Reduce max_steps from 15 to 10 to improve runtime

[apenab]

Problem

Current max_steps=15 can overrun straightforward samples and inflate latency.

Proposal

Reduce max_steps from 15 to 10 (or make it configurable with a lower default) to cut average runtime.

Expected Impact

• Estimated runtime improvement: ~20%
• Approximate per-sample runtime: 108.5s -> ~95s

Risk

• Yes, visible quality risk on complex samples that need deeper reasoning.

Implementation Considerations

• Keep max_steps configurable per experiment/profile.
• Consider dataset- or difficulty-aware overrides.
• Pair with evaluation guardrails to detect regressions quickly.

Acceptance Criteria

• max_steps is configurable and documented.
• Benchmarks compare 15 vs 10 with:
  • latency
  • token usage
  • correctness/grounding metrics
• Regression analysis identifies which sample types degrade most.
• Decision note documents whether 10 becomes default or remains an opt-in profile.