Add adaptive TP/FP/FN validation mosaic export (#2271)

- remove top-level cv2/annotator imports; lazy-load inside rendering functions
- remove save_result_images bool; save_directory_path is now keyword-only after metric_target
- drop hardcoded result/ subdirectory from benchmark output path
- propagate metric_target into _split_detections_by_outcome for correct OBB IoU dispatch
- add filename collision UserWarning in benchmark loop
- remove dead/unreachable combined None-check in _split_detections_by_outcome
- add Google-style docstrings to all 5 new private visualization functions
- add TestSplitDetectionsByOutcome covering 7 edge cases (empty inputs, cross-class, confidence-None)
- fix FP/FN pixel assertions to check interior box pixels rather than border/title regions
- fix benchmark_a_model.md: full panel names, add Visual Benchmarking section, update API examples

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: jirka <6035284+Borda@users.noreply.github.com>
Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com>

Author: 4 people

docs/how_to/benchmark_a_model.md

@@ -329,6 +329,20 @@ Here, predictions in purple are targets (ground truth), and predictions in teal
 
     See [annotator documentation](https://supervision.roboflow.com/latest/detection/annotators/) for even more options.
 
+## Visual Benchmarking
+
+To inspect where a model succeeds and fails, pass `save_directory_path` to `sv.ConfusionMatrix.benchmark(...)`. For every dataset image it writes a 2x2 result grid — `Ground Truth`, `True Positives`, `False Positives`, and `False Negatives` panels — directly into that directory, reusing the original image filenames. This makes it easy to skim through per-image outcomes alongside the aggregate confusion matrix.
+
+```python
+import supervision as sv
+
+confusion_matrix = sv.ConfusionMatrix.benchmark(
+    dataset=test_set,
+    callback=callback,
+    save_directory_path="./results",
+)
+```
+
 ## Benchmarking Metrics
 
 With multiple models, fine details matter. Visual inspection may not be enough. `supervision` provides a collection of metrics that help obtain precise numerical results of model performance.
@@ -462,7 +476,7 @@ Yes, if you want to evaluate their bounding boxes. Convert model outputs to `Det
 
 ### What is a ConfusionMatrix and how do I use it?
 
-`sv.ConfusionMatrix` visualizes true positives, false positives, and false negatives per class. Create one with `sv.ConfusionMatrix.from_detections(predictions=predictions, targets=targets, classes=classes, conf_threshold=0.5, iou_threshold=0.5)`, then call `metric.plot()` to render a heatmap.
+`sv.ConfusionMatrix` visualizes true positives, false positives, and false negatives per class. Create one with `sv.ConfusionMatrix.from_detections(predictions=predictions, targets=targets, classes=classes, conf_threshold=0.5, iou_threshold=0.5)`, then call `confusion_matrix.plot()` to render a heatmap. If you want per-image validation visualizations saved to disk, pass `save_directory_path="./results"` to `sv.ConfusionMatrix.benchmark(...)`; it will write 2x2 result grids directly into that directory using the original image filenames, with `Ground Truth`, `True Positives`, `False Positives`, and `False Negatives` panels.
 
 ## Author