diff --git a/FishMeasure/fish_video_weight_evaluation.py b/FishMeasure/fish_video_weight_evaluation.py
index 435a78a..811a0bb 100644
--- a/FishMeasure/fish_video_weight_evaluation.py
+++ b/FishMeasure/fish_video_weight_evaluation.py
@@ -10,6 +10,24 @@ import json
 import numpy as np
 import torch
 from pathlib import Path
+
+if not hasattr(argparse, "BooleanOptionalAction"):
+    class _BooleanOptionalAction(argparse.Action):
+        def __init__(self, option_strings, dest, default=None, type=None,
+                     choices=None, required=False, help=None, metavar=None):
+            _option_strings = []
+            for opt in option_strings:
+                _option_strings.append(opt)
+                if opt.startswith("--"):
+                    _option_strings.append("--no-" + opt[2:])
+            super().__init__(option_strings=_option_strings, dest=dest, nargs=0,
+                             default=default, type=type, choices=choices,
+                             required=required, help=help, metavar=metavar)
+
+        def __call__(self, parser, namespace, values, option_string=None):
+            setattr(namespace, self.dest, not option_string.startswith("--no-"))
+
+    argparse.BooleanOptionalAction = _BooleanOptionalAction
 from typing import List, Dict, Any, Optional, Tuple
 from ultralytics import YOLO
 from seg import init_models
@@ -20,6 +38,8 @@ import importlib
 import tempfile
 import subprocess
 
+_REPO_ROOT = Path(__file__).resolve().parent
+
 try:
     import pyzed.sl as sl
     ZED_AVAILABLE = True
@@ -936,7 +956,7 @@ def classify_pointcloud_array(classifier: torch.nn.Module, points: np.ndarray, c
         possible_paths = [
             Path(__file__).parent / "pointcloud_classifier" / "Pointnet_Pointnet2_pytorch" / "test_classification.py",
             Path(__file__).parent.parent / "pointcloud_classifier" / "Pointnet_Pointnet2_pytorch" / "test_classification.py",
-            Path("/home/ubuntu/projects/FishMeasure/pointcloud_classifier/Pointnet_Pointnet2_pytorch/test_classification.py"),
+            _REPO_ROOT / "pointcloud_classifier" / "Pointnet_Pointnet2_pytorch" / "test_classification.py",
         ]
         
         test_classification_path = None
@@ -2132,7 +2152,7 @@ def main():
         help="With --batch-svo-folder: find *.svo2 in all subfolders (output dirs use path like parent__child__file)",
     )
     parser.add_argument("--yolo-model", 
-                       default="/home/ubuntu/projects/FishMeasure/runs/train/fish_detection_20251127_104658/weights/best.pt",
+                       default=str(_REPO_ROOT / "runs/train/fish_detection_20251127_104658/weights/best.pt"),
                        help="YOLO model path")
     parser.add_argument("--conf", type=float, default=0.25, help="Confidence threshold")
     parser.add_argument("--imgsz", type=int, default=640, help="Image size")
@@ -2176,7 +2196,7 @@ def main():
     parser.add_argument("--run-weight-estimation", action="store_true",
                        help="After processing, run DGCNN weight estimation on saved point clouds (test_dgcnn_weight_estimator)")
     parser.add_argument("--weight-estimator-checkpoint", type=str, 
-                       default="/home/ubuntu/projects/FishMeasure/weight_estimator/runs/dgcnn_20260312_171043/best.pt",
+                       default=str(_REPO_ROOT / "weight_estimator/runs/dgcnn_20260312_171043/best.pt"),
                        help="Path to DGCNN weight estimator checkpoint (.pt)")
     parser.add_argument("--weight-topk-length", type=int, default=3,
                        help="Optional: length-weighted top-K for predict_cloud_folder (default: 3; set 0 to disable)")
diff --git a/FishMeasure/generate_video_with_labels.py b/FishMeasure/generate_video_with_labels.py
index 539a02c..4b326fa 100644
--- a/FishMeasure/generate_video_with_labels.py
+++ b/FishMeasure/generate_video_with_labels.py
@@ -167,22 +167,41 @@ def generate_video(
     imgsz: int = 640,
     frame_stride: int = 1,
     show_large: bool = False,
+    summary_weight_g: Optional[float] = None,
+    summary_length_mm: Optional[float] = None,
+    summary_star: bool = False,
+    output_video_name: Optional[str] = None,
+    sam_device: str = "cuda",
 ) -> Optional[Path]:
     if not ZED_AVAILABLE:
         print("ERROR: pyzed not available, cannot generate labeled video")
         return None
 
-    per_frame, summary_wg, summary_lmm, is_confident = _parse_weight_json(weight_json)
-    star_s = " *" if is_confident else ""
+    per_frame, parsed_summary_wg, parsed_summary_lmm, raw_confident = _parse_weight_json(weight_json)
+    if summary_weight_g is None:
+        summary_weight_g = parsed_summary_wg
+    if summary_length_mm is None:
+        summary_length_mm = parsed_summary_lmm
+    star_s = " *" if summary_star else ""
     print(f"  Per-frame predictions: {len(per_frame)} PLYs mapped")
-    print(f"  Summary: weight={summary_wg}g, length={summary_lmm}mm{star_s}")
+    print(
+        f"  Summary: weight={summary_weight_g}g, length={summary_length_mm}mm{star_s} "
+        f"(raw_confident={raw_confident})"
+    )
 
-    if not per_frame and summary_wg is None:
+    if not per_frame and summary_weight_g is None:
         print("  WARNING: No weight data in JSON, video will show '--'")
 
     from ultralytics import YOLO
     yolo = YOLO(yolo_model_path)
     class_names = yolo.names if hasattr(yolo, "names") else {}
+    from fish_video_weight_evaluation import (
+        create_segmentation_overlay,
+        load_sam_predictor_with_fallback,
+        segment_with_sam,
+    )
+    sam_predictor, eff_sam_device = load_sam_predictor_with_fallback(sam_device)
+    sam_torch_device = eff_sam_device
 
     from dataset.zed_reader import ZEDReader
     zed_reader = ZEDReader(svo_path=str(svo_path), camera_mode=False, use_yolo_detector=False)
@@ -218,6 +237,7 @@ def generate_video(
                 continue
 
             frame_number = idx + 1
+            frame_name = f"frame_{frame_number:06d}"
             if frame_number in per_frame:
                 cur_wg, cur_lmm = per_frame[frame_number]
                 last_wg = cur_wg
@@ -230,6 +250,7 @@ def generate_video(
             num_dets = len(results.boxes) if results.boxes is not None else 0
 
             left_disp = img.copy()
+            right_disp = img.copy()
             if num_dets > 0:
                 boxes = results.boxes.xyxy.cpu().numpy()
                 tids = (results.boxes.id.cpu().numpy().astype(int)
@@ -245,19 +266,31 @@ def generate_video(
                     cname = class_names.get(cid, "fish")
                     _draw_label_on_box(left_disp, box, tid, cname, cur_wg, cur_lmm)
 
-            if show_large or summary_wg is not None:
-                _draw_large_summary(left_disp, summary_wg, summary_lmm, is_confident)
+                try:
+                    masks = segment_with_sam(sam_predictor, img, boxes, sam_torch_device)
+                except Exception as e:
+                    print(f"  WARNING: SAM segmentation failed on {frame_name}: {e}")
+                    masks = []
+
+                if masks:
+                    right_disp = create_segmentation_overlay(img.copy(), masks)
+                    cv2.putText(right_disp, "Segmentation", (10, right_disp.shape[0] - 20),
+                                cv2.FONT_HERSHEY_SIMPLEX, 0.7, (255, 255, 255), 2, cv2.LINE_AA)
+                else:
+                    cv2.putText(right_disp, "Segmentation (failed)", (10, right_disp.shape[0] - 20),
+                                cv2.FONT_HERSHEY_SIMPLEX, 0.7, (0, 0, 255), 2, cv2.LINE_AA)
+            else:
+                cv2.putText(right_disp, "No detections", (10, right_disp.shape[0] - 20),
+                            cv2.FONT_HERSHEY_SIMPLEX, 0.7, (128, 128, 128), 2, cv2.LINE_AA)
+
+            if show_large or summary_weight_g is not None:
+                _draw_large_summary(left_disp, summary_weight_g, summary_length_mm, summary_star)
 
-            frame_name = f"frame_{frame_number:06d}"
             info = f"[{frame_number}] {frame_name} | Detections: {num_dets}"
             cv2.putText(left_disp, info, (10, 30), cv2.FONT_HERSHEY_SIMPLEX, 0.7, (0, 255, 0), 2, cv2.LINE_AA)
             cv2.putText(left_disp, "Detection", (10, left_disp.shape[0] - 20),
                         cv2.FONT_HERSHEY_SIMPLEX, 0.7, (255, 255, 255), 2, cv2.LINE_AA)
 
-            right_disp = img.copy()
-            cv2.putText(right_disp, "Original", (10, right_disp.shape[0] - 20),
-                        cv2.FONT_HERSHEY_SIMPLEX, 0.7, (255, 255, 255), 2, cv2.LINE_AA)
-
             combined = np.hstack([left_disp, right_disp])
             if num_dets > 0:
                 frames.append(combined)
@@ -275,7 +308,7 @@ def generate_video(
         print(f"  WARNING: No detection frames collected from {svo_name}")
         return None
 
-    video_path = images_dir / f"{svo_name}_preview.mp4"
+    video_path = images_dir / (output_video_name or f"{svo_name}_preview.mp4")
     h, w = frames[0].shape[:2]
     writer = cv2.VideoWriter(str(video_path), cv2.VideoWriter_fourcc(*"mp4v"), 10.0, (w, h))
     for f in frames:
@@ -295,7 +328,17 @@ def main():
     parser.add_argument("--conf", type=float, default=0.25)
     parser.add_argument("--imgsz", type=int, default=640)
     parser.add_argument("--frame-stride", type=int, default=1)
+    parser.add_argument("--sam-device", type=str, default="cuda")
     parser.add_argument("--show-large-labels-at-top-right", action="store_true")
+    parser.add_argument(
+        "--summary-star",
+        action=argparse.BooleanOptionalAction,
+        default=False,
+        help="Whether to draw * on the Final summary line; caller/DB is the source of truth.",
+    )
+    parser.add_argument("--summary-weight-g", type=float, default=None)
+    parser.add_argument("--summary-length-mm", type=float, default=None)
+    parser.add_argument("--output-video-name", type=str, default=None)
     args = parser.parse_args()
 
     svo = Path(args.svo).expanduser().resolve()
@@ -315,7 +358,12 @@ def main():
         conf=args.conf,
         imgsz=args.imgsz,
         frame_stride=args.frame_stride,
+        sam_device=args.sam_device,
         show_large=args.show_large_labels_at_top_right,
+        summary_weight_g=args.summary_weight_g,
+        summary_length_mm=args.summary_length_mm,
+        summary_star=bool(args.summary_star),
+        output_video_name=args.output_video_name,
     )
 
 
diff --git a/FishMeasure/predict_weigth_from_svo2.py b/FishMeasure/predict_weigth_from_svo2.py
index a039cb8..48a178b 100644
--- a/FishMeasure/predict_weigth_from_svo2.py
+++ b/FishMeasure/predict_weigth_from_svo2.py
@@ -25,6 +25,24 @@ import subprocess
 import sys
 import math
 from pathlib import Path
+
+if not hasattr(argparse, "BooleanOptionalAction"):
+    class _BooleanOptionalAction(argparse.Action):
+        def __init__(self, option_strings, dest, default=None, type=None,
+                     choices=None, required=False, help=None, metavar=None):
+            _option_strings = []
+            for opt in option_strings:
+                _option_strings.append(opt)
+                if opt.startswith("--"):
+                    _option_strings.append("--no-" + opt[2:])
+            super().__init__(option_strings=_option_strings, dest=dest, nargs=0,
+                             default=default, type=type, choices=choices,
+                             required=required, help=help, metavar=metavar)
+
+        def __call__(self, parser, namespace, values, option_string=None):
+            setattr(namespace, self.dest, not option_string.startswith("--no-"))
+
+    argparse.BooleanOptionalAction = _BooleanOptionalAction
 from typing import Any, Dict, List, Optional, Tuple
 
 import torch
@@ -136,11 +154,14 @@ def _run_generate_video_with_labels_subprocess(
         "--conf", str(args.conf),
         "--imgsz", str(args.imgsz),
         "--frame-stride", str(args.frame_stride),
+        "--sam-device", str(args.sam_device),
         "--weight-json", str(weight_json.expanduser().resolve()),
     ]
 
     if getattr(args, "show_large_labels_at_top_right", False):
         cmd.append("--show-large-labels-at-top-right")
+    if getattr(args, "summary_star", False):
+        cmd.append("--summary-star")
 
     print(f"Invoking generate_video_with_labels.py:\n  {' '.join(cmd)}")
     proc = subprocess.run(cmd, cwd=str(REPO_ROOT))
@@ -505,7 +526,7 @@ def main() -> None:
     parser.add_argument(
         "--yolo-model",
         type=str,
-        default="/home/ubuntu/projects/FishMeasure/runs/train/fish_detection_20251127_104658/weights/best.pt",
+        default=str(REPO_ROOT / "runs/train/fish_detection_20251127_104658/weights/best.pt"),
     )
     parser.add_argument(
         "--conf",
@@ -622,6 +643,12 @@ def main() -> None:
         action="store_true",
         help="Show large weight/length labels (10x font) at top right corner for real/camera generated videos.",
     )
+    parser.add_argument(
+        "--summary-star",
+        action=argparse.BooleanOptionalAction,
+        default=False,
+        help="Pass to generate_video_with_labels: whether the summary line should draw *.",
+    )
 
     args = parser.parse_args()
     if args.frame_stride < 1:
diff --git a/FishMeasure/runs/train/fish_detection_20251127_104658/F1_curve.png b/FishMeasure/runs/train/fish_detection_20251127_104658/F1_curve.png
deleted file mode 100644
index 4078b2f..0000000
Binary files a/FishMeasure/runs/train/fish_detection_20251127_104658/F1_curve.png and /dev/null differ
diff --git a/FishMeasure/runs/train/fish_detection_20251127_104658/PR_curve.png b/FishMeasure/runs/train/fish_detection_20251127_104658/PR_curve.png
deleted file mode 100644
index 608e97f..0000000
Binary files a/FishMeasure/runs/train/fish_detection_20251127_104658/PR_curve.png and /dev/null differ
diff --git a/FishMeasure/runs/train/fish_detection_20251127_104658/P_curve.png b/FishMeasure/runs/train/fish_detection_20251127_104658/P_curve.png
deleted file mode 100644
index bd1a3d4..0000000
Binary files a/FishMeasure/runs/train/fish_detection_20251127_104658/P_curve.png and /dev/null differ
diff --git a/FishMeasure/runs/train/fish_detection_20251127_104658/R_curve.png b/FishMeasure/runs/train/fish_detection_20251127_104658/R_curve.png
deleted file mode 100644
index 54c5f50..0000000
Binary files a/FishMeasure/runs/train/fish_detection_20251127_104658/R_curve.png and /dev/null differ
diff --git a/FishMeasure/runs/train/fish_detection_20251127_104658/args.yaml b/FishMeasure/runs/train/fish_detection_20251127_104658/args.yaml
deleted file mode 100644
index e00bcbf..0000000
--- a/FishMeasure/runs/train/fish_detection_20251127_104658/args.yaml
+++ /dev/null
@@ -1,106 +0,0 @@
-task: detect
-mode: train
-model: yolov8n.pt
-data: ./yolo_dataset/dataset.yaml
-epochs: 100
-time: null
-patience: 50
-batch: 96
-imgsz: 640
-save: true
-save_period: -1
-cache: false
-device: null
-workers: 8
-project: runs/train
-name: fish_detection_20251127_104658
-exist_ok: false
-pretrained: false
-optimizer: auto
-verbose: true
-seed: 0
-deterministic: true
-single_cls: false
-rect: false
-cos_lr: false
-close_mosaic: 10
-resume: false
-amp: true
-fraction: 1.0
-profile: false
-freeze: null
-multi_scale: false
-overlap_mask: true
-mask_ratio: 4
-dropout: 0.0
-val: true
-split: val
-save_json: false
-save_hybrid: false
-conf: null
-iou: 0.7
-max_det: 300
-half: false
-dnn: false
-plots: true
-source: null
-vid_stride: 1
-stream_buffer: false
-visualize: false
-augment: false
-agnostic_nms: false
-classes: null
-retina_masks: false
-embed: null
-show: false
-save_frames: false
-save_txt: false
-save_conf: false
-save_crop: false
-show_labels: true
-show_conf: true
-show_boxes: true
-line_width: null
-format: torchscript
-keras: false
-optimize: false
-int8: false
-dynamic: false
-simplify: true
-opset: null
-workspace: null
-nms: false
-lr0: 0.01
-lrf: 0.01
-momentum: 0.937
-weight_decay: 0.0005
-warmup_epochs: 3.0
-warmup_momentum: 0.8
-warmup_bias_lr: 0.1
-box: 1.75
-cls: 0.5
-dfl: 1.5
-pose: 12.0
-kobj: 1.0
-nbs: 64
-hsv_h: 0.015
-hsv_s: 0.7
-hsv_v: 0.4
-degrees: 10.0
-translate: 0.2
-scale: 0.5
-shear: 2.0
-perspective: 0.0
-flipud: 0.1
-fliplr: 0.5
-bgr: 0.0
-mosaic: 1.0
-mixup: 0.2
-copy_paste: 0.3
-copy_paste_mode: flip
-auto_augment: randaugment
-erasing: 0.4
-crop_fraction: 1.0
-cfg: null
-tracker: botsort.yaml
-save_dir: runs/train/fish_detection_20251127_104658
diff --git a/FishMeasure/runs/train/fish_detection_20251127_104658/confusion_matrix.png b/FishMeasure/runs/train/fish_detection_20251127_104658/confusion_matrix.png
deleted file mode 100644
index 1a15ed8..0000000
Binary files a/FishMeasure/runs/train/fish_detection_20251127_104658/confusion_matrix.png and /dev/null differ
diff --git a/FishMeasure/runs/train/fish_detection_20251127_104658/confusion_matrix_normalized.png b/FishMeasure/runs/train/fish_detection_20251127_104658/confusion_matrix_normalized.png
deleted file mode 100644
index e6fdc6a..0000000
Binary files a/FishMeasure/runs/train/fish_detection_20251127_104658/confusion_matrix_normalized.png and /dev/null differ
diff --git a/FishMeasure/runs/train/fish_detection_20251127_104658/events.out.tfevents.1764211621.svap-master.979906.0 b/FishMeasure/runs/train/fish_detection_20251127_104658/events.out.tfevents.1764211621.svap-master.979906.0
deleted file mode 100644
index 85bdec3..0000000
Binary files a/FishMeasure/runs/train/fish_detection_20251127_104658/events.out.tfevents.1764211621.svap-master.979906.0 and /dev/null differ
diff --git a/FishMeasure/runs/train/fish_detection_20251127_104658/labels.jpg b/FishMeasure/runs/train/fish_detection_20251127_104658/labels.jpg
deleted file mode 100644
index 2fb0236..0000000
Binary files a/FishMeasure/runs/train/fish_detection_20251127_104658/labels.jpg and /dev/null differ
diff --git a/FishMeasure/runs/train/fish_detection_20251127_104658/labels_correlogram.jpg b/FishMeasure/runs/train/fish_detection_20251127_104658/labels_correlogram.jpg
deleted file mode 100644
index ce8161b..0000000
Binary files a/FishMeasure/runs/train/fish_detection_20251127_104658/labels_correlogram.jpg and /dev/null differ
diff --git a/FishMeasure/runs/train/fish_detection_20251127_104658/results.csv b/FishMeasure/runs/train/fish_detection_20251127_104658/results.csv
deleted file mode 100644
index 68f34e5..0000000
--- a/FishMeasure/runs/train/fish_detection_20251127_104658/results.csv
+++ /dev/null
@@ -1,101 +0,0 @@
-epoch,time,train/box_loss,train/cls_loss,train/dfl_loss,metrics/precision(B),metrics/recall(B),metrics/mAP50(B),metrics/mAP50-95(B),val/box_loss,val/cls_loss,val/dfl_loss,lr/pg0,lr/pg1,lr/pg2
-1,10.3904,0.39855,2.68593,1.4848,1,0.29733,0.96537,0.56943,0.33362,2.56495,1.02233,0.00038,0.00038,0.00038
-2,17.5483,0.30028,1.35659,1.12817,1,0.49396,0.95678,0.49421,0.35142,2.66289,1.12329,0.000772278,0.000772278,0.000772278
-3,25.1797,0.29352,1.17269,1.13078,0.63354,0.70588,0.71908,0.35996,0.3706,1.94815,1.15705,0.00115664,0.00115664,0.00115664
-4,31.2751,0.29379,1.05877,1.14652,0.92901,0.76994,0.8997,0.51519,0.34598,1.68774,1.25469,0.00153307,0.00153307,0.00153307
-5,38.2345,0.30129,1.01013,1.16273,1,0.05809,0.11957,0.0573,0.56052,3.3705,2.25768,0.00190159,0.00190159,0.00190159
-6,44.1377,0.29631,0.9286,1.17263,0.77814,0.72549,0.77275,0.49905,0.30463,2.21664,1.22409,0.001901,0.001901,0.001901
-7,51.7031,0.28969,0.82499,1.16157,0.81899,0.82353,0.86643,0.50327,0.30846,1.06156,1.23743,0.0018812,0.0018812,0.0018812
-8,57.6974,0.28756,0.79748,1.16646,0.97759,0.85543,0.93703,0.54512,0.30945,1.3021,1.15419,0.0018614,0.0018614,0.0018614
-9,64.6096,0.28988,0.78404,1.16845,0.87355,0.54195,0.61651,0.44007,0.28541,1.8088,1.19646,0.0018416,0.0018416,0.0018416
-10,70.8722,0.28509,0.74823,1.17001,1,0.97714,0.9879,0.66339,0.27556,0.67389,1.10534,0.0018218,0.0018218,0.0018218
-11,77.6729,0.2811,0.71605,1.16472,0.97131,1,0.99327,0.68686,0.26039,0.53918,1.12841,0.001802,0.001802,0.001802
-12,83.6789,0.27956,0.71366,1.17215,0.94349,0.7451,0.81617,0.36975,0.44564,1.10381,1.5902,0.0017822,0.0017822,0.0017822
-13,90.5196,0.27954,0.70341,1.16105,0.98977,0.98039,0.99395,0.70985,0.26287,0.71862,1.08777,0.0017624,0.0017624,0.0017624
-14,96.2279,0.27389,0.69109,1.16035,0.99776,0.98039,0.99481,0.68656,0.27765,0.65545,1.09127,0.0017426,0.0017426,0.0017426
-15,103.274,0.27484,0.69505,1.16409,0.94274,0.96857,0.98537,0.6647,0.2812,0.56622,1.11215,0.0017228,0.0017228,0.0017228
-16,109.28,0.27709,0.68435,1.17154,0.99813,0.98039,0.99462,0.71965,0.24753,0.50567,1.06086,0.001703,0.001703,0.001703
-17,116.354,0.27127,0.65515,1.15472,0.99011,0.96078,0.99405,0.68376,0.27062,0.59393,1.10243,0.0016832,0.0016832,0.0016832
-18,121.994,0.27291,0.65216,1.1543,0.97318,0.98039,0.99153,0.72137,0.24675,0.73777,1.04403,0.0016634,0.0016634,0.0016634
-19,128.471,0.27609,0.64481,1.1633,0.99856,1,0.995,0.71822,0.2506,0.44838,1.04764,0.0016436,0.0016436,0.0016436
-20,134.908,0.26641,0.63605,1.13566,0.97925,0.98039,0.99423,0.73946,0.22498,0.4882,1.01436,0.0016238,0.0016238,0.0016238
-21,141.659,0.27002,0.63276,1.15338,0.99622,1,0.995,0.73392,0.24781,0.6929,1.06027,0.001604,0.001604,0.001604
-22,147.596,0.27016,0.62133,1.15598,1,0.71878,0.7985,0.38994,0.43809,1.15961,1.45966,0.0015842,0.0015842,0.0015842
-23,154.606,0.26856,0.63903,1.14822,0.99006,0.96078,0.98976,0.73202,0.23704,0.89525,1.02748,0.0015644,0.0015644,0.0015644
-24,160.874,0.26556,0.61805,1.14759,0.52381,0.21569,0.37119,0.23923,0.48324,4.161,1.79371,0.0015446,0.0015446,0.0015446
-25,167.63,0.26283,0.62175,1.14417,0.99771,0.98039,0.99481,0.54043,0.36833,0.62126,1.2961,0.0015248,0.0015248,0.0015248
-26,172.972,0.26277,0.62184,1.14693,0.97588,0.79352,0.838,0.60246,0.26723,1.16324,1.14688,0.001505,0.001505,0.001505
-27,180.886,0.25845,0.58391,1.1274,1,0.97466,0.99444,0.7358,0.23001,0.43927,1.03289,0.0014852,0.0014852,0.0014852
-28,186.475,0.25736,0.59062,1.13027,0.99938,1,0.995,0.77511,0.22035,0.41332,1.01175,0.0014654,0.0014654,0.0014654
-29,193.178,0.25997,0.59153,1.13884,0.99623,1,0.995,0.72041,0.2431,0.43747,1.04154,0.0014456,0.0014456,0.0014456
-30,198.402,0.26318,0.59378,1.14352,0.99588,0.98039,0.99086,0.71822,0.23253,0.59404,1.04641,0.0014258,0.0014258,0.0014258
-31,204.755,0.25794,0.57997,1.12954,0.99758,1,0.995,0.65546,0.2929,0.53569,1.17732,0.001406,0.001406,0.001406
-32,210.705,0.25981,0.59023,1.14197,0.99123,0.96078,0.97752,0.7113,0.24951,0.53056,1.04465,0.0013862,0.0013862,0.0013862
-33,217.738,0.25927,0.58286,1.1443,0.99828,1,0.995,0.76497,0.21878,0.37529,1.00619,0.0013664,0.0013664,0.0013664
-34,223.275,0.25433,0.55825,1.11915,0.98879,0.84314,0.93602,0.58228,0.34063,0.70089,1.23218,0.0013466,0.0013466,0.0013466
-35,230.18,0.25841,0.5804,1.13861,0.99727,0.98039,0.9925,0.75134,0.21602,0.45677,1.00342,0.0013268,0.0013268,0.0013268
-36,236.28,0.2532,0.57095,1.13238,0.9975,0.98039,0.99093,0.74316,0.24353,0.43215,1.05353,0.001307,0.001307,0.001307
-37,243.19,0.25301,0.57535,1.13453,1,0.99155,0.995,0.72838,0.24188,0.67456,1.0605,0.0012872,0.0012872,0.0012872
-38,249.356,0.25509,0.56332,1.13272,1,0.95882,0.99102,0.74599,0.21894,0.44255,1.00044,0.0012674,0.0012674,0.0012674
-39,256.308,0.25554,0.5739,1.13027,0.99595,1,0.995,0.75996,0.22547,0.50187,1.01792,0.0012476,0.0012476,0.0012476
-40,262.306,0.25143,0.55879,1.12012,0.99711,1,0.995,0.76673,0.21535,0.46243,1.00072,0.0012278,0.0012278,0.0012278
-41,268.917,0.2534,0.56009,1.12077,0.99082,1,0.995,0.73951,0.23347,0.66055,1.02986,0.001208,0.001208,0.001208
-42,274.828,0.2489,0.5661,1.11458,1,0.91885,0.98684,0.72466,0.22931,0.53634,1.02747,0.0011882,0.0011882,0.0011882
-43,281.555,0.25301,0.55236,1.13143,0.99738,1,0.995,0.74646,0.2298,0.49582,1.02689,0.0011684,0.0011684,0.0011684
-44,287.74,0.2486,0.53306,1.11907,1,0.97736,0.99364,0.73858,0.23603,0.49003,1.03041,0.0011486,0.0011486,0.0011486
-45,294.836,0.24944,0.54158,1.1268,0.99062,1,0.995,0.75756,0.23048,0.60132,1.04024,0.0011288,0.0011288,0.0011288
-46,301.451,0.24913,0.53379,1.1269,0.99596,0.98039,0.99048,0.77386,0.20906,0.54314,0.99925,0.001109,0.001109,0.001109
-47,308.242,0.24957,0.53985,1.12517,0.99593,1,0.995,0.73996,0.23001,0.52276,1.02912,0.0010892,0.0010892,0.0010892
-48,314.344,0.24351,0.53626,1.10608,1,0.99956,0.995,0.74934,0.23199,0.39499,1.03412,0.0010694,0.0010694,0.0010694
-49,321.241,0.24759,0.53893,1.11816,1,0.99844,0.995,0.74604,0.22845,0.39064,1.02153,0.0010496,0.0010496,0.0010496
-50,327.139,0.2464,0.53694,1.1138,0.99835,0.98039,0.99048,0.71125,0.24541,0.42938,1.07937,0.0010298,0.0010298,0.0010298
-51,333.821,0.24154,0.51898,1.10733,0.99567,0.90196,0.97517,0.62595,0.29596,0.61014,1.14221,0.00101,0.00101,0.00101
-52,339.882,0.24658,0.53043,1.11979,0.99546,0.98039,0.99323,0.65978,0.27159,0.52309,1.13272,0.0009902,0.0009902,0.0009902
-53,346.689,0.24636,0.52582,1.11253,0.99778,0.98039,0.99481,0.69862,0.26351,0.45512,1.08185,0.0009704,0.0009704,0.0009704
-54,352.945,0.24152,0.51738,1.10331,0.99739,1,0.995,0.72452,0.24969,0.40628,1.06735,0.0009506,0.0009506,0.0009506
-55,359.178,0.24263,0.50863,1.10604,0.99683,1,0.995,0.77739,0.21597,0.42763,1.01173,0.0009308,0.0009308,0.0009308
-56,365.965,0.24373,0.51169,1.10391,0.95858,0.90771,0.97651,0.63893,0.29309,0.63256,1.16514,0.000911,0.000911,0.000911
-57,372.475,0.24282,0.51188,1.10565,0.9915,0.98039,0.99054,0.75734,0.21952,0.56379,1.01633,0.0008912,0.0008912,0.0008912
-58,379.602,0.23912,0.50781,1.09877,1,0.95988,0.98778,0.72449,0.24914,0.46787,1.05661,0.0008714,0.0008714,0.0008714
-59,385.13,0.24179,0.51076,1.10568,1,0.95731,0.98289,0.71162,0.23833,0.6395,1.04927,0.0008516,0.0008516,0.0008516
-60,392.571,0.23757,0.50074,1.09636,0.99702,0.98039,0.99261,0.76777,0.20872,0.47718,0.99627,0.0008318,0.0008318,0.0008318
-61,397.864,0.23342,0.49493,1.09485,0.99568,0.98039,0.9931,0.76937,0.20784,0.45567,1.01043,0.000812,0.000812,0.000812
-62,405.506,0.23929,0.49675,1.10508,0.9953,0.96078,0.9827,0.72038,0.24622,0.5664,1.07435,0.0007922,0.0007922,0.0007922
-63,411.243,0.2385,0.49568,1.09486,0.99783,0.98039,0.99086,0.76087,0.22259,0.49003,1.01614,0.0007724,0.0007724,0.0007724
-64,418.044,0.23775,0.49129,1.08891,0.99544,0.98039,0.99054,0.77224,0.21441,0.51803,1.00319,0.0007526,0.0007526,0.0007526
-65,423.941,0.23656,0.49423,1.09174,0.97965,0.94412,0.98937,0.76534,0.21473,0.53841,0.99562,0.0007328,0.0007328,0.0007328
-66,430.931,0.23491,0.49057,1.08267,0.99223,0.98039,0.98955,0.76326,0.21453,0.64708,0.99581,0.000713,0.000713,0.000713
-67,436.831,0.23603,0.4875,1.08431,0.99649,0.98039,0.99427,0.76543,0.21578,0.47111,0.99195,0.0006932,0.0006932,0.0006932
-68,443.698,0.23371,0.48684,1.08189,1,0.97435,0.99379,0.77164,0.215,0.53772,0.99543,0.0006734,0.0006734,0.0006734
-69,449.871,0.23341,0.48312,1.0883,0.99289,0.98039,0.99379,0.78205,0.20299,0.46992,0.97694,0.0006536,0.0006536,0.0006536
-70,456.649,0.2335,0.48756,1.08284,0.99768,0.98039,0.9931,0.76909,0.21567,0.3792,0.99144,0.0006338,0.0006338,0.0006338
-71,462.607,0.22882,0.47871,1.07799,0.99471,0.98039,0.991,0.76463,0.22008,0.44229,1.00736,0.000614,0.000614,0.000614
-72,469.768,0.23333,0.48255,1.08936,0.99842,0.92157,0.95471,0.6835,0.25689,0.5887,1.10088,0.0005942,0.0005942,0.0005942
-73,476.152,0.22814,0.47147,1.07796,0.99695,0.96078,0.99179,0.75486,0.22212,0.48941,1.01634,0.0005744,0.0005744,0.0005744
-74,483.377,0.22914,0.47235,1.08384,1,0.97669,0.9901,0.77219,0.2153,0.39275,0.99558,0.0005546,0.0005546,0.0005546
-75,488.716,0.22821,0.47539,1.08204,0.99868,1,0.995,0.74383,0.23322,0.39742,1.03818,0.0005348,0.0005348,0.0005348
-76,496.052,0.22693,0.47841,1.0774,0.9966,0.98039,0.99427,0.77597,0.22186,0.42881,1.00906,0.000515,0.000515,0.000515
-77,501.774,0.2272,0.46429,1.07859,0.99656,0.98039,0.9925,0.7684,0.21909,0.4223,1.00152,0.0004952,0.0004952,0.0004952
-78,508.605,0.22715,0.46817,1.08007,1,0.99833,0.995,0.77597,0.20663,0.39658,0.98381,0.0004754,0.0004754,0.0004754
-79,514.483,0.22845,0.46984,1.08602,0.99476,0.98039,0.99297,0.76487,0.22367,0.53976,1.01338,0.0004556,0.0004556,0.0004556
-80,521.104,0.22489,0.45308,1.07029,0.99378,0.98039,0.99364,0.76591,0.22644,0.44335,1.01611,0.0004358,0.0004358,0.0004358
-81,526.831,0.22808,0.46696,1.08249,0.9984,0.98039,0.99043,0.76392,0.22464,0.40558,1.00919,0.000416,0.000416,0.000416
-82,534.261,0.22461,0.45758,1.07345,0.99706,0.98039,0.9918,0.76203,0.22683,0.37658,1.02904,0.0003962,0.0003962,0.0003962
-83,539.83,0.2224,0.45574,1.06465,0.9966,0.96078,0.98809,0.77183,0.22791,0.58994,1.00929,0.0003764,0.0003764,0.0003764
-84,547.068,0.22321,0.45489,1.07012,0.99534,0.98039,0.99218,0.77586,0.21211,0.4875,0.99631,0.0003566,0.0003566,0.0003566
-85,553.422,0.22324,0.44994,1.07037,0.99705,0.98039,0.99122,0.77203,0.21439,0.42988,1.00439,0.0003368,0.0003368,0.0003368
-86,560.244,0.22593,0.46254,1.07436,0.99551,0.98039,0.99323,0.7568,0.21436,0.39946,1.00288,0.000317,0.000317,0.000317
-87,566.207,0.22535,0.45462,1.07467,0.99768,0.98039,0.99427,0.77229,0.20736,0.36008,0.99378,0.0002972,0.0002972,0.0002972
-88,573.201,0.22031,0.44716,1.0635,0.99697,0.98039,0.99481,0.76489,0.21791,0.38311,1.01245,0.0002774,0.0002774,0.0002774
-89,579.084,0.22289,0.44359,1.06804,0.99621,0.98039,0.99162,0.77105,0.21459,0.41661,0.99051,0.0002576,0.0002576,0.0002576
-90,585.517,0.2213,0.4519,1.07424,0.99675,0.98039,0.99122,0.76516,0.21041,0.42833,0.99584,0.0002378,0.0002378,0.0002378
-91,597.175,0.19679,0.38958,1.03361,0.99618,0.98039,0.9925,0.76415,0.21165,0.41842,0.98503,0.000218,0.000218,0.000218
-92,603.128,0.19341,0.33723,1.01494,0.99613,0.98039,0.99285,0.76376,0.21456,0.42319,0.99363,0.0001982,0.0001982,0.0001982
-93,609.58,0.19245,0.34007,1.02201,0.99697,0.98039,0.99229,0.76732,0.21373,0.38196,0.99964,0.0001784,0.0001784,0.0001784
-94,614.558,0.18975,0.33377,1.01562,0.99702,0.98039,0.99285,0.76477,0.21757,0.39742,0.9974,0.0001586,0.0001586,0.0001586
-95,620.215,0.18952,0.32777,1.01532,0.9962,0.98039,0.99107,0.7668,0.21387,0.44539,1.0045,0.0001388,0.0001388,0.0001388
-96,625.105,0.1894,0.32969,1.01349,0.99492,0.98039,0.9913,0.76397,0.21124,0.43134,0.99738,0.000119,0.000119,0.000119
-97,630.785,0.18892,0.32829,1.00956,0.99236,0.98039,0.9908,0.7635,0.21104,0.4798,0.99472,9.92e-05,9.92e-05,9.92e-05
-98,636.519,0.18811,0.32653,1.00609,0.99603,0.98039,0.99261,0.77306,0.21321,0.40139,0.99426,7.94e-05,7.94e-05,7.94e-05
-99,642.554,0.1863,0.32067,1.00634,0.99516,0.98039,0.9935,0.75723,0.21665,0.41315,1.00323,5.96e-05,5.96e-05,5.96e-05
-100,647.259,0.18644,0.31964,1.00625,0.99453,0.98039,0.99297,0.76087,0.21668,0.43148,1.00482,3.98e-05,3.98e-05,3.98e-05
diff --git a/FishMeasure/runs/train/fish_detection_20251127_104658/results.png b/FishMeasure/runs/train/fish_detection_20251127_104658/results.png
deleted file mode 100644
index 7a204f6..0000000
Binary files a/FishMeasure/runs/train/fish_detection_20251127_104658/results.png and /dev/null differ
diff --git a/FishMeasure/runs/train/fish_detection_20251127_104658/train_batch0.jpg b/FishMeasure/runs/train/fish_detection_20251127_104658/train_batch0.jpg
deleted file mode 100644
index 6e29b85..0000000
Binary files a/FishMeasure/runs/train/fish_detection_20251127_104658/train_batch0.jpg and /dev/null differ
diff --git a/FishMeasure/runs/train/fish_detection_20251127_104658/train_batch1.jpg b/FishMeasure/runs/train/fish_detection_20251127_104658/train_batch1.jpg
deleted file mode 100644
index d47aae3..0000000
Binary files a/FishMeasure/runs/train/fish_detection_20251127_104658/train_batch1.jpg and /dev/null differ
diff --git a/FishMeasure/runs/train/fish_detection_20251127_104658/train_batch1800.jpg b/FishMeasure/runs/train/fish_detection_20251127_104658/train_batch1800.jpg
deleted file mode 100644
index 8ac9713..0000000
Binary files a/FishMeasure/runs/train/fish_detection_20251127_104658/train_batch1800.jpg and /dev/null differ
diff --git a/FishMeasure/runs/train/fish_detection_20251127_104658/train_batch1801.jpg b/FishMeasure/runs/train/fish_detection_20251127_104658/train_batch1801.jpg
deleted file mode 100644
index 84bdfd9..0000000
Binary files a/FishMeasure/runs/train/fish_detection_20251127_104658/train_batch1801.jpg and /dev/null differ
diff --git a/FishMeasure/runs/train/fish_detection_20251127_104658/train_batch1802.jpg b/FishMeasure/runs/train/fish_detection_20251127_104658/train_batch1802.jpg
deleted file mode 100644
index fda43c1..0000000
Binary files a/FishMeasure/runs/train/fish_detection_20251127_104658/train_batch1802.jpg and /dev/null differ
diff --git a/FishMeasure/runs/train/fish_detection_20251127_104658/train_batch2.jpg b/FishMeasure/runs/train/fish_detection_20251127_104658/train_batch2.jpg
deleted file mode 100644
index ef21dc9..0000000
Binary files a/FishMeasure/runs/train/fish_detection_20251127_104658/train_batch2.jpg and /dev/null differ
diff --git a/FishMeasure/runs/train/fish_detection_20251127_104658/val_batch0_labels.jpg b/FishMeasure/runs/train/fish_detection_20251127_104658/val_batch0_labels.jpg
deleted file mode 100644
index 3ec747a..0000000
Binary files a/FishMeasure/runs/train/fish_detection_20251127_104658/val_batch0_labels.jpg and /dev/null differ
diff --git a/FishMeasure/runs/train/fish_detection_20251127_104658/val_batch0_pred.jpg b/FishMeasure/runs/train/fish_detection_20251127_104658/val_batch0_pred.jpg
deleted file mode 100644
index be38b81..0000000
Binary files a/FishMeasure/runs/train/fish_detection_20251127_104658/val_batch0_pred.jpg and /dev/null differ
diff --git a/FishMeasure/runs/train/fish_detection_20251127_104658/weights/last.pt b/FishMeasure/runs/train/fish_detection_20251127_104658/weights/last.pt
deleted file mode 100644
index b132f28..0000000
Binary files a/FishMeasure/runs/train/fish_detection_20251127_104658/weights/last.pt and /dev/null differ
diff --git a/FishMeasure/weight_estimator/test_dgcnn_weight_estimator.py b/FishMeasure/weight_estimator/test_dgcnn_weight_estimator.py
index 12ff0d5..1e50eee 100644
--- a/FishMeasure/weight_estimator/test_dgcnn_weight_estimator.py
+++ b/FishMeasure/weight_estimator/test_dgcnn_weight_estimator.py
@@ -30,6 +30,24 @@ import re
 import sys
 import zlib
 from pathlib import Path
+
+if not hasattr(argparse, "BooleanOptionalAction"):
+    class _BooleanOptionalAction(argparse.Action):
+        def __init__(self, option_strings, dest, default=None, type=None,
+                     choices=None, required=False, help=None, metavar=None):
+            _option_strings = []
+            for opt in option_strings:
+                _option_strings.append(opt)
+                if opt.startswith("--"):
+                    _option_strings.append("--no-" + opt[2:])
+            super().__init__(option_strings=_option_strings, dest=dest, nargs=0,
+                             default=default, type=type, choices=choices,
+                             required=required, help=help, metavar=metavar)
+
+        def __call__(self, parser, namespace, values, option_string=None):
+            setattr(namespace, self.dest, not option_string.startswith("--no-"))
+
+    argparse.BooleanOptionalAction = _BooleanOptionalAction
 from typing import Any, Dict, List, Optional, Tuple
 
 import numpy as np
diff --git a/docs/conda-deploy.md b/docs/conda-deploy.md
new file mode 100644
index 0000000..26970bf
--- /dev/null
+++ b/docs/conda-deploy.md
@@ -0,0 +1,130 @@
+# 使用 Conda 单环境部署 FishServer（方案 A）
+
+本文说明在另一台 **Linux x86_64** 机器上，用 **一个 conda 环境** 安装 Fish API、FishMeasure、FishAction 所需依赖并启动服务。子进程与 uvicorn 共用同一 Python 解释器，**无需**在 `fish_api/.env` 中设置 `PYTHON_FISH_MEASURE` / `PYTHON_FISH_ACTION`（保持为空即可）。
+
+---
+
+## 前置条件
+
+- 已克隆本仓库，目录中包含 `fish_api/`、`FishMeasure/`、`FishAction/`。
+- **模型权重**：将原机器上的 `models/` 及 `.env` 中引用的 checkpoint 拷贝到新机器；空目录会导致推理失败。
+- **GPU（可选）**：若使用 `SAM_DEVICE=cuda` 或 PyTorch CUDA，需安装与 PyTorch 版本匹配的 **NVIDIA 驱动与 CUDA**。
+- **ZED SDK（处理 `.svo2` 时）**：在新机器安装 **Stereolabs ZED SDK** 及 Python API，版本尽量与现有开发/生产环境一致。
+- **ffmpeg**：用于视频转码；可用系统包（`apt install ffmpeg` 等），或使用仓库内 `scripts/setup_ffmpeg.sh`（需将脚本中的 `PROJECT_ROOT` 改为本机仓库路径）。
+
+---
+
+## 1. 创建并激活 conda 环境
+
+Fish API 要求 **Python ≥ 3.11**（见 `fish_api/pyproject.toml`）。
+
+```bash
+conda create -n fishserver python=3.11 -y
+conda activate fishserver
+```
+
+后续所有 `pip`/`python` 命令均在该环境中执行。
+
+---
+
+## 2. 安装 Fish API（FastAPI 网关）
+
+```bash
+cd /path/to/FishServer/fish_api
+pip install -e .
+```
+
+可选开发依赖（本地冒烟测试等）：
+
+```bash
+pip install -e ".[dev]"
+```
+
+---
+
+## 3. 在同一环境中安装 FishMeasure / FishAction 依赖
+
+**方案 A 的核心**：测量与行为脚本由 API 以子进程调用，默认使用 **当前 `python`（与 uvicorn 相同）**，因此 PyTorch、OpenCV、ZED、X3D 等必须**全部**装进 **`fishserver`** 这一环境。
+
+本仓库未提供单一的 `FishMeasure/requirements.txt`，请按以下方式之一对齐依赖：
+
+1. **从已有机器导出**（推荐）：在已能跑通测量的环境中执行 `pip freeze > requirements-frozen.txt`，在新环境中用 `pip install -r requirements-frozen.txt`（注意需人工剔除/替换与平台、CUDA 不兼容的包）。
+2. **按模块 README 与脚本逐步安装**：参考 `FishMeasure/README.md`、各子目录说明，以及 `FishAction/predict_video_x3d_3class.py` 的实际 import；至少安装 `FishAction/requirements.txt` 中的依赖，并补齐 PyTorch、torchvision、与 ZED 文档匹配的包。
+
+安装完成后，在 **`conda activate fishserver`** 下验证：
+
+```bash
+python -c "import torch; print(torch.__version__)"
+# 若使用 ZED：按 Stereolabs 文档验证 pyzed 等
+```
+
+---
+
+## 4. 配置 `fish_api/.env`
+
+从旧机器复制 `fish_api/.env` 到新机器，**全部路径改为本机绝对路径**，例如：
+
+- `PUBLIC_BASE_URL`：客户端可访问的 API 根地址，**不要**末尾 `/`。
+- `ACTION_WATCH_DIR`、`MEASURE_WATCH_DIR`：若启用目录监控，填本机目录。
+- `YOLO_MODEL`、`WEIGHT_CHECKPOINT`、`ACTION_CHECKPOINT` 等：指向本机真实文件。
+- `SAM_DEVICE`：`cuda` 或 `cpu`。
+
+**方案 A 下不要填写** `PYTHON_FISH_MEASURE`、`PYTHON_FISH_ACTION`（留空），让子进程使用当前环境的 `python`。
+
+---
+
+## 5. 与 `uv` 的优先级
+
+`fish_api/start_fresh.sh` 若检测到系统中有 `uv` 命令，会优先使用 `uv run`。若你希望**强制使用当前 conda 环境**：
+
+- 启动前执行 **`conda activate fishserver`**，并确认 `which python` / `which uvicorn` 指向 `fishserver`；或
+- 在不含 `uv` 的 PATH 下启动；或临时从 PATH 中移除 `uv`。
+
+若未安装 `uv`，脚本会自动使用 `python3` 与 `uvicorn`（依赖已激活的 conda）。
+
+---
+
+## 6. 启动服务
+
+在**仓库根目录**执行：
+
+```bash
+conda activate fishserver
+cd /path/to/FishServer
+bash scripts/start_fresh.sh
+```
+
+默认监听 `0.0.0.0:8000`。可覆盖：
+
+```bash
+HOST=0.0.0.0 PORT=8000 bash scripts/start_fresh.sh
+```
+
+可选环境变量（见 `fish_api/start_fresh.sh` 注释）：
+
+- `CLEAR_SQLITE_DATABASE=1`：清空 SQLite（谨慎）。
+- `CLEAR_MEASURE_OUTPUT=1` / `CLEAR_ACTION_OUTPUT=1`：清空中间输出目录。
+
+浏览器访问：`http://<主机>:8000/docs`。
+
+---
+
+## 7. 自检清单
+
+| 检查项 | 说明 |
+|--------|------|
+| 环境 | `conda activate fishserver` 后 `python -V` 为 3.11+ |
+| API 包 | 在 `fish_api` 目录下 `python -c "import app"` 无报错 |
+| 可执行文件 | `which uvicorn` 指向 `fishserver` 的 `bin` |
+| 权重 | `.env` 中模型路径在本机存在 |
+| 测量（SVO） | ZED 与 PyTorch 在同一 conda 环境中可用 |
+| ffmpeg | `ffmpeg -version` 成功（若走视频转码路径） |
+
+---
+
+## 8. 参考
+
+- 根目录 `README.md`、`fish_api/README.md`：环境变量表与启动说明。
+- `fish_api/start_fresh.sh`：启动前 `prestart_fresh` 与 uvicorn 参数。
+
+若需将 FishMeasure / FishAction 拆到**独立 conda 环境**，请在 `fish_api/.env` 中填写 `PYTHON_FISH_MEASURE` 与 `PYTHON_FISH_ACTION`（不在本文「方案 A」范围内）。
diff --git a/fish_api/README.md b/fish_api/README.md
index 7d93a11..50324a3 100644
--- a/fish_api/README.md
+++ b/fish_api/README.md
@@ -1,6 +1,6 @@
 # Recalculate Everything:
 
-CLEAR_SQLITE_DATABASE=1 CLEAR_MEASURE_OUTPUT=1 CLEAR_ACTION_OUTPUT=1 CLEAR_MEDIA=1 CLEAR_STREAM_TMP=1 bash scripts/start_fresh.sh
+CLEAR_SQLITE_DATABASE=1 CLEAR_MEASURE_OUTPUT=1 CLEAR_ACTION_OUTPUT=1 CLEAR_MEDIA=1 CLEAR_STREAM_TMP=1 LD_PRELOAD=/lib/aarch64-linux-gnu/libGLdispatch.so.0  bash scripts/start_fresh.sh
 
 
 
@@ -17,19 +17,62 @@ CLEAR_SQLITE_DATABASE=1 CLEAR_MEASURE_OUTPUT=1 CLEAR_ACTION_OUTPUT=1 CLEAR_MEDIA
 | `MEASURE_OUTPUT_ROOT` | 传给 `--save-output` 的目录 | `<repo>/fish_api/.data/measure_output` |
 | `YOLO_MODEL` / `WEIGHT_CHECKPOINT` / `ACTION_CHECKPOINT` | 模型路径 | 与仓库内脚本默认一致 |
 | `SAM_DEVICE` | `cuda` 或 `cpu` | `cuda` |
+| `MEASURE_FINAL_AGGREGATE_MODE` | 齐套后对各段 former 体重/体长聚合：`median` / `mean` / `trimmed_mean` | `median` |
 可在 `fish_api/.env` 中填写上述变量（`pydantic-settings` 会读取）。
 
 ## 安装与启动
 
 ```bash
 cd fish_api
-uv sync
-# 可选：包含 httpx，便于本地用 FastAPI TestClient 做冒烟测试
-# uv sync --group dev
+python3 -m pip install -e .   # 安装 pyproject 依赖（无 venv 时可用 --user 或系统 site-packages）
 ./scripts/start_fresh.sh   # 默认仅重置 client_id 投递进度，保留 SQLite 历史与快照
 # CLEAR_SQLITE_DATABASE=1 bash start_fresh.sh  # 需要时才彻底清 SQLite
-# 或：uv run uvicorn app.main:app --host 0.0.0.0 --port 8000（需自行 prestart）
+# python3 -m uvicorn app.main:app --host 0.0.0.0 --port 8000  # 需自行 prestart
 ```
+
+## SVO 输入与 ``measure_watch``（两种来源，同一套目录逻辑）
+
+后台 **[``measure_watch``](app/services/measure_watch.py)** 只认 **`MEASURE_WATCH_DIR`** 下一层的 **`fish{N}/`** 子目录及其中的 **``.svo2``**（见 ``iter_svo2_folders``）。与入口无关：
+
+| 方式 | 说明 |
+|------|------|
+| **ZED 分段录制** | 每次会话分配 ``fish_id``：取库表（``fish_id`` 与 ``output_dir`` 路径）、父目录下 ``fish``+数字 子目录名、以及 ``fish{N}/`` 下已有 ``.svo2`` 路径 四者编号的最大值再加 1（磁盘有数据而库未记时也不冲突）；文件写入 ``{MEASURE_WATCH_DIR}/fish{N}/``（若未配置 ``MEASURE_WATCH_DIR`` 则 ``{STREAM_TMP_DIR}/zed_svo2/fish{N}/``，此时**不会**启用 ``measure_watch``，除非把 ``MEASURE_WATCH_DIR`` 指到 ``…/ingest/zed_svo2``） |
+| **手工拷贝** | 将 ``.svo2`` 放入 ``MEASURE_WATCH_DIR/fish{N}/``（自建 ``fish{N}`` 即可） |
+
+**逐段测量与齐套 final**：对每个 ``.svo2`` 稳定后轮询跑 FishMeasure，写入 SQLite；服务端可在 ``calculation_log`` 中区分 segment/final。``GET /api/v1/biomass/real/camera/`` 的 ``data.result[]`` **仅含** ``id``、``type``、``length``、``weight``、``date``（与历史客户端约定一致）；按投递顺序先可能收到多段再收到聚合行。``video_left`` / ``video_right`` 规则不变。
+
+**ingest** 的 ``/api/v1/ingest/svo/...`` 为分块上传流程，落盘与上述 ``fish{N}`` 路径独立；``fish{N}`` 目录的测量以 ``measure_watch`` 为准。
+
+**ZED 相机**：**fish_api 启动/停止不会自动开关相机录制**；录制由独立服务或仓库根 ``start_recording.sh`` / ``zed_record_cli`` 等自行管理，写入 ``MEASURE_WATCH_DIR/fish{N}/`` 后由 ``measure_watch`` 消费。需要时也可调 ``/api/v1/zed/recording/start|stop``（不经过 uvicorn lifespan）。
+
+## ZED 分段录制 CLI
+
+每次 **开始/停止** 录制会在 SQLite 表 ``zed_recording_sessions`` 中记录一行（``fish_id``、``started_at`` / ``stopped_at``、``output_dir``；分配规则见上表）。
+
+在 `fish_api` 目录下执行（与 `app` 包路径一致；依赖已 `pip install -e .`）：
+
+```bash
+# 本机直连相机：前台录制，终端不关，Ctrl+C 结束（终端会打印分配的 fish_id）
+python3 -m app.zed_record_cli start [--segment-sec SEC]
+
+# 仅向已运行的 fish_api 发请求（不阻塞）
+python3 -m app.zed_record_cli start --remote [--segment-sec SEC]
+
+# 停止 / 查询状态（HTTP，需 uvicorn 已监听；基址见 FISH_API_BASE_URL 或 PUBLIC_BASE_URL）
+python3 -m app.zed_record_cli stop
+python3 -m app.zed_record_cli status
+```
+
+若已通过 `pip install -e .` 安装了控制台脚本，也可使用 `fish-zed-record`（与上表等价）。
+
+也可使用仓库根目录下的封装脚本：
+
+```bash
+# 在 FishServer 仓库根目录执行
+./start_recording.sh              # 本地前台；可选 --remote / --segment-sec
+./stop_recording.sh               # HTTP 停止（需 API 已启动）
+```
+
 ## Weight Rule (Current)
 
 最终体重 `pred_weight_g` 由以下规则链决定（按优先级从高到低）：
@@ -55,15 +98,15 @@ bash scripts/measure_debug.sh --fish-id 14
 
 # 方式 B：先进入 fish_api
 cd fish_api
-uv sync
+python3 -m pip install -e .   # 若尚未安装依赖
 # 默认：MEASURE_WATCH_DIR/fish{N}/ 下所有 .svo2 → 输出到 MEASURE_OUTPUT_ROOT/fish{N}（默认 fish_api/.data/measure_output/fish{N}）
-uv run python -m app.measure_debug_cli --fish-id 1
+python3 -m app.measure_debug_cli --fish-id 1
 
-# 或等价入口（须在 fish_api 目录）
-uv run fish-measure-debug --fish-id 1
+# 或等价入口（须在 fish_api 目录，且已 pip install -e .）
+fish-measure-debug --fish-id 1
 
 # 指定 SVO 目录或输出目录（在 fish_api 目录下）
-uv run python -m app.measure_debug_cli --batch-folder /path/to/fish1 --fish-id 1 --output-root /path/to/out
+python3 -m app.measure_debug_cli --batch-folder /path/to/fish1 --fish-id 1 --output-root /path/to/out
 ```
 
 结束后会在终端打印 `weight_prediction.json` 中的 `pred_weight_g`、`pred_weight_rule` 等摘要。
diff --git a/fish_api/app/compat.py b/fish_api/app/compat.py
new file mode 100644
index 0000000..fdf80f5
--- /dev/null
+++ b/fish_api/app/compat.py
@@ -0,0 +1,14 @@
+"""Python 3.8 兼容：提供 asyncio.to_thread 的等价实现。"""
+from __future__ import annotations
+
+import asyncio
+import functools
+from typing import Any, Callable, TypeVar
+
+T = TypeVar("T")
+
+
+async def to_thread(func: Callable[..., T], *args: Any, **kwargs: Any) -> T:
+    """asyncio.to_thread 的 Python 3.8 兼容版本。"""
+    loop = asyncio.get_running_loop()
+    return await loop.run_in_executor(None, functools.partial(func, *args, **kwargs))
diff --git a/fish_api/app/db.py b/fish_api/app/db.py
index a188297..5125096 100644
--- a/fish_api/app/db.py
+++ b/fish_api/app/db.py
@@ -9,6 +9,7 @@ from __future__ import annotations
 
 import json
 import math
+import re
 import shutil
 import sqlite3
 from datetime import datetime, timezone
@@ -23,7 +24,7 @@ DEFAULT_CLIENT_ID = "default"
 MAX_CLIENT_ID_LEN = 128
 
 # 客户端切片索引起缓存：记录每个 client_id 上次返回的切片索引（用于对齐 water/video 端点）
-_client_health_slice_index: dict[str, int] = {}
+_client_health_slice_index = {}  # type: Dict[str, int]
 
 
 def _parse_slice_index_from_source_path(source_path: Optional[str]) -> int:
@@ -117,6 +118,17 @@ def init_db(settings: Settings) -> None:
                 last_delivered_id INTEGER NOT NULL DEFAULT 0,
                 PRIMARY KEY (client_id, kind)
             );
+
+            CREATE TABLE IF NOT EXISTS zed_recording_sessions (
+                id INTEGER PRIMARY KEY AUTOINCREMENT,
+                fish_id INTEGER NOT NULL,
+                started_at TEXT NOT NULL,
+                stopped_at TEXT,
+                output_dir TEXT NOT NULL
+            );
+
+            CREATE INDEX IF NOT EXISTS idx_zed_sessions_fish_id
+                ON zed_recording_sessions(fish_id);
             """
         )
         _migrate_delivery_cursor_from_legacy(conn)
@@ -128,6 +140,147 @@ def init_db(settings: Settings) -> None:
         conn.close()
 
 
+_FISH_DIR_RE = re.compile(r"^fish(\d+)$")
+
+
+def _max_numeric_fish_folder_id(parent: Path) -> int:
+    """``parent`` 下名为 ``fish{N}``（N 为数字）的直接子目录中，最大的 N；无则 0。"""
+    if not parent.is_dir():
+        return 0
+    m = 0
+    try:
+        for p in parent.iterdir():
+            if not p.is_dir():
+                continue
+            mo = _FISH_DIR_RE.match(p.name)
+            if mo:
+                m = max(m, int(mo.group(1)))
+    except OSError:
+        return 0
+    return m
+
+
+def _max_fish_id_from_svo2_under_parent(parent: Path) -> int:
+    """扫描 ``parent`` 下 ``fish{{N}}/.../*.svo2``，从相对路径第一段取最大 N（库未记会话时仍能发现已有数据）。"""
+    if not parent.is_dir():
+        return 0
+    m = 0
+    try:
+        for p in parent.rglob("*.svo2"):
+            try:
+                rel = p.relative_to(parent)
+            except ValueError:
+                continue
+            if not rel.parts:
+                continue
+            mo = _FISH_DIR_RE.match(rel.parts[0])
+            if mo:
+                m = max(m, int(mo.group(1)))
+    except OSError:
+        return 0
+    return m
+
+
+def _max_fish_id_from_zed_sessions(conn: sqlite3.Connection) -> int:
+    """合并 ``fish_id`` 列与 ``output_dir`` 路径中出现的 ``fish{{N}}`` 段（防库与路径不一致）。"""
+    m = 0
+    for row in conn.execute("SELECT fish_id, output_dir FROM zed_recording_sessions"):
+        m = max(m, int(row["fish_id"]))
+        od = row["output_dir"]
+        if not od:
+            continue
+        try:
+            for part in Path(str(od)).parts:
+                mo = _FISH_DIR_RE.match(part)
+                if mo:
+                    m = max(m, int(mo.group(1)))
+        except (ValueError, OSError):
+            pass
+    return m
+
+
+def begin_zed_recording_session(settings: Settings) -> Tuple[int, int, Path]:
+    """为本次录制分配 ``fish_id`` 并写入 ``zed_recording_sessions``。
+
+    编号取库内会话、父目录下 ``fish``+数字 子目录、以及其下 ``.svo2`` 文件路径所反映编号三者的最大值再加 1，
+    避免目录里已有数据而库尚未记录时仍复用同一编号；若目标路径仍存在则顺延直至可用。
+
+    返回 ``(session_row_id, fish_id, output_dir)``。目录规则：
+    若配置了 ``MEASURE_WATCH_DIR`` 则为 ``{MEASURE_WATCH_DIR}/fish{N}``，
+    否则为 ``{STREAM_TMP_DIR}/zed_svo2/fish{N}``。
+    """
+    init_db(settings)
+    conn = _connect(settings.sqlite_path)
+    try:
+        conn.execute("BEGIN IMMEDIATE")
+        db_max = _max_fish_id_from_zed_sessions(conn)
+        if settings.measure_watch_dir is not None:
+            parent = settings.measure_watch_dir.resolve()
+        else:
+            parent = (settings.stream_tmp_dir / "zed_svo2").resolve()
+        fs_dir = _max_numeric_fish_folder_id(parent)
+        fs_svo = _max_fish_id_from_svo2_under_parent(parent)
+        fs_max = max(fs_dir, fs_svo)
+        fish_id = max(db_max, fs_max) + 1
+        ts = datetime.now(timezone.utc).isoformat()
+        output_dir = (parent / f"fish{fish_id}").resolve()
+        # 极端情况：并发或其它原因导致目录已存在，则顺延编号
+        for _ in range(10000):
+            if not output_dir.exists():
+                break
+            fish_id += 1
+            output_dir = (parent / f"fish{fish_id}").resolve()
+        else:
+            raise RuntimeError(
+                f"无法为 ZED 录制分配空闲 fish 目录（父目录 {parent}）"
+            )
+        output_dir.mkdir(parents=True, exist_ok=True)
+        cur = conn.execute(
+            """
+            INSERT INTO zed_recording_sessions (fish_id, started_at, stopped_at, output_dir)
+            VALUES (?, ?, NULL, ?)
+            """,
+            (fish_id, ts, str(output_dir)),
+        )
+        session_id = int(cur.lastrowid)
+        conn.commit()
+        return (session_id, fish_id, output_dir)
+    except Exception:
+        conn.rollback()
+        raise
+    finally:
+        conn.close()
+
+
+def mark_zed_recording_session_stopped(
+    settings: Settings, session_row_id: int
+) -> Optional[int]:
+    """将对应会话行的 ``stopped_at`` 置为当前时间；返回 ``fish_id``，未更新则 ``None``。"""
+    init_db(settings)
+    conn = _connect(settings.sqlite_path)
+    try:
+        ts = datetime.now(timezone.utc).isoformat()
+        cur = conn.execute(
+            """
+            UPDATE zed_recording_sessions
+            SET stopped_at = ?
+            WHERE id = ? AND stopped_at IS NULL
+            """,
+            (ts, session_row_id),
+        )
+        if cur.rowcount == 0:
+            conn.commit()
+            return None
+        row = conn.execute(
+            "SELECT fish_id FROM zed_recording_sessions WHERE id = ?",
+            (session_row_id,),
+        ).fetchone()
+        conn.commit()
+        return int(row["fish_id"]) if row else None
+    finally:
+        conn.close()
+
+
 def _migrate_add_client_id_column(conn: sqlite3.Connection) -> None:
     """为旧数据库添加 client_id 列（如果不存在）。"""
     row = conn.execute(
@@ -376,6 +529,37 @@ def list_all_measure_snapshots(settings: Settings) -> List[Dict[str, Any]]:
         conn.close()
 
 
+def list_all_health_snapshots(settings: Settings) -> List[Dict[str, Any]]:
+    """返回 ``health_snapshots`` 全部行（id 降序，最新在前），供调试接口使用。"""
+    init_db(settings)
+    conn = _connect(settings.sqlite_path)
+    try:
+        rows = conn.execute(
+            """
+            SELECT id, created_at, behavior_result, health_result,
+                   raw_class_en, error, source_path
+            FROM health_snapshots
+            ORDER BY id DESC
+            """
+        ).fetchall()
+        out: List[Dict[str, Any]] = []
+        for row in rows:
+            out.append(
+                {
+                    "id": row["id"],
+                    "created_at": row["created_at"],
+                    "behavior_result": row["behavior_result"] or "",
+                    "health_result": row["health_result"] or "",
+                    "raw_class_en": row["raw_class_en"] or "",
+                    "error": row["error"],
+                    "source_path": row["source_path"],
+                }
+            )
+        return out
+    finally:
+        conn.close()
+
+
 def get_latest_health(settings: Settings) -> HealthSnapshot:
     init_db(settings)
     conn = _connect(settings.sqlite_path)
diff --git a/fish_api/app/deps.py b/fish_api/app/deps.py
index 8d322d1..ff7e423 100644
--- a/fish_api/app/deps.py
+++ b/fish_api/app/deps.py
@@ -1,6 +1,6 @@
 from __future__ import annotations
 
-from typing import Annotated, Optional
+from typing import Optional
 
 from fastapi import Depends, Header, HTTPException
 
@@ -8,8 +8,8 @@ from app.settings import Settings, get_settings
 
 
 def require_ingest_auth(
-    settings: Annotated[Settings, Depends(get_settings)],
-    x_api_key: Annotated[Optional[str], Header(alias="X-API-Key")] = None,
+    settings: Settings = Depends(get_settings),
+    x_api_key: Optional[str] = Header(None, alias="X-API-Key"),
 ) -> None:
     expected = settings.ingest_api_key.strip()
     if not expected:
diff --git a/fish_api/app/logging_config.py b/fish_api/app/logging_config.py
index fac46f6..4bfb726 100644
--- a/fish_api/app/logging_config.py
+++ b/fish_api/app/logging_config.py
@@ -6,7 +6,7 @@ import json
 import logging
 import os
 import sys
-from typing import Any
+from typing import Any, Optional, Union
 
 from loguru import logger
 
@@ -17,7 +17,7 @@ def format_json_pretty(
     data: Any,
     *,
     indent: int = 2,
-    max_chars: int | None = 24_000,
+    max_chars: Optional[int] = 24_000,
 ) -> str:
     """将对象格式化为带缩进、保留中文的 JSON 字符串，供 loguru 多行输出。
 
@@ -37,7 +37,7 @@ class InterceptHandler(logging.Handler):
 
     def emit(self, record: logging.LogRecord) -> None:
         try:
-            level: str | int = logger.level(record.levelname).name
+            level = logger.level(record.levelname).name  # type: Union[str, int]
         except ValueError:
             level = record.levelno
         logger.opt(depth=6, exception=record.exc_info).log(level, record.getMessage())
diff --git a/fish_api/app/main.py b/fish_api/app/main.py
index 3442c84..2c171d2 100644
--- a/fish_api/app/main.py
+++ b/fish_api/app/main.py
@@ -2,13 +2,14 @@ from __future__ import annotations
 
 import asyncio
 from contextlib import asynccontextmanager
+from typing import List
 
 from fastapi import FastAPI
 
 from app.logging_config import setup_logging
 from app.media_static import MediaStaticFiles
 from app.db import init_db
-from app.routers import biomass, debug, ingest
+from app.routers import biomass, debug, ingest, zed
 from app.services.action_watch import run_action_watch_loop
 from app.services.measure_watch import run_measure_watch_loop
 from app.settings import get_settings
@@ -23,11 +24,12 @@ async def lifespan(app: FastAPI):
     init_db(s)
     s.media_root.mkdir(parents=True, exist_ok=True)
     s.stream_tmp_dir.mkdir(parents=True, exist_ok=True)
-    tasks: list[asyncio.Task[None]] = []
+    tasks = []  # type: List[asyncio.Task]
     if s.action_watch_dir is not None:
         tasks.append(asyncio.create_task(run_action_watch_loop(s)))
     if s.measure_watch_dir is not None:
         tasks.append(asyncio.create_task(run_measure_watch_loop(s)))
+
     yield
     for t in tasks:
         t.cancel()
@@ -42,6 +44,7 @@ app = FastAPI(title="Fish API", lifespan=lifespan)
 app.include_router(ingest.router)
 app.include_router(biomass.router)
 app.include_router(debug.router)
+app.include_router(zed.router)
 
 _settings = get_settings()
 _settings.media_root.mkdir(parents=True, exist_ok=True)
@@ -63,5 +66,6 @@ async def root():
         "biomass_water_video": "/api/v1/biomass/water/video/",
         "biomass_sonar_video": "/api/v1/biomass/sonar/video/",
         "debug_measure": "/api/v1/debug/meause",
-        "note": "若配置了 ACTION_WATCH_DIR / MEASURE_WATCH_DIR，启动后会后台监控对应目录。",
+        "zed_recording": "/api/v1/zed/recording/start|stop|status",
+        "note": "若配置了 ACTION_WATCH_DIR / MEASURE_WATCH_DIR，启动后会后台监控对应目录。ZED 分段录制由独立进程/脚本负责，不由 fish_api 启停；可选 HTTP /api/v1/zed/recording/start|stop|status。",
     }
diff --git a/fish_api/app/measure_debug_cli.py b/fish_api/app/measure_debug_cli.py
index ba703ac..4302333 100644
--- a/fish_api/app/measure_debug_cli.py
+++ b/fish_api/app/measure_debug_cli.py
@@ -1,6 +1,6 @@
-"""CLI：对单条 fish 跑与 fish_api / measure-watch 相同的 FishMeasure 子进程（不写 SQLite、不发布 /media）。
+"""CLI：对单条 fish 跑 FishMeasure 子进程（不写 SQLite、不发布 /media）。
 
-使用 ``run_measure_batch_subprocess``，与 ``run_full_measure_batch`` 中的推理步骤一致。
+使用 ``run_measure_batch_subprocess``（合并点云）。线上 ``measure_watch`` 为逐段 ``run_full_measure``，齐套后再 aggregate final。
 """
 
 from __future__ import annotations
diff --git a/fish_api/app/media_static.py b/fish_api/app/media_static.py
index ef11341..3edc6b8 100644
--- a/fish_api/app/media_static.py
+++ b/fish_api/app/media_static.py
@@ -8,7 +8,7 @@ from __future__ import annotations
 import os
 from mimetypes import guess_type
 from pathlib import Path
-from typing import Union
+from typing import Dict, List, Optional, Tuple, Union
 
 from starlette.datastructures import Headers
 from starlette.responses import FileResponse, Response, StreamingResponse
@@ -32,7 +32,7 @@ def _media_type_for_file(path: PathLike) -> str:
     return "application/octet-stream"
 
 
-def _parse_range_header(range_header: str, file_size: int) -> list[tuple[int, int]]:
+def _parse_range_header(range_header: str, file_size: int) -> List[Tuple[int, int]]:
     """解析 HTTP Range 头，返回 (start, end) 列表。"""
     if not range_header.startswith("bytes="):
         return []
@@ -72,10 +72,10 @@ class _RangeFileResponse(Response):
     def __init__(
         self,
         path: PathLike,
-        media_type: str | None = None,
-        stat_result: os.stat_result | None = None,
-        headers: dict | None = None,
-        range_header: str | None = None,
+        media_type: Optional[str] = None,
+        stat_result: Optional[os.stat_result] = None,
+        headers: Optional[Dict] = None,
+        range_header: Optional[str] = None,
     ):
         self.path = path
         self.stat_result = stat_result or os.stat(path)
diff --git a/fish_api/app/prestart_fresh.py b/fish_api/app/prestart_fresh.py
index 106d983..07d0165 100644
--- a/fish_api/app/prestart_fresh.py
+++ b/fish_api/app/prestart_fresh.py
@@ -16,12 +16,13 @@ from __future__ import annotations
 
 import os
 from pathlib import Path
+from typing import Optional
 
 from app.db import _safe_rm_tree, remove_sqlite_database_files, reset_delivery_client_progress
 from app.settings import get_settings
 
 
-def _rm_legacy_json(path: Path | None) -> None:
+def _rm_legacy_json(path: Optional[Path]) -> None:
     if path is None:
         return
     try:
diff --git a/fish_api/app/routers/biomass.py b/fish_api/app/routers/biomass.py
index a59e010..2d44461 100644
--- a/fish_api/app/routers/biomass.py
+++ b/fish_api/app/routers/biomass.py
@@ -1,6 +1,6 @@
 from __future__ import annotations
 
-from typing import Optional
+from typing import Any, Dict, List, Optional
 
 from fastapi import APIRouter, Depends, Header, Query
 from starlette.responses import JSONResponse
@@ -16,10 +16,27 @@ router = APIRouter(prefix="/api/v1/biomass", tags=["biomass"])
 HEADER_BIOMASS_NEW = "X-Fish-Biomass-New"
 
 
-def _new_headers(has_new: bool) -> dict[str, str]:
+def _new_headers(has_new: bool) -> Dict[str, str]:
     return {HEADER_BIOMASS_NEW: "1" if has_new else "0"}
 
 
+# GET /real/camera/ 的 data.result[] 仅含：id, type, length, weight, date（与客户端约定一致）
+_BIOMASS_CAMERA_RESULT_KEYS = ("id", "type", "length", "weight", "date")
+
+
+def _biomass_camera_result_rows(result: Any) -> List[Dict[str, Any]]:
+    if not isinstance(result, list):
+        return []
+    out: List[Dict[str, Any]] = []
+    for it in result:
+        if not isinstance(it, dict):
+            continue
+        row = {k: it[k] for k in _BIOMASS_CAMERA_RESULT_KEYS if k in it}
+        if row:
+            out.append(row)
+    return out
+
+
 def _resolve_client_id(
     x_fish_client_id: Optional[str] = Header(None, alias="X-Fish-Client-Id"),
     client_id: Optional[str] = Query(
@@ -69,7 +86,7 @@ async def get_real_camera(
         )
 
     payload: dict = {
-        "result": m.result,
+        "result": _biomass_camera_result_rows(m.result),
         "video_left": m.video_left,
         "video_right": m.video_right,
     }
diff --git a/fish_api/app/routers/debug.py b/fish_api/app/routers/debug.py
index 8d21f4e..2bbb8a4 100644
--- a/fish_api/app/routers/debug.py
+++ b/fish_api/app/routers/debug.py
@@ -4,7 +4,7 @@ from __future__ import annotations
 
 from fastapi import APIRouter, Depends
 
-from app.db import list_all_measure_snapshots
+from app.db import list_all_health_snapshots, list_all_measure_snapshots
 from app.settings import Settings, get_settings
 
 router = APIRouter(prefix="/api/v1/debug", tags=["debug"])
@@ -16,3 +16,11 @@ async def debug_list_measure_results(settings: Settings = Depends(get_settings))
     """列出 SQLite 中已保存的全部 FishMeasure 计算结果（含每条鱼的 result / pred / star 等）。"""
     items = list_all_measure_snapshots(settings)
     return {"count": len(items), "items": items}
+
+
+@router.get("/aciton")
+@router.get("/action")
+async def debug_list_action_results(settings: Settings = Depends(get_settings)):
+    """列出 SQLite 中已保存的全部 FishAction 健康结果（含切片 source_path）。"""
+    items = list_all_health_snapshots(settings)
+    return {"count": len(items), "items": items}
diff --git a/fish_api/app/routers/ingest.py b/fish_api/app/routers/ingest.py
index 068b09b..88ae8fd 100644
--- a/fish_api/app/routers/ingest.py
+++ b/fish_api/app/routers/ingest.py
@@ -5,6 +5,7 @@ from pathlib import Path
 
 from fastapi import APIRouter, BackgroundTasks, Depends, HTTPException, Request, Response
 
+from app.compat import to_thread
 from app.db import (
     health_snapshot_deliverable,
     measure_snapshot_deliverable,
@@ -30,9 +31,12 @@ async def _measure_job_serial(svo_path: Path, settings: Settings) -> None:
     async with app_state.measure_lock:
         app_state.measure_status = "running"
         try:
-            snap = await asyncio.to_thread(
-                measure_svc.run_full_measure, svo_path, settings
-            )
+
+            def _run():
+                with app_state.measure_thread_lock:
+                    return measure_svc.run_full_measure(svo_path, settings)
+
+            snap = await to_thread(_run)
             if measure_snapshot_deliverable(snap):
                 save_measure_snapshot(
                     settings, snap, source_path=str(svo_path.resolve())
@@ -47,7 +51,7 @@ async def _action_job_serial(mp4_path: Path, settings: Settings) -> None:
         app_state.action_status = "running"
         try:
             # 返回 (第一个快照, 所有切片快照列表)
-            first_snap, all_snaps = await asyncio.to_thread(
+            first_snap, all_snaps = await to_thread(
                 action_svc.run_full_action, mp4_path, settings
             )
 
diff --git a/fish_api/app/routers/zed.py b/fish_api/app/routers/zed.py
new file mode 100644
index 0000000..f0eba33
--- /dev/null
+++ b/fish_api/app/routers/zed.py
@@ -0,0 +1,125 @@
+"""ZED 分段录制 HTTP 控制（响应格式对齐 biomass：code/msg/data；不校验 ingest API Key）。"""
+
+from __future__ import annotations
+
+from typing import Any, Dict, Optional
+
+from fastapi import APIRouter, Body, Depends
+from pydantic import BaseModel, Field
+from starlette.responses import JSONResponse
+
+from app.services.zed_recording_control import (
+    start_zed_recording,
+    stop_zed_recording,
+    zed_recording_is_running,
+)
+from app.settings import Settings, get_settings
+from app.state import app_state
+
+router = APIRouter(prefix="/api/v1/zed/recording", tags=["zed-recording"])
+
+
+def _ok(data: Dict[str, Any], msg: str = "成功") -> JSONResponse:
+    return JSONResponse(
+        content={
+            "code": 200,
+            "msg": msg,
+            "data": data,
+        }
+    )
+
+
+def _err(
+    http_status: int,
+    *,
+    code: int,
+    msg: str,
+    data: Optional[Dict[str, Any]] = None,
+) -> JSONResponse:
+    body: Dict[str, Any] = {"code": code, "msg": msg}
+    if data is not None:
+        body["data"] = data
+    return JSONResponse(status_code=http_status, content=body)
+
+
+class ZedRecordingStartRequest(BaseModel):
+    """可选覆盖分段时长；``fish_id`` 由服务端每次启动在库中递增分配。"""
+
+    segment_sec: Optional[float] = Field(
+        None,
+        ge=1.0,
+        description="每段时长（秒）；缺省使用 ZED_SVO_SEGMENT_SEC",
+    )
+
+
+@router.get("/status")
+async def zed_recording_status() -> JSONResponse:
+    data: Dict[str, Any] = {"running": zed_recording_is_running()}
+    if app_state.zed_recording_fish_id is not None:
+        data["fish_id"] = app_state.zed_recording_fish_id
+    return _ok(data)
+
+
+@router.post("/start")
+async def zed_recording_start(
+    body: ZedRecordingStartRequest = Body(default_factory=ZedRecordingStartRequest),
+    settings: Settings = Depends(get_settings),
+) -> JSONResponse:
+    ok, msg, fish_id, _ = start_zed_recording(
+        settings,
+        segment_sec=body.segment_sec,
+    )
+    if not ok:
+        if msg == "already_running":
+            return _err(
+                409,
+                code=409,
+                msg="已在录制中",
+                data={"ok": False, "message": "已在录制中"},
+            )
+        if msg.startswith("session_db_error:"):
+            return _err(
+                500,
+                code=500,
+                msg="写入录制会话失败",
+                data={"ok": False, "message": msg},
+            )
+        return _err(
+            500,
+            code=500,
+            msg="启动录制失败",
+            data={"ok": False, "message": msg},
+        )
+
+    return _ok(
+        {
+            "ok": True,
+            "message": "录制开始",
+            "fish_id": fish_id,
+        }
+    )
+
+
+@router.post("/stop")
+async def zed_recording_stop(settings: Settings = Depends(get_settings)) -> JSONResponse:
+    ok, msg, fish_id = stop_zed_recording(settings)
+    if not ok:
+        if msg == "not_running":
+            return _ok({"ok": False, "message": "未在录制"}, msg="未在录制")
+        if msg == "stop_timeout":
+            return _err(
+                504,
+                code=504,
+                msg="停止超时",
+                data={"ok": False, "message": "停止超时"},
+            )
+        return _err(
+            500,
+            code=500,
+            msg="停止录制失败",
+            data={"ok": False, "message": msg},
+        )
+    out: Dict[str, Any] = {"ok": True, "message": "录制停止"}
+    if fish_id is not None:
+        out["fish_id"] = fish_id
+    return _ok(out)
diff --git a/fish_api/app/services/action.py b/fish_api/app/services/action.py
index 7718a70..ed3eb7a 100644
--- a/fish_api/app/services/action.py
+++ b/fish_api/app/services/action.py
@@ -6,7 +6,7 @@ import sys
 import tempfile
 from datetime import datetime, timezone
 from pathlib import Path
-from typing import List
+from typing import List, Tuple
 
 from app.logging_config import format_json_pretty
 from app.services.video_slice import get_video_duration, slice_video
@@ -102,7 +102,7 @@ def run_action_subprocess(mp4_path: Path, settings: Settings) -> str:
         Path(out_json).unlink(missing_ok=True)
 
 
-def run_full_action(mp4_path: Path, settings: Settings) -> tuple[HealthSnapshot, list[HealthSnapshot]]:
+def run_full_action(mp4_path: Path, settings: Settings) -> Tuple[HealthSnapshot, List[HealthSnapshot]]:
     """运行 FishAction 健康检测。如果视频较长，会自动切片后分别检测。
 
     每个切片被视为独立的视频，返回所有切片的结果列表。
@@ -112,7 +112,7 @@ def run_full_action(mp4_path: Path, settings: Settings) -> tuple[HealthSnapshot,
         settings: 应用配置
 
     Returns:
-        tuple[HealthSnapshot, list[HealthSnapshot]]: (第一个切片/完整视频的快照, 所有切片快照列表)
+        Tuple[HealthSnapshot, List[HealthSnapshot]]: (第一个切片/完整视频的快照, 所有切片快照列表)
             - 如果视频被切片：返回 (第一个切片, 所有切片列表)
             - 如果视频未被切片：返回 (完整视频快照, [完整视频快照])
     """
@@ -141,7 +141,7 @@ def run_full_action(mp4_path: Path, settings: Settings) -> tuple[HealthSnapshot,
             )
 
             # 处理每个切片
-            all_snaps: list[HealthSnapshot] = []
+            all_snaps = []  # type: List[HealthSnapshot]
             for i, slice_file in enumerate(slice_files):
                 start_time = i * DEFAULT_SLICE_DURATION
                 end_time = min(start_time + DEFAULT_SLICE_DURATION, duration)
diff --git a/fish_api/app/services/action_watch.py b/fish_api/app/services/action_watch.py
index a53fc6b..fc42ac5 100644
--- a/fish_api/app/services/action_watch.py
+++ b/fish_api/app/services/action_watch.py
@@ -2,10 +2,12 @@ from __future__ import annotations
 
 import asyncio
 from pathlib import Path
-from typing import Dict, Set
+from typing import Dict, List, Set, Tuple
 
 from loguru import logger
 
+from app.compat import to_thread
+
 from app.db import (
     add_watch_processed,
     health_snapshot_deliverable,
@@ -28,7 +30,7 @@ def _state_path(settings: Settings) -> Path:
     return settings.action_watch_dir / ".fishaction_watch_processed.json"
 
 
-def iter_mp4(watch_dir: Path, recursive: bool) -> list[Path]:
+def iter_mp4(watch_dir: Path, recursive: bool) -> List[Path]:
     if recursive:
         return sorted(
             p
@@ -56,7 +58,7 @@ async def _run_inference_and_state(
         app_state.action_status = "running"
         try:
             # 返回 (第一个快照, 所有切片快照列表)
-            first_snap, all_snaps = await asyncio.to_thread(
+            first_snap, all_snaps = await to_thread(
                 action_svc.run_full_action, mp4, settings
             )
 
@@ -98,7 +100,7 @@ async def _run_inference_and_state(
 async def watch_tick(
     settings: Settings,
     processed: Set[str],
-    stability: Dict[str, tuple[int, int]],
+    stability: Dict[str, Tuple[int, int]],
     state_file: Path,
 ) -> bool:
     """处理一轮目录扫描；若处理了至少一个文件返回 True。"""
@@ -149,7 +151,7 @@ async def run_action_watch_loop(settings: Settings) -> None:
         if settings.action_watch_use_state_file
         else set()
     )
-    stability: Dict[str, tuple[int, int]] = {}
+    stability = {}  # type: Dict[str, Tuple[int, int]]
 
     logger.info(
         "[action-watch] watching {} (poll={}s, stable_polls={}, state={} {})",
diff --git a/fish_api/app/services/measure.py b/fish_api/app/services/measure.py
index 52239f1..96e014e 100644
--- a/fish_api/app/services/measure.py
+++ b/fish_api/app/services/measure.py
@@ -3,6 +3,7 @@ from __future__ import annotations
 import json
 import math
 import os
+import statistics
 import re
 import shutil
 import subprocess
@@ -942,13 +943,208 @@ def _find_preview_videos(output_dir: Path) -> Tuple[Optional[Path], Optional[Pat
     return None, None
 
 
+def _snapshot_length_mm(snap: MeasureSnapshot) -> Optional[float]:
+    if snap.result and isinstance(snap.result[0], dict):
+        return _row_finite_field(snap.result[0], "length")
+    return None
+
+
+def _materialize_weight_json_for_svo(
+    svo_path: Path,
+    settings: Settings,
+    *,
+    fish_id: str,
+    temp_dir: Path,
+) -> Path:
+    root = settings.measure_output_root / f"fish{fish_id}"
+    candidate = root / svo_path.stem / "weight_prediction.json"
+    if candidate.is_file():
+        return candidate
+    data = _load_weight_json(svo_path, settings, output_root=root)
+    temp_dir.mkdir(parents=True, exist_ok=True)
+    materialized = temp_dir / f"{svo_path.stem}_weight_prediction.json"
+    materialized.write_text(
+        json.dumps(data, ensure_ascii=False, indent=2),
+        encoding="utf-8",
+    )
+    return materialized
+
+
+def _run_generate_video_with_labels_cli(
+    *,
+    settings: Settings,
+    svo_path: Path,
+    output_dir: Path,
+    weight_json: Path,
+    summary_weight_g: Optional[float],
+    summary_length_mm: Optional[float],
+    summary_star: bool,
+    output_video_name: str,
+) -> Path:
+    script = settings.fish_measure_root / "generate_video_with_labels.py"
+    if not script.is_file():
+        raise FileNotFoundError(f"Missing FishMeasure preview script: {script}")
+
+    output_dir.mkdir(parents=True, exist_ok=True)
+    cmd = [
+        _py_exe(settings),
+        str(script),
+        "--svo",
+        str(svo_path.resolve()),
+        "--save-output",
+        str(output_dir.resolve()),
+        "--weight-json",
+        str(weight_json.resolve()),
+        "--conf",
+        str(settings.measure_yolo_conf),
+        "--sam-device",
+        settings.sam_device,
+        "--output-video-name",
+        output_video_name,
+    ]
+    if settings.predict_show_large_labels_at_top_right:
+        cmd.append("--show-large-labels-at-top-right")
+    if summary_weight_g is not None and math.isfinite(summary_weight_g):
+        cmd.extend(["--summary-weight-g", str(float(summary_weight_g))])
+    if summary_length_mm is not None and math.isfinite(summary_length_mm):
+        cmd.extend(["--summary-length-mm", str(float(summary_length_mm))])
+    if summary_star:
+        cmd.append("--summary-star")
+
+    proc = run_subprocess_with_log(
+        cmd,
+        cwd=str(settings.fish_measure_root),
+        env=os.environ.copy(),
+        log_name="FishMeasure",
+        stream_to_logger=False,
+    )
+    if proc.returncode != 0:
+        err = proc.stdout or ""
+        raise RuntimeError(
+            f"generate_video_with_labels.py failed ({proc.returncode}): {err[-4000:]}"
+        )
+
+    video_path = output_dir / "images" / output_video_name
+    if not video_path.is_file():
+        raise FileNotFoundError(f"Expected generated preview video at {video_path}")
+    return video_path
+
+
+def _concat_preview_videos(inputs: List[Path], dst: Path) -> bool:
+    if not inputs:
+        return False
+    try:
+        import cv2
+    except Exception:
+        return False
+
+    writer = None
+    target_size: Optional[Tuple[int, int]] = None
+    fps = 10.0
+    wrote_any = False
+    try:
+        dst.parent.mkdir(parents=True, exist_ok=True)
+        for src in inputs:
+            if not src.is_file():
+                continue
+            cap = cv2.VideoCapture(str(src))
+            if not cap.isOpened():
+                continue
+            cur_fps = float(cap.get(cv2.CAP_PROP_FPS) or 0.0)
+            width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH) or 0)
+            height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT) or 0)
+            if width <= 0 or height <= 0:
+                cap.release()
+                continue
+            if writer is None:
+                target_size = (width, height)
+                fps = cur_fps if cur_fps > 0.0 else 10.0
+                writer = cv2.VideoWriter(
+                    str(dst),
+                    cv2.VideoWriter_fourcc(*"mp4v"),
+                    fps,
+                    target_size,
+                )
+            while True:
+                ok, frame = cap.read()
+                if not ok:
+                    break
+                if target_size is not None and (
+                    frame.shape[1] != target_size[0] or frame.shape[0] != target_size[1]
+                ):
+                    frame = cv2.resize(frame, target_size)
+                writer.write(frame)
+                wrote_any = True
+            cap.release()
+    finally:
+        if writer is not None:
+            writer.release()
+    return wrote_any and dst.is_file() and dst.stat().st_size > 0
+
+
+def generate_aggregate_preview_media(
+    contributing_svos: List[Path],
+    snap: MeasureSnapshot,
+    fish_id: str,
+    settings: Settings,
+    *,
+    final_key: str,
+) -> Tuple[str, str]:
+    """为 final 快照生成并发布串接预览视频，返回 (video_left, video_right)。"""
+    if not contributing_svos or not snap.result:
+        return "", ""
+
+    summary_weight_g = snap.pred if snap.pred is not None and math.isfinite(snap.pred) else None
+    summary_length_mm = _snapshot_length_mm(snap)
+    root = settings.measure_output_root / f"fish{fish_id}"
+    temp_root = root / "__aggregate_preview__" / _safe_media_prefix(final_key)
+    temp_root.mkdir(parents=True, exist_ok=True)
+
+    part_videos: List[Path] = []
+    weight_json_cache_dir = temp_root / "__weight_json_cache__"
+    for idx, svo in enumerate(contributing_svos, start=1):
+        part_dir = temp_root / f"part_{idx:02d}_{_safe_media_prefix(svo.stem)}"
+        weight_json = _materialize_weight_json_for_svo(
+            svo,
+            settings,
+            fish_id=fish_id,
+            temp_dir=weight_json_cache_dir,
+        )
+        part_video = _run_generate_video_with_labels_cli(
+            settings=settings,
+            svo_path=svo,
+            output_dir=part_dir,
+            weight_json=weight_json,
+            summary_weight_g=summary_weight_g,
+            summary_length_mm=summary_length_mm,
+            summary_star=bool(snap.star),
+            output_video_name=f"final_part_{idx:02d}.mp4",
+        )
+        part_videos.append(part_video)
+
+    if not part_videos:
+        return "", ""
+
+    final_preview = temp_root / "final_preview.mp4"
+    if not _concat_preview_videos(part_videos, final_preview):
+        raise RuntimeError(f"Failed to concatenate final preview videos into {final_preview}")
+
+    return _publish_media(
+        final_preview,
+        final_preview,
+        settings,
+        f"{datetime.now(timezone.utc).strftime('%Y%m%dT%H%M%S')}_final_fish{fish_id}_{_safe_media_prefix(final_key)}",
+    )
+
+
 def _split_sbs_video(src: Path, left_dst: Path, right_dst: Path) -> bool:
     """Split a side-by-side stereo video (W x H where W == 2*H_single) into left/right halves.
 
     Returns True if split succeeded, False otherwise (caller should fall back to copy).
     """
     ffmpeg_path = _get_ffmpeg_path()
-    ffprobe_path = str(Path(ffmpeg_path).parent / "ffprobe")
+    from app.services.video_slice import _get_ffprobe_path
+    ffprobe_path = _get_ffprobe_path()
     probe = subprocess.run(
         [
             ffprobe_path, "-v", "quiet", "-print_format", "json",
@@ -990,21 +1186,19 @@ def _split_sbs_video(src: Path, left_dst: Path, right_dst: Path) -> bool:
 
 
 def _get_ffmpeg_path() -> str:
-    """获取可用的 ffmpeg 路径。优先使用项目配置的 ffmpeg。"""
-    # 优先使用项目目录下的 ffmpeg
-    project_ffmpeg = Path("/home/ubuntu/projects/FishServer/tools/ffmpeg/bin/ffmpeg")
-    if project_ffmpeg.is_file():
-        return str(project_ffmpeg)
-    # 尝试系统路径
+    """获取可用的 ffmpeg 路径。优先使用 .env 中 FFMPEG_PATH 配置。"""
+    from app.settings import get_settings
+    configured = get_settings().ffmpeg_path.strip()
+    if configured and Path(configured).is_file():
+        return configured
     system_paths = ["/usr/bin/ffmpeg", "/usr/local/bin/ffmpeg"]
     for path in system_paths:
         if Path(path).is_file():
             return path
-    # 回退到 PATH 中的 ffmpeg
     return "ffmpeg"
 
 
-def _get_h264_encoder() -> tuple[str, list[str], str]:
+def _get_h264_encoder() -> Tuple[str, List[str], str]:
     """检测可用的H.264编码器，返回 (encoder_name, options, ffmpeg_path)。
 
     优先使用 libx264（纯软件，最可靠），硬件编码器需要实际测试才能确认可用。
@@ -1062,9 +1256,10 @@ def _transcode_with_x264(src: Path, dst: Path) -> bool:
         return False
 
     # 首先用 ffprobe 获取视频信息
+    from app.services.video_slice import _get_ffprobe_path
     try:
         probe = subprocess.run(
-            ["ffprobe", "-v", "quiet", "-print_format", "json", "-show_streams", str(src)],
+            [_get_ffprobe_path(), "-v", "quiet", "-print_format", "json", "-show_streams", str(src)],
             capture_output=True, text=True, timeout=10
         )
         import json as _json
@@ -1373,7 +1568,7 @@ def build_measure_snapshot(
         updated_at=datetime.now(timezone.utc),
         raw_prediction_path=str(root / svo_path.stem / "weight_prediction.json"),
         pred=pred_weight,
-        star=is_confident,
+        star=False,
         calculation_log=calc_log,
     )
 
@@ -1537,6 +1732,209 @@ def build_measure_snapshot_batch(
         updated_at=datetime.now(timezone.utc),
         raw_prediction_path=str(combined_json) if combined_json.is_file() else None,
         pred=pred_weight,
-        star=is_confident,
+        star=False,
         calculation_log=calc_log,
     )
+
+
+def _row_finite_field(d: dict, key: str) -> Optional[float]:
+    v = d.get(key)
+    if v is None:
+        return None
+    try:
+        x = float(v)
+        return x if math.isfinite(x) else None
+    except (TypeError, ValueError):
+        return None
+
+
+def tag_measure_snapshot_meta(
+    snap: MeasureSnapshot,
+    *,
+    measurement_phase: str,
+    fish_folder: str = "",
+    segment_source: str = "",
+) -> MeasureSnapshot:
+    """仅在 ``calculation_log`` 中记录 segment/final 元信息；不向 ``result`` 行追加字段（保持 biomass 接口约定）。"""
+    lines = [
+        f"# measurement_phase: {measurement_phase}",
+        f"# fish_folder: {fish_folder}",
+    ]
+    if segment_source:
+        lines.append(f"# segment_source: {segment_source}")
+    meta = "\n".join(lines) + "\n"
+    snap.calculation_log = meta + (snap.calculation_log or "")
+    return snap
+
+
+def _segment_weight_g(snap: MeasureSnapshot) -> Optional[float]:
+    if snap.pred is not None and math.isfinite(snap.pred):
+        return float(snap.pred)
+    if snap.result and isinstance(snap.result[0], dict):
+        return _row_finite_field(snap.result[0], "weight")
+    return None
+
+
+def _segment_length_mm(snap: MeasureSnapshot) -> Optional[float]:
+    if snap.result and isinstance(snap.result[0], dict):
+        return _row_finite_field(snap.result[0], "length")
+    return None
+
+
+def _aggregate_numeric(values: List[float], mode: str) -> float:
+    if not values:
+        return float("nan")
+    m = (mode or "median").strip().lower()
+    if m not in ("median", "mean", "trimmed_mean"):
+        m = "median"
+    if m == "mean":
+        return float(sum(values) / len(values))
+    if m == "trimmed_mean" and len(values) >= 3:
+        s = sorted(values)
+        core = s[1:-1]
+        return float(sum(core) / len(core))
+    return float(statistics.median(values))
+
+
+def _final_star_from_contributing_svos(
+    contributing_svos: List[Path], fish_id: str, settings: Settings
+) -> bool:
+    """与逐段曾用 ``all(s.star)`` 等价：每段对 dgcnn_summary 做 ``_star_confident_like_test_dgcnn``。"""
+    if not contributing_svos:
+        return False
+    root = settings.measure_output_root / f"fish{fish_id}"
+    for svo in contributing_svos:
+        try:
+            data = _load_weight_json(svo, settings, output_root=root)
+        except (FileNotFoundError, RuntimeError, OSError, ValueError):
+            return False
+        summary = data.get("dgcnn_summary") or data.get("weight_summary") or {}
+        if not _star_confident_like_test_dgcnn(summary):
+            return False
+    return True
+
+
+def reload_segment_snapshots_for_aggregate(
+    svo_list: List[Path], fish_id: str, settings: Settings
+) -> List[Tuple[Path, MeasureSnapshot]]:
+    """从已写入的 FishMeasure 输出目录重载各段快照（用于齐套 final 聚合，无需重跑子进程）。
+
+    逐段测量失败或未写出 ``weight_prediction.json`` 的段仍会被 watch 标为已处理；
+    此处跳过这些段，仅用成功段聚合，避免 final 整批因单段缺文件而失败。
+
+    返回 ``(svo_path, snapshot)`` 列表，与 ``build_measure_snapshot_aggregate(..., contributing_svos=...)`` 对齐。
+    """
+    root = settings.measure_output_root / f"fish{fish_id}"
+    out: List[Tuple[Path, MeasureSnapshot]] = []
+    for svo in sorted(svo_list, key=lambda x: str(x.resolve())):
+        try:
+            out.append((svo, build_measure_snapshot(svo, settings, output_root=root)))
+        except (FileNotFoundError, RuntimeError) as e:
+            logger.warning(
+                "[FishMeasure] final aggregate skip segment (no usable weight JSON) fish_id={} svo={}: {}",
+                fish_id,
+                svo.name,
+                e,
+            )
+    return out
+
+
+def build_measure_snapshot_aggregate(
+    segments: List[MeasureSnapshot],
+    fish_id: str,
+    settings: Settings,
+    *,
+    contributing_svos: List[Path],
+    fish_folder: str = "",
+    segment_source_paths: str = "",
+) -> MeasureSnapshot:
+    """从逐段 ``MeasureSnapshot`` 聚合为一条 final 快照（无预览视频）。
+
+    ``contributing_svos`` 须与 ``segments`` 等长且顺序一一对应（成功重载的各段 SVO），
+    用于按段 ``dgcnn_summary`` 计算 final ``star``。
+    """
+    if not segments:
+        return MeasureSnapshot(
+            result=[],
+            video_left="",
+            video_right="",
+            error="no_segments",
+        )
+    if len(contributing_svos) != len(segments):
+        raise ValueError(
+            f"contributing_svos length {len(contributing_svos)} != segments length {len(segments)}"
+        )
+
+    mode = (settings.measure_final_aggregate_mode or "median").strip().lower()
+    if mode not in ("median", "mean", "trimmed_mean"):
+        mode = "median"
+
+    ws: List[float] = [w for s in segments if (w := _segment_weight_g(s)) is not None]
+    ls = [ln for s in segments if (ln := _segment_length_mm(s)) is not None]
+    star_final = _final_star_from_contributing_svos(contributing_svos, fish_id, settings)
+
+    if not ws:
+        return MeasureSnapshot(
+            result=[],
+            video_left="",
+            video_right="",
+            error="aggregate_no_segment_weights",
+            calculation_log="# measurement_phase: final\n# error: no segment weights\n",
+        )
+
+    pred_w = _aggregate_numeric(ws, mode)
+    pred_len = _aggregate_numeric(ls, mode) if ls else float("nan")
+
+    if not math.isfinite(pred_w):
+        for s in segments:
+            w = _segment_weight_g(s)
+            if w is not None and math.isfinite(w):
+                pred_w = w
+                break
+    if not math.isfinite(pred_len):
+        for s in segments:
+            ln = _segment_length_mm(s)
+            if ln is not None and math.isfinite(ln):
+                pred_len = ln
+                break
+
+    fid = int(fish_id) if fish_id.isdigit() else 1
+    fish_type = "大黄鱼"
+    if segments[0].result and isinstance(segments[0].result[0], dict):
+        t = segments[0].result[0].get("type")
+        if isinstance(t, str) and t.strip():
+            fish_type = t.strip()
+
+    row: Dict[str, Any] = {
+        "id": fid,
+        "type": fish_type,
+        "weight": str(round(pred_w)) if math.isfinite(pred_w) else "",
+        "length": str(round(pred_len)) if math.isfinite(pred_len) else "",
+        "date": datetime.now(timezone.utc).strftime("%Y-%m-%d"),
+    }
+    result = [row]
+
+    lines = [
+        f"# aggregate_mode: {mode}",
+        f"# segment_count: {len(segments)}",
+        f"# segment_weights_g: {ws}",
+        f"# segment_lengths_mm: {ls}",
+    ]
+    calc_log = "\n".join(lines) + "\n# Aggregated from per-segment former results (measure_watch).\n"
+
+    snap = MeasureSnapshot(
+        result=result,
+        video_left="",
+        video_right="",
+        updated_at=datetime.now(timezone.utc),
+        raw_prediction_path=None,
+        pred=pred_w if math.isfinite(pred_w) else None,
+        star=star_final,
+        calculation_log=calc_log,
+    )
+    return tag_measure_snapshot_meta(
+        snap,
+        measurement_phase="final",
+        fish_folder=fish_folder,
+        segment_source=segment_source_paths,
+    )
diff --git a/fish_api/app/services/measure_watch.py b/fish_api/app/services/measure_watch.py
index cd08308..0aa0035 100644
--- a/fish_api/app/services/measure_watch.py
+++ b/fish_api/app/services/measure_watch.py
@@ -1,13 +1,16 @@
-"""后台轮询目录中的 .svo2，跑 FishMeasure，写入 SQLite（与 ingest 共用）。"""
+"""后台轮询目录中的 .svo2，逐段跑 FishMeasure，齐套后聚合 final（与 ingest 共用 SQLite）。"""
 
 from __future__ import annotations
 
 import asyncio
+import hashlib
 from pathlib import Path
-from typing import Dict, Set
+from typing import Dict, List, Set, Tuple
 
 from loguru import logger
 
+from app.compat import to_thread
+
 from app.db import (
     add_watch_processed,
     load_watch_processed,
@@ -30,143 +33,256 @@ def _state_path(settings: Settings) -> Path:
     return settings.measure_watch_dir / ".fishmeasure_watch_processed.json"
 
 
-def iter_svo2_folders(watch_dir: Path) -> list[tuple[list[Path], str]]:
+def iter_svo2_folders(watch_dir: Path) -> List[Tuple[List[Path], str]]:
     """扫描子文件夹，返回 (svo文件路径列表, fish_id) 列表。
 
-    文件夹命名格式为 fish{N}，如 fish1、fish2 等。
-    每个子文件夹可以包含多个 .svo2 文件（同一条鱼的多段视频），
-    这些 SVO 文件会被批量处理，点云合并后进行重量预测。
+    文件夹命名格式为 fish{N}。每个子文件夹内多个 .svo2 先逐段测量，齐套后再聚合 final。
     """
-    result: list[tuple[list[Path], str]] = []
+    result = []  # type: List[Tuple[List[Path], str]]
     if not watch_dir.is_dir():
         return result
 
     for entry in sorted(watch_dir.iterdir()):
         if not entry.is_dir():
             continue
-        # 从文件夹名提取 fish_id，格式为 fish{N}
         folder_name = entry.name
         if not folder_name.startswith("fish"):
             continue
         try:
-            fish_id = folder_name[4:]  # 去掉 "fish" 前缀
+            fish_id = folder_name[4:]
             if not fish_id.isdigit():
                 continue
         except (IndexError, ValueError):
             continue
 
-        # 在子文件夹中查找所有 .svo2 文件
         svo_files = sorted([
             p for p in entry.iterdir()
             if p.is_file() and p.suffix.lower() == ".svo2"
         ])
         if svo_files:
-            # 返回该文件夹中的所有 SVO 文件，它们将被批量处理
             result.append((svo_files, fish_id))
 
     return result
 
 
-async def _run_measure_and_state(
-    svo_list: list[Path],
+def _final_processed_key(fish_id: str, svo_list: List[Path]) -> str:
+    sig = "|".join(sorted(str(p.resolve()) for p in svo_list))
+    h = hashlib.sha256(sig.encode("utf-8")).hexdigest()[:24]
+    return f"__measure_final__fish{fish_id}:{h}"
+
+
+def _folder_size_tuple(svo_list: List[Path]) -> Tuple[Tuple[str, int], ...]:
+    out: List[Tuple[str, int]] = []
+    for p in sorted(svo_list, key=lambda x: str(x.resolve())):
+        try:
+            st = p.stat()
+            out.append((str(p.resolve()), int(st.st_size)))
+        except OSError:
+            return tuple()
+    return tuple(out)
+
+
+async def _run_single_svo_measure(
+    svo: Path,
     fish_id: str,
     settings: Settings,
     processed: Set[str],
     state_file: Path,
 ) -> None:
-    """批量处理同一条鱼的多个 SVO：合并点云后一次 DGCNN；解析与 test_dgcnn summary 对齐。
-    """
-    if not svo_list:
-        return
+    key = str(svo.resolve())
+    fish_folder = svo.parent.resolve()
+    fish_output_root = settings.measure_output_root / f"fish{fish_id}"
+    fish_output_root.mkdir(parents=True, exist_ok=True)
 
-    # 生成唯一的 key 列表（用于 processed 标记）
-    keys = [str(svo.resolve()) for svo in svo_list]
-    # 检查是否全部已处理
-    if all(key in processed for key in keys):
-        return
-
-    svo_names = ", ".join(svo.name for svo in svo_list)
-    logger.info("[measure-watch] batch inference for fish_id={}: {} SVO(s): {}",
-                fish_id, len(svo_list), svo_names)
+    logger.info(
+        "[measure-watch] segment inference fish_id={} svo={}",
+        fish_id,
+        svo.name,
+    )
 
     async with app_state.measure_lock:
         app_state.measure_status = "running"
         try:
-            # 使用 batch 模式处理所有 SVO，传入 fish_id 作为结果 id
-            snap = await asyncio.to_thread(
-                measure_svc.run_full_measure_batch, svo_list, settings, fish_id
+
+            def _run():
+                with app_state.measure_thread_lock:
+                    return measure_svc.run_full_measure(
+                        svo, settings, output_root=fish_output_root
+                    )
+
+            snap = await to_thread(_run)
+
+            snap = measure_svc.tag_measure_snapshot_meta(
+                snap,
+                measurement_phase="segment",
+                fish_folder=str(fish_folder),
+                segment_source=str(svo.resolve()),
             )
 
             if measure_snapshot_deliverable(snap):
-                # 保存结果，client_id=None 表示对所有客户端可见
-                # fish_id 只用于 result 中的 id 字段，不作为 client_id
-                source_paths = "|".join(keys)  # 合并所有 source_path
                 save_measure_snapshot(
-                    settings, snap, source_path=source_paths, client_id=None
+                    settings,
+                    snap,
+                    source_path=str(svo.resolve()),
+                    client_id=None,
                 )
             else:
                 logger.warning(
-                    "[measure-watch] no deliverable measure rows for fish_id={}, skip SQLite",
+                    "[measure-watch] no deliverable measure rows for fish_id={} svo={}, skip SQLite",
+                    fish_id,
+                    svo.name,
+                )
+
+            app_state.measure_status = "idle"
+            processed.add(key)
+            if settings.measure_watch_use_state_file:
+                add_watch_processed(settings, key, "measure")
+
+            r0 = snap.result[0] if snap.result else {}
+            logger.info(
+                "[measure-watch] segment done: fish_id={} svo={} weight={!r}",
+                fish_id,
+                svo.name,
+                r0.get("weight", ""),
+            )
+
+        except (RuntimeError, FileNotFoundError) as e:
+            logger.warning(
+                "[measure-watch] measure failed fish_id={} svo={}: {}",
+                fish_id,
+                svo.name,
+                e,
+            )
+            app_state.measure_status = "idle"
+            processed.add(key)
+            if settings.measure_watch_use_state_file:
+                add_watch_processed(settings, key, "measure")
+
+        except Exception as e:
+            logger.exception(
+                "[measure-watch] error fish_id={} svo={}: {}",
+                fish_id,
+                svo.name,
+                e,
+            )
+            app_state.measure_status = "idle"
+            processed.add(key)
+            if settings.measure_watch_use_state_file:
+                add_watch_processed(settings, key, "measure")
+
+
+async def _run_final_aggregate(
+    svo_list: List[Path],
+    fish_id: str,
+    settings: Settings,
+    processed: Set[str],
+    state_file: Path,
+    final_key: str,
+) -> None:
+    fish_folder = svo_list[0].parent.resolve()
+
+    logger.info(
+        "[measure-watch] final aggregate fish_id={} {} segment(s)",
+        fish_id,
+        len(svo_list),
+    )
+
+    async with app_state.measure_lock:
+        app_state.measure_status = "running"
+        try:
+
+            def _reload():
+                return measure_svc.reload_segment_snapshots_for_aggregate(
+                    svo_list, fish_id, settings
+                )
+
+            pairs = await to_thread(_reload)
+            contributing_svos = [p[0] for p in pairs]
+            segments = [p[1] for p in pairs]
+            paths_joined = "|".join(sorted(str(p.resolve()) for p in contributing_svos))
+
+            snap = measure_svc.build_measure_snapshot_aggregate(
+                segments,
+                fish_id,
+                settings,
+                contributing_svos=contributing_svos,
+                fish_folder=str(fish_folder),
+                segment_source_paths=paths_joined,
+            )
+
+            if measure_snapshot_deliverable(snap):
+                try:
+                    v_left, v_right = await to_thread(
+                        measure_svc.generate_aggregate_preview_media,
+                        contributing_svos,
+                        snap,
+                        fish_id,
+                        settings,
+                        final_key=final_key,
+                    )
+                    snap.video_left = v_left
+                    snap.video_right = v_right
+                except Exception as e:
+                    logger.warning(
+                        "[measure-watch] final preview generate failed fish_id={}: {}",
+                        fish_id,
+                        e,
+                    )
+                save_measure_snapshot(
+                    settings,
+                    snap,
+                    source_path=f"aggregate:{final_key}",
+                    client_id=None,
+                )
+            else:
+                logger.warning(
+                    "[measure-watch] final not deliverable for fish_id={}, skip SQLite",
                     fish_id,
                 )
 
             app_state.measure_status = "idle"
-
-            # 标记所有 SVO 为已处理
-            for key in keys:
-                processed.add(key)
-                if settings.measure_watch_use_state_file:
-                    add_watch_processed(settings, key, "measure")
+            processed.add(final_key)
+            if settings.measure_watch_use_state_file:
+                add_watch_processed(settings, final_key, "measure")
 
             r0 = snap.result[0] if snap.result else {}
-            w = r0.get("weight", "")
             logger.info(
-                "[measure-watch] done: fish_id={} SVOs={} weight={!r}",
-                fish_id, len(svo_list), w
+                "[measure-watch] final done: fish_id={} weight={!r}",
+                fish_id,
+                r0.get("weight", ""),
             )
 
-        except (RuntimeError, FileNotFoundError) as e:
-            logger.warning("[measure-watch] measure failed for fish_id={}: {}", fish_id, e)
-            app_state.measure_status = "idle"
-            for key in keys:
-                processed.add(key)
-                if settings.measure_watch_use_state_file:
-                    add_watch_processed(settings, key, "measure")
-
         except Exception as e:
-            logger.exception("[measure-watch] error on fish_id={}: {}", fish_id, e)
+            logger.exception(
+                "[measure-watch] final aggregate failed fish_id={}: {}",
+                fish_id,
+                e,
+            )
             app_state.measure_status = "idle"
-            for key in keys:
-                processed.add(key)
-                if settings.measure_watch_use_state_file:
-                    add_watch_processed(settings, key, "measure")
+            processed.add(final_key)
+            if settings.measure_watch_use_state_file:
+                add_watch_processed(settings, final_key, "measure")
 
 
 async def watch_tick(
     settings: Settings,
     processed: Set[str],
-    stability: Dict[str, tuple[int, int]],
+    stability: Dict[str, Tuple[int, int]],
+    final_stability: Dict[str, Tuple[Tuple[Tuple[str, int], ...], int]],
     state_file: Path,
 ) -> bool:
-    """处理一轮目录扫描；若处理了至少一个文件返回 True。
-
-    使用 batch 模式：同一条鱼（fish{N} 文件夹）下的所有 SVO 文件会被一起处理，
-    点云合并后进行重量预测（与 test_dgcnn.sh --batch-root 相同的逻辑）。
-    """
+    """逐段稳定即测量；同一 fish 目录全部段已处理且整体稳定后写 final。"""
     assert settings.measure_watch_dir is not None
     watch_dir = settings.measure_watch_dir
     did = False
     seen_keys: Set[str] = set()
 
-    # 使用新的子文件夹扫描方式，返回 (svo_list, fish_id)
     for svo_list, fish_id in iter_svo2_folders(watch_dir):
         if not svo_list:
             continue
 
-        # 为该 fish 文件夹中的所有 SVO 文件计算稳定性
-        # 只有当所有 SVO 都达到稳定轮询次数时才处理
-        all_stable = True
-        any_new = False
+        fish_folder = svo_list[0].parent
+        folder_key = str(fish_folder.resolve())
 
         for svo in svo_list:
             key = str(svo.resolve())
@@ -174,43 +290,57 @@ async def watch_tick(
 
             if key in processed:
                 continue
-            any_new = True
 
             try:
                 st = svo.stat()
             except OSError:
-                all_stable = False
                 continue
 
             size = int(st.st_size)
             if size <= 0:
                 stability.pop(key, None)
-                all_stable = False
                 continue
 
             last = stability.get(key)
             if last is None or last[0] != size:
                 stability[key] = (size, 1)
-                all_stable = False
             else:
                 _, cnt = last
                 cnt += 1
                 stability[key] = (size, cnt)
-                if cnt < settings.measure_watch_stable_polls:
-                    all_stable = False
+                if cnt >= settings.measure_watch_stable_polls:
+                    await _run_single_svo_measure(
+                        svo, fish_id, settings, processed, state_file
+                    )
+                    stability.pop(key, None)
+                    did = True
 
-        # 如果该文件夹下有新的 SVO 文件且全部达到稳定，则批量处理
-        if any_new and all_stable:
-            await _run_measure_and_state(svo_list, fish_id, settings, processed, state_file)
+        fk = _final_processed_key(fish_id, svo_list)
+        if fk in processed:
+            continue
 
-            # 清理已处理的 SVO 文件的稳定性记录
-            for svo in svo_list:
-                key = str(svo.resolve())
-                stability.pop(key, None)
+        if not all(str(p.resolve()) in processed for p in svo_list):
+            final_stability.pop(folder_key, None)
+            continue
 
-            did = True
+        tup = _folder_size_tuple(svo_list)
+        if not tup:
+            continue
+
+        prev = final_stability.get(folder_key)
+        if prev is None or prev[0] != tup:
+            final_stability[folder_key] = (tup, 1)
+        else:
+            _, c = prev
+            c += 1
+            final_stability[folder_key] = (tup, c)
+            if c >= settings.measure_watch_stable_polls:
+                await _run_final_aggregate(
+                    svo_list, fish_id, settings, processed, state_file, fk
+                )
+                final_stability.pop(folder_key, None)
+                did = True
 
-    # 清理不再看到的文件的稳定性记录
     for k in list(stability.keys()):
         if k not in seen_keys:
             del stability[k]
@@ -231,20 +361,24 @@ async def run_measure_watch_loop(settings: Settings) -> None:
         if settings.measure_watch_use_state_file
         else set()
     )
-    stability: Dict[str, tuple[int, int]] = {}
+    stability = {}  # type: Dict[str, Tuple[int, int]]
+    final_stability = {}  # type: Dict[str, Tuple[Tuple[Tuple[str, int], ...], int]]
 
     logger.info(
-        "[measure-watch] watching {} (poll={}s, stable_polls={}, state={} {})",
+        "[measure-watch] watching {} (poll={}s, stable_polls={}, aggregate={}, state={} {})",
         wd,
         settings.measure_watch_poll_interval,
         settings.measure_watch_stable_polls,
+        settings.measure_final_aggregate_mode,
         "on" if settings.measure_watch_use_state_file else "off",
         state_file if settings.measure_watch_use_state_file else "",
     )
 
     idle_warn_state = IdleWatchWarnState()
     while True:
-        did = await watch_tick(settings, processed, stability, state_file)
+        did = await watch_tick(
+            settings, processed, stability, final_stability, state_file
+        )
         maybe_warn_idle_watch(
             did_work=did,
             log_tag="measure-watch",
diff --git a/fish_api/app/services/sessions.py b/fish_api/app/services/sessions.py
index b155d66..c75009c 100644
--- a/fish_api/app/services/sessions.py
+++ b/fish_api/app/services/sessions.py
@@ -2,11 +2,12 @@ from __future__ import annotations
 
 import uuid
 from pathlib import Path
+from typing import Tuple
 
 PARTIAL_NAME = "upload.partial"
 
 
-def new_session_dir(base: Path) -> tuple[str, Path]:
+def new_session_dir(base: Path) -> Tuple[str, Path]:
     session_id = uuid.uuid4().hex
     d = base / session_id
     d.mkdir(parents=True, exist_ok=True)
diff --git a/fish_api/app/services/sonar_video.py b/fish_api/app/services/sonar_video.py
index de04b6e..3512185 100644
--- a/fish_api/app/services/sonar_video.py
+++ b/fish_api/app/services/sonar_video.py
@@ -10,8 +10,12 @@ import asyncio
 import shutil
 from pathlib import Path
 
+from typing import Optional, Tuple
+
 from loguru import logger
 
+from app.compat import to_thread
+
 from app.services.action_watch import iter_mp4
 from app.services.measure import transcode_src_to_h264_dst
 from app.settings import Settings
@@ -21,7 +25,7 @@ _publish_lock = asyncio.Lock()
 DEFAULT_CLIENT_ID = "default"
 
 # 源路径 + mtime，用于跳过重复转码
-_last_src_key: tuple[str, float] | None = None
+_last_src_key = None  # type: Optional[Tuple[str, float]]
 _cached_public_url: str = ""
 
 
@@ -37,7 +41,7 @@ def _safe_sonar_media_basename(raw: str) -> str:
     return Path(n).name or "biomass_sonar.mp4"
 
 
-def resolve_sonar_video_source(settings: Settings) -> Path | None:
+def resolve_sonar_video_source(settings: Settings) -> Optional[Path]:
     """优先 BIOMASS_SONAR_VIDEO_SOURCE；否则在 BIOMASS_SONAR_VIDEO_DIR 中取 mtime 最新的 .mp4。"""
     cfg = settings.biomass_sonar_video_source
     if cfg is not None:
@@ -71,13 +75,13 @@ async def _publish_video(
     tmp.unlink(missing_ok=True)
 
     try:
-        ok = await asyncio.to_thread(transcode_src_to_h264_dst, src, tmp)
+        ok = await to_thread(transcode_src_to_h264_dst, src, tmp)
         if ok and tmp.is_file() and tmp.stat().st_size > 0:
             tmp.replace(dst)
             logger.info("[sonar-video] published H.264: {} -> {}", src.name, dst.name)
         else:
             tmp.unlink(missing_ok=True)
-            await asyncio.to_thread(shutil.copy2, src, dst)
+            await to_thread(shutil.copy2, src, dst)
             logger.warning(
                 "[sonar-video] transcode failed, copied raw: {} -> {}",
                 src.name,
diff --git a/fish_api/app/services/video_slice.py b/fish_api/app/services/video_slice.py
index d7b8b34..a51596f 100644
--- a/fish_api/app/services/video_slice.py
+++ b/fish_api/app/services/video_slice.py
@@ -5,16 +5,17 @@ from __future__ import annotations
 import subprocess
 import tempfile
 from pathlib import Path
-from typing import List, Tuple
+from typing import List, Optional, Tuple
 
 from loguru import logger
 
 
 def _get_ffmpeg_path() -> str:
-    """获取可用的 ffmpeg 路径。"""
-    project_ffmpeg = Path("/home/ubuntu/projects/FishServer/tools/ffmpeg/bin/ffmpeg")
-    if project_ffmpeg.is_file():
-        return str(project_ffmpeg)
+    """获取可用的 ffmpeg 路径。优先使用 .env 中 FFMPEG_PATH 配置。"""
+    from app.settings import get_settings
+    configured = get_settings().ffmpeg_path.strip()
+    if configured and Path(configured).is_file():
+        return configured
     system_paths = ["/usr/bin/ffmpeg", "/usr/local/bin/ffmpeg"]
     for path in system_paths:
         if Path(path).is_file():
@@ -23,11 +24,15 @@ def _get_ffmpeg_path() -> str:
 
 
 def _get_ffprobe_path() -> str:
-    """获取可用的 ffprobe 路径。"""
+    """获取可用的 ffprobe 路径。从 ffmpeg 同目录推导，回退到系统 PATH。"""
     ffmpeg_path = Path(_get_ffmpeg_path())
-    ffprobe = ffmpeg_path.parent / "ffprobe"
-    if ffprobe.is_file():
-        return str(ffprobe)
+    if ffmpeg_path.parent.name:
+        sibling = ffmpeg_path.parent / "ffprobe"
+        if sibling.is_file():
+            return str(sibling)
+    for p in ["/usr/bin/ffprobe", "/usr/local/bin/ffprobe"]:
+        if Path(p).is_file():
+            return p
     return "ffprobe"
 
 
@@ -58,7 +63,7 @@ def get_video_duration(video_path: Path) -> float:
 def slice_video(
     video_path: Path,
     slice_duration: float = 10.0,
-    output_dir: Path | None = None,
+    output_dir: Optional[Path] = None,
 ) -> Tuple[List[Path], Path]:
     """将视频按固定时长切分为多个片段。
 
diff --git a/fish_api/app/services/water_video.py b/fish_api/app/services/water_video.py
index 0c2e55d..e6232fa 100644
--- a/fish_api/app/services/water_video.py
+++ b/fish_api/app/services/water_video.py
@@ -12,10 +12,12 @@ from __future__ import annotations
 import asyncio
 import shutil
 from pathlib import Path
-from typing import Dict, List
+from typing import Dict, List, Optional
 
 from loguru import logger
 
+from app.compat import to_thread
+
 from app.services.action_watch import iter_mp4
 from app.services.measure import transcode_src_to_h264_dst
 from app.services.video_slice import get_video_duration, slice_video
@@ -60,7 +62,7 @@ _global_slice_urls: List[str] = []
 _last_source_mtime: float = 0.0
 
 
-def resolve_water_video_source(settings: Settings) -> Path | None:
+def resolve_water_video_source(settings: Settings) -> Optional[Path]:
     """优先 BIOMASS_WATER_VIDEO_SOURCE；否则取 ACTION_WATCH_DIR 中 mtime 最新的 .mp4。"""
     cfg = settings.biomass_water_video_source
     if cfg is not None:
@@ -103,13 +105,13 @@ async def _publish_video(
     tmp.unlink(missing_ok=True)
 
     try:
-        ok = await asyncio.to_thread(transcode_src_to_h264_dst, src, tmp)
+        ok = await to_thread(transcode_src_to_h264_dst, src, tmp)
         if ok and tmp.is_file() and tmp.stat().st_size > 0:
             tmp.replace(dst)
             logger.info("[water-video] published H.264: {} -> {}", src.name, dst.name)
         else:
             tmp.unlink(missing_ok=True)
-            await asyncio.to_thread(shutil.copy2, src, dst)
+            await to_thread(shutil.copy2, src, dst)
             logger.warning(
                 "[water-video] transcode failed, copied raw: {} -> {}",
                 src.name,
diff --git a/fish_api/app/services/zed_recording_control.py b/fish_api/app/services/zed_recording_control.py
new file mode 100644
index 0000000..214c9a5
--- /dev/null
+++ b/fish_api/app/services/zed_recording_control.py
@@ -0,0 +1,132 @@
+"""ZED 分段录制线程的启动/停止（供 lifespan 与 HTTP API 共用）。"""
+
+from __future__ import annotations
+
+import threading
+from typing import Optional, Tuple
+
+from loguru import logger
+
+from app.db import (
+    begin_zed_recording_session,
+    mark_zed_recording_session_stopped,
+)
+from app.services.zed_svo_record import run_zed_svo_record_loop
+from app.settings import Settings
+from app.state import app_state
+
+
+def _cleanup_stale_zed_thread() -> None:
+    t = app_state.zed_recording_thread
+    if t is not None and not t.is_alive():
+        sid = app_state.zed_recording_session_row_id
+        app_state.zed_recording_thread = None
+        app_state.zed_recording_stop_event = None
+        if sid is not None:
+            try:
+                from app.settings import get_settings
+
+                mark_zed_recording_session_stopped(get_settings(), sid)
+            except Exception as e:
+                logger.warning("[zed-svo] 线程已退出但写库失败: {}", e)
+        app_state.zed_recording_session_row_id = None
+        app_state.zed_recording_fish_id = None
+
+
+def start_zed_recording(
+    settings: Settings,
+    *,
+    segment_sec: Optional[float] = None,
+) -> Tuple[bool, str, Optional[int], Optional[int]]:
+    """启动后台录制线程；已在运行则 ``(False, 'already_running', None, None)``。
+
+    每次启动在 SQLite ``zed_recording_sessions`` 中分配新 ``fish_id`` 并写入输出目录。
+
+    成功时返回 ``(True, 'started', fish_id, session_row_id)``。
+    """
+    with app_state.zed_recording_lock:
+        _cleanup_stale_zed_thread()
+        if (
+            app_state.zed_recording_thread is not None
+            and app_state.zed_recording_thread.is_alive()
+        ):
+            return False, "already_running", None, None
+
+        try:
+            session_row_id, fish_id, resolved = begin_zed_recording_session(settings)
+        except Exception as e:
+            logger.exception("[zed-svo] 分配会话失败")
+            return False, f"session_db_error:{e}", None, None
+
+        stop = threading.Event()
+
+        def _run() -> None:
+            run_zed_svo_record_loop(
+                settings,
+                stop,
+                output_dir=resolved,
+                segment_sec=segment_sec,
+            )
+
+        th = threading.Thread(
+            target=_run,
+            name="zed-svo-record",
+            daemon=True,
+        )
+        app_state.zed_recording_stop_event = stop
+        app_state.zed_recording_thread = th
+        app_state.zed_recording_session_row_id = session_row_id
+        app_state.zed_recording_fish_id = fish_id
+        th.start()
+        logger.info(
+            "[zed-svo] 录制线程已启动 session_id={} fish_id={} output_dir={}",
+            session_row_id,
+            fish_id,
+            resolved,
+        )
+        return True, "started", fish_id, session_row_id
+
+
+def stop_zed_recording(
+    settings: Settings, timeout_sec: float = 30.0
+) -> Tuple[bool, str, Optional[int]]:
+    """请求停止并 ``join`` 录制线程；成功停止后更新数据库并返回 ``fish_id``。"""
+    session_row_id: Optional[int] = None
+    with app_state.zed_recording_lock:
+        _cleanup_stale_zed_thread()
+        if app_state.zed_recording_thread is None:
+            return False, "not_running", None
+        if not app_state.zed_recording_thread.is_alive():
+            app_state.zed_recording_thread = None
+            app_state.zed_recording_stop_event = None
+            app_state.zed_recording_session_row_id = None
+            app_state.zed_recording_fish_id = None
+            return False, "not_running", None
+
+        session_row_id = app_state.zed_recording_session_row_id
+        ev = app_state.zed_recording_stop_event
+        th = app_state.zed_recording_thread
+        if ev is not None:
+            ev.set()
+        if th is not None:
+            th.join(timeout=timeout_sec)
+            if th.is_alive():
+                logger.warning("[zed-svo] stop join 超时，线程仍在运行")
+                return False, "stop_timeout", None
+
+        app_state.zed_recording_stop_event = None
+        app_state.zed_recording_thread = None
+        app_state.zed_recording_session_row_id = None
+        app_state.zed_recording_fish_id = None
+        logger.info("[zed-svo] 录制线程已停止")
+
+    fish_id: Optional[int] = None
+    if session_row_id is not None:
+        fish_id = mark_zed_recording_session_stopped(settings, session_row_id)
+    return True, "stopped", fish_id
+
+
+def zed_recording_is_running() -> bool:
+    _cleanup_stale_zed_thread()
+    t = app_state.zed_recording_thread
+    return t is not None and t.is_alive()
diff --git a/fish_api/app/services/zed_svo_record.py b/fish_api/app/services/zed_svo_record.py
new file mode 100644
index 0000000..b8febd6
--- /dev/null
+++ b/fish_api/app/services/zed_svo_record.py
@@ -0,0 +1,109 @@
+"""后台线程：单台 ZED 分段录制 .svo2（可选依赖 pyzed）。"""
+
+from __future__ import annotations
+
+import time
+from datetime import datetime, timezone
+from pathlib import Path
+from threading import Event
+from typing import Optional
+
+from loguru import logger
+
+from app.settings import Settings
+
+
+def _try_import_sl():
+    try:
+        import pyzed.sl as sl
+
+        return sl
+    except ImportError:
+        return None
+
+
+def run_zed_svo_record_loop(
+    settings: Settings,
+    stop_event: Event,
+    *,
+    output_dir: Optional[Path] = None,
+    segment_sec: Optional[float] = None,
+) -> None:
+    """阻塞运行直到 ``stop_event`` 置位；每段按 ``segment_sec`` 或配置轮换输出文件。
+
+    ``output_dir`` / ``segment_sec`` 非空时覆盖 ``settings`` 中对应逻辑。
+    """
+    sl = _try_import_sl()
+    if sl is None:
+        logger.warning(
+            "[zed-svo] pyzed 不可用，跳过 ZED 分段录制（请安装 ZED SDK / pyzed）。"
+        )
+        return
+
+    out_dir = output_dir if output_dir is not None else settings.zed_svo_record_dir
+    if out_dir is None:
+        logger.error("[zed-svo] zed_svo_record_dir 未解析，跳过录制")
+        return
+
+    out_dir = out_dir.expanduser().resolve()
+    out_dir.mkdir(parents=True, exist_ok=True)
+
+    zed = sl.Camera()
+    init = sl.InitParameters()
+    init.camera_resolution = sl.RESOLUTION.HD720
+    init.camera_fps = 30
+    init.depth_mode = sl.DEPTH_MODE.ULTRA
+    init.coordinate_units = sl.UNIT.MILLIMETER
+    if settings.zed_serial_number is not None:
+        init.set_from_serial_number(settings.zed_serial_number)
+
+    err_open = zed.open(init)
+    if err_open != sl.ERROR_CODE.SUCCESS:
+        logger.error("[zed-svo] 打开相机失败: {}", repr(err_open))
+        return
+
+    runtime = sl.RuntimeParameters()
+    seg_sec = float(
+        segment_sec if segment_sec is not None else settings.zed_svo_segment_sec
+    )
+    segment_index = 0
+
+    try:
+        while not stop_event.is_set():
+            ts = datetime.now(timezone.utc).strftime("%Y%m%dT%H%M%SZ")
+            out_path = out_dir / f"zed_{ts}_{segment_index:06d}.svo2"
+            segment_index += 1
+
+            rec = sl.RecordingParameters(
+                str(out_path.resolve()),
+                sl.SVO_COMPRESSION_MODE.H264,
+            )
+            err_rec = zed.enable_recording(rec)
+            if err_rec != sl.ERROR_CODE.SUCCESS:
+                logger.error(
+                    "[zed-svo] enable_recording 失败: {} path={}",
+                    repr(err_rec),
+                    out_path,
+                )
+                return
+
+            logger.info("[zed-svo] 开始写入 {}", out_path.name)
+
+            deadline = time.monotonic() + seg_sec
+            while time.monotonic() < deadline and not stop_event.is_set():
+                err = zed.grab(runtime)
+                if err != sl.ERROR_CODE.SUCCESS:
+                    if err == sl.ERROR_CODE.END_OF_SVOFILE_REACHED:
+                        break
+                    logger.warning("[zed-svo] grab 非成功: {}", repr(err))
+
+            zed.disable_recording()
+            logger.info("[zed-svo] 结束一段 {}", out_path.name)
+
+    finally:
+        try:
+            zed.disable_recording()
+        except Exception as e:
+            logger.debug("[zed-svo] disable_recording (finally): {}", e)
+        zed.close()
+        logger.info("[zed-svo] 相机已关闭")
diff --git a/fish_api/app/settings.py b/fish_api/app/settings.py
index 696ab28..d133d3a 100644
--- a/fish_api/app/settings.py
+++ b/fish_api/app/settings.py
@@ -56,6 +56,11 @@ class Settings(BaseSettings):
         default="http://127.0.0.1:8000",
         validation_alias=AliasChoices("PUBLIC_BASE_URL", "public_base_url"),
     )
+    #: ZED 录制 CLI（``fish-zed-record``）等访问本机 API 时的基址；未设时与 ``public_base_url`` 相同。**FISH_API_BASE_URL**
+    fish_api_base_url: str = Field(
+        default="",
+        validation_alias=AliasChoices("FISH_API_BASE_URL", "fish_api_base_url"),
+    )
 
     ingest_api_key: str = ""
 
@@ -84,6 +89,9 @@ class Settings(BaseSettings):
     python_fish_measure: str = ""
     python_fish_action: str = ""
 
+    #: ffmpeg 可执行文件路径；为空时按顺序尝试 tools/ffmpeg/bin/ffmpeg → 系统 PATH。**FFMPEG_PATH**
+    ffmpeg_path: str = ""
+
     #: SAM/CUDA 设备（cuda 或 cpu）
     sam_device: str = "cuda"
 
@@ -261,15 +269,57 @@ class Settings(BaseSettings):
     measure_watch_recursive: bool = False
     #: 状态管理：true=持久化到 SQLite（重启后记住），false=内存模式（重启后清空）
     measure_watch_use_state_file: bool = True
+    #: 齐套后对各段 former 体重/体长聚合方式：``median``、``mean``、``trimmed_mean``（至少 3 段时去头尾再均值）。**MEASURE_FINAL_AGGREGATE_MODE**
+    measure_final_aggregate_mode: str = Field(
+        default="median",
+        validation_alias=AliasChoices(
+            "MEASURE_FINAL_AGGREGATE_MODE", "measure_final_aggregate_mode"
+        ),
+    )
+
+    #: 分段 SVO2 输出目录；未设时：有 MEASURE_WATCH_DIR 则为 ``{MEASURE_WATCH_DIR}/fish{N}``（N 见 zed_svo_record_fish_id），否则为 ``{STREAM_TMP_DIR}/zed_svo2``。**ZED_SVO_RECORD_DIR**
+    zed_svo_record_dir: Optional[Path] = Field(
+        default=None,
+        validation_alias=AliasChoices("ZED_SVO_RECORD_DIR", "zed_svo_record_dir"),
+    )
+    #: 每段时长（秒），默认 300（5 分钟）。**ZED_SVO_SEGMENT_SEC**
+    zed_svo_segment_sec: float = Field(
+        default=300.0,
+        ge=1.0,
+        validation_alias=AliasChoices("ZED_SVO_SEGMENT_SEC", "zed_svo_segment_sec"),
+    )
+    #: 仅连接一台相机时可选；指定序列号打开对应设备。**ZED_SERIAL_NUMBER**
+    zed_serial_number: Optional[int] = Field(
+        default=None,
+        validation_alias=AliasChoices("ZED_SERIAL_NUMBER", "zed_serial_number"),
+    )
+    #: 与 MEASURE_WATCH_DIR 组合为 ``fish{N}``（默认 1）。**ZED_SVO_RECORD_FISH_ID**
+    zed_svo_record_fish_id: int = Field(
+        default=1,
+        ge=1,
+        validation_alias=AliasChoices(
+            "ZED_SVO_RECORD_FISH_ID", "zed_svo_record_fish_id"
+        ),
+    )
 
     default_fish_species: str = "大黄鱼"
 
+    @field_validator("zed_serial_number", mode="before")
+    @classmethod
+    def _zed_serial_empty_none(cls, v: object) -> object:
+        if v is None:
+            return None
+        if isinstance(v, str) and not v.strip():
+            return None
+        return v
+
     @field_validator(
         "action_watch_dir",
         "biomass_water_video_source",
         "biomass_sonar_video_source",
         "biomass_sonar_video_dir",
         "measure_watch_dir",
+        "zed_svo_record_dir",
         mode="before",
     )
     @classmethod
@@ -300,6 +350,25 @@ class Settings(BaseSettings):
                 / "checkpoints"
                 / "best_model.pth",
             )
+        if self.zed_svo_record_dir is None:
+            if self.measure_watch_dir is not None:
+                object.__setattr__(
+                    self,
+                    "zed_svo_record_dir",
+                    (
+                        self.measure_watch_dir / f"fish{self.zed_svo_record_fish_id}"
+                    ).resolve(),
+                )
+            else:
+                object.__setattr__(
+                    self,
+                    "zed_svo_record_dir",
+                    (self.stream_tmp_dir / "zed_svo2").resolve(),
+                )
+        else:
+            object.__setattr__(
+                self, "zed_svo_record_dir", self.zed_svo_record_dir.expanduser().resolve()
+            )
         return self
 
 
diff --git a/fish_api/app/state.py b/fish_api/app/state.py
index 46b724b..346ff97 100644
--- a/fish_api/app/state.py
+++ b/fish_api/app/state.py
@@ -1,6 +1,7 @@
 from __future__ import annotations
 
 import asyncio
+import threading
 from dataclasses import dataclass, field
 from datetime import datetime
 from typing import List, Optional
@@ -15,7 +16,7 @@ class MeasureSnapshot:
     error: Optional[str] = None
     raw_prediction_path: Optional[str] = None
     pred: Optional[float] = None  # 最终预测的体重值
-    star: bool = False  # 置信度标记，True 表示计算可信
+    star: bool = False  # DB 是唯一真相源：segment/batch 恒 False；final 行按参与聚合段的 dgcnn_summary 规则判定
     #: 与 FishMeasure ``test_dgcnn_weight_estimator.py`` 终端输出一致的体重推算过程文本
     calculation_log: Optional[str] = None
 
@@ -32,11 +33,21 @@ class HealthSnapshot:
 @dataclass
 class AppState:
     measure_lock: asyncio.Lock = field(default_factory=asyncio.Lock)
+    #: 与 ``run_full_measure`` 子进程串行化（跨 ``measure_watch`` / ingest 线程）
+    measure_thread_lock: threading.Lock = field(default_factory=threading.Lock)
     action_lock: asyncio.Lock = field(default_factory=asyncio.Lock)
 
     # job status for optional polling（业务结果见 SQLite）
     measure_status: str = "idle"
     action_status: str = "idle"
 
+    #: ZED 分段录制线程与协作停止事件（由 ``zed_recording_control`` 管理）
+    zed_recording_lock: threading.Lock = field(default_factory=threading.Lock)
+    zed_recording_stop_event: Optional[threading.Event] = None
+    zed_recording_thread: Optional[threading.Thread] = None
+    #: 当前会话（与 ``zed_recording_sessions`` 表对应）
+    zed_recording_session_row_id: Optional[int] = None
+    zed_recording_fish_id: Optional[int] = None
+
 
 app_state = AppState()
diff --git a/fish_api/app/subprocess_run.py b/fish_api/app/subprocess_run.py
index aa298a3..156ade6 100644
--- a/fish_api/app/subprocess_run.py
+++ b/fish_api/app/subprocess_run.py
@@ -29,6 +29,8 @@ def run_subprocess_with_log(
         stdout=subprocess.PIPE,
         stderr=subprocess.STDOUT,
         text=True,
+        encoding="utf-8",
+        errors="replace",
         bufsize=1,
     )
     lines: List[str] = []
diff --git a/fish_api/app/zed_record_cli.py b/fish_api/app/zed_record_cli.py
new file mode 100644
index 0000000..70676c8
--- /dev/null
+++ b/fish_api/app/zed_record_cli.py
@@ -0,0 +1,150 @@
+"""命令行：ZED 分段录制。
+
+- ``start``（默认）：本进程内 ``start_zed_recording``，前台阻塞直到 Ctrl+C 再 ``stop``（不经过 HTTP）。
+- ``start --remote``：``POST`` 已运行的 fish_api，不阻塞。
+- ``stop`` / ``status``：仅 HTTP，用于停/查 **由 uvicorn 进程托管** 的录制。
+
+HTTP 基址：环境变量 ``FISH_API_BASE_URL`` 或 ``PUBLIC_BASE_URL``，否则 ``http://127.0.0.1:8000``。
+ZED 路由不校验 ingest API Key，请求无需 ``X-API-Key``。
+
+每次启动由服务端分配 ``fish_id``（库表与 ``output_dir``、目标父目录下 ``fish``+数字 及 ``.svo2`` 路径综合）并写入 ``zed_recording_sessions``。
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import sys
+import time
+import urllib.error
+import urllib.request
+from typing import Any, Dict, Optional
+
+
+def _base_url() -> str:
+    """优先环境变量；否则从 ``fish_api/.env`` 经 ``get_settings()`` 读 ``FISH_API_BASE_URL`` / ``PUBLIC_BASE_URL``。"""
+    u = os.environ.get("FISH_API_BASE_URL") or os.environ.get("PUBLIC_BASE_URL")
+    if u:
+        return str(u).strip().rstrip("/")
+    from app.settings import get_settings
+
+    s = get_settings()
+    u2 = (s.fish_api_base_url or "").strip() or (s.public_base_url or "").strip()
+    if u2:
+        return u2.rstrip("/")
+    return "http://127.0.0.1:8000"
+
+
+def _http_request_json(method: str, path: str, body: Optional[Dict[str, Any]] = None) -> Dict[str, Any]:
+    url = _base_url() + path
+    if method == "GET":
+        req = urllib.request.Request(url, method="GET")
+    elif method == "POST":
+        payload = json.dumps(body if body is not None else {}).encode("utf-8")
+        req = urllib.request.Request(
+            url,
+            data=payload,
+            headers={"Content-Type": "application/json; charset=utf-8"},
+            method="POST",
+        )
+    else:
+        raise ValueError(method)
+
+    try:
+        with urllib.request.urlopen(req, timeout=120) as resp:
+            raw = resp.read().decode("utf-8")
+            if not raw.strip():
+                return {}
+            return json.loads(raw)
+    except urllib.error.HTTPError as e:
+        err_body = e.read().decode("utf-8", errors="replace")
+        print(f"HTTP {e.code}: {err_body}", file=sys.stderr)
+        raise SystemExit(1) from None
+    except urllib.error.URLError as e:
+        print(f"请求失败: {e}", file=sys.stderr)
+        raise SystemExit(1) from None
+
+
+def _cmd_start_remote(args: argparse.Namespace) -> None:
+    body: Dict[str, Any] = {}
+    if args.segment_sec is not None:
+        body["segment_sec"] = args.segment_sec
+    r = _http_request_json("POST", "/api/v1/zed/recording/start", body)
+    print(json.dumps(r, ensure_ascii=False, indent=2))
+
+
+def _cmd_start_local(args: argparse.Namespace) -> None:
+    from app.services.zed_recording_control import (
+        start_zed_recording,
+        stop_zed_recording,
+    )
+    from app.settings import get_settings
+
+    s = get_settings()
+    ok, msg, fish_id, _ = start_zed_recording(
+        s,
+        segment_sec=args.segment_sec,
+    )
+    if not ok:
+        print(f"启动失败: {msg}", file=sys.stderr)
+        raise SystemExit(1)
+
+    print(
+        f"fish_id={fish_id}，录制中，按 Ctrl+C 停止…",
+        flush=True,
+    )
+    try:
+        while True:
+            time.sleep(1.0)
+    except KeyboardInterrupt:
+        print("\n正在停止…", flush=True)
+    stop_ok, stop_msg, stop_fish = stop_zed_recording(s)
+    if stop_ok:
+        print(f"已退出。fish_id={stop_fish}", flush=True)
+    else:
+        print(f"停止结果: {stop_msg}", file=sys.stderr)
+
+
+def _cmd_stop(_args: argparse.Namespace) -> None:
+    r = _http_request_json("POST", "/api/v1/zed/recording/stop", {})
+    print(json.dumps(r, ensure_ascii=False, indent=2))
+
+
+def _cmd_status(_args: argparse.Namespace) -> None:
+    r = _http_request_json("GET", "/api/v1/zed/recording/status", None)
+    print(json.dumps(r, ensure_ascii=False, indent=2))
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(
+        description="ZED 分段录制：本地阻塞 start，或 HTTP 调用 fish_api",
+    )
+    sub = parser.add_subparsers(dest="cmd", required=True)
+
+    p_start = sub.add_parser("start", help="本地启动（阻塞至 Ctrl+C）或 --remote 仅请求 API")
+    p_start.add_argument(
+        "--remote",
+        action="store_true",
+        help="通过 HTTP 调用 fish_api，不阻塞本进程",
+    )
+    p_start.add_argument("--segment-sec", type=float, default=None, metavar="SEC")
+
+    sub.add_parser("stop", help="HTTP 停止 fish_api 进程中的录制")
+    sub.add_parser("status", help="HTTP 查询 fish_api 中的录制状态")
+
+    args = parser.parse_args()
+
+    if args.cmd == "start":
+        if args.remote:
+            _cmd_start_remote(args)
+        else:
+            _cmd_start_local(args)
+    elif args.cmd == "stop":
+        _cmd_stop(args)
+    elif args.cmd == "status":
+        _cmd_status(args)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/fish_api/pyproject.toml b/fish_api/pyproject.toml
index 85d19a2..2de57e5 100644
--- a/fish_api/pyproject.toml
+++ b/fish_api/pyproject.toml
@@ -17,3 +17,4 @@ dev = ["httpx>=0.28.1"]
 [project.scripts]
 fish-action-watch = "app.action_watch_cli:main"
 fish-measure-debug = "app.measure_debug_cli:main"
+fish-zed-record = "app.zed_record_cli:main"
diff --git a/fish_api/start_fresh.sh b/fish_api/start_fresh.sh
old mode 100755
new mode 100644
diff --git a/start_recording.sh b/start_recording.sh
new file mode 100755
index 0000000..6a82f78
--- /dev/null
+++ b/start_recording.sh
@@ -0,0 +1,19 @@
+#!/usr/bin/env bash
+# 独立运行：与 fish_api/uvicorn 的启动、停止无关。调用 ZED 录制 CLI：本地前台录制（Ctrl+C 停），或传 --remote 仅请求已运行的 fish_api。
+# 用法（在仓库根目录）：
+#   ./start_recording.sh
+#   ./start_recording.sh --remote
+#   ./start_recording.sh --segment-sec 300
+# 依赖：fish_api 已 pip install -e .；直连相机时需 pyzed。
+set -euo pipefail
+
+ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+cd "$ROOT/fish_api"
+
+if command -v uv >/dev/null 2>&1; then
+  PY=(uv run python)
+else
+  PY=(python3)
+fi
+
+exec "${PY[@]}" -m app.zed_record_cli start "$@"
diff --git a/stop_recording.sh b/stop_recording.sh
new file mode 100644
index 0000000..87809f3
--- /dev/null
+++ b/stop_recording.sh
@@ -0,0 +1,16 @@
+#!/usr/bin/env bash
+# 与 fish_api/uvicorn 的启动、停止无关；仅向已运行的 fish_api 发 HTTP stop（适用于曾用 /api/v1/zed/recording/start 或 start_recording.sh --remote 启录的情形）。
+# 基址：环境变量 FISH_API_BASE_URL 或 PUBLIC_BASE_URL（见 fish_api/.env）。
+# 用法（在仓库根目录）：./stop_recording.sh
+set -euo pipefail
+
+ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+cd "$ROOT/fish_api"
+
+if command -v uv >/dev/null 2>&1; then
+  PY=(uv run python)
+else
+  PY=(python3)
+fi
+
+exec "${PY[@]}" -m app.zed_record_cli stop "$@"