重组为 backend/clients/docs 三层结构，并清理 git 污染。

将后端迁入 backend/，完善根目录 .gitignore，删除误提交的 .mypy_cache 缓存文件。

Co-authored-by: Cursor <cursoragent@cursor.com>

backend/app/baked/__init__.py

@@ -0,0 +1,5 @@
+"""编译进代码的默认参数（非部署差异）；见 `algorithm` 与 `pipeline` 子模块。"""
+
+from app.baked import algorithm, pipeline
+
+__all__ = ["algorithm", "pipeline"]

backend/app/baked/algorithm.py

@@ -0,0 +1,121 @@
+"""YOLO 耗材/手部 + configured reference ActionFormer 实时段落检测默认配置。"""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+from app.algorithm_runner.reference_bundle_runtime import default_reference_bundle_dir
+
+_PACKAGE_DIR = Path(__file__).resolve().parent.parent
+
+
+def default_consumable_classifier_weights_path() -> str:
+    return str(_PACKAGE_DIR / "resources" / "consumable_classifier.pt")
+
+
+def default_consumable_classifier_labels_yaml_path() -> str:
+    return str(_PACKAGE_DIR / "resources" / "consumable_classifier_labels.yaml")
+
+
+def default_or_site_config_sample_path() -> str:
+    """站点配置示例：video_rtsp_urls + voice_or_room_bindings。"""
+    return str(_PACKAGE_DIR / "resources" / "or_site_config.sample.json")
+
+
+def default_actionformer_hand_det_weights_path() -> str:
+    return str(default_reference_bundle_dir() / "weights" / "hand_detect.pt")
+
+
+def default_actionformer_goodbad_weights_path() -> str:
+    return str(default_reference_bundle_dir() / "weights" / "goodbad_frame.pt")
+
+
+def default_actionformer_haocai_weights_path() -> str:
+    return str(default_reference_bundle_dir() / "weights" / "haocai_classify.pt")
+
+
+def default_actionformer_ckpt_path() -> str:
+    return str(default_reference_bundle_dir() / "weights" / "actionformer_epoch_045.pth.tar")
+
+
+def default_actionformer_tear_weights_path() -> str:
+    return str(default_reference_bundle_dir() / "weights" / "tear_classify.pt")
+
+
+# --- 耗材分类（YOLO-cls）---
+CONSUMABLE_CLASSIFIER_WEIGHTS: str = default_consumable_classifier_weights_path()
+CONSUMABLE_CLASSIFIER_IMGSZ: int = 224
+CONSUMABLE_CLASSIFIER_DEVICE: str = ""
+CONSUMABLE_CLASSIFIER_TOPK: int = 5
+CONSUMABLE_MIN_CLS_CONFIDENCE: float = 0.5
+CONSUMABLE_CLASSIFIER_LABELS_YAML_PATH: str = default_consumable_classifier_labels_yaml_path()
+CONSUMABLE_VISION_WINDOW_SEC: float = 15.0
+
+# --- 手部检测；全空路径则整帧分类 ---
+HAND_DETECTION_WEIGHTS: str = ""
+HAND_DETECTION_IMGSZ: int = 640
+HAND_DETECTION_CONF: float = 0.25
+HAND_DETECTION_PAD_RATIO: float = 0.30
+HAND_DETECTION_MIN_CROP_PX: int = 64
+HAND_DETECTION_DEVICE: str = ""
+
+
+# --- refs/5.15 ActionFormer 实时版（VideoSwin → ActionFormer + 手检/好坏帧/耗材投票）---
+# 默认开启并与常见 demo 主摄 id 对齐；若缺少配置引用包 weights 下权重，开录时该路会记录异常。
+ACTIONFORMER_ENABLED: bool = True
+# 检查期：单路算法固定拉流 or-cam-03（恢复按请求首路时请改回 "" 或按需配置）。
+ACTIONFORMER_PRIMARY_CAMERA_ID: str = "or-cam-03"
+
+# RTSP 帧率假设（用于 frame_idx → t_abs 推算与 ActionFormer fps 入参；线上需与摄像机一致）。
+ACTIONFORMER_ASSUMED_FPS: float = 25.0
+
+# 与 refs/5.15 ``run_haocai_actionformer_consumables_e2e`` 写出的 infer_single.yaml 同语义的 VideoSwin 参数：
+# clip_len 32, frame_stride 2, feat_stride_frames 32 → feat_num_frames=64。
+VIDEOSWIN_CLIP_LEN: int = 32
+VIDEOSWIN_FRAME_STRIDE: int = 2
+VIDEOSWIN_FEAT_STRIDE_FRAMES: int = 32
+VIDEOSWIN_INPUT_DIM: int = 768
+VIDEOSWIN_IMAGE_SIZE: int = 224
+
+# 模型权重（默认读取配置引用包 weights；文件名与 configs/default_config.yaml 对齐）
+ACTIONFORMER_CKPT: str = default_actionformer_ckpt_path()
+ACTIONFORMER_HAND_DET_WEIGHTS: str = default_actionformer_hand_det_weights_path()
+ACTIONFORMER_GOODBAD_WEIGHTS: str = default_actionformer_goodbad_weights_path()
+ACTIONFORMER_HAOCAI_WEIGHTS: str = default_actionformer_haocai_weights_path()
+ACTIONFORMER_TEAR_WEIGHTS: str = default_actionformer_tear_weights_path()
+ACTIONFORMER_NUM_CLASSES: int = 1  # 单类（Action）划段，与 e2e infer_single.yaml 对齐
+
+# 与 refs/5.15 e2e 默认对齐的运行时阈值
+ACTIONFORMER_DET_CONF: float = 0.5
+ACTIONFORMER_PAD_RATIO: float = 0.30
+ACTIONFORMER_DET_IMGSZ: int = 640
+ACTIONFORMER_GOODBAD_IMGSZ: int = 224
+ACTIONFORMER_HAOCAI_IMGSZ: int = 224
+ACTIONFORMER_GOOD_TOP1_CONF_THRESHOLD: float = 0.9
+ACTIONFORMER_HAOCAI_MIN_CONF: float = 0.8
+
+# ActionFormer 候选段 score 下限（不含等于）；与 e2e ``--af-min-score`` 默认对齐
+ACTIONFORMER_AF_MIN_SCORE: float = 0.1
+
+# 候选耗材 → softmax mask（按会话候选耗材索引集做 mask，避免命中候选外类别）
+ACTIONFORMER_WHITELIST_SOFTMAX: bool = True
+
+# 周期性 ActionFormer + 稳定段发布
+# 周期间隔（墙钟秒）：每过这么久才考虑跑一次 ActionFormer，避免逐帧打 forward
+ACTIONFORMER_PERIOD_SEC: float = 10.0
+# 至少累积多少特征才有第一次有效推理（32 帧/特征 @ 25 fps ≈ 1.28 秒/特征）
+ACTIONFORMER_MIN_FEATURES_FOR_INFER: int = 8
+# 段尾时间 + lookahead < now 才视为「冻结」，可以增量落库；越大越接近全量批处理。
+ACTIONFORMER_INCREMENTAL_STABLE_FLUSH: bool = True
+ACTIONFORMER_FREEZE_LOOKAHEAD_S: float = 12.0
+# 新旧配置兼容；未显式使用时以 ACTIONFORMER_FREEZE_LOOKAHEAD_S 为准。
+ACTIONFORMER_STABLE_FLUSH_MARGIN_S: float = ACTIONFORMER_FREEZE_LOOKAHEAD_S
+# 已发布段去重：新段与已发布段的较短段覆盖率超过该阈值时视为同一段。
+ACTIONFORMER_DEDUPE_OVERLAP_RATIO: float = 0.5
+
+# 设备：空字符串走 resolve_inference_device（cuda 优先、mac 上再 mps、否则 CPU）
+ACTIONFORMER_VIDEOSWIN_DEVICE: str = ""
+ACTIONFORMER_DEVICE: str = ""
+ACTIONFORMER_HAND_DET_DEVICE: str = ""
+ACTIONFORMER_GOODBAD_DEVICE: str = ""
+ACTIONFORMER_HAOCAI_DEVICE: str = ""

backend/app/baked/pipeline.py

@@ -0,0 +1,42 @@
+"""视频流、待确认、归档与可观测性默认可调参数（非 env）。"""
+
+# --- 手术录制 API 重试 ---
+SURGERY_RECORDING_MAX_ATTEMPTS: int = 3
+SURGERY_RECORDING_RETRY_DELAY_SECONDS: float = 1.0
+
+# --- RTSP 连接与抽帧、推理门控（不含 URL，URL 在 Settings）---
+# 运行时以 Settings.video_open_timeout_sec（环境变量 VIDEO_OPEN_TIMEOUT_SEC）为准；此处为未注入 Settings 时的回退默认。
+VIDEO_OPEN_TIMEOUT_SEC: float = 45.0
+VIDEO_READ_FAILURE_RECONNECT_THRESHOLD: int = 15
+VIDEO_RECONNECT_BACKOFF_SECONDS: float = 1.0
+VIDEO_INFERENCE_INTERVAL_SEC: float = 2.0
+VIDEO_INFERENCE_CONFIDENCE_THRESHOLD: float = 0.35
+VIDEO_AUTO_CONFIRM_CONFIDENCE: float = 0.9
+VIDEO_VOICE_CONFIRM_MIN_CONFIDENCE: float = 0.35
+VOICE_CONFIRMATION_ENABLED: bool = True
+# 语音「请确认」播报时最多念几个候选项：在模型 topK 不足时用术程耗材清单顺序补齐（补项置信度为 0）。
+VOICE_PROMPT_MAX_OPTIONS: int = 3
+VIDEO_VOICE_CONFIRM_DOCTOR_ID: str = "voice"
+VIDEO_DETAIL_COOLDOWN_SEC: float = 15.0
+VIDEO_JPEG_QUALITY: int = 85
+VIDEO_RESULT_DOCTOR_ID: str = "vision"
+VIDEO_LOG_INFERENCE_RESULTS: bool = False
+
+# --- 停录后归档重试 + durable fallback ---
+ARCHIVE_PERSIST_RETRY_INTERVAL_SECONDS: float = 30.0
+ARCHIVE_PERSIST_MAX_RETRIES: int = 12
+ARCHIVE_PERSIST_BACKOFF_CAP_SECONDS: float = 900.0
+ARCHIVE_PERSIST_DURABLE_FALLBACK_DIR: str = "logs/pending_archive"
+ARCHIVE_PERSIST_DURABLE_FALLBACK_ENABLED: bool = True
+
+# --- 消耗/语音 文本日志与终端展示 ---
+CONSUMPTION_TSV_LOG_ENABLED: bool = True
+CONSUMPTION_TSV_LOG_PATH: str = "logs/consumption_{surgery_id}.txt"
+CONSUMPTION_LOG_MARKDOWN_TERMINAL: bool = True
+CONSUMPTION_LOG_TIMEZONE: str = ""
+
+VOICE_FILE_LOG_ENABLED: bool = True
+VOICE_FILE_LOG_PATH: str = "logs/voice_{surgery_id}.txt"
+
+VOICE_UPLOAD_MAX_BYTES: int = 10 * 1024 * 1024
+VOICE_CONFIRM_MAX_FAILED_PARSE_ROUNDS: int = 2