bug(cdk): AgentCore runtime leaks ENIs that block cdk destroy; Memory undeletable while CREATING

### Observed (live, us-east-1, 2026-07-31, ADR-021 P1 verification run)

1. **`cdk destroy` of the agent stack hung >1h40m and ended `DELETE_FAILED`**: two service-managed `agentic_ai` ENIs (AgentCore runtime) remained attached to the platform VPC's private subnets after the runtime resource was deleted, blocking deletion of both private subnets and the runtime SG. The stack had to be left in `DELETE_FAILED` (zero-cost residue: VPC shell, 2 subnets, 1 SG) pending eventual ENI release and a delete retry.
2. **`AWS::BedrockAgentCore::Memory` cannot be deleted while in `CREATING`** — a teardown racing a slow Memory creation wedges.

### Impact

Any operator tearing down a deployment (ephemeral stacks, #72's scheduled cleanup) hits multi-hour destroy times or `DELETE_FAILED` states through no fault of their own.

### Proposal

- Investigate whether AgentCore ENI release can be awaited or forced (custom resource that waits/detaches on stack deletion, similar to VPC-lambda ENI cleanup patterns)
- Guard Memory deletion on reaching a stable state
- At minimum: document the expected destroy latency + retry procedure in the developer guide

Not specific to the Lambda MicroVMs backend — observed on the AgentCore substrate during unrelated verification. Refs #645.

Beginne damit, das AgentCore-Teardown-Verhalten in us-east-1 zu reproduzieren, und lies ADR-021 sowie den zugehörigen Kontext in #645. Vergleiche die vorgeschlagene ENI-Bereinigung und die Löschsperre für Memory mit dem dokumentierten Minimalfix; abgeschlossen bedeutet, dass der Destroy-Vorgang ohne DELETE_FAILED abgeschlossen wird oder der Entwicklerleitfaden die erwartete Latenz und die Schritte für Wiederholungsversuche eindeutig dokumentiert.

Vom Indexierungsmodell aus dem Issue-Text verfasst.