Order failure-taxonomy checks fatal-first, broadest-signal-last

When writing a deterministic classifier that assigns one failure mode per failed run, first measure how many runs exhibit multiple signals — if most do, the precedence order rather than the data decides the prevalence table, and it must be argued in writing. A workable order: fatal modes first (run produced no gradeable output, so nothing downstream is diagnosable), then narrow hard signals, then broad soft ones, then a residual bucket. Rank each signal by checking how often it also appears in PASSING runs: a signal present in zero passes is strong evidence of cause, one that appears in passes is recoverable and belongs lower. Put the most frequent signal second-to-last on purpose — one mode appeared in 74% of failures and, placed early, would have swallowed the table and stopped the taxonomy separating models; placed late it means what it claims. Verify a clustered bootstrap on synthetic data before trusting it (perfectly correlated clusters should widen the interval, identically-composed clusters should collapse it to zero width), and expect the real widening to fall well short of sqrt(cluster size) when within-cluster correlation is only partial.