Scientific and numerical correctness rules

Rule-pack rigor-foundry/1.13.0 adds the bounded scientific category (prefix SN). Findings are anchored needs-evidence candidates, not proof that a numerical result is wrong or a stochastic test is flaky.

Rules

Rule	Signal	Confidence
SN001-exact-float-equality-in-test	a Python test function compares a direct signed or unsigned float literal with == or !=	high
SN002-unseeded-stochastic-test	an explicitly imported supported random or numpy.random API is used before local deterministic seeding, or a supported generator is constructed without a seed	high

SN001 is deliberately narrower than type inference. It ignores integer, Decimal, container, and string comparisons, and explicit approx or isclose operands. Reviewers should replace accidental exactness with a justified absolute/relative tolerance, while preserving exact comparison when the test really specifies binary identity, parsing, or serialisation.

SN002 resolves explicit module and direct imports for a bounded set of common draw APIs. It treats random.seed, numpy.random.seed, random.Random, and numpy.random.default_rng as deterministic only when a non-None seed is supplied before the relevant draw. Module-level seeds, custom wrappers, fixtures, and arbitrary generator data flow are not inferred. Reviewers should seed locally before the first draw or prove that an external fixture establishes the exact replay contract.

Only tracked UTF-8 .py files classified as tests by repository policy or test naming are parsed. Invalid Python yields no SN candidate because the existing unparseable-test authenticity rule owns that condition. Evidence contains exact tracked-blob and line/file digests without copying source text.

The family contributes a portable control to the scientific-numerical-correctness audit domain. Both rules enter maturity probation; enforcement still requires adjudicated cross-repository precision, false-positive, and reviewer-effort evidence.