Add /test-2: Point cloud rendering from RGBD sprite sheets

[billybjork]

Summary

Introduces a new /test-2 route that explores an alternative rendering approach for RGBD sprite sheet animations. Instead of the mesh-based displacement technique used in /test, this implementation treats RGBD data as a point cloud, where each pixel becomes an independent 3D point. This eliminates mesh stretching at depth discontinuities while creating a visually compelling aesthetic where depth edges naturally dissolve into particle scatter.

Key Changes

• New route & template: routers/test_2.py and templates/test-2.html following existing patterns
• Point cloud renderer: Replaces DepthThumbnail mesh geometry with THREE.BufferGeometry + THREE.Points, using vertex/fragment shaders to sample RGBD atlases and position points in 3D space
• Enhanced camera control: Adds mouse parallax tracking (X/Y offset) combined with scroll-based tilt, with smooth lerp interpolation for immersive perspective shifts
• Mobile support: Optional device orientation (gyroscope) integration for mobile parallax
• Configurable point cloud styling:
  • Point density (stride-based sampling), size, shape (square/circle/soft), and attenuation
  • Depth-based sizing and opacity for visual depth cueing
  • Edge scatter effect at depth discontinuities using noise-based dispersion
  • Depth of field simulation and multiple color modes (RGB, depth-colorized, normal-colorized)
• Reuses existing systems: Animation loop, scroll-based frame advancement, asset loading, and control panel UI patterns from /test
• Updated main.py: Registers the new router

Implementation Details

• Vertex shader approach: UV grid buffer (static) + frame index uniform allows efficient frame changes without buffer updates; shaders sample both RGB and depth atlases to compute 3D positions and colors
• Performance: ~115K points per thumbnail at 0.5 density (640x360), well within Three.js budget for 60fps
• Visual direction: Creates a "living photograph dissolving into particles" aesthetic — smooth regions appear continuous while depth edges naturally scatter, reinforcing volumetric perception
• Mouse tracking: Normalized cursor position (-1 to 1) drives camera X/Y offset with configurable strength; smoothly returns to center on mouse leave
• No asset pipeline changes: Reuses all existing sprite sheets and metadata from static/test/rgbd-sprites/

https://claude.ai/code/session_01ESEahvkYKaUe1SzmYVebqc