Add per-batch retry for enrich step, update OPTIMIZE_INSTRUCTIONS

[neoneye]

Summary

• B1: Enrich step now retries failed batches by splitting in half instead of aborting the entire plan. Single-lever failures are logged and skipped. Should recover gpt-oss-20b's 0/5 failure caused by output token limits on large batches.
• B4: runner.py reports actual batches_succeeded count instead of hardcoding 1.
• D5: Added two new known problems to OPTIMIZE_INSTRUCTIONS: consequence echoing without elaboration (llama3.1 post-PR Include consequences and review in enrich batch prompt #451) and UUID cross-reference format inconsistency.

Identified by analysis/54_enrich_potential_levers assessment backlog.

Test plan

• Run enrich_potential_levers experiment with this PR across all models
• Verify gpt-oss-20b recovers plans via batch splitting (was 0/5 in analysis 53 and 54)
• Verify no regressions for models that already succeed (6/6 models at 100%)
• Check calls_succeeded in output metadata reflects actual batch count

🤖 Generated with Claude Code

[neoneye]

Self-Improve Iteration Result: CONDITIONAL

Experiment ran as analysis/55_enrich_potential_levers (runs 92–98, 7 models × 5 plans).

Improvements

• gpt-oss-20b recovered 2/5 plans (was 0/5) — batch splitting works, confirmed via usage_metrics (8192 token overflow triggered split into two sub-batches)
• calls_succeeded now accurate — reports 3–4 per plan instead of hardcoded 1
• No regressions for the 6 working models (all still 100%)

Issues

• 3 gpt-oss-20b plans timeout at 600s instead of failing fast (~30s before) — retry cascade with no depth limit wastes ~1800s per run
• Root cause unaddressed: BATCH_SIZE=5 is too large for gpt-oss-20b's context_window=3900

Required follow-ups

1. Adaptive BATCH_SIZE based on context_window (would likely recover all 5 plans)
2. Retry depth limit (max 1 split) + remaining-time guard
3. UUID fix at line 168 (full_lever_context_str still exposes UUIDs)

Full assessment: analysis/55_enrich_potential_levers/assessment.md

[neoneye]

not good enough, I'm closing it.