Support multiple datasets in eval command

[alexkroman]

Enable the assembly eval command to score multiple datasets in a single run, with each dataset loaded, scored, and reported separately.

Summary

The eval command now accepts multiple dataset arguments instead of a single dataset. Each dataset is evaluated independently and emits its own JSON payload (when --json is used), while failure counts are aggregated across all datasets in the exit message.

Key Changes

• Command signature: Changed dataset: str argument to datasets: list[str] to accept variadic dataset inputs
• Evaluation loop: Refactored run_evaluate() to iterate over multiple datasets, extracting the single-dataset logic into _evaluate_one()
• Failure aggregation: Accumulated failed and total counts across all datasets for the final error message
• Output format: Each dataset emits its own JSON object (under --json), with human-readable output showing a block per dataset
• Help text: Updated docstrings and CLI help to reflect multi-dataset support with examples
• Test coverage: Added four new tests covering multiple datasets with JSON output, human output, failure aggregation, and the requirement for at least one dataset

Implementation Details

• The _evaluate_one() function encapsulates the scoring logic for a single dataset and returns its payload
• run_evaluate() now orchestrates the loop, emitting each payload and accumulating failure statistics
• The error message format remains consistent: "{failed} of {total} items failed to transcribe." where totals span all datasets
• All existing flags (--limit, --split, --subset, --column) apply uniformly to every dataset in the run

https://claude.ai/code/session_01JmbkV2dVWv8FZxbgGVrPXu