If you generate 40,000 fair dice rolls, and use this to drive Nakowa's bet sequence, you never get anywhere near earning a profit of 1/2 the amount wagered no matter how many simulations you do. Shift this data all you like, but you'll never make it fit. If the file is correct and I didn't make a mistake, then this is a smoking gun.
Lastly, if you look at my earlier posts in this thread from this time yesterday, you will see that I was recommending people analyze the probability of Nakowa winning a session assuming he bets until he hits his target, or he busts through his large bankroll. I even made plots showing that Nakawa as "cake" was actually more likely to reach 4400 BTC before he bust his (assumed) 10,000 BTC bankroll. So we agree on this point.
On the first paragraph you clearly made an error somewhere - as he's not up by anything like half the amount wagered. Or even 5%.
On the second paragraph you seem to be arguing with yourself here. You agree that him winning 4400 before losing 10k was actually a (small) favourite - but then seem surprised that if you shift the rolls you end up with a different result. When if you shift the rolls you then need to shorten or extend the series until +4400 or -10000. At a glance I suspect something's more basically wrong - as your results graph seems way too close to hugging the expectation line anyway.
You still haven't explained why you chose to shift by 2 rather than by 1 or by 3.