I disagree. The act of choosing an outcome can reduce its entropy to potentially 0. Take an extreme case: Generate random 6 character passwords until the results are "123456". In both theoretical and practical terms, that password has an entropy of 0. Nothing about it is random. You picked a password and then used a random process to generate the password that you picked. By choosing an outcome, you reduce the "entropy space" by removing the regions that you would reject.
I haven't thought of that. Of course if a password is in a dictionary list then it doesn't matter how many times you changed words. You could keep changing until you come up with "correct horse battery stapler" by chance but that's also going to have zero entropy, but if there are five words instead of four and the fifth one is something nonsensical then that word is still going to carry entropy so that five-word diceware phrase is going to have 12.92 bits of entropy.
Can anyone verify if Diceware behave more similar to /dev/random or /dev/urandom ?
The diceware website as we know it is in JavaScript which means it's using whatever random number generator the OS and browser have, so this question really depends on which combination you run Diceware on.