Isolate Trackio from Hub model acquisition

training/run_fable_router_structural_smoke.py

@@ -435,13 +435,18 @@ def main() -> int:
     try:
         import trackio
 
+        # Structural-smoke evidence is persisted directly to the private Hub
+        # dataset.  Keep Trackio local here: Trackio's Space mode configures a
+        # Hub Bucket/Xet write context, which can leak into a subsequent large
+        # model download in the same process.  Full training uses a dedicated
+        # remote Trackio process after acquisition instead.
+        os.environ["TRACKIO_DIR"] = str(output / "trackio")
         trackio.init(
             project="autonoma-fable-router-smoke",
             name=f"{args.bank}-{stamp}",
-            space_id="Swordnael/trackio",
             config={"bank": args.bank, "platform": args.platform, "mode": "structural-smoke"},
         )
-        tracking = {"status": "started", "space": "Swordnael/trackio"}
+        tracking = {"status": "started_local", "directory": str(output / "trackio")}
         run(args, result)
         result["status"] = "structural_smoke_passed_nonrouting"
         result["passed"] = True
@@ -462,7 +467,7 @@ def main() -> int:
         result_path = output / "result.json"
         result_path.write_text(json.dumps(result, indent=2) + "\n", encoding="utf-8")
         try:
-            if tracking["status"] == "started":
+            if tracking["status"] == "started_local":
                 import trackio
 
                 trackio.finish()