XY-1156: Implement qmd candidate-replay comparability gate

[yvette-carlisle]

Summary

Implements the XY-1156 qmd candidate-replay comparability gate for the Docker-owned quantitative benchmark path.

Changes

• Adds scripts/materialize-qmd-candidate-replay-gate.py with schema elf.qmd_candidate_replay_comparability_gate/v1.
• Wires scripts/real-world-quantitative-docker.sh to emit qmd-candidate-replay-comparability-gate.json after the freshness manifest.
• Adds Rust integration tests for pass and blocker behavior, including the P1 false-positive cases found by skeptic review.
• Documents the gate contract and claim boundary in the quantitative benchmark spec and benchmarking evidence docs.

Gate Coverage

The gate blocks unless qmd has:

• typed product result state and measured pass state;
• every qmd product/per-query row mapped to the manifest corpus_id;
• held-out and leakage-audited status plus an audit manifest id;
• passing per-query replay rows with positive candidate and expected-relevance counts plus explicit qrels;
• aggregate replay fields that match the per-query replay rows;
• a matching freshness row with both valid runner image digest and product commit;
• no unqualified leaderboard claim.

Decodex / Manual Intervention

Decodex attempt xy-1156-attempt-1-1783069646 failed with no_effective_diff after the app-server timeout path. A second Decodex run was skipped because of Linear connector backoff. Manual intervention was used under the goal authorization, then this PR keeps the Decodex issue branch and authority XY-1156.

Validation

• python3 -m py_compile scripts/materialize-qmd-candidate-replay-gate.py
• cargo test -p elf-eval --test real_world_job_benchmark qmd_candidate_replay -- --nocapture - 6 passed
• python3 scripts/check-docs.py
• git diff --check
• cargo make checks - check, clippy, vstyle, and nextest: 437 passed, 92 skipped

Review

• Initial read-only skeptic: support_with_changes, 4 P1 false-positive findings.
• Follow-up read-only skeptic: pass, no remaining P0/P1 blockers.

Claim Boundary

This supports only a qualified qmd same-corpus candidate-replay comparability claim when the gate passes. It does not permit a broad product leaderboard or ELF-vs-qmd win claim.