The input is read via the lavfi device, and the tags are inserted by the blackdetect filter inside the filtergraph. So, they are present in the packets emitted by the filtergraph. Then, once decoded, they are also present in the decoded frames.
You're using two separate systems in parallel, which have no way to communicate timing info to each other. There's some delay on the video side, and some delay on the audio side, both pre- and post-USB, and the Mac has no idea how to put them back together. Normally, two-system sound setups are genlocked together to keep their internal clocks synchronized.