Reduce word counts, add anti-echoing, and use verbatim id guidance

[neoneye]

Summary

• Reduced word counts: description 80-100→50-70, synergy/conflict 40-60→20-40
• Anti-echoing (positive framing): "Add new insight beyond what consequences and review already state"
• Verbatim id guidance: Field description says "copy it verbatim from the prompt, without XML tags". System prompt says "copy the id verbatim from inside the tags — strip the XML tags but do not alter the id itself." Avoids "hexadecimal" which caused gpt-4o-mini to strip hyphens (Reduce word counts, add anti-echoing, and guide lever_id extraction #468), and avoids "uuid" which models interpret differently.

Supersedes #467 (XML-tag regression) and #468 (hyphen-stripping regression).

Test plan

• Verify all models return correct lever_id (no XML tags, no hyphen stripping)
• Verify gpt-oss-20b completes 5/5 plans
• Verify UUID contamination stays at 0
• Verify no model < 0.4x baseline field lengths
• Check assessment verdict before merging

🤖 Generated with Claude Code

[neoneye]

Self-improve iteration — analysis 68

Verdict: CONDITIONAL (near-YES)

All goals achieved:

• 100% success rate (35/35 plans, all 7 models)
• gpt-oss-20b: 5/5 (was 3/5), max 166s
• 0 UUID contamination in synergy/conflict
• 0 hyphen-stripping issues (gpt-4o-mini fixed)
• 0 XML-tag leakage (llama3.1 fixed)
• All 35/35 levers enriched for every model
• Field lengths normalized: avg desc 51 words, syn/conf 26 words

Minor residual: haiku gta_game produced 1 dummy_override lever_id (benign, discarded). 2 haiku noise errors total. All real levers unaffected.

Conditions: Remove errors.append for unknown_lever_id (noise cleanup) and refine field description wording.