Sure, they could assume every random-looking cracked password is a hash and try to crack each one, and conceivably discover your master password that way, but depending on your master password's entropy, that can be as unlikely as you need it to be. And the whole point of something like this is that you can use all your memory remembering an extremely high-entropy master password, rather than a large number of medium-entropy single-use passwords.
I'm also completely ignoring what would happen if that site were compromised. A little javascript snippet could just forward all passwords to the hacker's server. Even a browser extension could be compromised if it auto-updates. These are implementation details though that could be fixed/worked around.
Aside from the implementation details that you've raised, I'm not finding as many flaws as I expected in it.
I mean, using a password safe is no more inconvenient than having to go to a website. When set up, the safe can even be a one click auto-fill deal. I don't see any reason to take the added risk.
Imagine one of your hashes leaks. A cracker can notice your password was a hash of something and just run their normal attacks on it for leaked password + salt hashes. They don't know the salt exactly, but it will likely be part of the leaking site's url.
Now the cracker can generate your password for any account you have. Great.
Please use Keepass or another password safe to generate pseudo-random hashes. These were designed by experts.