override prototype in basic construction(like array) it's monkey patching, not best practices)

You could substitute the last for loop with a call to:

auto iter = odds.begin();
std::transform(array.begin(), array.end(), array.begin(), [&](const int& x) { return (x % 2) ? *iter++ : x; });

I believe that'd make it even cleaner.

supa smart

For Java，obveriously the ListNode class already have a proper equals() and hashCode() overriden though not shown in description, maybe it's better to mention that in the question.

In Haskell, the test contains an array which is not of type [Int], but it should be so if you look at the task and function's type declaration.
Also (-6850,[7003,-6850,7003]) is an even length array.

Great kata. During implementation of this kata the answer of the question could be found. What question?
Why do every, and all? methods in JS and Ruby repsectively return true on empty sets?

Thank you a lot.

Clean solution!

In my solution, I copied the entire input array to a separate vector and then the odd values to a separate vector, so that my last for loop only needed to go through the values that were already known to be odd. In your solution, your last for loop might take longer than my last for loop. Yet...my initial copy is probably less efficient than your copy_if(). I probably could have used copy_if() to clean up my solution.

There's a comparison of (signed) int to (unsigned) array.size()?

Minor 'nit' to pick: Modulus is expensive. x % 2 == x & 1.

Those are two different cases and you should read this note:

Note: the list may be null and can hold any type of value.

There is your problem.

what should I learn to be able to write code for this kata?

I still think it is a kata issue, the original list should not be mutated should be a requirement then :/