Depending on the size of the blob you might not want to use `WITHOUT ROWID`. I had a similar table with uuid keys and blobs of 40kb size on average and `WITHOUT ROWID` really tanked performance.
In Germany I spent exactly 0 dollars/euros on books in university. We got access to a huge amounts of ebooks through the university network and profs never required a specific one and would just recommend a few. One of these was always available. This free access was of course covered by the 250 euros per semester tuition…
But yes, standards and certifications are horrible.
I would understand if they didn't backup anything with FILE_ATTRIBUTE_RECALL_ON_DATA_ACCESS set (on Windows). Everything else makes their product worthless.
Since all input is run through a tokenizer, I would expect the tokenizer space doesn't change a lot between one trained on uncompressed vs one trained on compressed data.