V3 deterministic audio-parse capability pack. By-value connector composition (service_kind=cli, no separate ingredient): the local whisper CLI transcribing a recorded audio file (mp3 / wav / m4a / flac / ogg — whisper auto-decodes via ffmpeg) into plain text the model can read, via the audio.transcribe catalog operation. This is the media-toolkit parse-in flagship — audio -> text mirrors docling's document -> Markdown. Fixed-function and reproducible: transcription happens locally before any model sees the content. Write-tier but approval=never: the transcript is a faithful local file write into an engine-managed throwaway temp dir, ingested into data.file.received (CAS) and surfaced as result.file_ref; the temp dir is removed after capture, so the only durable copy is the content-addressed, Gateway-gated data.file record (output_capture). A downstream ai-* step reads the transcript via { file_ref } so the content never flows through op-step values. The source can be a data.file ref (the ffmpeg media.extract_audio handoff, or a Drive/manual download) or a local path (input_materialize). Long recordings (whisper runs an ML model) can exceed the 90s foreground cap; those ride a detached execution path when it lands. Requires the openai-whisper CLI (pip:openai-whisper) and ffmpeg on PATH (whisper shells out to ffmpeg to decode); the model weights download on first run. Part of the Lane-1 media toolkit alongside ffmpeg (extract audio) and imagemagick (convert images).