Rolling a die 500 times to look for bias is grossly insufficient. You need thousands of flips to exclude a small bias from a coin which only has two possible outcomes (which is why it is faster to not bother and just use a Von Neumann approach). For a die with (presumably) 6 possible outcomes you are looking at tens of thousands of rolls.

You're sure about that? As soon as you deviate from a uniform distribution, then min entropy becomes more import than Shannon entropy. What's the min entropy of your biased die? If you don't know what I'm talking about, then how do you know your system is safe?

Ian Coleman did not create BIP39. You might also be interested in this discussion of bias on his GitHub: Bias in dice based entropy

The coin flip method is better because it is faster, it is simpler, there is no possibility to introduce various biases such as modulo bias, and most importantly, it is verifiably random.