Post-reboot visualization and bounded runtime validation¶
Date: 2026-09-08. Source snapshot: e160a4888d2e69eec581258a8b25647f0d6da7b1.
Bounded runtime checks: PASS for both runtimes. Driver 580.173.02 is now
available after the user's server reboot. four_scene11 is the original paired
runtime used by the existing visualization commands. publication12 was rebuilt
from the uploaded selected-source branch before reboot and was not rebuilt or
modified for these checks. Each runtime's complete installation hash manifest
was verified; every test process checked loaded NovaPhy and CUDA library paths.
Executed checks¶
| Check | four_scene11 | publication12 |
|---|---|---|
| Multi-material, viscous, snow and beam Viser | All four: HTTP 200, ten updates, normal exit | All four: HTTP 200, ten updates, normal exit |
| Focused lifecycle/diagnostic/multi-material/beam regressions | 27 passed; memcheck 0 errors | 27 passed; memcheck 0 errors |
| Four actual scene sequences, two frames each | 4 passed; memcheck 0 errors | 4 passed; memcheck 0 errors |
| Multi-material / viscous / snow / beam long sequences | 300 / 300 / 300 / 1200 frames, 4 passed | 300 / 300 / 300 / 1200 frames, 4 passed |
Viser used the documented solver flags, including snow compression and beam stress
display, with local test ports and NOVAPHY_VIZ_QUIT_AFTER=10. HTTP and actual
frame updates were checked; no browser screenshot/pixel assertion was made. All
test viewers exited after ten frames and are not left serving in the background.
Long tests reused the frozen test_scene_metrics.py observer. At the recorded
checkpoints, positions, velocities, velocity gradients, elastic strain, stress
and Jp were finite float32 arrays. Reads did not modify state or solver settings.
The following values describe the final sampled state, not strict physics gates.
| Runtime | Scene | Frames | Max speed (m/s) | Max absolute stress component (Pa) | Fe determinant min / max |
|---|---|---|---|---|---|
| four_scene11 | multi | 300 | 0.182016 | 311844 | 1 / 1 |
| four_scene11 | viscous | 300 | 0.879424 | 0 | 1 / 1 |
| four_scene11 | snow | 300 | 14.5337 | 279180 | 0.226788 / 12.9896 |
| four_scene11 | beam | 1200 | 0.275008 | 2.05919e+06 | 0.909108 / 1.0765 |
| publication12 | multi | 300 | 0.182016 | 311844 | 1 / 1 |
| publication12 | viscous | 300 | 0.879424 | 0 | 1 / 1 |
| publication12 | snow | 300 | 14.7853 | 370120 | 0.178945 / 4.16501 |
| publication12 | beam | 1200 | 0.275008 | 2.05919e+06 | 0.909108 / 1.0765 |
Scope and evidence¶
The test harness reused the full frozen delivery's test files, including the external R1 lifecycle test and scene metrics observer; this is not a claim that all historical harness/fixture files are in this selected-source branch. The JSON records test source hashes, paired library hashes, log/metric hashes, actual demo flags, test counts and sampled-state summaries. Raw logs and measurements remain in the independent local publication evidence directory, outside Git.
The two runtimes ran on separate idle RTX 3090 GPUs (0 and 1). Timing is diagnostic, not a benchmark or a controlled numerical equivalence comparison. Snow's known repeatability/long-trajectory differences, PIC-B strict mismatch, deferred two-way candidate and historical full-engine failures remain open. Passing sampled finite-state checks does not establish Newton equivalence, all-frame boundedness, indefinite stability or whole-engine memory safety. Full pytest execution and strict Newton reference comparisons were not rerun here. Clean-machine wheel/ editable-install portability remains uncertified.
The existing project-local visualization commands and solver flags remain valid; no engine code, package selection, original runtime, shared conda or default environment was changed. This update only adds verification results to the personal snapshot branch; it does not merge upstream or create a PR.
Machine-readable results: post_reboot_validation.json. Historical findings: snapshot README.