The vocabulary-size memory example assumes the tokenizer and embedding table agree on the token-to-row mapping. That compatibility matters beyond matching the number of rows: a tokenizer with reordered IDs can fit the same matrix shape while retrieving the wrong learned vectors.

I would package the tokenizer revision with the checkpoint and add a tiny known-token-to-ID check when loading custom assets. Extending a vocabulary also needs an explicit embedding resize and initialization step. This turns the table lookup explanation into a useful deployment invariant without requiring readers to understand the later Transformer layers first.