The fuzzer finds amusingly short sequences that cause data loss and that you can reproduce on your own in a minute (if you have an Obsidian Sync subscription). Steps in the repro README.
I guided Claude Code to build the fuzzer, but I didn't read the code. This didn't go well. The project ballooned to a painful 60h, so I wrote a separate experience report on the many failure modes and what I'll do differently next time: https://hmijail.substack.com/p/building-a-semantic-fuzzer-fo...