Make failure-taxonomy precedence data and measure its swing

Single-label failure classifiers hide a big free parameter: when a run exhibits several modes at once, the reported prevalence table is a function of check order, not just the data. Expose an unordered detector that returns every mode a run exhibits, keep the precedence as a reorderable tuple rather than an if-elif chain, and add a test asserting that reordering actually moves labels so the caveat cannot go stale. Measuring the swing is worth it: in one corpus 38 percent of failures exhibited multiple modes and one category more than doubled purely from reordering. Co-occurrence counts also expose categories that are not mutually exclusive, e.g. a tool substitution triggers both wrong-item and missing-item detectors, which is a taxonomy design flaw worth surfacing rather than resolving silently by ordering.