fix: Stop the PyTorch backend from using pure greedy decoding

Live-tested against the deployed bf16 model: greedy decoding (do_sample=
False) reproducibly derived the wrong closed-loop coefficients for a plain
zeta=0.6/wn=1.4 step response -- 2.56/2.1952 instead of the correct
1.68/1.96, identically in 3/3 runs, plus mislabeled the result "critically
damped" (that's zeta=1, not 0.6). Switched to temperature=0.2 sampling,
matching what the GGUF and MLX backends already use.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

controlai_agent/orchestrator.py

@@ -735,8 +735,14 @@ class ControlAIAgent:
                 outputs = self.model.generate(
                     **inputs,
                     max_new_tokens=max_tokens,
-                    do_sample=False,
-                    temperature=None,
+                    # Pure greedy (do_sample=False) reproducibly derived the
+                    # WRONG closed-loop coefficients for a plain zeta/wn step
+                    # response on this exact model (2.56/2.1952 instead of the
+                    # correct 1.68/1.96, in 3/3 identical runs on deployed
+                    # bf16) -- low-temperature sampling, matching the GGUF/MLX
+                    # backends, is the fix being tested for that failure.
+                    do_sample=True,
+                    temperature=0.2,
                     top_p=None,
                     top_k=None,
                     repetition_penalty=REPETITION_PENALTY,