Here's a possible practical idea: 1) run the full contents of HEAD into a space-separated list of tokens considered "okay". 2) dump out the full history into

Here's a possible practical idea: 1) run the full contents of HEAD into a space-separated list of tokens considered "okay". 2) dump out the full history into a space-separated list of tokens, filter out everything in the "okay" list, and list what's left. You might want to set up some sort of incremental regex filter thing to chew through the list efficiently. But if you implicitly trust HEAD as "incontrovertibly okay" this might filter out a lot of tokens for you.