Add replicated GPT-5.6 Sol, Terra, and Luna evaluations

[antonkayai]

Summary

• register GPT-5.6-Terra and GPT-5.6-Luna in the Codex evaluation configuration
• publish three complete, auditable full-corpus result packages each for GPT-5.6-Sol, Terra, and Luna
• report arithmetic means and sample standard deviations across matched runs, with variability available in leaderboard metric hints
• verify prediction hashes, input fingerprints, report scores, costs, and aggregate statistics from the retained artifacts
• update the leaderboard export, release documentation, and regression tests for replicated results

Verification

• PYTHONPATH=. .venv/bin/pytest -q
• .venv/bin/python benchmarks/check_replicate_summary.py
• .venv/bin/python benchmarks/export_leaderboard_space.py --overwrite
• git diff --check

[Shchoholiev]

Please add complete, auditable artifacts for all three GPT-5.6-Terra runs and all three GPT-5.6-Luna runs (six full-corpus executions total) before merging.

The PR currently retains full predictions, evaluation_report.json, run_metadata.json, and per_sample_status.tsv only for each model's selected run 1. The replicate directories retain aggregate summary.json files, while runs 2 and 3 remain ignored scratch artifacts. That means reviewers cannot independently reproduce those four runs or calculate three-run averages for the Structural, Scale-control, tier, family, stressor, and per-document metrics shown by the leaderboard.

Requested changes:

1. Preserve each of the three Terra runs and three Luna runs as separate complete result packages, including all 32 per-document predictions, evaluation_report.json with detailed_results and scorer breakdowns, run_metadata.json, and per_sample_status.tsv (plus the Markdown report if that is part of the release convention).
2. Record enough provenance to verify that all six runs used the same 32 transcripts, contracts, prompts, model/effort settings, isolation policy, and scorer. Retain input/contract fingerprints, execution status and attempt counts, timestamps, CLI versions, and prediction hashes.
3. Link each replicate-summary entry to its exact saved result package.
4. Add verification that recomputes every report from the retained predictions and checks that the three-run summary aggregates match the underlying reports.

This evidence is required before presenting Terra or Luna as three-run leaderboard results. A summary-only aggregate is not enough to audit the full leaderboard row.

[antonkayai]

Added the requested audit trail in 503b15cc:

• all six Terra/Luna executions now have complete 32-document prediction packages, reports, metadata, and status tables
• summaries link each run to its result directory and report SHA
• check_replicate_summary.py verifies fingerprints, prediction hashes, protocol metadata, costs, aggregates, and replays every report against ground truth
• the leaderboard now reports three-run mean ± sample SD, not a selected run

Local verification and both GitHub checks pass.

[Shchoholiev]

Updated leaderboard

[Shchoholiev]

Verified the retained Sol, Terra, and Luna run packages, provenance, replayed reports, aggregate checks, leaderboard export, and current test suite. The requested auditability changes are addressed.