No such guarantee should be inferred from deleting source files. NIST defines machine unlearning as removing training-data influence from a trained model; that describes a technical task distinct from deleting a stored copy. NIST glossary.
The FTC's finalised Everalbum settlement included deletion of models and algorithms developed using specified user material. It is an example of a remedy tied to a particular enforcement matter, not a general contractual rule for every licence. FTC.
Ask separately about source files, caches, embeddings, evaluation sets, training checkpoints and deployed models. Require the supplier to identify what can be removed, what testing demonstrates removal, retained legal records and any continuing permissions. Negotiate the residual risk before training starts rather than relying on a generic termination clause.