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 at exactly 8192 tokens, causing 0/5 plan failures
• C1 — Adaptive batch size: Probes model num_output metadata; uses batch_size=2 when below 16384 tokens as a safety net for output-constrained models
• C2 — Guarded retry: On batch failure, split once (depth limit=1) within a 300s time budget, then skip. Prevents the unbounded retry cascade that caused 3×600s timeouts in PR Add per-batch retry for enrich step, update OPTIMIZE_INSTRUCTIONS #452
• B4 — Accurate batch counting: Report actual batches_succeeded instead of hardcoded 1 in runner.py
• D5 — OPTIMIZE_INSTRUCTIONS: Document consequence-echoing and UUID format inconsistency as known problems

Test plan

• Run self_improve experiment across 7 models
• Verify gpt-oss-20b recovers to 5/5 plan success (was 0/5 before)
• Verify no 600s timeouts for gpt-oss-20b (should complete in ~300-400s)
• Verify all other models remain at 100% success rate
• Verify calls_succeeded reports accurate batch counts
• Check assessment verdict is YES before merging

🤖 Generated with Claude Code

[neoneye]

Self-improve iteration — analysis 56

Verdict: CONDITIONAL

Wins:

• Guarded retry: 4.3× reduction in wasted runner time (gpt-oss-20b ~2224s → ~520s)
• max_tokens bump: 2/5 gpt-oss-20b plans produce complete JSON (no more truncation at 8192)
• All 6 other models: 100% success, unchanged quality

Bug found — C1/C2 interaction:
max_tokens=128000 inflated num_output to 128000, making adaptive batch check (num_output < 16384) always false for gpt-oss-20b. The model still runs with batch_size=5 instead of the intended batch_size=2.

gpt-oss-20b: 2/5 functional success (unchanged from PR #452, but failure mode improved from 600s timeout → fast empty_response)

Follow-up needed: C4 — fix threshold to also check context_window (stays at 3900 regardless of max_tokens)

[neoneye]

Closing — C1/C2 interaction bug makes adaptive batch size ineffective for the target model. Will rework.