Just a couple points, from a programming point of view.
1) The seeds are not meaningful unless you can actually exercise the code and produce the same throws. Unless you know for certain the exact bytes running on the server, you can't be sure what you are seeing. There are many possible cheats here, but they are complex in nature for an unskilled programmer. I won't say more, but there are many possibilities.
2) You can still cheat without mucking with the random numbers by favorable rounding of the thresholds, etc. A ton of little nicks can add up and if caught ... "Thank you for finding a bug!"
The problem is proof in the crypto world is only hash deep. You can wrap a cheat in proof and it would still cheat.
So you are saying it possible that I test 100000 rolls with their code and the seed (which is partly provided by me) in question which all turn out exactly as they did on the server, but they still cheat every 1millionth roll?
Even if thats true - which I doubt - it still needs only a single person to test 1million instead of 100k rolls and they are done with it. E.g. paradocks (#1 rolls on coinichiwa) has as of now made 3,929,072 rolls. Even if only half of them have been verified the scam would be public by now. The users total wagered is ~4.8 BTC which is 122 Satoshi on average. Lets say your code would pass all the tests and cheat every millionth roll, that user would be short 488 satoshi.
Again I know provably fair is not 100% because there are other factors but the roll, but to say that there are possible loopholes is very unlikely. Id like to see stats like those for primedice, but aparently Stunna keeps them hidden or I am to stupid to find them. The site has run 3 billion rolls now. Someone would have caught a cheat.
[1] https://bitcointalksearch.org/topic/m.10174134