Honor exact output lengths in replay requests

src/aituner/trace.py

@@ -59,7 +59,15 @@ def resolve_window_record(study: StudySpec, *, study_spec_path: Path) -> WindowR
             raise TraceError(f"window {study.trace.window_id} does not define trace_file")
         trace_path = Path(trace_file)
         if not trace_path.is_absolute():
-            trace_path = (windows_path.parent / trace_path).resolve()
+            candidate = (windows_path.parent / trace_path).resolve()
+            if candidate.exists():
+                trace_path = candidate
+            else:
+                parts = trace_path.parts
+                if parts and parts[0] == "trace_windows":
+                    trace_path = (windows_path.parent / Path(*parts[1:])).resolve()
+                else:
+                    trace_path = candidate
         return WindowRecord(
             window_id=study.trace.window_id,
             trace_path=trace_path,
@@ -154,6 +162,7 @@ def load_trace_requests(study: StudySpec, *, study_spec_path: Path) -> tuple[Win
             }
             completion_tokens = _coerce_completion_tokens(row)
             if completion_tokens is not None:
+                body["min_tokens"] = completion_tokens
                 body["max_tokens"] = completion_tokens
             temperature = row.get("temperature")
             if isinstance(temperature, (int, float)) and not isinstance(temperature, bool):