Move UUID to end of lever details, add positive framing and exact-count

[neoneye]

Summary

• Move UUID to end: Change Lever ID: {uuid} (first line) to (internal reference: {uuid}) (last line). Models tend to copy identifiers from the beginning of text blocks. Full UUID kept for reliable matching — no mapping code changes needed.
• Positive framing: "refer to levers by their name — for example, write 'Policy Advocacy Strategy', not an identifier"
• Exact-count instruction: "Return exactly N characterizations — one per lever, no more, no fewer"
• OPTIMIZE_INSTRUCTIONS: Document lessons from PRs Strip UUIDs from full lever context string in enrich step #457-462 (negative prohibitions backfire, short prefixes/integers break matching, full UUIDs at end of block is the sweet spot)

Test plan

• Verify llama3.1 UUID contamination drops from 15 (analysis 60 baseline)
• Verify all levers still correctly enriched (UUID matching unchanged)
• Verify haiku errors drop from 7 (analysis 60 baseline)
• Verify no content regressions
• Check assessment verdict before merging

🤖 Generated with Claude Code

[neoneye]

Self-improve iteration — analysis 64

Verdict: NO

Moving UUID to end of lever block broke structured-output matching for small models. They use positional heuristics — the first item in a block becomes the lever_id value.

• llama3.1: 35/35 → 0/35 levers (returned lever names as lever_id)
• gpt-5-nano: 35/35 → 10/35 levers
• Total: 245 → 181 characterized levers (–26%)
• 152 errors (was 7)

UUID position matters for matching. Must stay as the first item in each lever block.