Testing

296 tests across 17 files. They are design guards, not coverage theatre — several encode an argument that would otherwise be lost:

• •mean vs median — five posts and one at 100k, proving a mean baseline would switch outlier detection off permanently
• •contrast against every gradient stop, not just the first; it caught two real failures
• •verticalFill — because "everything is inside the artboard" passes for the broken reflow too
• •the gates stay quiet — no finding below MIN_TOTAL, none from a group below MIN_GROUP
• •the demo’s two Story posts have excellent numbers on purpose, so a test can prove the app refuses to conclude from them
• •deletions do not resurrect across a second sync round trip