Since I only vaguely remembered getting it to send to the network, and had not (at the time) found the source I wrote back in 2020, I actually had to reverse engineer my own engineering.
Jogging the dial would send 36 bytes of binary data to port 12123 on the network's broadcast address. This ended up being jtlvi (Just TLV it!), a general-purpose TLV standard I wrote years ago which would be XKCD 927 except I'm sure I'm the only one who has ever used it.
Tag 0 was 8 bytes of what looked like random data. Tag 1 was always 0x1ebf. Tag 2 was the event; I worked out 1 for step up, 2 for step down, 3 for single press, 4 for double, 5 for long. For step events, Tag 3 would be set to the number of steps seen during the event window.
I was later able to find the actual firmware source I modified, and the mystery of Tags 0 and 1 were solved. Tag 0 was... literally 8 bytes of random data. No idea why, but that's what the source was doing. Tag 1 was hardcoded to 0x1ebf; maybe I intended that to be a node ID.
https://forge.colobox.com/rfinnie/jtlvi/