Adaptive batch size, guarded retry, and max_tokens bump for enrich step

[neoneye]

Summary

• max_tokens bump: gpt-oss-20b from 8192 → 128000 in baseline.json — the old limit truncated structured JSON output
• Adaptive batch size: Probes model context_window metadata (not num_output which just reflects max_tokens config); uses batch_size=2 when context_window < 6000. gpt-oss-20b has context_window=3900 so it gets batch_size=2
• Guarded retry: On batch failure, split once (depth limit=1) within a 300s time budget, then skip. Prevents unbounded retry cascades
• Accurate batch counting: Report actual batches_succeeded instead of hardcoded 1 in runner.py
• OPTIMIZE_INSTRUCTIONS: Document consequence-echoing and UUID format inconsistency as known problems

Supersedes #453 which had a bug where num_output (inflated by the max_tokens bump) disabled the adaptive batch size for gpt-oss-20b.

Test plan

• Run self_improve experiment across 7 models
• Verify gpt-oss-20b uses batch_size=2 (check logs for "Adaptive batch_size=2")
• Verify gpt-oss-20b success rate improves (was 0/5 at baseline)
• Verify all other models remain at 100% success rate
• Check assessment verdict before merging

🤖 Generated with Claude Code

[neoneye]

Self-improve iteration — analysis 57

Verdict: CONDITIONAL

Wins:

• Adaptive batch size works correctly: gpt-oss-20b gets batch_size=2 (context_window=3900 < 6000)
• Guarded retry works: failing plans abort in <2s (300× faster than pre-PR 600s timeouts)
• Accurate batch counting: calls_succeeded now reports actual counts
• Error tracking: errors array persisted in raw JSON for offline diagnosis

Regression — max_tokens too high:
max_tokens=128000 leaves only 131072 − 128000 = 3072 tokens for input. Each lever needs ~3900 input tokens → every batch fails with BadRequestError. 3/5 gpt-oss-20b plans silently produce zero levers while reporting status: "ok".

gpt-oss-20b: 2/5 functional success (same as before, but failure mode changed from timeout → silent empty output)

Fix needed: Reduce max_tokens to ≤65536 and add "context_window": 131072 to gpt-oss-20b config

[neoneye]

Closing — max_tokens=128000 leaves only 3072 input tokens, causing BadRequestError for 3/5 gpt-oss-20b plans. Will rework with max_tokens=65536 and explicit context_window=131072.