Durable QML training checkpoints

Purpose and scope

Defines a separate checkpoint artifact for stopping training and continuing from the same retained state. Keeping only the best parameters does not preserve optimizer history or dataset identity.

Core design rules

• nm-training-checkpoint@0.1 carries source and language version, objective/dataset identity, GD state, parameters and histories.
• The reader rejects unknown fields, inconsistent histories, mismatched parameter names and integrity-digest mismatches.
• Resume-boundary evaluation, cancellation and persistence behavior are recorded explicitly.

Example from the original

This example illustrates the design recorded in the original. It is not by itself a claim of executable or stable support; check required options and the current version.

The original does not provide a short source example for this topic. Read its detailed data contracts, evaluation rules and review gates in the original-document view.

Limits and interpretation

• The checkpoint format does not silently accept arbitrary optimizer state or future versions.
• Continuation without the retained reproducibility mode and frozen dataset split is not the same experiment.

Status and implementation boundary

The source describes a runtime-experimental artifact and continuation API; it adds no source grammar. Verify separately that durable storage is actually configured.