feat: support NVIDIA DGX Spark edge deployments with local and external model routing

Summary

Support NVIDIA DGX Spark edge deployments so vLLM Semantic Router can run efficiently on compact NVIDIA edge/workstation systems and route between local models and external models/providers.

Parent roadmap: #2287 Related: #1138, #2331, #2334, #2358, #2359, #2361 Related code: src/vllm-sr/cli, deploy/local, deploy/kubernetes/dynamo, deploy/kubernetes/llm-d, deploy/kubernetes/routing-strategies, deploy/recipes, src/fleet-sim/fleet_sim/hardware, src/semantic-router/pkg/config, src/semantic-router/pkg/modelruntime, src/semantic-router/pkg/modelinventory

Motivation

DGX Spark-style systems are positioned as compact AI supercomputers for local AI development and edge deployment. They are not the same operational target as large multi-node NVIDIA clusters, but they can run meaningful local models and still need routing, privacy, fallback, tool, and external-provider policies.

NVIDIA's DGX Spark documentation describes a Grace Blackwell architecture system with integrated GPU/CPU, 20-core Arm processor, 128 GB unified memory, compact form factor, and support for large local models: https://docs.nvidia.com/dgx/dgx-spark/hardware.html

Scope

Design and implement support for:

• DGX Spark local/edge deployment profiles;
• local model + external provider routing recipes;
• NVIDIA/Arm64 container/runtime considerations;
• hardware profile entries for fleet-sim/capacity planning;
• model capability and local resource discovery or documented manual configuration;
• lightweight dashboard/storage/observability defaults for edge use;
• fallback behavior when local models are overloaded, unavailable, or lack capability.

Non-goals

• Do not require DGX Spark-specific code paths in generic routing logic if standard NVIDIA/Arm64 profiles are sufficient.
• Do not replace existing Kubernetes/NVIDIA production-stack integrations.
• Do not route local-only/private traffic externally without explicit policy.

Acceptance criteria

• A documented DGX Spark edge deployment profile exists.
• At least one reference recipe routes between local and external models.
• Fleet-sim or docs can represent DGX Spark-class capacity assumptions.
• Packaging/runtime notes cover NVIDIA/Arm64 or DGX Spark-specific constraints.
• Local overload/unavailable fallback behavior is explicit and testable.

Validation

make agent-report ENV=cpu CHANGED_FILES="src/vllm-sr deploy/local deploy/kubernetes deploy/recipes src/fleet-sim src/semantic-router/pkg/config src/semantic-router/pkg/modelruntime src/semantic-router/pkg/modelinventory website/docs"
make agent-lint CHANGED_FILES="src/vllm-sr deploy/local deploy/kubernetes deploy/recipes src/fleet-sim src/semantic-router/pkg/config src/semantic-router/pkg/modelruntime src/semantic-router/pkg/modelinventory website/docs"
make agent-ci-gate CHANGED_FILES="src/vllm-sr deploy/local deploy/kubernetes deploy/recipes src/fleet-sim src/semantic-router/pkg/config src/semantic-router/pkg/modelruntime src/semantic-router/pkg/modelinventory website/docs"