Generate and send a real RBAC-hard-fail fixture to the EvalPort spec (discussion #51) #85

Context

In discussion #51, we offered
@adhabnr-ux a real conformance fixture for their
EvalPort interchange spec: an actual Result
document emitted from a genuine aobench run that RBAC-hard-fails, rather than a
hand-authored example.

What to do

1. Find or construct a task/trace combination that RBAC-hard-fails for a legitimate
   reason — an agent actually attempting a forbidden tool call or crossing a role boundary.
   Do not use AIOPS_USR_001 or PERF_USR_001 — those hard-fail because of the
   "own-job" premise mismatch tracked in #63,
   which is a corpus defect, not a legitimate policy violation, and would misrepresent what
   a real hard fail looks like to a third party building against it.
2. Run it, capture the real BenchmarkResult / grader output, and verify by hand that
   hard_fail=True with a genuine hard_fail_reason and real dimension_scores.
3. Write a small generation script (in the style of scripts/audit_perf_res_002.py) so the
   fixture is regeneratable against a pinned AOBench version rather than a one-off paste.
4. Open it as a PR against evalport (external repo — needs Mohsen's own review/approval
   before opening, since it's outward-facing to a project we don't own).
5. Reply on discussion Idea: aobench-openeval-adapter for portable eval interchange (EvalPort) #51 linking the PR.

Acceptance criteria

• The fixture comes from an actual aobench run, not hand-authored
• The RBAC hard fail is genuine, not one of the two known own-job defects
• A generation script accompanies it, pinned to an AOBench version
• PR opened against evalport with Mohsen's review before submission
• Discussion Idea: aobench-openeval-adapter for portable eval interchange (EvalPort) #51 updated with the link