The historical simulations are not yet a validated matched code comparison. The fixes below do not retroactively correct existing simulations. Published meshes and MP4s are visualizations, not complete raw data backups.
Eight native grid codes passed 24 initial output tests: Athena++, AthenaPK, Athena 4.2, FLASH 4.8, Flash-X, Enzo, Enzo-E and RAMSES, each at χ = 10, 100 and 1000 on a 64 × 32 × 32 grid. Tests checked density, velocity, pressure, domain and cell count. The four particle code IC generators passed 12 separate input field tests; their evolved solver results were not validated here.
The corrected prescription has zero velocity inside 1.3 cloud radii and constant wind velocity outside, with the existing tanh density edge. Grønnow et al. (2018), section 2 specifies a sharp velocity boundary, not a tanh velocity law. This is not an exact reproduction of its entire setup or of Braspenning et al.
Analysis now reads physical parameters from input files, handles FLASH's radius and AthenaPK's units, and normalizes dense mass to its fixed initial value. The rectangular RAMSES reader now includes all native cells. Historical mass curves are withheld from the viewer until corrected native data diagnostics exist.
The destructive Athena++ rerun path was replaced, and the raw data cleanup daemon was disabled. The new runner preserves raw outputs and uses distinct versioned run IDs.
Performance update: Short matched setting tests on this PC reduced Athena++ wall time by 15% at level 5 and 18% at level 6; AthenaPK level 5 took 30% less time. Larger GPU blocks also reduced sampled level 6 GPU memory use from roughly 15.3 to 12.8 GiB. Full field test comparisons passed, including the native tracers. These are short tests, not full production timings or new completed comparison runs. Measured evidence, optimized launch scripts and reuse guidance.
Second performance pass: The current profiles now use a separately validated LTO build at CPU level 5 (6.2% less time in repeated tests) and faster lossless snapshot encoding at GPU level 6 (12.4% less time in repeated output including tests). Decoded GPU fields matched exactly. The CPU level 6 candidate was slower and was rejected; its previous settings remain. These improvements are not universal or full run accuracy claims. Second pass evidence and decisions.
Extended tracking tests: With the existing conditions unchanged, late restarts at all three chi values, MPI rank count checks and per step frame displacement checks pass. Three higher resolution L4 tests each reached 5 t_cc with 101 native snapshots. However, chi=100 and chi=1000 retain only 98.4% and 95.4% of their initial tracer mass. Material retention is still not fully validated. Measured results, scope and retained data locations.
Follow up repairs and native tests: Analysis now rechecks physical metadata against native inputs, rather than trusting a shared group label. The launchers also check Windows backing volume space, and two additional destructive legacy queues were retired. 42 Python tests and two viewer JavaScript checks pass. Full report, scripts, logs and test evidence.
Cooling correction: All nine retained cooling labelled inputs omitted the cooling enable switch, and their native logs report the cooling parameters unused. These are not radiative results. Their physical temperature units and Schure table also do not implement the requested Sutherland-Dopita setup. The viewer labels them explicitly; the old cooling generator now refuses to write another such input. Physical cooling parameters still need to be agreed and tested.
Tracking prototype: A separate Athena++ executable passed tracking off and restart tests. Six level 3 integration tests reached 5 t_cc with 101 native outputs each. Nevertheless, the tracked chi=1000 test retained only 88.9% of initial tracer mass in the box. The full tracer/buffer scheme and a material retaining comparison are not complete; the prototype is not in production.
Windows C: had about 35.6 GiB free at this check. Large replacement launches are blocked by conservative raw data retention budgets. No native output was deleted in this follow up. The viewer now preserves later measured times beyond t=5 and labels Mach variants separately; this cannot restore previously deleted native fields or frames that were never exported. Time labels now follow the displayed mesh while downloads are pending, and comparison panes show their own measured times. An isolated browser test verified the loading behavior and a real streamed final frame without JavaScript errors.
Equal boundary conditions, passive tracers and effective particle resolution across every code; complete corrected production runs at every level; complete measured frame cadence for all historical runs; validated cooling and Galilean frame variants. These remain necessary before drawing a matched code scientific conclusion.
Corrected scripts and test evidence · Methods, formats, directory layout and limitations · 3D viewer · Studio