warnings

Anything worth reporting from evaluating the study, as ModaWarning rather than as text.

Kept typed because a caller holding only a snapshot is the ordinary case — a user interface is expected to read results from here rather than from the model that produced them — and each warning names the metric it concerns and what happened to it. Flattening to messages would leave such a caller able to display a warning but not to act on one: it could not group by metric, tell a tied metric from a domain that was not applied, or offer the remedy that fits. The text is still one step away, through the message of each warning.