first commit

docs/contact-depth-geometry.md

@@ -0,0 +1,18 @@
+# Contact depth placement
+
+Depth uses normalized values where 0 is behind and 1 is in front. For each installed `DataTexture`, the app scans the red channel once and places the photo and depth mesh at
+
+```text
+contentBaseZ = glassZ - 0.04 - (maxDepth - depthBias) * depthScale
+```
+
+The depth mesh keeps a Z scale of 1, so its highest vertex remains 0.04 world units behind the glass as the depth scale changes. The XY photo size and pan are multiplied by `(cameraZ - contentBaseZ) / cameraZ` to keep the flat photo footprint stable as it moves closer. A flat depth map stays at the near-flat placement `glassZ - 0.04`.
+
+Mosaic, Reeded, and Frosted trace refracted rays to the same `contentBaseZ` plane. The Mosaic color-region atlas shares that uniform. Existing glass surface normals and Mosaic height cap are unchanged; no extra slab or optical surface is added.
+
+New still photos without paired manual depth start the existing Depth Anything V2 Small provider automatically. Videos and the paired local sample remain unchanged. Automatic depth keeps the current view mode; the manual AI button still switches to No Glass for parallax inspection. The model does not detect or crop faces, and the highest depth point may be a paw or another foreground part, so general uploads keep their chosen framing.
+
+## Verification
+
+- `npm run build` passed (`tsc -b` and Vite production build); Vite reports the existing large-chunk advisory.
+- Glass-off sample check at depth scale 0.25 reported `contentBaseZ=0.315`, `maxFrontZ=0.560`, and `glassGap=0.040`. At scale 0, the placement formula keeps the front at 0.560 with the base at 0.560. Scale 0.4 keeps the same front clearance by moving the base farther back.