feat(backend): PolicyRunner ONNX inference wrapper

Backend/PolicyRunner.cs

@@ -0,0 +1,52 @@
+using Microsoft.ML.OnnxRuntime;
+using Microsoft.ML.OnnxRuntime.Tensors;
+
+namespace Backend;
+
+/// <summary>
+/// Loads a PPO policy ONNX model once and provides deterministic action selection.
+/// Thread-safe: <see cref="InferenceSession"/> is safe for concurrent Run() calls.
+/// </summary>
+public sealed class PolicyRunner : IDisposable
+{
+    public const int ObsDim = 7;
+    public const int NActions = 4;
+
+    private readonly InferenceSession _session;
+    private readonly string _inputName;
+
+    public PolicyRunner(string onnxPath)
+    {
+        if (!File.Exists(onnxPath))
+            throw new FileNotFoundException($"ONNX model not found at {onnxPath}");
+        _session = new InferenceSession(onnxPath);
+        _inputName = _session.InputMetadata.Keys.First();
+    }
+
+    /// <summary>Run one inference and return the argmax action index.</summary>
+    public int SelectAction(ReadOnlySpan<float> obs)
+    {
+        if (obs.Length != ObsDim)
+            throw new ArgumentException($"expected obs of length {ObsDim}, got {obs.Length}", nameof(obs));
+
+        var tensor = new DenseTensor<float>(new[] { 1, ObsDim });
+        for (int i = 0; i < ObsDim; i++) tensor[0, i] = obs[i];
+
+        using var results = _session.Run(new[]
+        {
+            NamedOnnxValue.CreateFromTensor(_inputName, tensor)
+        });
+
+        var logits = results.First().AsEnumerable<float>().ToArray();
+        // argmax
+        int best = 0;
+        float bestVal = logits[0];
+        for (int i = 1; i < logits.Length; i++)
+        {
+            if (logits[i] > bestVal) { best = i; bestVal = logits[i]; }
+        }
+        return best;
+    }
+
+    public void Dispose() => _session.Dispose();
+}