feat: add ConversationAudioObserver for mic and agent audio

[renal128]

What

• Introduce ConversationAudioObserver for observing decoded PCM from agent output or the local microphone.
• Add durable observer APIs to ConversationClient:
  • addAgentAudioObserver / removeAgentAudioObserver
  • addMicAudioObserver / removeMicAudioObserver
• Keep LiveKit AudioRenderer / AudioTrack fully internal behind AudioObserverRegistry.
• Reattach client observers for each new session and on mid-session track changes.

Why

Apps need raw audio for visualization, recording, or custom analysis without:

• LiveKit types in the public SDK surface
• main-actor audio processing
• SwiftUI churn from high-rate @Published updates

Usage

final class Recorder: ConversationAudioObserver, @unchecked Sendable {
    func didReceive(_ buffer: AVAudioPCMBuffer) {
        // Copy samples synchronously, then process them away from this callback.
    }
}

let client = ConversationClient()
let recorder = Recorder()

client.addMicAudioObserver(recorder)
client.addAgentAudioObserver(recorder)

Behavior

• didReceive(_:) receives a borrowed, read-only buffer on the audio callback path.
• Consumers own any buffering, copying, analysis, or visualization work outside the callback.
• Cross-session durability is owned by ConversationClient; Conversation itself is single-use.
• Use separate observers when you need independent state per audio stream.

Testing

• swift test --filter 'AudioObserverRegistryTests|ConversationAudioObserverTests|ConversationClientTests'

---

Note

Medium Risk
New public audio callback API runs on the real-time audio path and changes teardown ordering around WebRTC disconnect; behavior is covered by unit tests but misuse in observers could affect audio performance.

Overview
Adds a public ConversationAudioObserver protocol so apps can tap decoded PCM from agent output and local mic for visualization, recording, or analysis without LiveKit types on the SDK surface.

ConversationClient exposes durable add/remove APIs for agent and mic observers; registrations are re-bound on each new session when bind runs. Conversation owns per-session AudioObserverRegistry instances that bridge observers to internal LiveKit AudioRenderer adapters, reattach on track swaps, and reset on teardown.

WebRTCConnectionManager surfaces onTracksChanged from LiveKit publish/subscribe/unpublish events and types tracks as AudioTrackProtocol internally. Session end / startup failure paths tear down observers before disconnect() and use isTearingDown so stale track callbacks cannot re-attach after cleanup.

^{Reviewed by Cursor Bugbot for commit ffd0362. Bugbot is set up for automated code reviews on this repo. Configure here.}

[jacksonh]

This looks very useful.

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, have a team admin enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 27ed8e2. Configure here.}