These results concern the new upstream contribution, separate from the historical MetaFlow deployment screenshots in the main evidence record.

All commands ran on Node 24.21.0 on macOS.

• Generator: npm test passed (851 passed, 1 skipped); explicit TEST_WEBGPU=1 node --import tsx --test --test-force-exit test/voxel-tiles.test.mjs passed 12/12 tests. Formatting, lint, typecheck, publint and Typedoc passed. Typedoc emitted 13 pre-existing warnings; publint emitted an existing metadata suggestion.
• Generator assertions cover decoded core occupancy against monolithic output, rotated cross-boundary Gaussians, zero-overlap seams, shared grid alignment, transforms, empty tiles/gaps, capacities, failed reads and failed binary writes. A final fixture-only enlargement was followed by another 12/12 CPU/GPU run.
• Viewer: build, typecheck, formatting, lint, publint and all 64/64 unit tests passed. Behavior tests include request deduplication, neighborhood changes, delayed results, eviction/cancellation, retries, destroy, absent coverage, occupied overlap union, legacy format loading, walking safeguards and overlay allocation/disposal.

GitHub-hosted CI has not run its jobs yet: the generator workflow and viewer workflow both report action_required, pending upstream approval. The viewer Vercel preview reports Authorization required to deploy and requests authorization from a PlayCanvas team member. These are external authorization gates, not passing hosted checks or observed build failures.

A headed Chrome session at 1280 × 900 on macOS loaded the 2,220-Gaussian synthetic fixture generated by the new writer. The browser reported SuperSplat Viewer 1.37.0 / Engine 2.23.0. A small local harness used the public createViewer, observable state, toggleWalk, setMoveInput and destroy APIs; the unchanged public HTML query-parameter entry point was loaded separately too.

• The manifest contains six standard voxel 1.1 tiles. At the initial camera [-2, 2, 0], four nearby JSON/bin pairs loaded successfully; the two farther tiles were not eagerly requested. Moving to the next core loaded its neighbors.
• On WebGPU, walk became available, settled on the floor, crossed the X seam at -0.4, and stopped at the wall at approximately [0.20006, 1.95, 0] after forward input. The wall is centered at X 0.8; the camera did not pass through it.
• On WebGL2, the same traversal stopped at the same wall position. walkAllowed remained true and hasCollisionOverlay was false, preserving the existing backend behavior.
• The single-voxel comparison rendered its overlay and stopped at the wall with the same movement input.
• WebGPU overlay and heatmap rendered. Hiding the overlay and destroying the viewer completed without console errors or warnings; destruction left zero canvas elements. Automated lifecycle tests verify tile and GPU resource cleanup, since a screenshot cannot establish GPU allocation lifetime.

Tiled collision overlay at the initial camera:

Single-voxel comparison at the same initial camera:

Heatmap of the active tiles:

This fixture establishes a working producer/consumer path and one seam/wall traversal, not exhaustive walking validation of the original large scenes. Slow/failing requests, retries, stale results and missing coverage were exercised in deterministic behavior tests, not by claiming a full live-network stress test.

The debug viewer composites one independently ray-marched result per tile. Overlap can be brighter, and farther-tile geometry is not globally depth-merged behind nearer tiles. This is a documented draft limitation of visualization; collision queries combine occupied voxels as a union.

The generator still performs repeated source scans and only supports surface voxelization. In-place replacement is not transactional, and eager source decoders or prior actions can retain scene memory. The 149M-Gaussian source in #343 was not rerun as part of these checks. No whole-scene memory/performance improvement measurement is claimed here.