To me, the more relevant “technical” observation is that all the samples on this page are bad choices, including OP’s, yours, mine, and XKCD’s because they have now all been publicly mentioned.
To keep the story self-contained, I was reflecting back to the referenced xkcd, and showing how their example horse correct battery staple fits into the table, with a slight onion adjustment to accommodate a 64-bit table.
But you are correct that since horse correct battery staple is in the Have I Been Pwned database, its effective (as opposed to mathematical) entropy caps out at a lesser strength. But that cap is not zero because there is no expectation that horse... would be tried first. I suspect a better estimate for the cap would be 30 bits because horse... occurs 11 times in HIBP’s database of 15 billion powns ( 30 ~= log2(15b/11) ).
As an aside, if one adopts the philosophy of checking substrings against HIBP, every passphrase, including accurate cocoa heroics impulsive last would cap out at zero (or 30) bits because all of the individual diceware words seem to have been pwned. To my knowledge, this philosophy has not been adopted.