refactor: methodology + principles via per-campaign CLAUDE.md and auto-memory

[sriumcp]

TL;DR

Move the campaign methodology and accumulated principles out of "re-sent every prompt" into the target repo's .nous/<run-id>/CLAUDE.md. Claude Code auto-loads it on every session. Auto-memory persists per-iteration learnings across sessions.

Why this matters

Right now Nous re-injects the campaign brief, methodology, and accumulated principles into every claude -p call. Claude Code already has two persistence mechanisms loaded automatically every session: CLAUDE.md and auto-memory. Using them both reduces token spend AND simplifies the orchestrator (less prompt assembly).

What's already shipped

• feat: compact handoff from designer to executor agent #52 made the handoff compact. This issue extends the principle: stop manually re-injecting things Claude Code can load itself.
• Pairs with Agentic Systems Experimentation and Discovery (ASED) #2 (caching): Agentic Systems Experimentation and Discovery (ASED) #2 wins within a session, this wins across sessions.

Proposed approach

1. On nous run, write the campaign methodology + current principles into <repo>/.nous/<run-id>/CLAUDE.md.
2. Pass --add-dir <repo>/.nous/<run-id> (or the SDK equivalent) so Claude Code auto-loads it.
3. After each iteration, regenerate the CLAUDE.md with updated principles and last-iteration handoff (deterministic Python — no LLM).
4. Use auto-memory for longer-tail learnings ("at high saturation, binary gating loses discriminatory power vs ordinal scheduling") that persist across runs.
5. The orchestrator's prompts/methodology/*.md stays as the source of truth, but it's installed into CLAUDE.md per-campaign rather than re-sent every call.

Acceptance criteria

• <repo>/.nous/<run-id>/CLAUDE.md is created on nous run and updated each iteration.
• Iteration N+1 prompts are measurably smaller than iteration N's were before this change.
• No regressions on the existing test suite.

---

Part of #120.

[sriumcp]

Closed by #153 (squash-merge 3e69b85, merged into reflective 2026-05-24). The merged commit message contains an explicit Closes #131. GitHub auto-close didn't fire because the PR targets reflective rather than the repo's default branch.