feat: run Practical Coding against SkillsBench

[Hubujiu]

Summary

• add benchmarks/run_external.ps1 as a one-command external benchmark entrypoint
• add a SkillsBench v1.1 adapter using BenchFlow's versioned dataset path
• compare the same codex-acp / gpt-5.6-luna / medium configuration as no-skill vs a custom Skill directory containing only Practical Coding
• dynamically select all SkillsBench v1.1 software-engineering tasks for the standard profile; smoke uses three tasks and full uses the complete versioned roster
• run the SkillsBench oracle before model calls by default
• run separate paired repetitions and alternate arm order to reduce order/cache bias
• require n>=3 plus complete healthy task/repetition pairs for stable external evidence
• report pass-rate lift, mean-reward lift, wins/losses/ties, normalized gain, per-task rates, and task-cluster-bootstrap 95% CIs
• preserve raw BenchFlow jobs, exact commands, dataset/source pins, adapter hash, and Practical Skill bundle hash
• add unit tests and CI coverage

Canonical stable run

pwsh -File benchmarks/run_external.ps1 `
  -Benchmark skillsbench `
  -Profile standard `
  -Runs 3 `
  -Workers 3 `
  -RequireStableRanking

Evidence boundary

This is a Practical-owned custom-Skill ablation on the immutable skillsbench@1.1 dataset. It intentionally does not use the benchmark's per-task curated Skills, so it should not be presented as an official SkillsBench leaderboard submission. It answers the narrower question: does mounting Practical Coding improve the same Codex/Luna agent on external software-engineering tasks?