Reading Isn't Proof

When two or more implementations share one contract, a code read that concludes "they agree" is a hypothesis, not a result. Do not report "there's a test gap, but no defect" and stop. Write the executable comparison, run it, and let it decide.

If a gap is worth naming out loud, it is worth the ~30 minutes the battery costs.

One implementation is out of scope, and that is not a hedge. A promise with a single implementation is a definition, not a claim that can diverge — this rule does not license writing batteries for everything. Two is the threshold, and a fake plus the real thing it stands in for is the best case of two, because that pair is what everyone else's tests are silently trusting. Where a shared battery for the contract already exists, extend it rather than fork it.

The rule