feat(agent): Phase 4 — Voice Agent WebSocket client (#152)

[dg-coreylweathers]

Phase 4 — Voice Agent WebSocket client

Stacked on #166 (Phase 3). Base is feat/phase-3-tts-ws-selfhosted; GitHub retargets to main as the stack merges. This is the strategic bet from the parity assessment.

What's included

• New agent feature + module. Deepgram::agent() → Agent; start() / start_at_url() open a live session at wss://agent.deepgram.com/v1/agent/converse, returning an AgentHandle (send Settings, UpdateSpeak/UpdateThink/UpdatePrompt, InjectUserMessage/InjectAgentMessage, FunctionCallResponse, KeepAlive, and binary audio) and an AgentEventStream (futures::Stream) of typed AgentResponse events (Welcome, SettingsApplied, ConversationText, UserStartedSpeaking, AgentThinking, FunctionCallRequest, AgentStartedSpeaking, AgentAudioDone, History, Error, Warning, mid-session *Updated acks, Unknown).
• Fully typed Settings: inline or saved (Uuid) configs; listen/think/speak provider settings (Deepgram STT V1/V2; OpenAI/Anthropic/AWS Bedrock/Google/Groq think; Deepgram/ElevenLabs/Cartesia/OpenAI/AWS Polly speak); audio I/O; conversation history. All event/response types #[non_exhaustive]; unknown messages preserved.
• Examples simple_agent, function_calling.

Provenance & scope

Ported and adapted from the maintainer's feat/agent-websocket branch — the live WebSocket client only. The REST saved-configuration / variables / think-models namespace is deferred per the assessment ("deferred until the live Voice Agent client exists").

Verification (Rust 1.97 container)

• build / clippy -D warnings / fmt / cargo test --all --all-features (207 lib + 141 doctests) ✅
• per-feature cargo check incl. agent-only ✅; doc build -D warnings ✅
• cargo semver-checks: no new breaking changes (only the inherited Phase 1 Extra metadata support #130 breaks).
• Live-verified end-to-end against production: connected, Welcome, Settings applied, received greeting audio + ConversationText "Hello! How can I help today?" + History + AgentAudioDone. Also verified the session runs cleanly under a current_thread runtime (no hang).

Note (fix beyond the port)

The ported agent worker had the same latent busy-spin/current-thread-hang as the Phase 3 TTS-WS worker (kept selecting on the drained input channel after close). Fixed here identically (stop selecting on message_rx once input closes; shared handle_agent_inbound helper) and verified under current_thread.

Closes #152

[dg-coreylweathers]

Review addressed. Worker fix verified correct by the review (no busy-spin, no dropped messages, ordering/Ping/close semantics preserved; verified live under current_thread). Fixed the one SHOULD-FIX: four [bracket] intra-doc links pointed at listen/speak-gated items and broke cargo doc --no-default-features --features agent -D warnings; demoted them to plain code-spans (agent-only doc now builds clean). Also documented the send-then-drain backpressure contract on AgentHandle. The untagged-AgentResponse degrade-to-Unknown and shared bounded-channel caveats are intentional/forward-compat and consistent with the existing Flux client.