User reports all 34 cells classified as 미인식 with score 0.00 even when the grid was correctly cropped. Multiple compounding issues: 1. _is_empty required mean<60 (dark) AND std<14. HDR/bright captures produce pinkish empty slots with mean ~150-180, so even empty cells fell through to template matching. Drop the mean check; uniformity alone (std<18 grayscale, std<22 per-channel) is the real signal. 2. Score 0.00 across the board strongly suggests templates list was empty (only path that returns exactly 0.0). Track per-bucket load counts (slabs_ok/fail, artifacts_ok/fail) and surface them in the GUI status bar so a CDN failure is immediately visible. Currently no signal at all on download failure. 3. min_score 0.55 was tuned against simulator-clean renders. Real game captures have decorative cell borders, stack-count badges in corners, HDR shader effects. Lower to 0.35 and inset cell crops by 16% on each side before matching to skip the decorative frame. 4. Add 디버그 저장 button + dump_debug() that saves screenshot.png, bbox_crop.png, cells/<row>-<col>.png, and report.txt with top-3 matches per cell to %LOCALAPPDATA%/sephiria_inv/debug/<timestamp>/. Lets us iterate on tuning from real captures without round-tripping raw screenshots through chat each time.
10 KiB
10 KiB