Add UUID prohibition and exact-count instruction to enrich prompts

[neoneye]

Summary

• UUID prohibition: Add "Do NOT include any Lever ID, UUID, or identifier string" to system prompt. Targets llama3.1's 15 residual same-batch UUID occurrences (from lever_details_for_prompt)
• Exact-count instruction: Add "Return exactly N characterizations — one per lever, no more, no fewer" to user prompt. Targets haiku's 7 fabricated extra entries from analysis 60
• OPTIMIZE_INSTRUCTIONS: Update stale UUID entry to reflect post-PR Strip UUIDs from full lever context string in enrich step #457 state, document haiku's extra-characterization behavior

Test plan

• Verify llama3.1 UUID contamination drops from 15 toward 0
• Verify haiku unknown_lever_id errors drop from 7 toward 0
• Verify all models still enrich all levers (no drops from exact-count constraint)
• Check assessment verdict before merging

🤖 Generated with Claude Code

[neoneye]

Self-improve iteration — analysis 61

Verdict: CONDITIONAL

Haiku exact-count instruction: improved (–71%)

• 7 → 2 unknown_lever_id errors. gta_game fully cleaned (5→0). All real levers correctly enriched.

UUID prohibition: backfired

• The "Do NOT include any Lever ID, UUID" instruction triggered negative-priming in llama3.1 — sovereign_identity went from 0 to 20 UUID occurrences. Net llama3.1 UUIDs: 29 → 29 (zero improvement).
• This violates the documented best practice in identify_potential_levers.py:80-82: "Do NOT add explicit prohibitions naming banned phrases — small models treat the prohibition text as a template."

Recommendation: Close this PR. Replace the negative prohibition with a post-process regex strip (model-agnostic, guaranteed) + positive framing ("refer to levers by NAME only — for example, write 'Policy Advocacy Strategy'").

[neoneye]

Closing — UUID prohibition backfired (negative-instruction priming). Will use short 6-char lever ID prefixes instead of full UUIDs in batch prompt.