[piggy-penguin]

Summary

Implements the next v0.6.0 immediate action in docs/roadmap/current.md: define the large-asset benchmark corpus.

• tests/benchmark/corpus.json declares four tiers, each with a recorded license:

  • fixture: the committed synthetic fixtures, one per format.
  • corpus: the six committed CC0 8,192-Gaussian subsets.
  • scale: a generated ladder of 1,048,576 / 4,194,304 / 7,995,392 Gaussians in PLY, SPZ, and SOG. The top step is the largest whole number of tiles under the 8M import ceiling.
  • real: full-size exports kept outside the repository. These are the four retained CC0 source exports, the local Postshot cactus files, and the Graphdeco garden/bicycle, which are download-on-demand and non-commercial.

  The manifest also sets provisional resource targets: metadata-only read and CanRead at most 0.25 s, and peak resident memory at most 0.35 GiB + 0.60 GiB per million Gaussians.

• scripts/benchmark_corpus.py has four commands:

  • list and verify.
  • generate: tiles the committed CC0 yashica-t4-top8192.ply on a grid and encodes all three formats from the same rows. It checks the output against the recorded hashes: file SHA-256 for PLY and SOG, and the decompressed-payload SHA-256 for SPZ.
  • run: measures each asset through benchmark_import.py, one process per asset, and gates the result on the expected count, SH degree, and targets.

• The SOG fixture generator gains webp_lossless_planar. It produces the same VP8L bitstream as the reference encoder, built from per-channel bytes so it works on multi-million-texel planes. The committed fixtures regenerate byte-identical.

• CTest, composed root build only: benchmark_corpus_manifest runs verify, and benchmark_corpus_ci measures the fixture, corpus, and 1M rows. Both run in the workspace-ctest-pr-* cells.

• Fix: on Linux, benchmark_import.py now reads peak memory from VmHWM. ru_maxrss starts from the parent's RSS at fork and survives exec, so it inflated the peak when a large parent launched the benchmark.

• Docs: new docs/reference/BENCHMARK_CORPUS.md. The docs index, BUILDING.md, WORKSPACE.md, PERFORMANCE_BASELINES.md, the roadmap, the backlog, and the CHANGELOG are updated.

Why?

The roadmap's step 2 is to "define and measure large-asset validation targets". No redistributable real asset reaches millions of Gaussians, so this adds a deterministic generated ladder up to the import ceiling. It sits beside the real references that design policy §12.1 names.

Validation

• python scripts/benchmark_corpus.py verify: 26 assets, 0 problems. The planar encoder matches the reference encoder at four image sizes.
• generate --tier scale: all nine assets generated in about 2 minutes. A second generate --ci into a fresh directory reproduced the recorded hashes.
• The nine scale assets were decoded locally on Linux with the three bundles' C++ decoders, outside USD. All decode with 0 CheckCloudContract violations, and decoded positions agree across formats within quantization.
• plugins/gaussian-sog/tools/generate_fixtures.py regenerates the fixtures with no diff.
• The run plumbing (gating, report, temp-dir cleanup) was exercised against a stub pxr. OpenUSD is not available in this environment, so the real benchmark_corpus_ci measurement is left to the hosted workspace CTest cells on this PR.

Checklist

• I linked related context or explained why there is no issue.
• I ran the relevant tests and checks, or described what remains.
• I updated documentation or tests when behavior changed.
• I did not add private or unlicensed assets. Generated assets are derived from a committed CC0 subset and are never committed.

Reviewer notes

• The targets are provisional until hosted rows exist. If a cell fails a memory or time bound, that is a measurement to record first, not a reason to loosen the bound silently.
• The scale tier represents costs that grow with Gaussian count. It does not represent costs that grow with source bytes: SPZ repeats deflate well, and SOG planes use the uncompressed flat-code VP8L. BENCHMARK_CORPUS.md spells this out.

---

Generated by Claude Code