feat: parallelize arm execution via SDK subagents (one subagent per arm × seed)

[sriumcp]

TL;DR

After DESIGN produces experiment_plan.yaml, fan out arms to subagents in parallel, each in its own git worktree. Reduce wall-clock 2–4× on multi-arm iterations and eliminate the 90-minute "single mega-session" failure mode.

Why this matters

mech-design-enforcement iter-2 ran 8 conditions × 3 seeds = 24 simulations sequentially in one Sonnet session. That 2.5-hour session was the proximate cause of the connection drops the user complained about on 5/18 — and when one of those connection drops triggered --max-cli-retries 10, a second worktree spawned while the first was still alive, leading to two executors racing on the same iter-2/results/ directory.

Subagents share the parent's CLAUDE.md / skills / auto-memory, so per-arm context is small. Each arm is independent (different seed / different policy / same patched binary). Worktree-per-arm gives clean isolation.

What's already shipped

• worktree.py already manages per-iteration worktrees.
• feat: retry transient claude -p subprocess errors with exponential backoff #71 + feat: pre-flight check + retry everything with failure persistence #111 handle transient failures gracefully — but per-arm subagents fail in smaller, more recoverable units.

Proposed approach

1. Parse experiment_plan.yaml into a list of (arm, seed, command) units.
2. For each unit, spawn a subagent via the SDK with:
   • Read access to iter-N/inputs/ and the patched binary.
   • Write access to iter-N/results/<arm>/<seed>/.
   • Worktree isolation (see docs: add Claude-based PR creation workflow #13).
3. Cap concurrency by a campaign-level max_parallel_arms: setting (default = min(CPU, 4)).
4. Parent agent awaits all subagents, then runs the existing deterministic merge into findings.json and principle_updates.json.
5. On partial failure, retry only the failed units (not the whole iteration).

Acceptance criteria

• examples/campaign-best-of-field.yaml runs end-to-end with max_parallel_arms: 4 in significantly less wall-clock than the serial baseline.
• Failure of one arm does not abort the iteration; it surfaces in the retry log and merges as arm_status: "failed" in findings.json.
• No two subagents ever write to the same results/ subpath.

Out of scope

• Cross-iteration parallelism (still one iteration at a time).
• Background-agents / multi-campaign supervision (see Phase 1: Schemas + templates + orchestrator skeleton + protocol doc #14).

---

Part of #120.

[sriumcp]

Closed by #153 (squash-merge 3e69b85, merged into reflective 2026-05-24). The merged commit message contains an explicit Closes #123. GitHub auto-close didn't fire because the PR targets reflective rather than the repo's default branch.