Add GPT-5.5 n=3 results and refresh leaderboard

[Shchoholiev]

Summary

• add three independent GPT-5.5 full-corpus runs with all 96 saved predictions, reports, metadata, and per-sample statuses
• add an aggregate summary with arithmetic means, sample standard deviations, usage, costs, result paths, and report hashes
• replace the GPT-5.5 leaderboard single-run row with the three-run aggregate
• extend the generic replicate checker, release consistency checks, exporter tests, and documentation

GPT-5.5 n=3 result

Metric	Mean ± sample SD
Exact-record recall	96.6% ± 1.5 pp
Complete documents	6.3 ± 0.6 / 32
Field micro-F1	99.3% ± 0.3 pp
Structural exact recall	89.5% ± 5.1 pp
Scale-control exact recall	99.4% ± 0.1 pp
API-equivalent full-run cost	$43.19 ± $3.34

The three runs used the same 32 released OCR transcripts, field contracts, prompt, GPT-5.5 model, xhigh effort, Codex CLI 0.146.0, four workers per run, 272K context cap, and repository-denied ephemeral isolation as the GPT-5.6 replicate rows. All 96 document jobs completed on their first attempt. Run 1 was selected as the reference before evaluation; the aggregate does not select the best run.

Leaderboard change

The GPT-5.5 row now reports n=3 · arithmetic mean and uses:

• $43.19 full-run cost
• 96.6% exact recall
• 6/32 complete documents in the table; the info popover retains the exact 6.3/32 ± 0.6 mean and sample SD
• 89.5% structural recall
• 99.4% scale-control recall
• 99.3% field F1

The paper's earlier single-run GPT-5.5 result remains unchanged; this PR updates the current leaderboard and repository summary table only.

Verification

• python -m pytest -q — 203 passed
• python benchmarks/check_replicate_summary.py — verified 4 summaries and their full run packages
• fresh leaderboard export plus JSON assertion for the GPT-5.5 n=3 row
• git diff --cached --check
• staged artifact audit: 96 predictions, 3 reports, 3 metadata files, 3 status tables, 0 transient logs