feat: Auto-purge old run records with retainRuns option

[coji]

Problem

durably_runs records accumulate indefinitely. While individual records are small, over time (e.g., hourly cron jobs running for months) the table grows unbounded. Currently, the only way to remove old runs is manually calling deleteRun().

This is separate from step data cleanup (#87, cleanupSteps option) — step output is the large data and is now cleaned up automatically. Run records are lightweight but should still have an optional retention policy.

Proposal

const durably = createDurably({
  dialect,
  // Auto-delete terminal runs older than 30 days
  retainRuns: '30d',
})

• Only deletes runs in terminal states (completed, failed, cancelled)
• Runs in pending or running are never auto-deleted
• Cleanup runs periodically (e.g., once per polling cycle or on a separate interval)
• Default: no retention (keep all runs, backward compatible)

Alternatives

• Explicit purge API: await durably.purgeRuns({ olderThan: '30d', status: ['completed'] }) — useful as a manual escape hatch regardless of whether auto-purge is implemented
• Both: retainRuns for automatic + purgeRuns() for manual control

Context

• Spun off from feat: Automatic step data cleanup for completed runs #87 (step data cleanup)
• Run records are small (~200-500 bytes each) so this is lower priority than step cleanup
• Mainly relevant for long-running production deployments with frequent job triggers

[coji]

現状整理

• step データの自動クリーンアップは cleanupSteps オプションで対応済み（feat: Automatic step data cleanup for completed runs #87）
• run レコード自体は軽量（200-500 bytes/件）なので、現時点で急ぎではない
• deleteRun() で手動削除は可能

今後の方針

実際に長期運用して run の蓄積が問題になったタイミングで着手する。

実装順としては：

1. purgeRuns() API を先に入れる — 手動でまとめて削除できる escape hatch。小さくて安全
2. retainRuns オプション — 自動パージ。purgeRuns() の上に薄くかぶせる形で実装

まずは実プロジェクトでの運用データ（どれくらいのペースで溜まるか、いつ問題になるか）を見てから判断します。