v0.7.0a13

What's Changed

• fix(tests): repair hangs and regressions from a13 hardening commit
• fix: pre-release hardening for a13 — hooks index, Gemini state, inject_image
• docs: add changelog entry for v0.7.0a13
• fix(tests): add APIConnectionError stub to OpenAI/Azure/vLLM mock modules
• refactor: lift callback dispatch into RealtimeVoiceProvider base class
• fix: Gemini inject_text/inject_image 1007 disconnect during active audio
• fix: guard against non-str deltas in delegation and supervisor streaming
• fix: OpenAI/vLLM/Azure provider resilience — fast fail on connection errors
• fix: mock has_hooks as sync MagicMock in turn detection tests
• perf: skip hook dispatch when no hooks registered for voice/realtime
• fix: cancel directive now interrupts streaming mid-generation
• style: ruff format _realtime_tool_recovery.py
• fix: lint errors — contextlib.suppress and line length
• fix: suppress ty unresolved-attribute for cross-mixin method call
• feat: add tool-call-in-text recovery and inject_text sanitization
• fix: make idx_participants_channel non-unique for group room support
• feat: pass server-side RTCConfiguration through to FastRTC Stream
• fix: _build_context loads only MESSAGE events via get_conversation()
• debug: add detailed logging for Gemini inject_text and disconnect errors
• fix: queue image injection during pending Gemini tool calls
• feat: add inject_image to RealtimeProvider for multimodal voice sessions

Full Changelog: v0.7.0a12...v0.7.0a13