"""
全局配置 - 服务器版（192.168.77.249, ~/顾工交接/wzj）
模型/训练代码目录与框架根同级（旧 ~/wzj/wzj → 上级 ~/wzj；现 ~/顾工交接/wzj → 上级 ~/顾工交接）。
"""
import os

PROJECT_ROOT = os.path.dirname(os.path.abspath(__file__))
# 框架根的上一级：模型目录(正在使用的模型_.../、20260415_.../)与之同级。
# 用 dirname(PROJECT_ROOT) 自适应，避免迁移路径后硬编码失效。
WZJ_ROOT = os.path.dirname(PROJECT_ROOT)

# ==================== 模型路径 ====================
YOLO_MODEL_PATH = os.path.join(WZJ_ROOT, "20260415_球星卡和宝可梦模型数据处理和训练代码/yolov11n_card_seg01.onnx")
DINOV2_MODEL_PATH = os.path.join(WZJ_ROOT, "正在使用的模型_3球星卡_1宝可梦/dinov2_base_retrieval_392_PokemonCN04")
# 评级卡检测模型（YOLO26 card_v1，检测封装壳上的评级公司标签 PSA/BGS/CGC/SGC）
# ultralytics 目录在框架树内，用 PROJECT_ROOT 相对路径（本地镜像 D:\顾工交接\wzj 与服务器 ~/wzj/wzj 都自洽）
GRADING_MODEL_PATH = os.path.join(PROJECT_ROOT, "ultralytics/runs/detect/card_v1/weights/best.pt")

# ==================== 数据目录 ====================
DATA_DIR = os.path.join(PROJECT_ROOT, "data")
GALLERY_IMG_DIR = os.path.join(DATA_DIR, "gallery_images")
QUERY_IMG_DIR = os.path.join(DATA_DIR, "query_images")
GALLERY_FEATURES_PATH = os.path.join(DATA_DIR, "gallery_features.npy")
GALLERY_META_PATH = os.path.join(DATA_DIR, "gallery_meta.json")
RESULT_PATH = os.path.join(DATA_DIR, "match_results.json")

# CSV 数据源（服务器读不了PG/CH，从本地导出CSV传入）
CARD_MASTER_CSV = os.path.join(DATA_DIR, "card_master.csv")
TRANSACTIONS_CSV = os.path.join(DATA_DIR, "transactions.csv")

# ==================== 数据库配置（服务器网络隔离，暂不可用，留作参考）====================
PG_CONFIG = {
    "host": "100.64.0.10", "port": 25432,
    "user": "readonlyuser", "password": "Pass2026", "database": "hs_sync_data",
}
PG_TABLE = "public.cards_master_v2"
CLICKHOUSE_CONFIG = {
    "host": "192.168.31.233", "port": 8123,
    "username": "card_transactions_ro", "password": "CardTxReadOnly2026!",
    "database": "card_transactions",
}
CLICKHOUSE_TABLE = "card_transactions_unified_sync"

# ==================== MinIO（交易图，已验证凭据）====================
MINIO_CONFIG = {
    "endpoint": "127.0.0.1:9000",       # 服务器本地访问
    "access_key": "minioadmin",
    "secret_key": "minioadmin",
    "secure": False,
    "bucket": "reverse.search",
}

# C 端 /match_fields 拍摄图落盘（249 MinIO grading 桶，与树莓派同源凭据；
# SDK 走本机 127.0.0.1，对外回显用 public_base）
CAPP_MINIO = {
    "endpoint": "127.0.0.1:9000",
    "public_base": "http://192.168.77.249:9000",
    "access_key": "pZEwCGnpNN05KPnmC2Yh",
    "secret_key": "KfJRuWiv9pVxhIMcFqbkv8hZT9SnNTZ6LPx592D4",
    "secure": False,
    "bucket": "grading",
    "prefix": "capp_img_data",
}


# ==================== 模型参数 ====================
YOLO_CONF_THRESHOLD = 0.25
YOLO_IMG_SIZE = 640
DINOV2_FEATURE_DIM = 768
GRADING_CONF_THRESHOLD = 0.25   # 评级公司标签检测置信度
GRADING_IMG_SIZE = 640

# ==================== 检索参数 ====================
TOP_K = 5
SIMILARITY_THRESHOLD = 0.50  # 100张人工核验：sim>=0.50 几乎无误，定为高/低置信界限(2026-09-02从0.60同步158生产值)
GALLERY_FILTER = None
DOWNLOAD_WORKERS = 16
FEATURE_BATCH_SIZE = 32


def ensure_dirs():
    for d in [DATA_DIR, GALLERY_IMG_DIR, QUERY_IMG_DIR]:
        os.makedirs(d, exist_ok=True)


# ===== 双区级联（2026-08 从本地同步追加，供 cascade_match / 生产入库）=====
CARD_CROP_MODE = os.environ.get("CARD_CROP_MODE", "seg_warp")
CARD_SEG_MODEL_PATH = os.path.join(PROJECT_ROOT, "ultralytics/runs/segment/card_seg_pn_v2/weights/best.pt")
YOLO_MODEL_PATH_LEGACY = YOLO_MODEL_PATH  # 保留旧 A4 onnx 路径
# 查询侧默认走 card_seg_v2；旧 serve 若依赖 YOLO_MODEL_PATH=onnx 请用 CARD_CROP_MODE=bbox_legacy
if CARD_CROP_MODE == "seg_warp" and os.path.exists(CARD_SEG_MODEL_PATH):
    YOLO_MODEL_PATH = CARD_SEG_MODEL_PATH
UPPER_MODEL_PATH = os.path.join(PROJECT_ROOT, "upper_model_output_v0904/best_upper_half_model.pth")
LOWER_MODEL_PATH = os.path.join(PROJECT_ROOT, "layer3_bg_model_output_v0904/best_layer3_bottom_model.pth")
UPPER_HALF_CROPS_DIR = os.path.join(PROJECT_ROOT, "_upper_half_crops")
GALLERY_UPPER_FEATURES_PATH = os.path.join(DATA_DIR, "gallery_v0904/gallery_upper_features.npy")
GALLERY_LOWER_FEATURES_PATH = os.path.join(DATA_DIR, "gallery_v0904/gallery_lower_features.npy")
GALLERY_DUAL_META_PATH = os.path.join(DATA_DIR, "gallery_v0904/gallery_dual_meta.json")
CARD_MASTER_ALL_CSV = os.path.join(DATA_DIR, "card_master_all.csv")
DUAL_FEATURE_DIM = 1024
CASCADE_TOP_K_RECALL = 30
CASCADE_ALPHA = 0.5
CASCADE_BACKEND = "gpu"  # 2026-09-03 切 GPU 常驻图库召回(bench Top-5 与 milvus 全等,40ms→0.4ms/卡);回滚改回 milvus  # 249 默认走本机 Milvus 双区库
MILVUS_HOST = "127.0.0.1"
MILVUS_PORT = "19530"
MILVUS_UPPER_COLLECTION = "pokemon_dual_upper_dinov2l_1024"
MILVUS_LOWER_COLLECTION = "pokemon_dual_lower_dinov2l_1024"

# ==================== 8020 生产接口（serve_card_match_v2）排队/缓存 ====================
# 全部支持环境变量覆盖；本地调试与 249 生产用同一套默认值
V2_QUEUE_SIZE = int(os.environ.get("V2_QUEUE_SIZE", "32"))          # 任务队列上限，满即 429
V2_WORKERS = int(os.environ.get("V2_WORKERS", "0"))                 # 推理 worker 数；0=自动(=可见GPU数,至少1)；>GPU数=每卡多worker(压测后再开)
V2_WAIT_TIMEOUT = float(os.environ.get("V2_WAIT_TIMEOUT", "60"))    # handler 等结果超时(秒)，超时 504
V2_CACHE_MAX = int(os.environ.get("V2_CACHE_MAX", "256"))           # 内容 md5 → 结果缓存条数（LRU）
V2_CACHE_TTL = float(os.environ.get("V2_CACHE_TTL", "900"))         # 缓存 TTL(秒)
V2_MAX_CONTENT_MB = int(os.environ.get("V2_MAX_CONTENT_MB", "20"))  # 上传图大小上限(MB)

# ==================== 卡牌分割近距 mask 合并（yolo_detector.detect_and_crop_all）====================
# 两实例中心（mask 质心，回退 bbox 中心）距离 <= CARD_MERGE_DIST_PX 视为同一张卡的重复分割，
# 保留 conf 高者；CARD_MERGE_DIST_PX<=0 关闭合并。
# CARD_MERGE_MIN_IOU：可选 IoU 门槛（>0 启用：距离近且 IoU 达标才合并，防误杀紧贴的不同卡）；
# 0 = 纯距离口径（默认，与 8020 接口定稿一致）。
CARD_MERGE_DIST_PX = float(os.environ.get("CARD_MERGE_DIST_PX", "50"))
CARD_MERGE_MIN_IOU = float(os.environ.get("CARD_MERGE_MIN_IOU", "0"))