Your 'razor' theory doesn't make a straight line when those 'test ballots accidentally left on servers' heavily skewed toward the super left AOC style candidate who was polling WAY behind. Flat numbers wouldn't be shifting totals in a direction.
Looks more like idiots trying to brute force things when they no longer have access to the subtler Dominion algorithm.
Not so long ago when officials F'ed Up this bad at any job they would resign to show a modicum of responsibility to remove the spectre of hijinx and corruption from the process. Now they just say 'Oops, my bad' and keep on trucking. People unwilling to take responsibility should never be trusted ever.
Yes, the test data IMO should have been roughly equal so the outcome can be verified easier. If you know one candidate has 1,000 votes to be input, and another has 1,001, then you will expect an exact count to result in candidate #2 winning for every test. If you make the numbers more like 1,000 vs. 10,000, and there is a "glitch" that only counts 9,990 for #2, the end result is he still wins. That's a less obvious flag something's wrong than a tie or loss for #2.
Tests require a set of votes to ensure each candidate can win, each candidate can lose, and each candidate can be in a tie. The end results should be the first checkpoint to be met. After that, the individual vote counts have to match the inputs.
You also have to have different sets of votes so the Democrat wins, the Republican wins, and the third party wins, just to weed out "errors" when ballots are tallied.
Then there's the "bad" ballot corrections. Most states allow less than 1% of ballots to be corrected as bad ballots. Last election, we saw as much as 60% of ballots being assigned to humans to correct. On its face, the machines failed their stated requirement. If the ballots are being messed up by mail-in voters, then they should be tossed. If it's a problem with the scanners or printing process, then the company failed to deliver. Either way, the bad ballots should have been segregated and reviewed, then analyzed for the causes, before being allowed to be counted.
The test plan would include more cases than I listed, but you get the idea. Testers should be familiar enough with the systems from a design, implementation and operation standpoint that they can anticipate what kinds of problems they should test for. Once the tests are declared "good", they then need to introduce test cases that contain errors, to see how the system handles them.
I've done software and system level testing on large computer systems and networks my whole career. Being a tester is boring at times, but the more time you spend developing the plan, the less time you spend fixing bugs after release.