Reproduce the public synthetic normalization input counterexample

These files run an original synthetic case through selected public SOL-ExecBench
input generation and compute_error_stats code. No SOL dataset, model, private
judge, cloud account, Modal client or CUDA compiler is required.

Files
  lab-research-affine-cuda-source.tar.gz: exact frozen 26-file public payload.
  lab-research-affine-cuda-runtime-lock.tar.gz: 63 unmodified files required by
    the existing locked installer (scripts, lock, requirements and metadata).
  lab-research-affine-cuda-result.json: unmodified retained CUDA observation.
  lab-research-affine-cuda-provenance.json: interpretation, attribution and limits.

Runtime prerequisites
  Linux x86_64, glibc >= 2.28, existing CPython 3.12.x with venv/ensurepip.
  The dependency lock selects 30 exact wheels, including Torch 2.9.0+cu130 and
  Pydantic 2.12.5. It verifies package versions, hashes and dependency metadata;
  it does not pin your OS, Python patch release or NVIDIA host driver.
  Installation downloads about 2.74 GB of compressed wheels from pinned official
  URLs and needs additional unmeasured unpacked disk space. It performs CPU-side
  imports, not a CUDA calculation. The same lock is used for CPU composition and
  CUDA execution; this is not a floating CPU-only Torch installation.
  CUDA mode additionally needs one visible compatible NVIDIA GPU and driver.
  The retained run used Tesla T4, driver 610.57.04, Python 3.12.10. Other hosts are
  not claimed validated. The driver requires actual NVIDIA driver-build metadata
  from /proc or NVML and records it separately from CUDA API compatibility.

Download the two archives, this instruction file, and the checksum file into
the current directory, then run:

  sha256sum -c lab-research-affine-cuda-reproduction.sha256
  mkdir sol-affine-repro
  SOL_DIR="$(pwd)/sol-affine-repro"
  mkdir "$SOL_DIR/payload"
  tar -xzf lab-research-affine-cuda-source.tar.gz -C "$SOL_DIR/payload"
  tar -xzf lab-research-affine-cuda-runtime-lock.tar.gz -C "$SOL_DIR"

Choose the absolute path to your existing Python 3.12 executable; the target
runtime directory must not exist. This command installs the pinned wheel closure:

  sh "$SOL_DIR/runtime-lock/install_runtime.sh" /absolute/path/to/python3.12 "$SOL_DIR/runtime"

Choose the mode you want to reproduce. Each output directory must be new; retained
results are never overwritten. CPU composition alone is not CUDA evidence:

  "$SOL_DIR/runtime/bin/python" -I "$SOL_DIR/payload/run_validation.py" --mode cpu-composition --output "$SOL_DIR/cpu-attempt"

For one visible GPU (select the appropriate local device index):

  CUDA_VISIBLE_DEVICES=0 "$SOL_DIR/runtime/bin/python" -I "$SOL_DIR/payload/run_validation.py" --mode cuda --output "$SOL_DIR/cuda-attempt"

Read the attempt's RESULT.json, events.jsonl and worker.stderr.txt. Successful
CUDA status is cuda_observations_match_expected: 14 observations, six incorrect
variants accepted before and rejected after the input patch, copied-reference
control accepted both times. Failures and incomplete metadata remain explicit.
The worker has a 120-second limit and bounded local termination/reaping; it does
not allocate, stop or bill cloud resources. Host/resource cleanup remains yours.

Scope
  One original 4x64 float32 fixture and seed42. The x tensor is held identical
  before/after within each backend run. CPU and CUDA RNG streams are not asserted
  bit-identical. The positive control is reference.clone(), not an independent
  implementation. This is not full benchmark coverage, a speed measurement,
  training evidence or production environment qualification.
  An earlier CUDA invocation lost its result during output transport. The public
  JSON is the separate retained run; the provenance file discloses that history.

Source and notices
  Upstream: NVIDIA/SOL-ExecBench at a9fa0804c793d438e70850c33fe34426e66d53dd.
  Original report: https://github.com/NVIDIA/SOL-ExecBench/issues/25 by zzw4257.
  Nodus follow-up: https://github.com/NVIDIA/SOL-ExecBench/issues/25#issuecomment-6734823293
  upstream/LICENSE and upstream/THIRD_PARTY_NOTICES.txt are included byte-for-byte.
  The modified io_after.py preserves the Apache-2.0 copyright/license header and
  explicitly identifies Nodus's input-generation changes. The checker is unchanged.
  SOURCE-MANIFEST.json records source digests and two original local build paths;
  these paths are historical provenance, not runtime dependencies.
  RUNTIME-INPUT.json and lock metadata preserve their preparation-time statements.
  They are not current run receipts; the public result/provenance describe the
  actual retained CUDA execution. No wheel binaries are included in these archives.