Streamed plays the file past the microphone-less audio graph and measures loudness as it goes, so memory stays flat no matter how long or how large the round is. That is what makes a 1 GB 4K round workable. Direct decodes the whole file at once at a reduced sample rate — exact and repeatable, but it holds the file in memory, so it is only offered for smaller clips.
Either way the audio is downmixed to mono, band-limited to 120–4000 Hz (where bag impacts live, and where room rumble and hiss are not), and reduced to one loudness number every 10 ms.
No impact detection, no combinations, no tempo, no coaching. Those are steps 3 and 4. This step exists to prove the phone can get clean audio out of a long round at all.
Filming tip: switch the iPhone to 1080p (Settings → Camera → Record Video). Same coaching quality, roughly a quarter the file size, smoother everywhere.