So I had reported a bug in Snort nostr:npub1v0lxxxxutpvrelsksy8cdhgfux9l6a42hsj2qzquu2zk7vc9qnkszrqj49 about threads being out of order frequently and I think I now understand what's going on.
When replying, most clients copy all the p-tags to notify all people that were mentioned before of the new event.
Some clients appear to also copy the e-tags verbatime.
Alice: id1, no tags
Bob: id2, tag p-id1--reply
Carol: id3, tag p-id1--reply, tag p-id2
Carol's client might "magically" show id3 as a reply to id2 because the positional first p-tag gets interpreted as the thread's "root" even though it's marked "reply" as per nip-10.
Clients that recognize marked p-tags will recognize id3 as a reply to id1.