Lane B (M2 + M8 geometry): COLMAP reconstruction, geometry math, graceful degradation
Implements spec M2 (frame sampling, COLMAP orchestration, TXT-model parsing, scene
normalization, pose interpolation, PLY + DB export) and the M8 geometry functions.
geometry.py (M8): slerp_pose (shortest-arc, double-cover), ray_from_pixel (COLMAP
+y-down back-projection), triangulate_rays (closest-point + parallel guard),
nearest_point_on_ray (in-front radius cylinder). Frozen colmap_to_threejs untouched.
frames.py (M2): variance-of-Laplacian sharpness + windowed sharpest-frame sampling.
sfm.py (M2): hand-written COLMAP images/cameras/points3D parsers; normalize_scene
(centroid->0, camera sphere r->10, up->+Y) as one similarity transform over points+poses;
interpolate_poses (slerp+lerp, no extrapolation); build-aware COLMAP CLI orchestration
(3.x/4.x option detection, CPU SIFT, single-camera-per-folder, undistort->TXT, largest
component by images.bin header count); run_reconstruct with full graceful degradation
(COLMAP absent / <60% / <2 videos / no frames / missing files -> poses left untouched).
Tests: 61 pass (24 foundation + 37 lane-B) with independent oracles (scipy Slerp,
projection-inverse, closed-form geometry) and every degradation/DB-safety branch covered.
Validated end-to-end against real COLMAP 4.1.0 (6 components, largest selected, weak ->
graceful degradation, exit 0).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
bb542960c2