feat: explore-then-synthesize design phase via Explore subagents

[sriumcp]

TL;DR

DESIGN today asks one Opus session to do BOTH codebase exploration AND bundle synthesis. Split it: spawn 3–4 parallel Explore subagents (cheap, read-only) to map the system, then a small Opus call synthesizes the bundle from their reports.

Why this matters

This is the standard Claude Code pattern for "broad codebase exploration that takes more than 3 queries" — and it's literally how this initiative was scoped. Opus on exploration is expensive overkill; Sonnet/Haiku Explore subagents in parallel are faster and cheaper.

What's already shipped

• prompts/methodology/design.md exists and works. This issue restructures how it's executed, not what it produces.

Proposed approach

1. Refactor DESIGN into two stages:
   • Stage A (parallel Explore): Spawn N read-only subagents (subagent_type="Explore"), each scoped to one aspect: metrics, knobs, prior findings, related principles. Run them concurrently.
   • Stage B (Opus synthesis): Single Opus call that reads the explorer reports + principles.json and writes problem.md + bundle.yaml.
2. Use Plan agent type for Stage B if you want plan-mode discipline.
3. With Phase 1: Data contracts + protocol doc #6 (MCP) live, one explorer can query the principle index across all prior campaigns.

Acceptance criteria

• DESIGN cost (tokens + dollars) drops by ≥ 30% on a representative campaign.
• DESIGN wall-clock improves on multi-aspect explorations.
• Bundle quality (subjectively rated by hand on 3 example campaigns) is no worse than the current single-Opus design.

Out of scope

• Changing the schema of bundle.yaml or problem.md — same artifacts, different production process.

---

Part of #120.

[sriumcp]

Closed by #153 (squash-merge 3e69b85, merged into reflective 2026-05-24). The merged commit message contains an explicit Closes #132. GitHub auto-close didn't fire because the PR targets reflective rather than the repo's default branch.