feat(eval): add reproducible research campaign runbook

[vincentkoc]

What does this PR do?

Adds a reusable ShellBench research campaign skill plus native planning and
post-run analysis support for retained ten-task r0 qualification followed by
reproducible n=3 through n=6 benchmark matrices.

Why?

The native planner previously rejected repetition zero, treated every task
subset as a repair, limited full repetitions to 1-3, and produced identical run
labels across reasoning levels. That makes a safe family/harness qualification
stage impossible and leaves model identity, installed harness versions,
turn/tool usage, and cost provenance spread across raw artifacts.

This change makes those requirements explicit and machine-auditable before a
campaign is called clean.

Related trace publishing work remains in
#48. The runbook uses that verified
uploader when available and documents a private AWS CLI fallback without
embedding a bucket or credentials.

Changes

• Add the shellbench-research-runbook skill with:
  • retained, non-scoring ten-task r0 gates per harness/model-family route
  • private checkpoint/final trace retention and S3 publication
  • strict observed agent model and judge routing checks
  • requested-versus-installed harness version auditing
  • n=3 qualification followed by n=6 total independent repetitions
  • every provider-supported non-maximum reasoning level
  • gpt-5.6-sol at high as the default judge
  • exact, estimated, and unavailable cost provenance rules
• Add reasoning-specific run labels, repeatable harness/model filters, explicit
  judge model metadata, and configurable positive repetition counts.
• Add explicit r0 planning with exactly ten named tasks, one harness, one
  representative model, recorded family provenance, and automatic leaderboard
  exclusion.
• Allow native job repetitions from r0 through r6 and preserve r0 exclusion
  metadata through retries and remote manifests.
• Add research_audit.py exports for task traces, run model identity, turn
  usage, tool calls, installed harness versions, and cost provenance.
• Add focused planner and research-audit coverage plus a changelog entry.

Tests

• python -m pytest -q tests/test_native_eval_runner.py tests/test_native_eval_fleet.py tests/test_native_eval_research_audit.py
  • focused runner, fleet, and research audit suite: 80 passed locally
• python -m ruff check scripts/native_eval tests/test_native_eval_runner.py tests/test_native_eval_research_audit.py
• python -m ruff check clawbench app.py scripts tests && python -m pytest -q
  • Blacksmith Testbox tbx_01kypnawgaghyp5syryqspbqqs
  • 428 passed, 5 skipped
  • https://github.com/openclaw/shellbench/actions/runs/30442238460