Reload the Selected Checkpoint and Emit Interpretability Artifacts

[RMKruse]

Parent

#35

What to build

Type: AFK.

Make the optimizer-trained selected checkpoint a trustworthy inference boundary. Reload it without rerunning optimization, reproduce saved predictions, and generate the complete interpretability surface from restored state: Local Shape Contributions, parameter-specific Graph Contributions, and Graph Gate Diagnostics.

Acceptance criteria

• The selected checkpoint can be loaded through a stable inference-facing interface without starting training.
• Checkpoint provenance records resolved configuration, seed, optimizer settings, selected epoch, selection criterion, environment details, and training history.
• Restored-checkpoint predictions reproduce the saved selected-candidate predictions within a declared numerical tolerance.
• Prediction, Local Shape Contribution, Graph Contribution, and Graph Gate artifacts are generated from restored checkpoint state.
• Artifacts distinguish mu and sigma and retain the non-causal Graph Gate Diagnostic wording.
• Tests cover checkpoint round-tripping, provenance completeness, prediction equivalence, and restored interpretability artifacts.

Blocked by

• Train an Optimizer-Trained GraphNAMLSS Synthetic Tracer Bullet #38