[functionstackx]

Changes

Add cmax gpu-fingerprint, an experimental, opt-in collector for numerical
GPU instance signatures. The P3 atomic-arrival-order kernel is adapted from
divagr18/gpu-instance-fingerprint,
with the upstream MIT notice retained in the distributed module.

• Require --acknowledge-gpu-work and the existing host-role guard before work.
  Run on one selected CUDA-visible device inside the operator's allocation.
• Keep collection outside all audit profiles and security grading. No remote
  dispatch, automatic dependency installation, network upload, or machine
  configuration changes.
• Bound work and run CUDA in a subprocess with a timeout. Save a new mode-0600
  JSON file; reject existing paths and remove incomplete results.
• Record protocol/kernel provenance, device model and software metadata, launch
  settings, replicate index and per-epoch rank vectors. Do not collect UUIDs,
  serials, hostnames or SM IDs.
• Add CPU-only tests, a run guide and a hardware-validation checklist. NumPy is
  a test/development dependency; CuPy remains optional and is imported only by
  the explicitly started collector.

Scope and limits

This is a collection prototype, not an identity verifier. It does not include
the P1/P2 channels, passive torch.index_add_ experiment, enrollment database,
classifier, match threshold or fleet score.

The kernel and workload sequence differ from upstream, so this implementation
has a separate protocol ID. Upstream's reported accuracy does not qualify this
code. The upstream restart data describes container restarts, not physical
host reboot validation. A finite forgery sweep is not proof against software
replay or hostile-provider output substitution.

The draft must remain experimental until hardware testing measures repeatability,
same-model separation and false accepts/rejects under held-out sessions, load,
driver changes and separately tested restart/reboot conditions.

Validation

• Full local suite: 2,390 tests and 1,306 subtests passed on Python 3.14.8.
• 38 new fingerprint tests passed, including mocked device/rank collection,
  host-role refusal, explicit acknowledgment, worker timeout/interruption,
  output permissions, overwrite/symlink rejection and missing CuPy.
• CLI and host-role regression tests passed with the new tests.
• Wheel build passed and includes cmax/gpu_fingerprint.py with its MIT notice.
• Public-data check and git diff --check passed.
• GitHub Actions were still running when the PR was opened; local results are
  separate from CI status.
• No CUDA device is available in this environment. The kernel has not been
  compiled or executed on GPU hardware; no accuracy or stability claim is made.

Example

# Inside an authorized GPU allocation; install the CuPy build for your CUDA version.
python -m pip install cupy-cuda12x
cmax gpu-fingerprint --acknowledge-gpu-work \
  --device 0 --replicate 0 --output gpu-a-r0.json

Fingerprint files can identify hardware even without UUIDs. Keep them private;
do not attach raw signatures to this public PR.