refactor: adopt adaptive engineering rigor model

[Hubujiu]

Summary

Replaces the peer Event Router model with an adaptive-rigor control model:

• Core remains the minimum always-on policy.
• Decision becomes a pre-execution gate only when a material unresolved choice blocks or changes the next safe action.
• Direct becomes the default execution state rather than a route/module.
• Debugging and Implementation become alternative execution-rigor escalations, not mandatory sequential stages.
• Retrieval remains orthogonal and is benchmarked with minimum-sufficient / maximum-reasonable cost bounds instead of one exact label.
• Context isolation explicitly handles the fact that loaded references cannot be unloaded textually.

Benchmark changes

• Adds benchmarks/adaptive_rigor.py and canonical runner v2.1 while retaining the v2.0 execution core for historical interpretability.
• Replaces the v1.2 classifier contract with DECISION, EXECUTION, and Retrieval cost bounds.
• Adds four transition regressions: Decision→Direct, Decision→Implementation, Debugging→Direct, Debugging→Implementation.
• Adds matching Native Behavior cases for settled decisions and diagnosed failures.
• Keeps all committed v1.0–v1.2 result directories unchanged as historical evidence.

Evidence discipline

No v1.3 model score is claimed in this PR before a fresh run. The accepted v1.2 baseline remains commit 88382d2b0c00fa278067a5933bbcacc86f46b56e.