Customize an agent prompt

Stub for SPEC-38.11. Substantial content lands in a follow-up batch.

Each agent's system prompt lives in a sibling .md file beside its agent code: src/efterlev/agents/gap_prompt.md, documentation_prompt.md, remediation_prompt.md. Customizing means editing the prompt and re-running.

What discipline is non-negotiable across any prompt customization:

• Per-run fence-nonce rules. Every prompt wraps cited evidence in <evidence_NONCE id="sha256:...">...</evidence_NONCE> fences with a per-run nonce. The post-generation validator rejects output that cites IDs not present as fences. Removing this is removing the hallucination defense.
• DRAFT marker. Every Claim carries a requires_review=Literal[True] invariant. The prompt must instruct the LLM to produce output that lands inside this invariant; you can't relax to Literal[False] via prompt engineering — the type system catches it.
• Citation discipline. Agents cite by content-addressed ID, not by paraphrase. The Gap Agent's prompt instructs the model to cite specific evidence records when classifying; weakening this regresses provenance integrity.

After editing, run the e2e harness:

ANTHROPIC_API_KEY=... uv run python scripts/e2e_smoke.py

If checks fail, the prompt change broke something. Read .e2e-results/<ts>/summary.md for the specific failure.