To lend credence to this as being more than some weird Lua-side quirk in how the above game is programmed, I can voice three prior times I experienced this bug:
About a year ago, we encountered a similar bug in Robloxian Highschool. After investigation into the server DataModel state, the client DataModel state, and the state of the actual scripts involved, we concluded that replication of reparenting a model must be breaking. If I recall correctly, we were reparenting a model from Workspace to ReplicatedStorage.
This bug was very inconsistent: some servers had it, most did not. In the RHS situation, we reparented the model on startup. If this bug is triggered by parenting a model while a client is connecting, then it explains why it happened rarely in RHS’s case: we weren’t parenting it on client connection, so a client had to connect at the right time by coincidence to trigger this bug. We were also unable to reproduce this in Studio, likely largely owing to the fact that we mostly tested with one player in Studio.
I have seen this bug in one other RedManta game and one personal project but I forget the details of both of those situations. I was not able to pin down an accurate repro in any of these situations, but I was able to isolate the bug to an internal replication issue with parenting objects near startup in all of them. The difficulty in reproing in all these cases likely came from it being a race condition on client connect, not on server startup, meaning that the parenting and client connection had to coincide to trigger this bug.
I wanted to post this bug in the past, but I always lacked a solid repro, a good theory as to the cause, and any solid place to test as a repro. Any time this happened, it happened rarely and we couldn’t leave the bug in the game as a repro. These anecdotes aren’t much better, but they certainly suggest that it’s an internal engine issue and not a Lua-side bug in all of these games.