Use a faster model for subcalls (gpt-4o) to reduce latency

[apenab]

Problem

Subcalls currently use a stronger but slower model (gpt-5.1), which increases per-sample latency.

Proposal

Use a faster model for subcalls (e.g., gpt-4o) while keeping the baseline/final reasoning path on gpt-5.1.

Suggested split:

• Baseline/final pass: gpt-5.1
• Subcalls/tool-like intermediate calls: gpt-4o

Expected Impact

• Estimated latency improvement: ~40-50% for subcall-heavy flows.
• Approximate per-sample runtime: 108.5s -> 60-70s (workload dependent).

Risk

• Yes, quality risk: possible ~5-10% drop in correctness/grounding.

Implementation Considerations

• Add model selection by call type in config.
• Support easy rollback to single-model mode.
• Add per-call tracing to measure where latency is actually saved.

Acceptance Criteria

• Config supports separate models for baseline/final vs subcalls.
• Benchmark report compares before/after on:
  • latency per sample
  • token usage/cost
  • correctness
  • grounding/evidence metrics
• Quality delta is explicitly documented and approved before default rollout.
• One-command rollback path is available.