refactor: 统一耗材视觉算法并扩展语音确认至全量候选清单

- 以 ConsumableVisionAlgorithmService 替代 consumable_classifier 与 tear_action；
  可选手部检测权重，未配置时全帧分类；时间窗众数与 Excel 白名单配置。
- 语音待确认：ASR 先匹配 pending topk，再匹配本台 candidate_consumables；
  记账 item_id 与 vision 一致使用 name_to_code。
- 更新 config、Compose、.env.example、依赖（pandas/openpyxl）与测试。

Made-with: Cursor

docker-compose.prod.yml

@@ -36,9 +36,12 @@ services:
       CONSUMABLE_CLASSIFIER_IMGSZ: ${CONSUMABLE_CLASSIFIER_IMGSZ:-224}
       CONSUMABLE_CLASSIFIER_DEVICE: ${CONSUMABLE_CLASSIFIER_DEVICE:-}
       CONSUMABLE_CLASSIFIER_TOPK: ${CONSUMABLE_CLASSIFIER_TOPK:-5}
-      TEAR_ACTION_IMGSZ: ${TEAR_ACTION_IMGSZ:-224}
-      TEAR_ACTION_DEVICE: ${TEAR_ACTION_DEVICE:-}
-      TEAR_ACTION_TOPK: ${TEAR_ACTION_TOPK:-5}
+      CONSUMABLE_MIN_CLS_CONFIDENCE: ${CONSUMABLE_MIN_CLS_CONFIDENCE:-0.5}
+      CONSUMABLE_VISION_WINDOW_SEC: ${CONSUMABLE_VISION_WINDOW_SEC:-15}
+      CONSUMABLE_CATALOG_XLSX_PATH: ${CONSUMABLE_CATALOG_XLSX_PATH:-}
+      HAND_DETECTION_WEIGHTS: ${HAND_DETECTION_WEIGHTS:-}
+      HAND_DETECTION_IMGSZ: ${HAND_DETECTION_IMGSZ:-640}
+      HAND_DETECTION_DEVICE: ${HAND_DETECTION_DEVICE:-}
       # Video backends (RTSP / optional Hikvision SDK) — see docs/video-backends.md
       VIDEO_DEFAULT_BACKEND: ${VIDEO_DEFAULT_BACKEND:-rtsp}
       VIDEO_RTSP_URL_TEMPLATE: ${VIDEO_RTSP_URL_TEMPLATE:-}