No, we do not break sentences. Data is stored via OP_HASH160 exactly the way you enter it, it's not possible for a single bit to be different.
Some sites like cryptograffiti chooses to break up scripts, when there is no accepted standard to do so (it should be joined together). Rather, they invented an arbitrary jigsaw-like format for re-arranging scripts.. that is going to look gibberish if you do it the traditional way (join output scripts).
That's why they're not suitable for archival purposes, you have to rely on their website for their esoteric ordering algo. If you 'cat {the blockchain} | string -n 20', you'll get unsorted gibberish with anything encoded using their service. If you use blockchain-db, you get them in order
Furthermore, their service can't force the change address to be at the bottom. There's gibberish in the middle of these messages, and that makes file storage impossible. We've patched our bitcoin core node to always place the change output at the end, and the OP_HASH160 is paddled with terminal 0x00s so you actually know where the file ends.
Also, Blockchain-DB lets you store raw bytes, eg binary files, encrypted files, small images, compressed audio files, etc.
As you said, blockchain.info notes are centralised and are not suitable for archival.
TL;DR: Blockchain-DB.com is the only archival-grade service that implements message embedding via OP_HASH160 correctly.