LAMMPS 30Sep2026 SLLOD: 500-atom Lennard-Jones shear

Fresh complete calculation: bash run.sh
Provided outputs only: bash analyze.sh
Fresh outputs again: bash analyze.sh work

Two sequential actual Slurm jobs: one P100, one V10016. Each reserves one GPU
and two CPU cores; one MPI/one thread. Two hardware platforms with Kokkos, two shear rates, two
seeds, plus four P100 GPU variants = 12 protocols. TC and V10032 are excluded.
LAMMPS binaries/modules are already installed. Queue/account permissions apply.

Units lj, rho .8442, epsilon=sigma=mass=1, cutoff2.5 and energy-shift yes.
FCC5^3=500 atoms; dt.001, 500 warmup steps, 3000 shear steps, 31 saved frames.
Warmup is brief; do not interpret the starting state as fully equilibrated.
These are teaching trajectories, not converged steady-state viscosity.

With peculiar no: velocities contain streaming, remap v, temp/deform removes
the prescribed affine contribution; check the measured velocity profile.
With peculiar yes: stored velocities already thermal, remap none, use temp;
do not subtract the streaming contribution again. Kick yes adds initial flow
once, so do not simultaneously add a manual velocity ramp. fix deform N=0 is
used with the reversible SLLOD integrator introduced in 11Feb2026.

The supplied eight base source profiles were recomputed using native LAMMPS rerun from
the saved coordinates and velocities, without integrating any additional
steps. The actual original source input is retained: some base ave/chunk formats
lacked a leading separator. New worker.py uses format " %.16g" directly.
Reference outputs are supplied results, not newly executed simulations.

Original mixed-precision GPU trajectories had cut-off force discontinuity
differences above the unchanged source tolerance and are not qualified public
examples. This packet uses double-precision Kokkos consistently.

Independent pair sums verify shifted LJ energy, full atomic forces, virial
pressure, thermal/raw temperatures, stress sum, bin counts and velocities.
Public checks use the same physical tolerances as the source. Fresh chaotic
GPU trajectories may differ and figures need not be byte-identical. Each
fresh trajectory must satisfy its own physical identities. Source-reference
figures are regenerated from immutable provided data and checked identically.
No fitted constants, smoothed stress, hidden clipping, software binaries,
credentials, scheduler receipts or private user data are included.
