Runs local speaker diarization and speech recognition on Apple Silicon (macOS 26+, CPython 3.12) with CPU Nemotron and MLX Qwen3-ASR, preserving the original timeline, speaker ranges, and raw transcript. Use it when an agent needs a staged prepare-edit, semantic-cleanup, and finalize-edit flow to turn a recording into readable text; speaker names use caller-provided mappings or text-based inference, and uncertain entries stay marked.