Strip UUIDs from full lever context string in enrich step

[neoneye]

Summary

• Remove lever_id UUIDs from the full-list context string provided to the LLM in each batch
• Models were copying UUIDs into synergy_text and conflict_text (llama3.1 89%, gemini 34%, gpt-5-nano 8-char truncated)
• The per-batch prompt still includes lever_id in lever_details_for_prompt so the LLM can return structured characterizations by ID
• One-line change: f"- {lever.lever_id}: {lever.name}" → f"- {lever.name}"

Documented in OPTIMIZE_INSTRUCTIONS since analysis 54. Highest priority content quality fix from analysis 59 backlog.

Test plan

• Verify llama3.1 UUID contamination drops from 89% toward <5%
• Verify gemini UUID contamination drops from 34% toward 0%
• Verify lever enrichment still succeeds (lever_id still in per-batch prompt)
• Verify no increase in phantom lever names
• Check assessment verdict before merging

🤖 Generated with Claude Code

[neoneye]

Self-improve iteration — analysis 60

Verdict: CONDITIONAL

Wins:

• UUID contamination: 105 → 15 occurrences (–86%)
• gemini: 30 → 0 UUIDs (fully cleaned)
• gpt-oss-20b: 42 → 0 UUIDs (fully cleaned for completed plans)
• llama3.1 phantom lever IDs: fully resolved (3 → 0 errors, 35/35 levers enriched)
• Content quality improved: clean lever-name references in gemini synergy/conflict text

Residual:

• llama3.1: 15 UUID occurrences remain (same-batch, from lever_details_for_prompt)
• haiku: 7 new unknown_lever_id errors (fabricated extra entries, all real levers correctly enriched — benign noise)
• gpt-oss-20b: 2/5 success (network noise, not PR-related)

Recommended follow-up: Add UUID prohibition to system prompt + exact-count instruction to user prompt