# Shared stage 06 visualization — V3

Stage 06 is now one shared section per scene, placed after reconstruction and setup and before the two framework columns. It exposes three visual steps: untextured 3D body motion; the transport of all prepared colored foreground points; and the existing source-camera RGB preview. Stages 07 and 08 retain the separate native RealWonder and depth-bridge conditions and generated results.

Four new CPU replay clips use the 25 committed bridge poses at 16 Hz (0–1.5 s), without physical integration or model inference. The lamp view uses its box collision proxy; the persimmon view uses full source rigid meshes. The material view uses every prepared foreground point and its original color. New 3D viewpoints are fixed, diagnostic and equally scaled, with the source-defined support plane shown only as a finite display patch. They do not replace the source camera or conceal its cropping.

The three-view controls pause and seek to the same saved frame, or start them together at their original cadence. All historical videos remain available. The separate native preview has been moved into the shared section’s provenance disclosure; it retains its actual 81-frame / 50-Hz contract. Its complete state history was not saved, so it is not silently rebound to the bridge replay.

Rigid mesh replay follows the solver’s source mesh recentering and saved body poses. Colored-point replay follows `pose(p,q) @ inverse(initial_render_transform)`, exactly as the completed source renderer does. This retains the existing initial-placement/render-binding distinction, including case offsets. No retrospective alignment, new texture, camera reconstruction or scientific correction has been introduced. The raw Genesis viewer recording remains unavailable (`debug=false`).

Per-scene `SHARED_REPLAY_V3.json` records exact files, frames, transforms and numerical checks. The former playback-publication handoff and V1/V2 results remain historical. This revision changes the visualization structure and adds saved-state replays; the existing scientific results are unchanged.

Numerical replay check: NumPy and CPU Torch FP32 homogeneous point transport differ by at most 0 for lamp and 4.76837158203125e-7 for persimmon, in prepared reconstruction units. All four new movies decode to 25 frames at 16 Hz and 960×720. Both complete foreground banks are used: 19,394 lamp points and 37,111 persimmon points. Existing pose arrays, input colors and historical camera RGB were not changed.

## Publication verification

Local and live Chrome checks passed all **20/20 videos** and both three-view groups: simultaneous playback and frame-12 alignment at 0.75 s. All 89 linked files returned HTTP 200 with matching content, allowing only the known Hugging Face variables script in HTML. No source/model results were modified. New CPU replay movies total 4 clips / 100 rendered frames, each 960×720, H.264/yuv420p, 16 Hz.