Add BLADE analysis skill

[jmabry]

Why

Issue #1585 asks for a public BLADE analysis skill so agents can turn NeMo Gym rollout artifacts into evidence-backed benchmark reports and improvement recommendations.

This PR adds that skill without making benchmark-specific examples part of the default agent context.

What Changed

• Add nemo-gym-blade-analysis for both Claude and Codex agents.
• Document the BLADE-ready package shape:
  • D1: analysis skill
  • D2: rollout data
  • D3: golden report package with metrics sidecars and anchor facts for the universal blade-judge
• Add a generic BLADE benchmark build guide covering artifact inventory, pass@k analysis, task bucketing, workflow funnels, root-cause taxonomy, evidence rules, anchor facts, shallow baselines, and calibration expectations.
• Add a public blade_toolkit.py helper for package-shape validation, draft anchor-fact extraction, shallow baseline generation, scoring, and local calibration checks when external BLADE tooling is not available.
• Add an opt-in original CVDP reference. It points to the existing public CVDP example rollout files already in resources_servers/cvdp/data/ and includes only the Nemotron 3 Super golden report, metrics, and anchor facts as optional reference material.
• Keep CVDP artifacts clearly labeled and non-default: agents should load them only when the user asks for a concrete CVDP example or needs help understanding a completed report shape.
• Add unit coverage for toolkit behavior, including shallow-baseline leakage checks and Claude/Codex toolkit copy synchronization.

CI And Tooling

The repo lint job previously depended on cloning hook repositories during pre-commit setup. This PR replaces those hook-repo clones with uv-managed local hooks, keeps Ruff pinned to the existing hook version, and updates the lint workflow to run pre-commit through uv run --no-project.

Validation

• git diff --check
• uv run --no-project --with pre-commit==3.6.0 pre-commit run --all-files --show-diff-on-failure --color=always
• uv run --extra dev pytest tests/unit_tests/test_blade_toolkit.py -q
• uv run --extra dev ng_dev_test — 239 passed, coverage 96.15%
• Parsed bundled JSON artifacts with uv run python -m json.tool.
• Ran blade_toolkit.py make-shallow and calibrate on the bundled CVDP golden artifacts; the golden report scored high against itself and the shallow baseline stayed below the target threshold.
• GitHub Actions are passing for DCO, lint, copyright, secrets, unit tests, and all server shards.

Closes #1585.
Supersedes #1584.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.