This Kata is a real Minefuck :)

I also tried to generate zero as a sorted list of index tupel, based on the length of that index in a dictonary, that contained (index):{applicable values for indexposition}. The idea was to make good decisions early; at the cost of potentially no marginal information gain - since we jump through the sudoku. The latter proved to be true: the midlevel values are traversed at very high cost. The number of iterations is by far greater, if zero is sorted this way (average traversals about 20000).

Any other idea to reduce the iterations?
I had one of the test cases take about 77000 steps. all others were reasonably within the range of 300 - 4000 with focus on the lower bound.

I just edited the devastation function with help of best practice - didnt see it before, but Blind4Basics had a much cleaner version of what i previously had to clean up. Thank you for putting my nose to it.

6 is the expected value of the previous test, for that test it's -1. If you're printing the input, it's shown before the test result, not after.

Test data No.7 for Python seems to be incorrect. The correct output for this data set "range(100)" is -1 (stated in the test function) while the checking system said the correct answer should be 6. What may be the reason here?

Read this
You're not using L (as a list) at all, and why are you using the same var in your for loop and your list comprehension?
It's not a kata issue when it's a problem with your code, use Question label next time.