deploy push for beverage (deploy)

miner.py

@@ -25,25 +25,6 @@ OLD v7 (kept for context, see miner.py.v7_backup_*):
   - alfred-aligned: conf=0.55, iou=0.5, TTA=union-then-NMS-then-boost
   - yolo26n FP32 (~10.3 MB)
 """
-
-Changes from v4:
-  - Input resolution 640 -> 1280. Stadium-CCTV frames have 30-70 px objects
-    that go below the detector head's effective stride at 640. At 1280 they
-    stay detectable. Confirmed: 0 detections at 640 vs >=1 at 1280 on the
-    actual challenge-61525 frame from manako.scoredata.me.
-  - cv2.INTER_CUBIC for upscale (linear for downscale): preserves small-object
-    pixels when scaling 1408x768 -> 1280-letterbox.
-  - TTA enabled: horizontal flip + per-class hard NMS merge + confidence boost
-    (each kept box gets max conf among overlapping cluster). Boost pushes
-    consensus detections higher in the rank-ordered conf curve, helping mAP.
-  - Cross-class dedup at IoU 0.7 (after per-class NMS, in single-frame path).
-  - ort.preload_dlls() instead of manual nvidia-* ctypes preload.
-  - Single postproc handler (no raw fallback) since our export is e2e [1,300,6].
-
-v7 model is trained natively in validator class order [cup, bottle, can] on
-merged_v7_aug (38k v1+OI images + 10k CCTV-degraded augmentations, 30% ratio),
-so cls_remap is identity. Compare to v5 emu which used [1,2,0] remap.
-"""
 import math
 from pathlib import Path