docs(benchmarks): Add production and Kimi results

[gricha]

Publish four standardized security-review corpus results for the current
production model stack, Kimi K2.6, and Kimi K3 at low and high effort. The
production row uses Warden 0.46.2 on Pi with Grok 4.5 high for scanning and
GPT 5.6 Luna high for auxiliary calls; it records 37 of 86 known findings.

Add a source-linked production profile and require stable benchmark data to
contain exact matches for both model lanes. This keeps the matrix from
labeling a stale run or a primary-only match as current production. The
committed result files are sanitized summaries; raw JSONL remains withheld
pending sensitive-data review.

All six production shards completed without repair. The final auxiliary merge
dropped two exact scan-time corpus matches, so the published 37/86 is
intentionally the end-to-end system score rather than primary-model recall.
The Kimi rows preserve the completed and repaired shard provenance needed to
interpret their costs.

Validation passed with pnpm lint, pnpm build, pnpm test,
pnpm --filter warden-docs check, and git diff --check.

[sentry-junior]

Looks good to me. The production profile matches the pinned getsentry/.github warden.toml, the new validator correctly requires a stable row for both model lanes, and the published numbers line up with the committed result files (including the July vs current overlap and the Kimi turn/token figures). Nice call documenting the aux-merge drop so 37/86 reads as an end-to-end system score.