Changes

• Replace eval() in feature expressions and disk expression caching with a restricted AST interpreter. Preserve registered operators and safe parameter syntax, including bounded containers, argument unpacking, indexing/slicing, and scalar conditions.
• Preserve quoted strings and necessary token boundaries during expression preprocessing and cache-key normalization.
• Replace TRA model and performance-graph name evaluation with explicit mappings.
• Require explicit, strictly boolean trusted=True for file-module imports, defaulting to refusal. Keep legacy module aliases for restoring trusted older file-based models.
• Update affected TRA/LightGBM examples, add regression coverage, and connect migration guidance to documentation entry points and actionable errors.

Compatibility and migration

Package imports and standard Alpha158/Alpha360 expressions remain supported without opt-in. File-based components must explicitly authorize their imports:

task:
  model:
    class: MyModel
    module_path: custom_modules/model.py
    trusted: true
    kwargs: {}

Direct imports use get_module_by_module_path(path, trusted=True). Each nested file component needs its own consent; there is no global grant. Top-level trusted controls module loading, while kwargs.trusted remains a constructor argument.

Configurations and imported code must still be trusted. This is not a sandbox or directory-containment policy, and module-import consent does not enable unrestricted expressions or authorize artifact deserialization. Artifact-loading changes are handled separately in #2339.

See the configuration execution migration guide for supported syntax, custom extensions, and old-model recovery.