refactor(fan-flames): replace per-task spec review with per-wave spec-informed peer review

[muloka]

Summary

• Replace Phase 4 (per-task spec reviewers) with cargo test gate + batched spec-informed peer review (~1 agent per 300 lines changed, change-reviewer type)
• Remove Phase 6 (VERIFY) — last wave's review inherently covers the combined result since prior waves are already squashed into @
• Collapse --skip-spec-review + --skip-review into single --skip-review flag
• Replace fan-flames-spec-reviewer.md with fan-flames-wave-reviewer.md (combined spec+quality prompt)

Why

A/B benchmarked fan-flames vs superpowers subagent-driven-development. Per-task spec reviewers cost ~45-50k tokens each (dominated by ~35k agent infrastructure tax) and found nothing that passing tests didn't already prove. Batched peer review at end-of-run cost 82k total (3 agents) vs 282k for 6 per-task spec reviews, and found real issues. Providing task specs to reviewers eliminates hallucinations (4 false findings in spec-blind review → 0 with specs as ground truth).

Test plan

• UAT in /tmp/fan-flames-uat — A/B benchmark complete
• Run fan-flames on a multi-wave plan and verify review phase dispatches batched reviewers
• Verify --skip-review skips the entire REVIEW phase

🤖 Generated with Claude Code