Support was never redundant

Writing a demo of destructible terrain, the plan was to narrate something satisfying: two columns hold up a mezzanine, a character destroys one of them, and the mezzanine holds — because the other column is still there. “Destroy one column and the mezzanine still has the other,” the narration said. Then the demo ran, and the engine collapsed the mezzanine anyway.

The narration wasn’t decoration. It was a claim about behavior, written down where it could be checked, and the check failed. Support wasn’t redundant the way the demo assumed. Losing any supporter collapsed the element it held up, which meant two columns behaved exactly like one: destroy either and the mezzanine came down. The model had no concept of “enough support remaining,” only “any support present.” A structure with two legs was, as far as the engine was concerned, exactly as fragile as a structure with one.

None of the existing tests had caught this, because none of them had tried to narrate the redundant case. They exercised single supporters failing and single supporters holding, which is a different question from whether a second supporter changes the outcome. The gap wasn’t in coverage of outcomes; it was in coverage of a specific structural claim that only showed up once someone tried to say, out loud, what should happen and why.

That’s the general shape of it: an example forces a narration, and a narration is a falsifiable claim in a way that a green test suite alone isn’t. Writing “the mezzanine still has the other column” and watching it turn out false was more informative than any number of passing assertions about columns in isolation.

The fix models each supported element with its own support requirement, defaulting to needing just one remaining supporter, so a two-column mezzanine now genuinely tolerates losing either column on its own. The example that broke first is the one that now demonstrates it works.