ARC-AGI-3 tests AI across novel environments, not puzzles. Here’s what 1% score means for the tools you’re building with.