Add GET /models/{name}/{version}/metadata endpoint to expose runtime model info

[AK11105]

What problem does this solve?

There is no API to inspect what the engine knows about a deployed model at runtime — its framework, expected input shape, routing strategy, executor, or whether it's currently loaded in the cache. This information exists across multiple internal components but is not surfaced anywhere.

Proposed solution

New endpoint: GET /models/{name}/{version}/metadata

Response:

{
  "name": "sentiment",
  "version": "v1",
  "framework": "sklearn",
  "load_format": "joblib",
  "device": "cpu",
  "input_hint": "raw text string",
  "output_hint": "integer class label",
  "sample_input": "great movie",
  "executor": "cpu",
  "routing_strategy": "static",
  "loaded": true,
  "artifact_size_mb": 2.1
}

Data sources, in priority order:

1. deployment.json in the model directory (written by inference-engine package) — provides framework, load_format, device, input_hint, output_hint, sample_input, artifact_size_mb
2. ExecutionPolicy config — provides executor
3. RoutingService config — provides routing_strategy
4. Registry cache state — provides loaded

Fields absent from deployment.json are returned as null. The endpoint does not fail if deployment.json is absent — it returns what it can from the other sources.

Alternatives considered

Reading deployment.json directly from disk. This doesn't include runtime state (loaded, routing_strategy) and requires filesystem access from the client.

Area

Model loading / registry