fix(core): make evaluation errors visible in results and reports

[dpaluy]

Closes #4

Problem

Evaluator#call rescues judge errors into nil scores, and downstream everything silently drops them: Result#overall compacts nils, Result#pass? can return true on partial data, Report#metric_stats/summary hide errored metrics, and Report#worst sorts nil-overall results as if they were the best. A judge outage that kills 5 of 6 metrics still shows green in CI.

Changes

• Result#errors returns per-metric error messages (derived from details entries with an :error key); Result#valid? is true when there are none
• Result#pass? now fails closed: returns false when any metric errored, regardless of the surviving metrics' scores
• Result#to_h includes errors
• Report#summary appends an error line when errors occurred, e.g. Errors: 3 metric errors across 2 samples
• JSON export (serializable_hash) includes per-metric error counts under errors
• Report#worst now ranks results with no valid scores as worst instead of best
• Version bumped to 0.3.0 with CHANGELOG entry (pass? and worst are behavior changes)

overall intentionally stays as the mean of valid scores, a partial score is still informative; only the gate requires completeness.

Verification

• bundle exec rake test: 136 runs, 301 assertions, 0 failures (9 new regression tests: all-nil scores, partial failure with high overall, error-free results, non-Hash details, report summary/worst/JSON with errors)
• bundle exec rubocop: no offenses
• End-to-end repro through Evaluator with a metric raising JudgeError: {failing: nil, passing: 1.0} now yields pass? == false, appears in Report#failures, and the summary shows the error count line

Implementation delegated to Codex, reviewed and verified with Claude Code.

CodeFactory builder