Deferred from the 2026-08-24 OnnxRLAdapter session: support recurrent (RNN/GRU/LSTM) ONNX policies.
Current adapter is stateless per compute() call. Recurrent policies need a state box mirroring the KalmanFilter state_x_hat mechanism — the tracer's attr-diff state detection already handles this pattern (array attr whose id() changes during the call = recurrent state).
Only worth doing when an actual recurrent policy shows up.
Priority: P4 — deferred until needed.
Deferred from the 2026-08-24 OnnxRLAdapter session: support recurrent (RNN/GRU/LSTM) ONNX policies.
Current adapter is stateless per compute() call. Recurrent policies need a state box mirroring the KalmanFilter state_x_hat mechanism — the tracer's attr-diff state detection already handles this pattern (array attr whose id() changes during the call = recurrent state).
Only worth doing when an actual recurrent policy shows up.
Priority: P4 — deferred until needed.