APIM AI Gateway: Wire intelligent-routing fragment for multi-backend failover

[mishraomp]

Summary

The intelligent-routing.xml policy fragment is registered in APIM (stacks/apim/main.tf) and documented in params/apim/README.md as "Multi-backend routing (future)", but is never included in any policy template.

Gap

Currently each tenant has exactly one backend per service type - no failover. If the shared AI Foundry Hub throttles or becomes unavailable, there is no fallback route. The intelligent-routing fragment already implements priority-based backend selection with throttle awareness and random load balancing within priority tiers.

Proposed Implementation

1. Introduce a routes configuration in tenant tfvars to define multiple backends with priority/weight for OpenAI (and optionally other services)
2. Include the intelligent-routing fragment in the per-tenant OpenAI path routing when multiple routes are configured
3. Create corresponding backend pool resources in Terraform for load-balanced backends
4. Update integration tests to verify failover behavior

When to Implement

This becomes critical when:

• Multi-region deployment is planned
• PTU + consumption spillover is needed
• High-availability SLA requirements increase

Severity

MEDIUM - Foundation exists but isn't wired. Value increases significantly with multi-region or PTU+consumption spillover plans.

Context

Identified during APIM multi-tenancy and AI gateway policy gap analysis (Feb 2026).