roomkit 0.20.0

Added

• Ephemeral tool-call events. The tool loops publish TOOL_CALL_START /
  TOOL_CALL_END events so callers can surface tool activity live.
• Anthropic prompt caching. Explicit cache breakpoints on the stable
  request prefix cut input-token cost on multi-turn conversations.
• Gemini cached-token usage. Implicitly-cached input tokens are now
  reported in usage.

Changed

• Vendored, gradio-free WebRTC transport. The WebRTC transport is
  vendored under roomkit.webrtc (extracted from fastrtc 0.0.34); the
  fastrtc extra now pulls the transport's own deps (aiortc, av, librosa,
  pydub, anyio) instead of the upstream fastrtc package and its gradio 5.x
  / pillow<12 constraints, so the default install is gradio-free.

Fixed

• OpenAI Realtime reconfigure is in-band. reconfigure sends a partial
  session.update instead of tearing down and reconnecting, so the
  conversation and the in-flight tool call survive — Tool Search and skill
  activation work over OpenAI Realtime.
• Gemini parallel tool calls are replayed signed, never as thought parts.
• ICE connection timeout raised 30s → 60s so a client reachable only over
  a slow TURN relay (strict NAT) can connect before the timeout fires.
• read_stored_result paging. Pages carry more content per round while
  staying under the re-eviction bound even for worst-case JSON escaping, so a
  large evicted result reads back in a few rounds without looping.