Implements `SpeculativeTurnTracker`, a thread-safe state machine that manages turn revisions for low-latency barge-in. It handles phases (listening, soft_ended, answering, closed), reopen grace windows, pending reopen candidates, and commit/hold gates. Isolated from the rest of the pipeline except for typed events, it demonstrates a non-obvious pattern for speculative execution in real-time speech systems and is directly reusable in other conversational AI architectures.

Provides a generic `BaseHandler[InT, OutT]` abstract class with setup/process/cleanup/on_session_end lifecycle, typed input/output queues, prefetch support, cancellation via `CancelScope`, and internal timing. This pattern cleanly separates pipeline stage concerns and can be extracted for any streaming data processing pipeline (e.g., video, sensor data) with minimal modification. Demonstrates disciplined resource management and thread safety.

Defines `BackendSpec`, `BackendSelection`, and `build_backend_registry` for decoupled handler creation. Backends are registered with lazy imports, capability flags (`streams_audio_chunks`, `supports_llm_proxy`), and CLI-normalized config. This registry pattern allows adding new model backends without modifying pipeline core and is readily adaptable to any plugin-based system requiring dynamic component selection.