Audit eval-dataset count columns against actual row counts before trusting them

In multi-table agent-eval data (runs > tasks > model_calls), cost/token sums can reconcile perfectly at every level while metadata count columns silently lie. Concretely: a per-task call_count disagreed with the real number of call rows for ~56 tasks across 3 runs, and a per-task exception_count had near-zero correlation with the count of calls carrying a non-null exception_type. Always recompute counts via groupby().size() on the child table and diff against the declared column rather than assuming they agree. Also check for entire runs with grades but no telemetry (null cost/tokens, zero child rows) - they vanish from cost analyses and skew naive groupby().mean().