version 3.2.0

[3.2.0]

Stability, correctness, and packaging release. No new user-facing features; the
public API (Shield.fast/balanced/strict/secure/paranoid, protect_input,
protect_output, protect_stream, protect_tool_call) is unchanged.

Fixed

• Critical — Shield.secure() ran with zero ML models. Duplicate factory-method
  definitions shadowed the correct presets, so the "maximum security" preset silently
  loaded no ML ensemble at all. Removed the shadowing duplicates; secure() (and every
  preset) now loads the intended models and matches the documentation.
• Shield.paranoid() loaded no models — it referenced a non-existent xgboost.pkl.
  It now uses the trained ensemble.
• Shield.balanced() now matches the documented behavior (patterns + session tracking).
• Dropped gradient_boosting from the shipped ensemble. Its pickle is incompatible
  with current NumPy and failed to load. The ensemble is now three models — Random Forest,
  Logistic Regression, LinearSVC. Gradient boosting will return, retrained, in a future release.