feat: ranked top-3 queen candidates with probability shown - honest AI, user verifies

fix: dynamic inference size (640/1280/1920 by image dimensions) - small thumbnails were being upscaled too aggressively

tune: imgsz=1920 + lower thresholds to catch more bees on high-res user photos

feat: 5 sample photos with verified queen detection + disclaimer about operating conditions

fix: cascade prompt emphasises color + size, more candidates, lower bee threshold

feat: VLM cascade for queen verification (YOLO finds candidates, Qwen-VL confirms via grid comparison)

feat: lower queen threshold to 0.05, show top-confidence per class in narrative

fix: distinguish empty detections from missing weights; lower per-class thresholds for retrained model

feat: ship retrained YOLO weights (queen mAP 0.992) + smarter class canonicalization

polish: remove emojis and em dashes from sources and docs

feat: class aliasing + varroa support, ready for hendricks_ricky weights

fix: per-class confidence thresholds, queen-only label, mites clearly marked as VLM estimate

fix: custom annotation — thin per-class boxes, only label queen+drone, no bee text

fix: force YOLO onto CPU (ZeroGPU rejects CUDA ops outside @gpu )

debug: log detector path/size at first call + retry instead of caching failure

feat: YOLO + Qwen pipeline with graceful fallback when weights absent