Run a model bake-off on title-register extraction

A bake-off is only fair if everything except the model and the effort level is held still: the same registers, the same schema, the same system prompt, no tools, and a scorer whose rules were written before any output was seen. The generator writes each register and its gold record from the same data, so the gold can't disagree with the text. The decision that matters most is the unit of comparison: expected cost per register, which adds the cost of a wrong answer to the model's own cost, because at any realistic failure cost the failures dominate.