feat: ai.poly:voice — WebRTC voice calling SDK

[Zakariya1057]

Overview

Adds ai.poly:voice (the :polyvoice module) — live two-way WebRTC voice calls to a PolyAI
agent, the on-device counterpart to ai.poly:messaging. Ships as a separate artifact so chat-only apps
stay lean (the call path pulls in the native libwebrtc audio engine).

This also bumps the SDK to 0.9.0 — tagging v0.9.0 after merge publishes ai.poly:messaging:0.9.0
and the first-ever ai.poly:voice:0.9.0.

What's new

:polyvoice module — hexagonal ports/adapters (mirrors :polymessaging):

• A pure CallCoordinator state machine (auth → session → link → ICE → signaling → peer → offer),
  signaling wire codec, trickle-ICE buffering (both directions), connection-timeout + disconnect-grace,
  signaling-drop reconnect with backoff (1s/2s/4s ×3), idempotent teardown.
• libwebrtc peer (Unified Plan, deadlock-safe native teardown) + OkHttp REST/WS adapters.
• Audio output — accessory-aware by default: the call follows a connected wired/Bluetooth headset
  automatically and auto-switches when you plug/unplug one mid-call, falling back to speaker (or
  earpiece). A picker (call.audio + setAudioDevice) lets users pin any output and return to
  automatic via setAudioDevice(null) — API 31+ communication-device model + the legacy SCO path, with
  live change events. All route mutation is serialized under a lock against the coordinator thread.
• Audio-focus interruptions: transient loss → mute/restore (call stays connected); permanent loss /
  incoming phone call → end with PolyError.Voice.Interrupted.
• Credentials: Configuration.apiKey (X-Token) + a required, distinct VoiceOptions.webrtcToken
  (the gateway auth token). Session create reports platform=android + a detected device_type
  (mobile/tablet), consistent with ai.poly:messaging.
• Public API mirrors :polymessaging idioms — Kotlin suspend and Java Callback/Executor
  overloads on everything. Ships consumer R8 keep rules for the JNI surface.

:polymessaging — additive only: PolyError.Voice.Disconnected + Interrupted; PolyCall KDoc
now points at the voice artifact. ABI otherwise unchanged.

Examples reorganized into examples/chat/{compose,views} and examples/voice/{compose,views} — a
tap-to-call demo with the audio-output picker (incl. an Auto option) and a microphone foreground
service (+ wake lock) so calls survive backgrounding.

Docs — a dedicated polyvoice/README.md guide (permissions, credentials,
audio output, interruptions, backgrounding, R8); the root README keeps a concise voice quickstart that
links to it.

Testing

• JVM unit tests in :polyvoice over port fakes (signaling codec, the full coordinator state
  machine, ICE buffering, timeouts, disconnect-grace + reconnect, mute, audio-device selection, focus
  interruptions, platform/device_type on session create) + the :polymessaging suite; both
  apiChecks green. CI runs build test lint apiCheck.
• On-device: two-way audio, accessory-aware routing + mid-call switching (speaker / earpiece /
  Bluetooth), incoming-call interruption (ends with Interrupted), and backgrounded-call survival
  (foreground service + wake lock) all verified.
• The accessory-aware routing was reviewed for cross-thread safety — all route mutation is serialized
  under a lock.

Notes

• The voice examples use YOUR_API_KEY / YOUR_WEBRTC_TOKEN placeholders (set them in the
  PolyVoice.call(...) block) — no committed credentials.
• The chat examples move (examples/compose|views → examples/chat/...) shows as git renames — most of
  the file count is moves; the substantive surface is polyvoice/ + examples/voice/.