First of all, good that you like to improve your site.
Technically the cold wallet address is not yet proven. You should sign a message with it so others can verify it. It would be rather foolish to give a random address, so I do think it is from you, still would be better.
Besides that, with a 0.5% HE you are risking at 40x kelly with that max payout with a 100 BTC bankroll. Basically a highroller can easily win more than 100 coins and it seems that you cannot pay him out then. That would be obviously bad. IMO you should be offering around 0.25-0.5 BTC max payout dynamically based on the bankroll. Basically: [0.5x or 1x] * [house edge] * [bankroll]. I am not an expert in kelly criterion and/or risk of ruin theory, but pretty sure that is the practical side of it.
The provably fair implementation seems a lot better, but is unfortunately not really good yet. Since I would repeat myself from another thread I will just copy this here:
When a player changes the clientseed, you give a new serverseed. This means you could calculate bad results based on these seeds. The idea of a clientseed is that there is a variable in the calculation which the site doesn't know in advance and therefor cannot predict the results in advance. 2 options:
1. When someone sets the clientseed, you should keep the same hash (and don't reset the nonce or else player can cheat.) And then there should be a separate button to "request new serverseed" that also resets the nonce.
2. OR show the "next serverseed hash" already when setting the new clientseed (that is actually what PD does.)
To give a practical example of how you (as site owner) could cheat: if a player is always betting 2x at high. You could generate a serverseed that has relatively a lot of low numbers based on the clientseed that you already got. Or some players like to bet at 9900x multipliers, you could easily generate a serverseed that doesn't hit that (in first hundred thousand bets.)
I know that might seem silly, but technically it could be possible and it is a reason why the implementation is not yet really good.
Besides that I only had a quick look so not sure if I missed something.