It's not, because if you run the inverse function (votes that Biden lost to Trump), you might find even higher discrepancies there than for Trump->Biden. Check the screenshot in my original post. I've also included the altered script that anyone can try on all of the datasets if they replace the original python file.
I've taken a look at the script and the data. First had to rewrite the script a bit to actually make it readable.
After that I made another function that compares the inverse (so the votes that "Biden lost to Trump") and here is an example. Findfraud_readable() is the original function, findfraud_readable_reverse() is the Trump -> Biden function.
The script is not flawed. However, the source data is. The vote share is in a 0.123 form, so there are only 3 decimal digits to the entire vote share. If you convert it into readable percentage, it would be 12.3%. The vote share percentage is not precise enough to base any wild assumptions on it, because the tenth of a percent rounding error on a million+ vote count easily causes 1000+ votes discrepancy (and on the whole state sample which iterates the entire election night, 15k+ may come up very easily). The function could be altered to let's say ignore a 0.2% discrepancy, but it still doesn't eliminate the rounding error which causes the vote count on either candidate to be way too imprecise on let's say more than 1000 votes total.
To list an actual example, let's assume 1 111 111 votes total.
Trump's share is 50.4%, Biden's share is 47.3%. That would make the Trump vote count total 561 999 votes and Biden total 525 555 votes.
Now add 10k votes to the total to make it 1 121 111 votes, make Trump's share 50.1%, Biden's share 47.6%. That would make the Trump total 561 676 votes and Biden total 533 648 votes. Seems like about 300 votes for Trump have disappeared.
The problem here is the actual Trump share value is 50.10% or perhaps 50.149% that was lost by the rounding error which would make the Trump total vote count 562 225 and (assuming the other candidates' share did not change during this period) Biden's share is 47.551% which would net him total 533 099 votes.
Here is the altered script that anyone can try, just replace the original python file, and call findfraud_both() function the same way as the original one.
It makes perfect sense if you look at it from a critical standpoint. A "swing state" is a state that after x number of updates swings to either candidate's advantage (instead of gaining linear traction like let's say the state of Indiana for Trump), making it much more prone to the rounding errors, since the votes are more balanced across the two candidates, causing larger discrepancies after even 0.1% of the votes is redistributed either way.