Simulate decoding strategies

The sampling functions never see the model: each takes one step's logits and a random generator and returns a token id, so they can be checked against the lesson's worked example with no model at all. The model sits behind a two-method interface (start a prompt, advance by one token), implemented once by a real Hugging Face model and once by a tiny fake for dry runs. The decision that matters most is fixing the list of factual errors before the full run and saying it is a proxy: a regex can tell you a sample calls the TA6 a rental form, but it cannot tell you that garbled text is wrong, so the error column has to be read next to the garbled column.