vitals 0.3.0

New features

• generate_structured() extracts structured data from model responses
  via ellmer::parallel_chat_structured(), analogous to how generate()
  wraps parallel_chat() (#153).

• model_graded_qa() now encourages brevity in its default instructions
  (#197). This reduces the tendency of model-graded scorers to "talk themselves
  out of" a reasonable score.

Log viewer

• Updated the vendored Inspect Log Viewer to version 0.3.161 (#194).

• Task IDs now follow Inspect's task_identifier format
  (task_name/model/hash), including the model name and a hash of
  solver/scorer arguments. This ensures evals with different models or
  arguments appear as separate log viewer entries rather than being
  collapsed as "retries."

• The home page now includes all of the metadata associated with the eval.

• Model events in the log no longer hardcode max_tokens = 4096. The
  logged value now reflects the provider's actual setting, and the field
  is omitted when unset (#213).

Bug fixes

• Accuracy calculation for ordered factor scores with more than two
  levels (e.g. I < P < C) no longer inflates partial-credit scores
  when the highest grade is absent from results.