Audit eval datasets by cross-joining FK integrity before analysis

When profiling a multi-file eval corpus (tasks/traces/rubrics/judge-labels), do not stop at per-file key inventories — the defects live in the joins and in fields that contradict each other. Concretely: count unique primary keys vs row counts to find verbatim duplicate rows; re-derive each trace error flag yourself from the tool registry rather than trusting the recorded result_status/error field (here 22 calls were missing a required arg yet recorded as ok, while the JSON-string args field was truncated mid-literal on 76 others); and check whether a difficulty/label column is fully determined by another column (difficulty was a perfect proxy for gold-sequence length, so the two are confounded). Also compute per-group row counts for any model/grader dimension before believing calibration numbers — a 45-run human subset had 24/20/1 across three models. A single Python pass with collections.Counter over tuple(sorted(d.keys())) per record cheaply reveals variant record schemas (e.g. tool_call steps carrying 7 keys vs final steps carrying 3).