This attack is still done on device. It just clones the NAND back to "0 strikes" after each 6 attempts.
This attack doesn't extract the memory and doesn't decode externally. It just copies NANDs.
Why is this significant? Because it means you can't do extraction in parallel, you still have to go through all the codes sequentially on the device.
It defeats the significant portions of the backoff. It defeats the erase after n failures. It's a very significant attack.
But no one said this type of attack was impossible. I personally read about variants on this attack while the controversy was going on. I even posited it myself. I believe Apple even addressed it claiming that this attack wasn't possible on later iPhones due to a change in how the failure count is stored.