Fix weekly eval gate replay comparison

[uncfreak1255-code]

Summary

• Add gbrain eval gate --drill for privacy-safe regression drivers: query hashes, counts, Jaccard, top-1 status, and latency only.
• Make regression gate replay compare current results at each row's captured result count by default.
• Document --drill and --compare-limit N|captured in the eval bench guide.

Proof

• bun test test/eval-replay.test.ts
• bun test test/eval-gate.test.ts
• bun run typecheck
• Live source-checkout receipt: ~/.gbrain/eval-receipts/gbrain-weekly-quality-streak/20260619-gate-drill-clean-pr.json
• Live result: verdict pass, rows 31/31, mean_jaccard 0.868663594470046, top1 0.8387096774193549, no breaches.

Notes

• bun run ci:local:diff was not run because gitleaks is not installed on PATH; Docker is available.
• The prior installed weekly loop still has streak 0; rerun after merge/install before updating the streak.
• The doctor JSONB progress work is intentionally not included in this PR.