calibration: persist counterfactual-replay results to audit_events so prompt drift is a trend, not a per-PR comment

[JSONbored]

Context

Parent: #8757. The calibration-advisory CI check posts counterfactual-replay (judge-prompt base-vs-head) results as PR comments only — nothing durable, so there is no prompt-quality trend over time and nothing a future dashboard can aggregate. persistThresholdBacktestRuns (src/services/threshold-backtest-run.ts) is the established pattern: structured result -> recordAuditEvent with a dedicated event type.

Requirements

• Persist each replay run's structured comparison (not rendered Markdown) under a new sibling event type, keyed by repo+PR, from the CI script's existing D1-write path (mirror how the corpus-export script already reads/writes).
• Additive; the PR comment stays.
• Fail-open: persistence failure must never fail the advisory CI job (memory of the advisory-jobs-fail-open rule).

Deliverables

• The persistence write + event type.
• Tests for the persisted shape + the fail-open path.

Test Coverage Requirements

99%+ patch coverage, branch-counted on the persistence logic (CI-script thin-IO exemption applies per repo precedent).

[linear-code]

JSO-6054

[JSONbored]

Filed on a stale audit snapshot — this is already fully implemented: scripts/counterfactual-replay-core.ts builds the audit_events INSERT (event type calibration.counterfactual_backtest_run, structured metadata.comparison + headSha/baseSha/promptVersions/modelSpec/scored), the CLI exposes it behind --persist, and .github/workflows/calibration-advisory.yml already passes --persist --remote --db loopover on both the logic-backtest and replay steps (lines ~133/~263), fail-open exactly as this issue required. The track-record rollup (#8763, PR #8788) already aggregates these rows. Nothing left to build.