Total forge attempts (approved + rejected combined).
Attempts the judge approved.
Attempts the judge or shape validator rejected.
Sum of confidence across approved attempts. Divide by approved for avg.
Count of unique tool names seen this run (union of approved + rejected). A tool rejected then re-forged under the same name counts once.
Count of unique tool names that landed approved at least once this run.
Count of unique tool names that were ONLY rejected (never approved) this run. The retry loop did not recover these. Actionable signal for "real quality failures" vs retry churn.
Histogram of rejection reasons, classified via classifyForgeRejection. Keys match ForgeRejectionCategory.
Per-run forge reliability rollup snapshot.