kohulan/orthonym-skills/test-gate

test-gate

The rule

A test earns its place only if some wrong code makes it fail. For every test you write or review, name the cheapest wrong code that a real bug could produce, and check that the test fails on it. If a plausible wrong version still passes, the test is not done.

Gate for a new test

Answer all four before adding it. A blank answer means do not add it yet.

1. Behaviour: the observable behaviour or contract it protects.
2. Regression: the credible bug that makes it fail. For a bug fix, run the test on the

pre-fix code and watch it fail for that reason.

1. Owner: why existing tests miss that bug. One contract has one primary test at the

strongest boundary; extend a parametrised case before adding a near-duplicate.

1. Seam: if it needs an export, flag or hook that no production caller uses, test through

the real entry point instead and drop the seam.

Wrong code to try

Try each row that applies, in both directions.

A guard is two contracts. "Refuses when unhealthy" also passes on a guard that refuses everything, which is a total outage. A fail-closed check needs a test that the healthy path still gets through. That test is required, not optional.

A check that loops over outputs passes on no output. "Every span is valid" holds for an empty list. Add inputs that must produce output, with exact expected values.

Signs a test cannot fail

• No assertion, or try/except: pass where pytest.raises belongs.
• The expected value comes from the code under test, or is imported from it.
• A mock or fixture already supplies the asserted answer: a backend mock that returns the

cached value, or an autouse fixture that pre-seeds state.

• A "concurrent" test whose callers never overlap.
• A name that promises more than its input exercises.
• A helper that exists only for the test.

Review procedure

1. Copy the tree to a fresh directory of your own (mktemp -d); a shared scratch directory

can hold another run's files. Check that the copy's tests import the copy, not an editable install of the real tree (print <pkg>.__file__). Leave the real tree as you found it.

1. For each test and fixture, pick the wrong versions from the table, in both directions.
2. Run each wrong version in the copy. A test that still passes on one is not done: fix it and

return to this step. "Would fail" counts only after you have seen it fail.

1. Report one line per test and fixture: keep, change or drop; the wrong versions you ran and

whether the test failed on each; the fix, for anything that is not a keep.

Related: prove-invariant for deriving a property the output must satisfy.