Hi, author here.

I'd be interested in hearing how to accurately estimate passphrase strength assuming the attacker has perfect knowledge (ie, that you used Wordentropy to generate your passphrase, approximate length, etc).

In reality, though, no attacker will have this. Your password hash will be brute forced along with thousands or millions of others, so I think it's accurate to look at the password as a semi-random bitstream. In that case length trumps all other characteristics when it comes to security.

It depends on the attacker. It is possible (though far less likely in the grand scheme) to be targeted as an individual. In that case, an attacker could use compromised data from weaker services to derive information about your other passwords.

This is one of the reasons I don't recommend patterns as with the one mentioned elsewhere in this thread. While something like "[category] [words] [name of service]" might aid memorization, the pattern will leak if a weakly-protected secret is compromised, allowing an attacker to approach more important secrets.

It's also much more costly because the attacker cannot rely on the birthday paradox anymore to reduce the cost per bit of entropy.

Also, it depends on how popular the scheme becomes, you never know! :-D