Add daily moderation quality statistics with late-event repair

[Szer]

Persist correction-based TP/TN/FP/FN counts for fast dashboard queries, separating actual pipeline totals, counterfactual ML, and matched ML/LLM comparisons by recorded deployment. Keep abstained, unresolved, and excluded cases visible.

The scheduled job atomically replaces today and the previous 14 UTC days, plus older days affected by edits, reviewer corrections, or unbans. An event-insert trigger queues work in the event transaction; resolving pending IDs into durable dirty days and acknowledging those IDs is atomic. This handles out-of-order commits without a high-water event-ID checkpoint, and failed rebuilds retain their dirty markers. All verdict interpretation remains in F#.

Authenticated POST /quality-history runs the daily refresh; ?day=YYYY-MM-DD rebuilds one day. Explicit ?mode=backfill fills older missing days in resumable batches (default seven), separately from the daily job. Includes service-role grants, dashboard SQL with provisional/freshness fields, and structured runtime logging with a warning above the 30-second daily budget.

Validation: Release build and nine PostgreSQL integration tests passed, covering corrected counts, reversal, unban, old edits, today's provisional counts, backfill boundaries, late commits with lower IDs, rollback, failed-rebuild retry, and endpoint authentication. Mandatory repository hooks passed.

Local synthetic benchmark using the actual job: 20,148 messages / 49,006 events across 15 days; daily refresh with 2,729 pending events took 1.15–1.45 seconds, and processing the full 49,006-event backlog took 1.71 seconds. Repeated aggregates matched the evaluator. These are local measurements; production end-to-end runtime remains to be verified after deployment.