Clearly clever, but isn't it costy as well?

Before to even be a programming challenge, this kata is a total riddle as for directives.
Reading if X[i] > X[j] then X[i] = X[i] - X[j], one expects a left to right parsing, and only for immediate neighbour. But then one has to decrypt the example. Let's try with arrow in other direction:

X = [6, 9, 21]
X_1 = [6, 9, 12] # <- X_1[2] = X[2] - X[1] = 21 - 9
X_2 = [6, 9, 6]  # <- X_2[2] = X_1[2] - X_1[0] = 12 - 6
X_3 = [6, 3, 6]  # <- X_3[1] = X_2[1] - X_2[0] = 9 - 6
X_4 = [6, 3, 3]  # <- X_4[2] = X_3[2] - X_3[1] = 6 - 3
X_5 = [3, 3, 3]  # <- X_5[1] = X_4[0] - X_4[1] = 6 - 3

It seems we first substract 2 from 3, then 1 from 3, then 1 from 2, then 2 from 3, but lastly we substract 2 from 1. Not left to right, not immediate neighbours, so what? Deep right to left parsing? Greater from next greater? Greater from lower? Strictly none of those. Let's assume it's a deep right to left as far as possible then ending as possible. But what if [21, 9, 6]? Plus three poor test cases, can't help a lot.

I really like puzzles, and this one could be great as a puzzle if it was clearly stated. If it's not a puzzle, then some things should be clarified, as a lot of people already noticed in discussion. Personnaly, I feel that tagging it as puzzle and replacing misleading instructions with some more examples/tests, would been more interesting… Then one can try it only when in "puzzle solving" state of mind.

Thanks for the (re)discovery!!