The answer is simple: it does not.

I am both delighted and frustrated by this kata after completing it in Python.
I am glad I found one fast solution myself and a number of slow ones.
I am glad that I learnt that in Python positive and negative power notations have dramatically different performance, that Py integers do not have precision and that things like float128 are platform-dependent fiction.
I am glad that I have gone through numPy docs and now understand what it is and how to use it (and even solved with it), as well as that even numPy may have a super-counterintuitive location of a seemingly super-necessary functionality.
And I am terribly frustrated that in the end I see the solutions out there that just should not be passing by timeout, it's basically my slow code ignoring all the above, which I just cannot explain.
Very enlightening task.
I do not mar it as spoiler, though this one has some hints.

Why is this not being upvoted? IMO the cleanest and the most straightforward to understand.
Oh yes I know why. Because someone wrote a bedsheet of code above and nobody scrolls here :)

It should be "...else x-2 * y...". or "for c in roman.reverse()"

I suggest to have the "bad source" with unmatched brackets in tests. That way the [ and ] operations processing would be tested for not going out of the code bounds. I feel that my current solution would fail this way.

Because if 1 OP is 1 cycle, by jumping PAST you are saving 1 cycle, otherwise wasted.

Doesn't counting the positions of brackets make this more a compiler than an interpreter? :)

Wait, you are creating a new StringBuilder EACH TIME you are flipping the bit? This solution deserves a special un-optimization reward.

I did somewhat similar, but this can be simplified, since tempWord.Remove() returns bool. So you could essentially foreach(){if!Remove => NOPE}