Cross-platform replay of the frozen assessor returned INVALID_SET because
recomputed eps(128) differed from the Horsea receipt by one ULP. Compare
only that derived float within 4 ULPs (close_ulps); hashes, case identity,
other metadata and scientific thresholds stay exact. Tests accept the
observed one/two-ULP differences and reject 1e-12 relative, 2x, string,
NaN and missing values.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
