first commit

2026-10-06 12:16:46 +09:00
commit 619d980098
415 changed files with 22283 additions and 0 deletions
@@ -0,0 +1,36 @@
+# Surface reflection browser verification
+
+2026-09-23, local Vite preview at 1280×720 unless noted. The primary comparison used dog1 after the same cached AI relative-depth generation completed: Final / Mosaic / 조각 유리, Grid 13, curvature +1, gloss 0.30, light 0.60/0.80, depth scale 0.25, camera offset 0.50, zoom/pan 1/0/0.
+
+## Same-condition comparison
+
+- `before-default.png`: pre-edit shader, centered view.
+- `after-default.png`: final shader, centered view.
+- `before-right.png`, `after-right.png`: equivalent pointer-drag probe; these are qualitative because camera smoothing does not guarantee identical final coordinates.
+- The centered glass-pane crop changed by a mean absolute `0.3106/255` per RGB channel, maximum `14`, over `13.85%` of channels. The top-right quadrant contained nearly all of the visible change (`1.2244/255` mean absolute difference); the other three quadrants were below `0.013/255`.
+- Compared with gloss 0, the pre-edit default gloss affected `12.80%` of crop channels with `0.6569/255` mean absolute difference; the final default gloss affected `15.81%` with `0.9015/255`. This is a broader but deliberately subtle highlight, not uniform pane brightening.
+
+The dog’s two dark eye planes and nose plane remained visible. Grass and face were not visibly washed out. The centered difference is too subtle to prove visual improvement by metric alone; `after-glass-only.png` confirms that the widened reflection remains localized around the common top-right light rather than becoming a constant white overlay.
+
+## Preservation checks
+
+- `before-gloss0.png` and `after-gloss0.png` are pixel-identical in the pane crop: mean absolute difference `0`, maximum `0`, changed channels `0`.
+- `after-gloss040.png`: no obvious face/grass washout at gloss 0.40.
+- `after-curvature0.png`, `after-curvature-minus1.png`: flat and concave cases rendered without reflection spill; the dark facial features remained identifiable. Curvature +1 was the default comparison.
+- `after-camera090-right.png`: camera range 0.90 plus horizontal drag showed no reflection outside the pane, bright rim, or exposed background gap.
+- `after-portrait.png`: 390×844 viewport retained the frame/pane fit and dark eye/nose planes. The temporary viewport override was reset.
+- `after-indoor.png`: dog2 with completed AI depth rendered without reflection washout or pane-edge leakage.
+- Reeded and Frosted were selected and rendered, then Mosaic was restored. No runtime errors appeared.
+
+The console contained only the already documented Three.js shader warning `X3595` about gradient instructions in a varying loop. No new error was observed.
+
+## Final fresh verification
+
+After both regression-review fix rounds, the controller reran the current workspace state. `scripts/test-surface-reflection.mjs` and the six existing regression scripts all exited successfully. `npm.cmd run build` completed TypeScript and Vite production output successfully; the existing main-chunk size advisory remained.
+
+## Limits
+
+- The requested intermediate “broad reflection only, before edge-weight separation” GPU screenshot was not captured; the exact pre-edit and final states are preserved instead.
+- The pointer-drag screenshots do not record an exact camera coordinate and therefore are qualitative.
+- Continuous-video flicker, physical mobile GPU performance, and real-device touch/gyro were not measured.
+- The browser comparison establishes no obvious regression and a small widened reflection footprint. Whether the subtle default-strength change is aesthetically strong enough remains a product judgment, not an automated pass condition.