feat: expand benchmark matrix and external evidence model

[Hubujiu]

Summary

• expand Router from 16 to 28 public regression cases, with new cases covering all six routes
• expand Decision from 4 to 10 full-profile two-turn scenarios (6 in standard)
• expand Debug from 4 to 12 full-profile cases (8 in standard)
• add eight new deterministic shared-boundary Debug fixtures covering header normalization, tenant cache isolation, pagination, duration units, blank-row parsing, inventory invariants, zero TTL semantics, and nullable sorting
• add oracle fixes for every new Debug case and CI tests proving each seed fails while its oracle passes
• keep the core runner separate from the evolving case catalog via run_catalog.py + case_catalog.py
• update standard/full matrix counts and reproduction docs
• document the external benchmark landscape: SkillsBench, LangChain skill benchmarks, SWE-bench, Terminal-Bench, FeatureBench, Senior SWE-Bench, ProgramBench
• explicitly separate public regression evidence from external benchmarks and a future private held-out set

Current matrix

• standard: 9 Delivery / 28 Router / 6 Decision / 8 Debug, n=3 by default (222 cells without previous/no-Skill arms)
• full: 18 Delivery / 28 Router / 10 Decision / 12 Debug, n=3 by default (324 cells without previous/no-Skill arms)

Benchmark-design rationale

The new cases are deliberately different failure mechanisms and routing boundaries rather than paraphrases of the cases that already influenced SKILL.md. Public cases remain regression tests, not proof of unseen-task generalization. The next external-validation target is SkillsBench because it directly measures with-Skills vs without-Skills under paired deterministic evaluation.