Make loader and scorer share one defect definition, then test the seam

When a pipeline reports data-quality counts at load time AND derives the same quantities during scoring, do not implement the check twice: put the per-item classifier in one module and have both call it, then add a test asserting the load report totals equal the sum over scored rows. The failure this prevents is silent disagreement where both numbers look plausible alone. Two patterns that make it stick: emit counts and rates together with the denominator (rate times n equals count, assert it exactly with math.isclose) so totals and cross-item means never diverge, and lock the output row key set in a module-level EXPECTED_ROW_KEYS constant so adding a field fails a test until someone declares its consumer. Also stamp rows with a SCORER_VERSION constant, otherwise two result files with identical columns but different methodology are indistinguishable later.