feat(routing): add cognitive roles (debugging/orchestration/evaluation) and fix critique tradeoff

[Joi]

Summary

Three new cognitive roles for high-stakes work, plus a correction to the critique role's model-vs-thinking-budget tradeoff. Applied to anthropic.yaml, balanced.yaml, and quality.yaml.

Why these roles

The current matrix has reasoning (deep analysis), critique (finding flaws), and creative (generative). It lacks targeted roles for:

• debugging — hypothesis-driven investigation of failures (bug-hunter, session-analyst, incident analysis). Different from reasoning: more iterative, more evidence-driven, less open-ended.
• orchestration — root-session coordination across multiple agents. Needs strong planning + judgment but not the same "think hard" depth as reasoning.
• evaluation — comparing parallel agent outputs, judging quality across candidates. The "which of these is better" task that needs Opus-level discrimination.

All three default to Opus because the cost of a wrong call in these roles compounds across the session (a bad debugging hypothesis cascades; a bad orchestration decision wastes parallel work; a bad evaluation picks the worse output).

Why critique changes

Current: Sonnet + reasoning_effort: xhigh.
New: Opus + reasoning_effort: high.

The hypothesis was that thinking budget could compensate for model strength. In practice, xhigh produces longer Sonnet outputs, not higher-quality outputs. Critique is a discrimination task — it benefits more from Opus's stronger judgment than from extra thinking on a weaker model. reasoning_effort is orthogonal to capability.

Why writing changes

Added reasoning_effort: medium to the writing role for coherence across long outputs. Long-form content (documentation, marketing, case studies) benefits from medium thinking — not high (which slows it down without proportional gain), not none (which produces less-coherent long outputs).

The companion proposal to remove reasoning_effort from creative was a no-op — it already has none.

Files changed

• routing/anthropic.yaml — single-provider, Opus across all new roles
• routing/balanced.yaml — multi-provider, mirrors existing reasoning chain
• routing/quality.yaml — multi-provider, mirrors existing reasoning chain

updated: bumped to 2026-05-08 in all three.

Verification

• pytest tests/ — 5/5 pass
• yaml.safe_load() — all three files parse
• Role count: 13 → 16 in each file
• Programmatic verification: new roles present, critique is Opus+high, writing has reasoning_effort: medium

References

• Source issue: joi-90y (filed 2026-04-29)
• Aligns with strategic routing decision documented in dotfiles 2026-04-30: "root session uses Opus, critique role and reviewers Opus, Sonnet for generative leaf nodes"