import-time RNG consumption in MNIST_CNN's default Cnn() argument invalidated the numerical reference

[rz4]

Symptom

tests/test_valiadation_tests.py::test_mnist_first_drift_losses_match_reference fails deterministically:

step	actual	reference
2	6.850990295410156	7.514564037322998
3	5.589588880538940	5.548082828521729

The failure is reproducible across test runs and only appears on machines with the
MNIST raw data downloaded. The test self-skips otherwise (@pytest.mark.slow + data
check), which is only caught if CI is run with MNIST downloaded.

Importance

Reproducibility. For any harness built on this framework: RNG state consumed at import time silently shifts every
subsequent seeded draw (i.e. weight initialization, data shuffling and sampling, augmentation/drift
simulation parameters, dropout, and stochastic detector behavior), so identically seeded
experiments can produce different metrics, drift timing, and losses depending on nothing more than
import order or whether a module was already cached.

Most likely happened when lazy imports of examples was introduced.

Root cause (found by Claude Fable 5.1; verbatim)

git bisect between 992691c (PR #73, where
tests/references/mnist_first_drift_ewc_losses.csv was recorded) and current main
identifies b95da75 (PR #82, "Make import statements for model classes conditional
based on the dataset name") as the first bad commit. The loss values it produces are
bit-for-bit identical to today's — no later commit changed the numerics further.

PR #82 didn't change any math. The divergence comes from an RNG-ordering interaction:

• examples/mnist/model.py defines
  MNIST_CNN.__init__(self, cfg, model: nn.Module = Cnn()). The default Cnn() is
  constructed at module import time, and its weight initialization consumes draws from
  torch's global RNG. (Verified: seeding torch to 1337 and importing
  examples.mnist.model changes the next torch.rand(1) from 0.0783… to 0.7181…; no
  other module in the import chain consumes RNG.)
• Before Make import statements for model classes conditional based on the dataset name #82, examples/utils.py imported all example model modules eagerly, so this
  happened before the test called torch.manual_seed(seed) — the seeded stream was
  untouched, giving the reference values.
• After Make import statements for model classes conditional based on the dataset name #82, the import happens lazily inside get_example(), after seeding. The
  Cnn() construction now burns RNG draws post-seed, shifting every subsequent draw
  (affine drift parameters, dataloader shuffling, dropout) and therefore all logged CL
  losses. The random weights themselves are irrelevant — they're overwritten by
  mnist.pth; only the RNG side effect matters.

Suggested Work

• The MNIST test is an end-to-end numerical regression test. Fix the example's RNG ordering, then regenerate the
  reference CSV.
• Make sure other templates do not have import-time RNG side effects.
• Assert that importing each example model module leaves torch.get_rng_state() (and NumPy/Python RNG state) unchanged, so no example performs RNG-consuming work at import time.