warm up windows cache for fair benchmark

tests/benchmarks/bench_candidate_graph.py

@@ -43,6 +43,22 @@ def seg_data():
     return _generate_segmentation()
 
 
+@pytest.fixture(scope="module")
+def _warm_spatial_compile():
+    """Warm the spatial_graph rtree JIT compile before the timed benchmark.
+
+    Building a SolutionTracks with segmentation creates a GraphArrayView, which builds a
+    spatial_graph rtree whose Cython module is JIT-compiled by witty on first use and
+    cached process-wide. That compile is a one-time cost (seconds-to-tens-of-seconds on
+    a cold Windows runner) that would otherwise land inside the timed region of
+    test_graph_to_solution. Trigger it here (untimed) on a tiny graph with the same
+    dimensionality so the benchmark measures tracking work, not compilation.
+    """
+    tiny = _generate_segmentation(num_frames=2, frame_shape=(64, 64), cells_per_frame=2)
+    graph = compute_graph_from_seg(tiny, MAX_EDGE_DISTANCE, iou=True)
+    SolutionTracks(graph, time_attr="t", pos_attr="pos")
+
+
 def test_compute_graph_from_seg(benchmark, seg_data):
     benchmark.pedantic(
         compute_graph_from_seg,
@@ -53,12 +69,14 @@ def test_compute_graph_from_seg(benchmark, seg_data):
     )
 
 
-def test_graph_to_solution(benchmark, seg_data):
+def test_graph_to_solution(benchmark, seg_data, _warm_spatial_compile):
     """Benchmark candidate graph -> SolutionTracks (tracklet/lineage assignment).
 
     Candidate-graph construction is benchmarked separately above and is built here in
     (untimed) setup, so only the SolutionTracks construction -- dominated by
     TrackAnnotator._assign_tracklet_ids and _assign_lineage_ids -- is measured.
+    The _warm_spatial_compile fixture ensures the one-time spatial_graph rtree JIT
+    compile is not measured here.
     """
 
     def setup():