fix(agent): back off failed VM reconcile sweeps

[branarakic]

Summary

• let failed VM reconcile passes reach the per-CG coalescer after they are logged
• suppress queued/live-nudge retries for one configured sweep interval after a failure
• retain immediate coalescing and trailing runs for successful reconciliation
• add deterministic coverage for failure cooldown and retry at the boundary

Mainnet evidence

10.0.6 cores repeatedly report 30-second Blazegraph deadlines while reconciling the same on-chain graphs (foodie-network on Gnosis and sports on Base). Several nodes produced roughly 15 VM timeout warnings in a 15-minute sample.

runVmReconcileForCg currently catches the timeout and resolves normally. The coalescer therefore cannot distinguish failure from success: a KACG event queued during the expensive scan immediately launches the trailing scan, and later live events can keep retriggering it despite the periodic safety net.

After this change, a failed CG is cooled down for DKG_VM_RECONCILE_INTERVAL_MS (10 minutes on the inspected mainnet units). The periodic sweep retries when that window expires. Successful live-event reconciliation is unchanged and remains immediate.

This is orchestration-only: it does not delete subscriptions or graph data and does not advance a failed reconciliation cursor.

Verification

• pnpm --filter @origintrail-official/dkg-agent exec vitest run --config vitest.unit.config.ts test/chain-reconciler.test.ts
• pnpm --filter @origintrail-official/dkg-agent exec tsc --noEmit
• pnpm --filter @origintrail-official/dkg-agent test:unit (52 files, 593 tests)