CelerisLab

Result: PASS. First local execution only; no retry or tolerance change.
15 core tests + 5 supplemental tests passed, plus 1 open-corner negative-control
method (identity/SRT, seeds 1/2, 1/8 coarse steps). The supplemental suite includes
the v0.2-to-v0.3 diagnostic-field delta check; this is not physical-field equality.

Scope: fixed doubly periodic 16x16 coarse grid, one static 2:1 patch covering
[4,12)^2, 16x16 active fine cells with halo 4. Evidence concerns transport,
corner handling, validity, ownership, conservation and bookkeeping only.
It does not validate interface accuracy, viscosity/stress transfer, long-time
stability, production GPU equivalence, dynamic AMR or Jet integration.

Static review completed before execution; no substantive objection within this
scope. Checked arrival_f phase, even integer tick and +2 advancement, collision
order, both fine hops, ghost validity/poisoning, eight uniquely keyed corner-pass
links, coalescence ownership and budget semantics. bookkeeping_check subtracts
interface_residual and is not independent conservation evidence. Core tests
provide a separate scalar transport/SRT oracle and active-leaf totals. Negative
control sums leaves outside the reference helpers, but uses reference masks and
collision budgets; do not describe every aspect as independent.

Delivery: five Python files were moved from review staging into the parent
multigrid directory without changing bytes. cpu_reference_v0.2.py is retained
only as the supplemental diagnostic comparison fixture, loaded explicitly.
Seven delivered artifact SHA256 values match the supplied originals both before
and after execution. No reference algorithm, test assertion or tolerance edits.
Grok's original README and stdout remain unchanged alongside the local evidence.

Local environment: /home/frank14f/anaconda3/envs/pycuda_3_10/bin/python
Python 3.10.13; NumPy 1.26.4; Linux 6.8.0-138-generic x86_64, glibc 2.35.
Grok reported Python 3.13.5 / NumPy 2.5.3; both original and local outputs are kept.
Process-only environment:
OPENBLAS_NUM_THREADS=1
PYTHONDONTWRITEBYTECODE=1
PYTHONNOUSERSITE=1
PYTHONPATH=/home/frank14f/CelerisLab/tests/validation/multigrid
CPU_REFERENCE_PATH=/home/frank14f/CelerisLab/tests/validation/multigrid/cpu_reference.py
Working directory: /home/frank14f/CelerisLab/tests/validation/multigrid
Each script was executed sequentially by the following command, replacing SCRIPT
with test_cpu_reference.py, test_tron_additions.py, then
test_open_corner_negative_control.py (absolute paths in results_horsea.json):
/usr/bin/prlimit --as=268435456 --fsize=2097152 --core=0 -- /home/frank14f/anaconda3/envs/pycuda_3_10/bin/python -B SCRIPT
The stdlib-only parent runner had the same hard limits. stdout and stderr of all
three scripts went to one shared stdout_horsea.txt; any nonzero exit would have
stopped subsequent execution. Per-case step ceiling of 8 was verified statically;
no grid, seed or step expansion. Peak child RSS across the run: 37.5 MiB.
Combined local stdout/stderr plus command metadata: 9013 bytes (<2 MiB).

Observed evidence: identity versus scalar oracle max difference 0; SRT max
2.776e-16. Eight-step SRT mass drift -1.535e-12 passes the unchanged delivered
mass-scaled 1e-14 tolerance (approximately 2.56e-12 absolute). Open-corner SRT
maximum drift reaches 2.078e-3; closed negative-control drift stays <=5.684e-14.
The eight-step mass tolerance is not an absolute 1e-14 bound.

Increment: only new files under tests/validation/multigrid. No src, old test,
MULTIGRID_VERIFICATION_STRATEGY.md, DynamisLab, dependency, environment or Git
commit/push changes. No CelerisLab/Simulation/pycuda imports or GPU/solver runs.
The original 269-file hash baseline and Git status are in baseline.json; final
verification and new-file sizes/line counts/SHA256 are in manifest.json. Ordinary
git diff omits these untracked additions; the manifest describes this increment.
No generated bytecode or duplicate candidate copies were retained.

Stopped after this authorized test-only integration for joint review.