README: document reflect-pad DPT swap (0.646 -> 0.314 RMSE on 5050-frame flat val)

README.md

@@ -96,6 +96,8 @@ A DPT multi-scale decoder was trained and evaluated alongside the linear probe d
 | Linear probe | 0.520 | 0.303 | 0.860 |
 | DPT multi-scale (current) | **0.480** | **0.219** | **0.872** |
 
+The DPT decoder's 3×3 convolutions originally used zero padding, which compounded boundary artifacts across the four fusion stages and required a post-hoc border crop at inference. Switching the convolutions in `_ResidualConvUnit`, `_DPTReassemble`, and the final head to reflection padding removes the artifact at its source, so the crop is no longer needed. Measured on the project's 5050-frame NYUv2 flat val split (a different protocol than the Eigen test used above), the swap drops RMSE from 0.646 (legacy: zero pad + crop) to 0.314 (current: reflect pad, no crop).
+
 ### Segmentation Decoder Comparison (Development Reference)
 
 The same DPT multi-scale architecture that improved depth by 8% was also trained and evaluated for semantic segmentation on ADE20K, using the same reassemble + fusion approach with a 150-class classification head and cross-entropy loss. The DPT segmentation decoder did not improve over the linear probe.