benchmarks: record repeated-investigation pre-fix failure

[anusbutt]

Outcome

Records the completed repeated-investigation benchmark as a pre-fix baseline. The result is explicitly classified as a failure: GraphKeeper retrieved the stored grounded claims first, but the session still repeated substantial source and Git-history investigation.

The change is evidence-only under benchmarks/results/repeated-investigation-pre-fix-2026-08-14/. It does not modify GraphKeeper implementation, schemas, templates, or application behavior.

Evidence

Included in the result folder:

• identical control and GraphKeeper task prompts
• paired session outputs and investigation-step records
• exact GraphKeeper entity, claims, producing run, and captured evidence
• a short factual README stating the pre-fix failure

Focused verification performed before commit:

• all three captured JSON files parsed successfully
• the README was checked for the explicit FAIL result and no-product-change statement
• the commit contains only the eight benchmark evidence files

Full build and test suites were not run because this PR adds documentation and captured benchmark data only.

Checklist

• Documentation reflects the benchmark result accurately.
• Schema compatibility is unchanged.
• Append-only/evidence guarantees remain unchanged.
• npm run build passes — not run; no implementation changes.
• npm run typecheck passes — not run; no TypeScript changes.
• npm test passes with no hidden skips — not run; evidence-only change.
• npm run package:smoke passes — not run; package contents are unchanged.
• No secrets, credentials, telemetry, generated archives, or unrelated edits are included.

Compatibility and recovery

No runtime, platform, diagnostic, schema, hook, or package behavior changes. Rollback is removal or revert of the single evidence commit; no manual recovery is required.