feat: add model tiering for agents

[mariuszs]

Summary

Assigns AI models to agents based on cognitive demand, instead of uniformly inheriting the parent model (Opus).

Conservative approach — Sonnet only where risk is minimal:

Tier	Model	Agents
Thinking	Opus (inherit)	All 20 reasoning agents (planning, spec creation, architecture, code review, analysis)
Execution	Sonnet	task-group-implementer (follows detailed plan), test-suite-runner (runs commands, reports results)
Classification	Haiku	task-classifier (new), project-analyzer (already haiku)

Why

• Cost: Opus tokens are ~5x more expensive than Sonnet. The implementer is the biggest token consumer — it writes all the code. Moving it to Sonnet gives the most savings.
• Speed: Sonnet is faster — implementation and test phases become snappier.
• No quality loss: Both Sonnet agents receive explicit instructions (detailed plan with file paths and steps, or test commands to run). They execute, not reason.
• Conservative scope: Initially considered 11 agents on Sonnet, narrowed to 2 after reviewing risk without a model escalation mechanism. Only agents with pure execution profiles made the cut.

Future considerations

• Model escalation: If an agent returns BLOCKED, automatically retry with a more capable model (pattern from superpowers project)
• Expand Sonnet tier: With escalation in place, more agents can safely move to Sonnet
• Haiku for implementation: With sufficiently detailed plans, even Haiku could handle simple tasks