I don’t think it’s an over optimization. You would need to make N relay subscriptions sequentially (N being the number of events in the thread) instead of one relay subscription in the worst case to even be able to populate your event graph in the first place. It’s highly inefficient, but also as you described it, probably necessary because it’s easy to not conform to NIP-10 properly.
Discussion
It would require clients to use another filter field, but the relay I'm working on will be capable of serving an entire thread with 1 subscription (or potentially a DVM). It may be useful enough for some clients to adopt it.
In that scheme, you can return a whole thread from the root note or from any branch whether clients use NIP-10 or not. And serve new notes that appear in thread over time.
Maybe it's the same problem nostr:nprofile1qqszw70nm86zcl0wzlcwd0xulzdgl82e95v78vdm6fl0rnlarfle35gpzemhxue69uhhyetvv9ujumn0wd68ytnzv9hxgqg4waehxw309aex2mrp0yhx6mmnw3ezuur4vgq3qamnwvaz7tmwdaehgu3wwa5kuegncs9v7 was having