Thanks for your quick response, and also for being constructive

I think I'm getting tired of other people and their conspiracies, and therefore I'm quick to assume (incorrectly sometimes) that everybody is being irrational.

[I love conspiracies sometimes, and Bitcoin is too important to be based on naivete. We should verify, not trust]

Anyway, to return to the topic:

Inscriptions require *two* transactions usually (or maybe always). Witness data can be attached only to inputs, not outputs. So, in order to attach some witness data (indirectly) on an output, you have to have a second transaction - taking that output as an input - and then attaching the witness data to the second transaction.

I mention this complexity because it might have been another factor in the topic. The Inscriptions folks might have been willing to do this. But maybe more recent use cases require a single-transaction approach and therefore the witness data is unusable and therefore the choice is between a large op_return or unspendable outputs.

[Again, I might be missing some details here. I think I know a lot of these details, but I'm kinda inferring things from multiple different blog posts. I would be happier if a really high quality document about all of this already existed]

My guess is that the truth is something like this: Witness data is the cheapest way to get arbitrary data in the block. But, as it requires two transactions - not just one - it is sometimes impossible for certain use cases. The single-transaction use cases therefore have to choose between a large op_return and unspendable outputs. Reluctantly, in this case we must large op_returns in the mempool if we are serious about miner decentralization.