fix(analyzer): report result coverage and reject misaligned inputs

[plox-sumit]

Fixes #195.

Analyzers skipped failed grader results without saying so. With three samples where two grading calls failed, DistributionAnalyzer reported a mean of 1.0 and AccuracyAnalyzer reported 100% accuracy, and nothing in the returned object showed that only one result was used. The validation analyzers also paired dataset and grader_results with zip, so a shorter results list was silently truncated instead of raising.

What changed:

• BaseAnalyzer gets two helpers. _validate_alignment raises ValueError when the two lists have different lengths. _coverage_metadata adds total_samples, valid_results, error_results, skipped_results and coverage_rate to the result metadata, extends the explanation when coverage is below 100%, and logs a warning.
• The seven validation analyzers (accuracy, precision, recall, F1, false negative, false positive, correlation) call the length check after their existing empty-input guard and report coverage.
• DistributionAnalyzer reports coverage. Its total_samples now means results submitted, not results used, which is what the name says.
• PairwiseAnalyzer checks alignment and sets total_comparisons to the number of comparisons it actually analyzed. The submitted count is still available as total_samples.
• F1ScoreAnalyzer never excluded failed results: a result without predictions counts as an empty prediction set and adds false negatives. That behavior is unchanged. Its valid_results therefore counts every result that entered the calculation, and error_results still tells you how many of those were errors.
• ConsistencyAnalyzer is untouched. It also pairs with zip, but test_analyze_mismatched_lengths asserts that truncation on purpose, so I left it for you to decide.

Failed results are still excluded from the metric itself, so existing numbers do not change when every call succeeds. The error_policy option suggested in the issue is left out to keep this small; it can be a follow-up if there is interest.

Tests: added to each analyzer's existing test file (plus a new tests/analyzer/test_pairwise_analyzer.py): the example from the issue, a full-coverage run with the explanation unchanged, a length-mismatch case for every validation analyzer and pairwise, the F1 error case, and the pairwise total. All 66 analyzer tests pass, and the full suite passes apart from three pre-existing Windows symlink tests in tests/harness/test_sandbox.py. black, isort, flake8 and pylint pass with the repo's pre-commit config.

🤖 Generated with Claude Code