I also remember when nostr:npub1qqqqqqyz0la2jjl752yv8h7wgs3v098mh9nztd4nr6gynaef6uqqt0n47m proposed the "d" tag for replacing events based on that parameter instead of based only on the author's public key. I couldn't see any use case at all for it for many months, and the use cases he cited were not very convincing to me either.
Then suddenly both I and many others started thinking about way too many things in terms of "parameterized replaceables", culminating with nostr:npub1q3sle0kvfsehgsuexttt3ugjd8xdklxfwwkh559wxckmzddywnws6cd26p's proposal that even poll votes and likes should be such "addressable" events. At that point I realized things had gone too far and now I see a healthy move towards doing more things as "normal" events, such as kind:20 photos and kind:21 video events, nostr:npub1v0lxxxxutpvrelsksy8cdhgfux9l6a42hsj2qzquu2zk7vc9qnkszrqj49 has also recently proposed getting rid of the notion of replaceables. nostr:npub1acg6thl5psv62405rljzkj8spesceyfz2c32udakc2ak0dmvfeyse9p35c, nostr:npub1xtscya34g58tk0z605fvr788k263gsu6cy9x0mhnm87echrgufzsevkk5s and Clojure lover nostr:npub1jlrs53pkdfjnts29kveljul2sm0actt6n8dxrrzqcersttvcuv3qdjynqn have also spoken against replaceables in the past if I'm not mistaken, even though I think both had a short period of love for them too. Me too, recently, every time I have to deal with addressable events in the context of querying handcrafted databases, feel their pain.
Most things can be "normal" actually, even if they have to change. If you control where they're being publish you can always delete them. But also dealing with events as "facts in the past" helps. If you liked something at some point that doesn't mean you can't dislike it at some other point in the future, and so on.
On the other hand, because of practical concerns since we don't live in the clouds of Rich Hickey, I think all user metadata stuff and all NIP-51 lists should be replaceable still, because these things are updated rarily and then fetched thousands of times by many apps and clients everywhere. Telling all these readers to fetch dozens of events for each target author-kind combination and compute the result locally would be too much of a burden.
