perf: cache static methodology prompts via SDK system blocks (`cache_control: ephemeral`)

[sriumcp]

TL;DR

Mark the 266-line design.md and 199-line execute_analyze.md as ephemeral-cached system blocks on every SDK call. Expected reduction: 30–50% input tokens once warm.

Why this matters

Every claude -p invocation today re-sends the full methodology prompt uncached. Across a 5-iteration campaign that's:

• design.md (~266 lines × 5 iterations) + execute_analyze.md (~199 lines × 5 iterations) = ~2300 lines of methodology re-tokenized 10 times.
• Plus per-iteration handoff (~150 lines, growing) and principles.json (up to 26 entries on mech-design-enforcement).

The Anthropic prompt cache has a 5-minute TTL. Within a single phase the cache should hit; designer→executor transitions are typically < 5 minutes.

What's already shipped

• feat: compact handoff from designer to executor agent #52 introduced compact handoff (variable context). This issue handles the static part — the methodology prompts themselves.
• feat: add token/cost tracking to both dispatchers #41 added token/cost tracking. This will let us measure the cache hit rate empirically.

Proposed approach

1. After Agentic Strategy Evolution: a three-loop methodology for optimizing multi-layer policy spaces #1 lands, restructure the prompt assembly so the methodology is a system block with cache_control: {"type": "ephemeral"} and the per-iteration context (handoff, principles) is in the user message (uncached).
2. Verify cache hits via SDK usage.cache_read_input_tokens and emit it to llm_metrics.jsonl.
3. Add a nous cost --cache-stats flag showing cache hit rate per campaign.

Acceptance criteria

• After two successive phase calls within 5 minutes, the second shows non-zero cache_read_input_tokens.
• On a representative 5-iteration campaign, total input tokens decrease by ≥ 25% vs the pre-change baseline.
• nous cost --cache-stats exists and is documented.

Notes for implementers

• Keep the user message as the pivot for cache-busting per iteration; never bury per-iter content inside the cached system block.
• The SDK uses an array of system blocks; only mark the methodology ones as cached.

---

Part of #120.

[sriumcp]

Closed by #153 (squash-merge 3e69b85, merged into reflective 2026-05-24). The merged commit message contains an explicit Closes #122. GitHub auto-close didn't fire because the PR targets reflective rather than the repo's default branch.