Implement durable human approval workflows

Why it matters

Approval can take hours, processes restart, and decisions arrive twice, so a pending flag in memory is not an approval gate.

What to practise

Implement a Python approval workflow that maps raw confidence to observed accuracy before choosing review, keeps workflow state and append-only audit events in Azure Cosmos DB, and carries review requests and decisions on Azure Service Bus queues. Resume a decision exactly once with optimistic concurrency, return duplicate decisions without acting, and escalate reviews whose deadline has expired. Demonstrate approved, rejected, overridden and escalated records, and confirm that nothing executes from a pending or rejected state. The exercise ends by mapping its design to the Durable Functions human-interaction pattern.

How it fits

This is 2.4.6 built by hand: persisted gates, bounded waits, and authorized overrides. Compare it with a framework-managed checkpoint that resumes from a recorded request, and note which replay and race handling the application now owns. The escalation thresholds apply 1.1.6.