SOG writer hangs indefinitely when output has >1 SH band (works with -H 1; SPZ output of identical input completes in seconds)

[nickk123456]

Summary

splat-transform's SOG writer hangs in the SH-clustering step when the output keeps more than 1 SH band. The same input with the same flags but output format .spz instead of .sog completes successfully in ~60 seconds. So the conversion math (decimate, NaN filter, Morton reorder, SH retention) is fine — the hang is specific to the SOG writer path when SH bands > 1.

Version & environment

• @playcanvas/splat-transform@2.1.1
• GitHub Actions ubuntu-latest runner
• Node.js 20
• CPU mode (-g cpu)

Reproduction

Input: a 430 MB Brush-trained PLY with ~5M Gaussians. Decimated to 1.4M (or 750K) for output via -F.

# These three all hang past 90 minutes and had to be cancelled:
npx splat-transform -g cpu -i 2 input.ply -F 1400000 -N -M -H 3 output.sog
npx splat-transform -g cpu -i 2 input.ply -F 1400000 -N -M -H 2 output.sog
npx splat-transform -g cpu -i 2 input.ply -F  750000 -N -M -H 3 output.sog

# Identical input + flags, output format changed to .spz — completes in 62s, 23.67 MB:
npx splat-transform -g cpu -i 2 input.ply -F 1400000 -N -M -H 3 output.spz

# Adding -H 1 to the SOG runs also completes fast (~5-10 min, ~12 MB):
npx splat-transform -g cpu -i 2 input.ply -F 1400000 -N -M -H 1 output.sog

Expected

SOG output should complete in comparable time to SPZ output, since both formats represent the same underlying clustered data.

Observed

Output	SH bands	Splats	Result
.sog	3	1.4M	hung >90 min, cancelled
.sog	2	1.4M	hung >90 min, cancelled
.sog	3	750K	hung >90 min, cancelled
.sog	1	1.4M	✅ ~5-10 min, ~12 MB
.spz	3	1.4M	✅ 62 sec, 23.67 MB

GH Actions runs (public, logs viewable):

• 1.4M / 3SH / SOG (cancelled after >90 min): https://github.com/nickk123456/formchasers/actions/runs/25999710696
• 1.4M / 2SH / SOG (cancelled after >90 min): https://github.com/nickk123456/formchasers/actions/runs/25999711717
• 750K / 3SH / SOG (cancelled after >90 min): https://github.com/nickk123456/formchasers/actions/runs/25999712576
• 1.4M / 3SH / SPZ (succeeded in 62s): https://github.com/nickk123456/formchasers/actions/runs/26001551381

Test workflow that produced these runs: https://github.com/nickk123456/formchasers/blob/main/.github/workflows/test-multi-sh-conversion.yml

Impact

We can't ship multi-SH SOGs to our Gaussian splat gallery — viewers see matte materials instead of the real specular shine on chrome, paint, glass. Worked around for now by switching the high-quality tier output format to SPZ.

Happy to provide the source PLY (publicly accessible URL in the workflow logs) if it helps reproduce.