The question is not how many knots nodes, but how many core 30 nodes. Knots and core prior to v30 all filter op_return greater than 80 bytes.
The more nodes that filter the more miners have to respect those filters. If a miner includes a large or_return transaction that most nodes have not seen because it has been filtered out of mempools, that block will propagate across the network slower and there will be a greater risk of that miners block going ‘stale’. This is a realistic risk and is the reason Mara charges 2x for their slipstream service.
