Skip to content

four_scene11 selected-source snapshot

Historical record for snapshot/four-scene11-20260908 at 1444351. The separate PR preparation branch now includes upstream main and installed import/packaging fixes; see PR preparation and fresh validation. Results below remain attributed to the original snapshot and runtimes.

This branch preserves the selected four-scene integration before reconciliation with upstream. It is a development snapshot, not a release or a passing upstream integration. No upstream merge or default-runtime switch is included.

Included source

Starting commit: 56789eb9dcedd0b8bf7b0d0feaa46b9f006f3d6a.

  • 181 changed core, binding, demo and build files from the frozen four_scene11 tree.
  • 91 frozen vendor/warp_native files, including their license. These are source dependencies used by the native CUDA build. Some beam helpers retain generated or adapted Newton/Warp code; this is not a claim of an entirely Warp-free implementation.
  • 122 old paths removed: 121 test/helper relocations paired with their new paths, and one old MPM-local SVD header superseded by the common math header.
  • Necessary test-layout documentation and collection support, plus the coloring, constitutive-diagnostic and beam-binding regression tests associated with the preserved changes. Their directly imported elasticity helper and dynamically loaded coloring helper are included.
  • Version provenance, selected-file hashes and known validation limits.

The 121 relocated files retain the frozen final content; some also contain pre-existing content changes. See test_renames.json and selected_files.json. This branch does not carry all of the research scripts and fixtures from the full local archive. Unchanged files remain from the starting commit, rather than being copied from an unrelated worktree.

The accepted scene candidates are multi-material visual e467, viscous QS, snow GP and beam validated CR velocity, built on the warmstart06 memory fixes. The paused two-way candidate and held beam postprocess_connected candidate were not promoted. Inherited two-way APIs/demos remain part of the base engine; their presence does not certify the deferred two-way candidate.

Evidence and known failures

Historical tests of the full local integration (2026-09-07):

  • four_scene11: 27 focused regressions and four two-frame scene sequences passed Compute Sanitizer memcheck with zero reported errors.
  • four_scene11: multi-material, viscous and snow ran 300 frames; beam ran 1200 frames. Sampled state fields were finite float32. Four Viser viewers ran ten frames, exited normally and returned HTTP 200; no browser-pixel assertion.
  • four_scene11: 55 API/CPU-math tests passed.
  • Full pytest was run on the four-package beam10 candidate before the final formatting rebuild: 5051 passed, 530 failed, 97 errors, 4337 skipped, 4 xfailed. These results are not a fresh full-suite run of this selected-source branch.

Unresolved algorithm and framework issues are intentionally retained:

  1. PIC-B strict comparison first regressed at the viscous common-dependency increment: max absolute difference 0.0099679, max relative difference 0.13173506. The precise formula-level cause remains open.
  2. Snow is not bitwise repeatable, including repeat runs of the same binary. Long-run stress and deformation ranges also differ. Finite sampled state does not certify strict Newton trajectory or indefinite stability.
  3. Two-way stress differences remain unresolved; the paused two-way package was not promoted. Three reference-side frozen local arithmetic comparisons also remain red.
  4. Existing float32/state/backend/capture conformance gaps and full-suite failures are not waived. Some research tests still require external fixtures or reference environments not included here.
  5. Some historical scene scripts inject the source path for direct-script execution. This conflicts with the migration guide's general source-path policy and remains recorded rather than silently refactored in a source-preservation commit.
  6. Native Warp dependency acceptance versus older module-specific “no Warp” contracts requires explicit architecture review before upstream promotion.

No tolerances, expected results or solver formulas were changed to make this snapshot pass. Fresh publication checks are recorded in publication_checks.json: the selected source rebuilt successfully, 11 CPU binding tests passed, and 1957 tests collected without errors. Collection is not test execution. One missing dynamic test helper was found and copied unchanged from the frozen source. The optional full coloring research probe is outside this snapshot validation scope.

Building and running

See the repository installation guide and build guide for the normal CMake/Python build workflow. Enable MPM CUDA explicitly and select the target CUDA architecture. The historical integration used Linux x86_64, Python 3.11, GCC 11, CUDA 12.6 and sm86, with IPC disabled. New machines must supply their own Python, CUDA, vcpkg dependencies and compatible driver; no prebuilt runtime or virtual environment is included in Git.

The original novaviz commands relied on a private batch directory and installation manifest. They are not universal commands for this branch. Standard package/demo installation portability has not been certified by the historical private-runtime tests, especially the demos / tools.mpm import layout. This snapshot preserves source without claiming a new portable distribution.

At initial publication preparation, the host NVIDIA driver was unavailable. After the user rebooted the server on 2026-09-08, driver 580.173.02 became usable. Both the original four_scene11 runtime and the publication12 runtime rebuilt from this uploaded source passed the bounded post-reboot checks: four Viser scenes with HTTP 200 and ten frame updates each; 27 focused regressions and four two-frame scene sequences under Compute Sanitizer with zero reported errors; and the existing 300/300/300/1200-frame sampled-state sequences. See post_reboot_validation.md and its machine-readable evidence summary. This does not close the known strict Newton, full-engine or clean-machine packaging gaps. No engine code, runtime library, driver or shared conda installation was changed during the recheck.

Preserved archive

The full local delivery archive, private runtimes and failed-run evidence remain unchanged outside Git. Archive and historical core/library hashes are recorded in provenance.json. Selected scene observations are in scene_observations.json; beam helper origin mappings are in beam_helper_sources.json.

The archive, runtime .so files, build outputs, virtual environments, caches and large temporary measurement arrays are deliberately not part of this branch.