v1.2.3

Fixed

• The bundled TF-IDF model did not work on Python 3.10. It had been saved by
  scikit-learn 1.8.0, which only exists for Python >= 3.11. On 3.10 pip
  resolves scikit-learn 1.7.x: the pickle still loaded, predict_proba raised
  'LogisticRegression' object has no attribute 'multi_class' (1.8 removed
  that parameter), the canary check refused the model, and the ML layer ran
  silently disabled behind a single log warning. The same weights are now
  re-serialized with scikit-learn 1.7.2 and verified to score identically on
  1.5.2, 1.6.1, 1.7.2 and 1.8.0 (the file is 2.8 MB instead of 4.3 MB because
  the vocabulary indexes are plain ints). integrity.json carries the new
  hash, the bundle's metadata records sklearn_version and python_version,
  and tests/unit/test_bundled_model.py turns a model that fails to load into
  a red CI job on every Python in the matrix. The ml-light extra now requires
  scikit-learn>=1.5: 1.4 cannot read the bundle (it could not read the old
  one either, it just failed quietly).
• scikit-learn's version-mismatch warning on model load now goes to the
  zugashield logger at INFO instead of stderr. The bundled model is saved
  with the oldest supported scikit-learn on purpose, so the warning fires on
  most installs; the hash and canary checks are the real verification.

Security

• Every Hugging Face Hub download in the ML scripts is pinned to a commit hash
  (zugashield/ml/hub_pins.py): the 9 training datasets in train_tfidf.py,
  the 10 dataset loads in benchmark_configs.py, the corpus in distill.py,
  and the ONNX model and tokenizer fetched by zugashield-ml download. An
  unpinned download trains or ships whatever the repo owner has pushed since.
  Bandit flagged 12 of these as B615 (CWE-494) and the weekly Security Scan was
  red; train_tfidf.py's 9 were invisible to Bandit because load_dataset is
  passed in as an argument there. python -m zugashield.ml.hub_pins compares
  each pin with the Hub's current main.
• The Security Scan workflow now runs on pushes to master. It was wired to a
  main branch that does not exist, so it only ever ran on the Monday schedule.