Adaptive batch size, guarded retry, and OpenRouter config fixes for enrich step

[neoneye]

Summary

• OpenRouter config fixes: Add explicit context_window for all 4 OpenRouter models (131072/1048576/128000/40960) — LlamaIndex defaults to 3900 for all OpenRouter models without this
• gpt-oss-20b max_tokens: 8192→65536 (balanced: 65536 input + 65536 output within 131K context)
• Adaptive batch size: batch_size=2 when context_window < 3000 (threshold below OpenRouter's 3900 fallback to avoid false positives)
• Guarded retry: Split once (depth=1) within 300s budget, then skip
• Error tracking: Persist errors in raw JSON output for offline diagnosis
• Accurate batch counting: batches_succeeded instead of hardcoded 1
• OPTIMIZE_INSTRUCTIONS: Document 4 new known problems including OpenRouter metadata fallback

Supersedes #453, #454, #455 which each had config or threshold bugs.

Test plan

• Verify all 7 models use batch_size=5 (no model has context_window < 3000)
• Verify gpt-oss-20b achieves 5/5 plan success with valid output
• Verify no BadRequestError in error arrays
• Verify qwen3-30b, gpt-4o-mini, gemini-flash use calls_succeeded=2 (ceil(7/5)=2, not 3-4)
• Check assessment verdict before merging

🤖 Generated with Claude Code

[neoneye]

Self-improve iteration — analysis 59

Verdict: CONDITIONAL

Wins (all 6 targeted issues resolved):

• gpt-oss-20b: 0/5 → 4/5 plans (+11.4pp overall success rate, largest single-iteration gain)
• OpenRouter context_window: all 4 models now report real values (131K/1M/128K/41K)
• Adaptive batch size: SMALL_CONTEXT_THRESHOLD=3000 correctly avoids false positives on OpenRouter's 3900 fallback
• Guarded retry: implemented, not triggered (all batches succeeded first attempt)
• Error tracking: surfaced pre-existing llama3.1 phantom lever ID bug (3 levers silently unenriched)
• Accurate batch counting: working correctly

Remaining issues (not regressions):

• gpt-oss-20b parasomnia timeout at 600s (model throughput limitation, not config issue — gta_game took 243s vs 21-47s for other models)
• llama3.1 phantom lever IDs: 3/35 levers unenriched due to truncated UUIDs (pre-existing, now visible thanks to error tracking)

No content regressions introduced.