feat: support AMD AI PC edge deployments with local and external model routing

Summary

Define and validate a supported AMD AI PC edge profile for running vLLM Semantic Router and routing between explicitly configured local and external logical models.

Parent roadmap: #2287
Related: #1138, #2294, #2331, #2334, #2358, #2372

Audited current-main baseline (2026-07-15)

• The repository's supported local flows are make vllm-sr-dev plus vllm-sr serve --image-pull-policy never, with VLLM_SR_PLATFORM=amd/--platform amd for the AMD image path.
• deploy/amd/, deploy/local/, recipes, provider models, and model inventory exist, but they do not establish tested Ryzen AI/NPU support.
• Current CLI implementation uses container_run_command.py and container_images.py. The previously cited docker_run_command.py and docker_images.py paths do not exist.
• AMD's official Ryzen AI Max+ 395 product page documents CPU/iGPU/NPU platform characteristics, but hardware specifications alone do not prove compatibility with the router's Candle/ONNX/container runtimes.
• Current main does not provide a first-class upstream-error ordered fallback contract; #2294 tracks model failover.

Scope

1. Define the supported hardware/software matrix separately for CPU, Radeon iGPU/ROCm where available, and Ryzen AI NPU. Unsupported accelerators must be explicit.
2. Reuse the canonical local image/serve flow; add a named edge profile only through existing CLI/container configuration seams.
3. Provide a maintained recipe with local and external provider models, explicit local-only/privacy decisions, and ordinary policy-based escalation.
4. Define memory/model-size, architecture, quantization, runtime, driver, OS, and container constraints for each tested path.
5. Add capability discovery where reliable; otherwise require validated manual configuration and surface “unknown/unsupported” honestly.
6. Use lightweight, bounded defaults for single-node storage, observability, and dashboard services.
7. Capture reproducible validation receipts with device, OS/kernel, driver/runtime, image digest, model artifacts, config, and test results.
8. Coordinate true upstream-error failover/fallback behavior with #2294 instead of implying it already exists.

Non-goals

• Do not claim NPU, iGPU, or unified-memory acceleration without an exercised runtime path.
• Do not assume data-center ROCm settings apply unchanged to AI PCs.
• Do not require every device to support the same model size or feature set.
• Do not send local-only/private traffic to external providers.
• Do not invent a second local serve path.

Acceptance criteria

• A support matrix distinguishes tested, experimental, and unsupported CPU/iGPU/NPU paths.
• The canonical local image/serve flow works for at least one documented AI PC configuration.
• One maintained recipe demonstrates local/external policy without violating local-only constraints.
• Resource limits, failure behavior, and unsupported capability diagnostics are explicit.
• Validation receipts are reproducible and do not contain secrets.
• Privacy decisions are covered by probes/eval/E2E tests.
• Any upstream-error model fallback depends on the contract in #2294.

Likely change surfaces

src/vllm-sr/cli/container_run_command.py, container_images.py, service_defaults.py, deploy/amd/, deploy/local/, deploy/recipes/, src/semantic-router/pkg/config/, modelruntime/, modelinventory/, docs, and tests.

Validation entrypoint

make agent-report ENV=amd CHANGED_FILES="<space-separated changed files>"

Follow the reported AMD/local-image and affected E2E gates.