The /help says it decodes only 3 specific lengths of base32 strings, but instead it's looking at all other lengths greater than 8 which are multiples of 8 without containing the '=' padding character, to see if they match the pattern of base32 encoding.
That is correct. The comments in my code state that it should check for >=16 and multiples of 8 instead of just 16/24/32. I cannot remember why as I implemented this feature three years ago and researched it at that time. You will need to research this, and the Google Authenticator format, yourself I'm afraid.
