feat(detect): 全量推理可视化，预测框与真值框画在一起

scripts/visualize_detections.py：在整个数据集上推理，绿=真值、蓝=命中的预测、
红=没对上真值的预测，按结果类型分目录（ok / missed / extra / none），再配一份
index.csv（逐图的真值数、预测数、命中数、最高置信度）与 summary.json（逐类 TP/FP/FN）。
--only errors 只输出有问题的图。

匹配逻辑放在 engine/detection.py：按置信度贪心、IoU 阈值、类别需一致，与 mAP 的规则
一致，所以图上看到的漏检误检和指标算的是同一回事。same_class=False 可以单看定位准不准
——定位对但分类错，和完全没找到是两种问题。

train 与 val 分开放：训练集的图模型见过，那部分结果不能当泛化能力看。

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

tests/test_visualize_detections.py

@@ -0,0 +1,89 @@
+"""可视化脚本里不依赖模型的部分。"""
+
+import sys
+from pathlib import Path
+
+import pytest
+
+pytest.importorskip("yaml")
+import yaml  # noqa: E402
+
+sys.path.insert(0, "scripts")
+import visualize_detections as vd  # noqa: E402
+
+from engine.detection import Box, from_yolo, match_boxes  # noqa: E402
+
+
+def _match(predictions, truth):
+    return match_boxes(predictions, truth)
+
+
+@pytest.mark.parametrize(
+    "truth,predictions,expected",
+    [
+        ([], [], "none"),                                                    # 负样本判对
+        ([from_yolo(0, .5, .5, .2, .2)], [], "missed"),                      # 完全没检出
+        ([], [Box(0, .4, .4, .6, .6, .9)], "extra"),                         # 背景上凭空框出
+        ([from_yolo(0, .5, .5, .2, .2)], [Box(0, .41, .41, .59, .59, .9)], "ok"),
+    ],
+)
+def test_outcome_categories(truth, predictions, expected):
+    assert vd.outcome(len(truth), len(predictions), _match(predictions, truth)) == expected
+
+
+def test_missed_takes_priority_over_extra():
+    """一张图同时有漏检和误检时归为 missed——漏检更该先看。"""
+    truth = [from_yolo(0, .2, .2, .1, .1)]
+    predictions = [Box(0, .7, .7, .9, .9, .9)]
+    assert vd.outcome(1, 1, _match(predictions, truth)) == "missed"
+
+
+def test_label_path_swaps_images_for_labels():
+    image = Path("/data/film/images/val/blackdot_0001.png")
+    assert vd.label_path_for(image) == Path("/data/film/labels/val/blackdot_0001.txt")
+
+
+def test_label_path_uses_the_last_images_segment():
+    """路径里多处出现 images 时，只换最靠后的那一段。"""
+    image = Path("/srv/images/film/images/train/a.png")
+    assert vd.label_path_for(image) == Path("/srv/images/film/labels/train/a.txt")
+
+
+@pytest.fixture
+def dataset(tmp_path):
+    for split in ("train", "val"):
+        (tmp_path / "images" / split).mkdir(parents=True)
+        (tmp_path / "images" / split / f"{split}_0001.png").write_bytes(b"")
+    data_yaml = tmp_path / "data.yaml"
+    data_yaml.write_text(
+        yaml.safe_dump({
+            "path": str(tmp_path), "train": "images/train", "val": "images/val",
+            "nc": 2, "names": {0: "blackdot", 1: "grease"},
+        }),
+        encoding="utf-8",
+    )
+    return data_yaml
+
+
+def test_load_dataset(dataset):
+    classes, splits = vd.load_dataset(dataset)
+    assert classes == ["blackdot", "grease"]
+    assert sorted(splits) == ["train", "val"]
+    assert splits["val"][0].name == "val_0001.png"
+
+
+def test_load_dataset_skips_a_split_pointing_at_the_same_directory(dataset):
+    """未划分的数据集 train 与 val 指向同一目录，不该把同一批图跑两遍。"""
+    meta = yaml.safe_load(dataset.read_text(encoding="utf-8"))
+    meta["val"] = meta["train"]
+    dataset.write_text(yaml.safe_dump(meta), encoding="utf-8")
+    _, splits = vd.load_dataset(dataset)
+    assert list(splits) == ["train"]
+
+
+def test_load_dataset_reports_a_missing_directory(dataset):
+    meta = yaml.safe_load(dataset.read_text(encoding="utf-8"))
+    meta["val"] = "images/nope"
+    dataset.write_text(yaml.safe_dump(meta), encoding="utf-8")
+    with pytest.raises(FileNotFoundError, match="不存在"):
+        vd.load_dataset(dataset)