This is all kinda off-topic, but true randomness is possible to generate via radioactive decay. By measuring how long it takes for an unstable atom to decay, you can generate truly unpredictable, random numbers.
As for RNG being both predictable and unpredictable:
Imagine you roll two 6-sided dice, one colored red and the other colored blue. There are 36 possible distinct combinations, 6 for the red die * 6 for the blue die, and each roll is essentially independent of the previous rolls (assuming stuff like ideal dice, perfect unpredictable rolls, etc.)
But if you record your results over a long period of time, you find that the sum of the numbers on the two dice is most likely to be 7. In fact, with an infinite number of trials, we’d expect that 7 would be the sum 1/6 of the time.
What seems to be tripping you up here is the Gambler’s Fallacy: while reversion to the mean is the tendency given a sufficiently large number of rolls, each individual roll is still unpredictable. After flipping a coin and getting 5 heads in a row, I’m not more likely to get a tails instead of a heads on my next flip, even though the probability of flipping 6 heads in a row is tiny: 1/64. That’s because the sequence of 5 heads and then a tail is equally tiny: 1/64.
So ultimately, random number generation is truly random, and you can consistently produce values that are impossible to predict. Given a sufficiently large number of trials, you can ensure relatively uniform distribution, but extremely unlikely but possible events can still happen. I’m not privy to how DW produces gacha rolls, but they most likely generate some big random numbers between 1 and 100,000 or so, and then look up each value in a big table, where each number is tied to a specific value (Mapo Tofu, Dragon’s Meridian, Arjuna). Then, they return the results to you in your daily dose of salt.