fix(eval): stabilize native harness results and traces

[vincentkoc]

Summary

Consolidates the still-unmerged child work from #43, #44, and #45 onto the landed native runner from #42. The original stacked branches cannot be force-updated under repository protection after #42 was squash-merged.

Changes

• preserve explicit GPT reasoning effort and add GPT-5.6 Luna/Terra catalog IDs
• preserve real OpenClaw, Hermes, and Codex trajectories
• isolate agent and judge reasoning configuration
• prevent OpenClaw onboarding bootstrap contamination
• recover active/degraded Crabbox leases and retain zero-result failure checkpoints
• support targeted canonical repair overlays and canonical task identities
• scope parity claims to the exact validated route; legacy Codex GPT-5.5 compatibility is limited to the fixed 20-task calibration
• emit original-versus-repaired score sensitivity reports

Validation

• pytest -q tests/test_native_eval*.py: 99 passed
• focused parity/repair suite: 91 passed
• Ruff passes on native runner and test surfaces
• git diff --check passes
• all 28 replayed commits carry valid signatures
• prior 93-run campaign: 10,788/10,788 task results, zero infra rows
• the 117-run low/medium/high matrix extension is running separately with verified periodic checkpoints

Stack cleanup

This PR supersedes #43, #44, and #45 without discarding their commit history. Result publications remain model-scoped in ShellBench/public-tasks.