Targets rebuilt so the picture a person reviews is the picture the model is trained and scored on. Every V4.1 target is RGBA and PIL's convert("RGB") -- used by the benchmark, the training image loader, the vLLM inference path and core/vision.py -- drops alpha and keeps the RGB stored underneath, while a browser composites.
- 61 targets hid a neighbouring widget under the mask; those pixels are now white. On them the pool's own reference render scored SSIM 0.5532 against the stored target and 0.6990 against the flattened one, against a pool mean of 0.7664. White was measured against the alternatives: white 0.7163, unchanged 0.6161, black 0.5557.
- 108 targets are cropped to the bounding box of their non-transparent pixels; a transparent margin is what the capture left, not part of the design. Opaque white margin is kept.
- side_info regenerated from the new pixels with w2c-bench 1.2.0 on CPU; sha256 updated; eval dropped as a stale ground-truth cache; flattened_from records the source…
