I would not think that would be a surprise. The code generates all 7040 solutions, with 880 unique arrangements of numbers, the other 6160 being the same arrangement but rotated/mirrored in some combination. There is no reason to believe that the code would generate all the unique combinations first.
My interest in code that generates only the 880 unique solutions was to see if it would speed up the code by reducing the number of calculations. Once you have the unique solutions, the rotation/mirroring can be done by having eight versions of the seq array that represent all the possible rearrangements.
That would be quite easy. Once the code has finished, copy the contents of the serial monitor to a text file, then take the first number of each row (1, 12, 13, and 8 in the example) and search for any occurrence of those numbers in a horizontal row (" 1 12 13 8"). Since the code starts with 1 in the upper left, all answers with a 1 in that position will be generated before advancing to 2 or any high number, making it unlikely the matching solution will be very far away.