Add replay time scaling for smoke tuning

src/aituner/spec.py

@@ -152,6 +152,7 @@ class TraceSpec:
     max_concurrency: int
     max_requests_per_probe: int | None = None
     synthetic_prompt_cap_tokens: int | None = None
+    replay_time_scale: float = 1.0
 
     @classmethod
     def from_dict(cls, data: Mapping[str, Any]) -> "TraceSpec":
@@ -172,6 +173,9 @@ class TraceSpec:
             synthetic_prompt_cap_tokens=(
                 int(synthetic_prompt_cap) if synthetic_prompt_cap is not None else None
             ),
+            replay_time_scale=_require_float(
+                data.get("replay_time_scale", 1.0), context="trace.replay_time_scale"
+            ),
         )
 
 

src/aituner/trace.py

@@ -134,6 +134,18 @@ def _downsample_requests(
 
 def load_trace_requests(study: StudySpec, *, study_spec_path: Path) -> tuple[WindowRecord, list[TraceRequest]]:
     window = resolve_window_record(study, study_spec_path=study_spec_path)
+    time_scale = float(study.trace.replay_time_scale)
+    if time_scale <= 0:
+        raise TraceError("trace.replay_time_scale must be > 0")
+    if time_scale != 1.0:
+        window = WindowRecord(
+            window_id=window.window_id,
+            trace_path=window.trace_path,
+            trace_type=window.trace_type,
+            window_start=window.window_start * time_scale,
+            window_end=window.window_end * time_scale,
+            source_payload=dict(window.source_payload),
+        )
     requests: list[TraceRequest] = []
     with window.trace_path.open("r", encoding="utf-8") as handle:
         for idx, raw in enumerate(handle):
@@ -181,7 +193,7 @@ def load_trace_requests(study: StudySpec, *, study_spec_path: Path) -> tuple[Win
             requests.append(
                 TraceRequest(
                     row_id=str(row.get("request_id") or row.get("id") or idx),
-                    arrival_s=float(timestamp),
+                    arrival_s=float(timestamp) * time_scale,
                     sampling_u=float(sampling_u),
                     body=body,
                     prompt_tokens_hint=prompt_tokens_hint,