ADR written to docs/ADR-324.md.

Key deviation from the plan: I recommend one call site after the if-else (line ~2852) instead of two inline calls. The existing skill_analyze_outcome block at line 2854 already follows this exact pattern — it checks exit_code in one place to handle both outcomes. Duplicating the JSON structure across two branches is unnecessary.

I also flagged a bug in the plan: the jq snippet reuses --arg it for both issue_type and iterations, which would cause a collision. The implementation must use distinct argument names.