Adaptive batch size, guarded retry, and gpt-oss-20b config fix for enrich step

[neoneye]

Summary

• gpt-oss-20b config fix: max_tokens 8192→65536 (balanced: 65536 input headroom), add explicit context_window=131072. Previous attempts: 128000 left only 3072 input tokens causing BadRequestError
• Adaptive batch size: batch_size=2 when context_window < 6000 (probes model metadata, no API call)
• Guarded retry: On batch failure, 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 consequence-echoing, UUID format inconsistency, and max_tokens overflow

Supersedes #453 (num_output bug) and #454 (max_tokens=128000 too high).

Test plan

• Verify gpt-oss-20b uses batch_size=5 (context_window=131072 > 6000 threshold)
• Verify gpt-oss-20b achieves 5/5 plan success (was 0/5 at baseline)
• Verify no BadRequestError in error arrays
• Verify all other models remain at 100% success rate
• Check assessment verdict before merging

🤖 Generated with Claude Code

[neoneye]

Self-improve iteration — analysis 58

Verdict: CONDITIONAL

Wins:

• gpt-oss-20b config fix works: all 5 plans produce valid enriched lever output (was 0/5 before)
• Error tracking works: llama3.1 UUID mutations now visible (4 errors in run 13)
• OPTIMIZE_INSTRUCTIONS: 3 new entries documented

Regression — adaptive batch_size=2 fires for wrong models:
LlamaIndex's OpenRouter class reports context_window=3900 universally for ALL OpenRouter models. Since 3900 < SMALL_CONTEXT_THRESHOLD(6000), qwen3-30b, gpt-4o-mini, and gemini-flash all get batch_size=2 (3-4 API calls/plan instead of 1-2). The adaptive feature now protects nothing and penalizes 3 unrelated models.

gpt-oss-20b: 5/5 valid output, but 3/5 plan timeout (600s)
The enrichment completes and writes correct output, but LLMExecutor's provider failover (6-9 attempts) exhausts the 600s plan budget. Different failure mode than before — the data is there, just the status is wrong.

Fix needed: Lower SMALL_CONTEXT_THRESHOLD from 6000 to 3000 (OpenRouter's 3900 falls above threshold)

[neoneye]

Closing — adaptive batch_size=2 incorrectly triggers for all OpenRouter models (context_window=3900 fallback). Will rework with SMALL_CONTEXT_THRESHOLD=3000 and explicit context_window for OpenRouter models.