriptionEvent {
    Chunk(TranscriptionChunk),
    StreamEnded { error: Option<TranscriptionError> },
}

Reproduction

1. Configure AWS Transcribe with invalid credentials or wrong region
2. Start voice session
3. Observe: Audio sink closes immediately with no explanation
4. Expected: Receive error like "AWS authentication failed" or "Invalid region"

---

This issue also applies to the stream completion path - there's no way to distinguish between "transcription completed normally" vs "stream failed with error".