Add benchmark workflow

[Joshdb-Amazon]

What

Adds an opt-in CI gate that scores this repo's own skills/agents by running them
through kiro-cli in Docker and grading the output against per-scenario rubrics, plus
a new neuron-dev-contribution-validation skill/agent (from the internal repo) that checks a contribution against the NAD contribution model. Upstream currently has no .github/, benchmark/, or tests/.

Why

Gives maintainers a reproducible, on-demand way to verify that a contributed
skill/agent actually behaves as intended before merging, without running untrusted code
automatically or exposing secrets on every push.

How it works

Grader (benchmark/)

• run_skill_tests.py loads scenarios from tests/registry.json (the source of truth
  for which scenarios are active/gated), runs each in its own Docker container, and
  scores combined = 0.75·pattern + 0.25·judge. pattern is proportional
  expectedStrings/forbiddenStrings matching; judge is an LLM verdict against the
  scenario rubric. A scenario passes at combined ≥ 0.90; the gate fails below --gate
  (default 90%).
• Reliability: no infra-error retry by design (a hung agent is skill behavior worth
  measuring, not masking), plus salvage-on-timeout (score a completed-but-hung run's
  output rather than discarding it).
• shell/{docker,common}.sh + environment/Dockerfile build the run image
  (python:3.11-slim + kiro-cli).

Workflow (.github/workflows/benchmark-external.yml)

• A maintainer reviews a PR and applies the run-benchmark label to opt it in.
  Applying a label requires the repo's Triage role or above, so fork contributors (no
  role on this repo) can't trigger it themselves. A later push does not re-run it — the
  label must be re-applied (re-review).
• Uses pull_request_target so the run can read KIRO_API_KEY even on fork PRs (a
  plain pull_request run gets no secrets on forks), with permissions: contents: read to drop the write-capable GITHUB_TOKEN.
• Also supports manual workflow_dispatch from the Actions tab (write access
  required), which checks out the selected branch directly.

New skill/agent (neuron-dev-contribution-validation)

• Validates a contribution's namespace, directory structure, NEURON_METADATA.md
  fields, agent-spec formats, and minimum test counts, reporting PASS/FAIL per check.
  Ships with a Claude Code .md agent + Kiro .agent-spec.json, and one active gated
  test.

Repo prerequisites

• A KIRO_API_KEY Actions secret — authenticates both the agent and the LLM judge.
• A run-benchmark label.

Security note for reviewers / labelers

The labeled job checks out and runs the PR's untrusted contributed code in an
environment that holds KIRO_API_KEY. contents: read restricts only the
GITHUB_TOKEN — it does not shield KIRO_API_KEY from that code. Applying the
label is vouching for the diff, so only label PRs whose changes you've reviewed.

Testing

python benchmark/run_skill_tests.py --nad-root=. --subset=all --judge --gate=90

• Confirmed the gate runs only on the labeled event, not on push/opened.

Notes

• The LLM judge is non-deterministic, so the pass rate can vary run-to-run near the
  threshold independent of any code change (documented in benchmark/README.md).