Add reproducible scorer reports for RCA benchmarks

[RitwijParmar]

Summary

This updates the 04-04-2026 RCA scorer so benchmark runs can produce an auditable run artifact instead of only per-output score files.

Changes:

• add --model and RCA_SCORER_MODEL support so the judge model can be changed without editing code
• add --dry-run to validate output discovery and answer-key coverage without making model calls
• add --report-json with per-eval records, statuses, scores, errors, and aggregate summaries
• add --strict so CI/manual runs can fail when an output cannot be scored
• print aggregate summaries after each run

Why

Right now the scorer writes per-eval score files, while the aggregate benchmark tables are maintained separately. That makes it harder to reproduce a scoring run, inspect missing/empty/error cases, or attach a single machine-readable artifact to a benchmark result.

Verification

I intentionally used dry-run mode so no model calls were made:

• /Users/ritwij/.bun/bin/bunx tsx scorer.ts --help
• /Users/ritwij/.bun/bin/bunx tsx scorer.ts foam --dry-run --report-json=/tmp/foam-dry-run-report.json

Dry-run result:

• 22 Foam outputs discovered
• 22 answer keys found
• 0 missing answers
• 0 empty outputs
• 0 errors
• JSON report written successfully

Notes

No generated reports, scores, lockfiles, or install artifacts are included in this PR.