feat: restore evidence-triggered event router

[Hubujiu]

Summary

• replace the rejected E0-E3 / R0-R3 specialist tree with the compact Core/Direct plus Debugging, Decision, and Implementation event router
• keep retrieval orthogonal: targeted, bounded lexical, structural, and bounded exhaustive/authoritative evidence
• keep requirements interviewing explicit-only and remove unsupported specialist runtime leaves
• preserve the rejected experiment and its lessons under evolution/
• publish current-only v1.5 n=3 public and held-out evidence

Release result

Candidate runtime: 30ac7e70b425b3f02f7bf4e21cb4809f0e4d6c2c

• public matrix: 294/294 determinate
• Delivery: 54/54
• Debug: 40/42 (correctness 42/42; sibling safety 40/42)
• Decision: 29/30
• Native Behavior: 52/54
• public event reasoning: 113/114
• public Router after contract adjudication: 107/114 (reasoning 113/114, retrieval 108/114)
• held-out: 61/66 quality cells, 18/22 tasks stable-pass across 3 frozen repositories
• held-out traces: 66/66 valid; 0/66 spontaneous requirements-interview activation

The recorded raw Router score was 99/114. Three frozen Retrieval expectations contradicted the active STRUCTURAL relationship-mapping definition; the adjudicated count changes expectations only, not model output. Held-out active reasoning expectations were also decoupled from rejected capability_path metadata. Both corrections have contract tests and are documented in benchmarks/results/v1.5/.

No no-skill, Ponytail, combined-skill, or prior-version arm was rerun. Historical v1.2 is an offline, non-paired reference only.

Validation

• python -m unittest discover -s benchmarks -p 'test_*.py' — 55 passed
• pwsh -NoProfile -File benchmarks/run.ps1 -SelfTest — 39 passed
• pwsh -NoProfile -File benchmarks/run.ps1 -ProgressiveSelfTest — 8 passed
• public full current-only n=3 — 294 cells, 0 indeterminate
• held-out current-only n=3 — 66 cells, 0 indeterminate
• PR CI validates SKILL.md with skills-ref and reruns the harness tests on Ubuntu

Evidence boundary

The report preserves every residual failure and its adjudication. No runtime wording was changed after the final n=3 run, and no case-specific noun was added to improve a benchmark score.