Derived training arms should extend contracts without weakening frozen arms

General lesson: when adding a derived synthetic training arm to a hash-pinned experiment, preserve the existing builder, validator, release-key set, and literal recipe digests. Put the new synthetic profile in a separate deterministic builder/validator, dispatch by an explicit arm, and give only that arm a distinct recipe identity plus exact source-lineage and review bindings. Validate exact per-tool counts and arguments, not only aggregate class counts; emit row/token digests and measured target histograms in the receipt. This applies when old arms must remain reproducible. It does not prove data quality or model improvement.