Include consequences and review in enrich batch prompt

[neoneye]

Summary

• The enrich step's batch prompt only provided lever_id, name, and options to the LLM, but omitted consequences and review — the two fields that document the lever's effects and primary trade-off.
• This forced the model to infer descriptions and conflict analysis from the name alone, producing vague or ungrounded enrichment.
• Adds both fields to the per-lever details in the batch prompt so the LLM can write grounded description, synergy_text, and conflict_text.

Identified by analysis/53_enrich_potential_levers synthesis as the #1 ranked improvement direction (B3 + I4).

Test plan

• Run enrich_potential_levers experiment with this PR across all models
• Compare assessment against analysis 53 baseline — expect more specific descriptions grounded in lever consequences/trade-offs
• Verify no increase in validation failures (change is input-only, output schema unchanged)

🤖 Generated with Claude Code

[neoneye]

Self-Improve Iteration Result: YES (keep)

Experiment ran as analysis/54_enrich_potential_levers (runs 85–91, 7 models × 5 plans).

Improvements

• gpt-5-nano template artifact eliminated — "Purpose: / Objectives: / Key success metrics:" sub-header structure gone from all 5 plans
• Conflict texts now grounded — multiple models echo review language in conflict_text instead of fabricating from lever name alone
• qwen3-30b synergy/conflict improved — from 0.58–0.69× to 0.74–0.76× baseline

Regressions (minor)

• llama3.1 description brevity — dropped from 0.80× to 0.65× baseline (echoes consequences verbatim instead of elaborating)
• qwen3-30b UUID format nondeterministic — pre-existing B2 bug surfaced more visibly
• Single CJK character leak in qwen3-30b (one occurrence, not systematic)

Unchanged

• Success rate: 30/35 (gpt-oss-20b still 0/5 — separate B1 fix needed)
• Input token cost: +18–20% (expected trade-off)

Full assessment: analysis/54_enrich_potential_levers/assessment.md