Phase 5: API code quality - ModelInspector, trainer mixins, registry modernization

[w4nderlust]

Summary

Phase 5 of the Ludwig modernization: break up god objects and improve code quality.

1. ModelInspector

Extracts model introspection from the 2400-line LudwigModel into a focused class:

from ludwig.model_inspector import ModelInspector

inspector = ModelInspector(model, config, metadata)
weights = inspector.collect_weights(['linear1.weight'])
summary = inspector.model_summary()
importance = inspector.feature_importance_proxy()

Provides: weight collection, model summary (param counts, layer types, model size), and feature importance estimation.

2. Trainer Mixins

Composable mixins for cross-cutting training concerns:

• CheckpointMixin: checkpoint save/restore decision logic
• EarlyStoppingMixin: early stopping based on validation metrics
• MetricsMixin: metric formatting and logging
• BatchSizeTuningMixin: automatic batch size search
• ProfilingMixin: wall-clock timing for training operations

3. Registry Modernization

Added to the existing Registry class:

• unregister(name): remove registered items (useful for testing)
• get_default(): get the default-registered item
• list_registered(): list all names excluding default key aliases
• Improved docstrings and type annotations

Test plan

• 14 new tests for ModelInspector and Registry
• 1155 existing tests pass (0 regressions)
• Pre-commit all clean
• CI