Agent Speak v0.2.0

Agent Speak 0.2.0 adds observable, cancellable playback and opt-in audio
mixing while preserving the local, policy-controlled MCP boundary.

Highlights:

• report queued, active, completed, cancelled, and failed playback without
  exposing spoken text or local paths;
• cancel one playback or stop the complete queue through a private local
  control channel;
• opt into bounded mix playback with one shared output stream and
  proportional gain headroom;
• discover layered system, user, and working-directory profiles when
  --config is absent, while an explicit --config remains a complete,
  unmerged profile;
• expose structured device and system-voice inventories for local UIs; and
• update the Pocket TTS example for the verified XN Q8 provider profile.

Existing v0.1 profiles remain valid. Mixing is disabled unless mix is added
to playback.allowed_concurrency; an omitted maximum_mix_streams defaults to
2. Release binaries and tags remain unsigned, with SHA-256 files published
beside every archive.

Full Changelog: v0.1.0...v0.2.0