feat(evaluation): add DeepSeek multi-agent incremental ablation

[Atlas2005]

Summary

• add a PIT-safe incremental ablation comparing:
  • non-LLM baseline
  • single-agent advisory
  • independent multi-agent advisory
  • research-committee multi-agent advisory
• reuse the existing DeepSeek advisory and research committee infrastructure
• add offline replay support using deterministic cached advisory outputs
• record decision changes, fallback/invalid-output ratios, token usage, estimated API cost, and net performance after API cost
• keep the runner offline by default with zero live requests unless explicitly enabled

Validation

• 63 focused tests passed
• 1607 full-suite tests passed
• PR121 full-A artifact replay completed successfully
• 5 common OOS folds
• 5 PIT-safe decision packets
• 4 expected ablation arms
• 0 physical model calls in offline replay
• committee provenance recorded
• offline replay validation passed

Offline replay result

The deterministic synthetic advisory replay produced:

• non-LLM compounded return: 16.81%
• single-agent compounded return: 20.36%
• multi-agent compounded return: 21.65%
• benchmark compounded return: 31.86%

The final recommendation remains no_ai_arm because no AI arm met the evaluation thresholds and the replay used synthetic advisory evidence. No profitability claim is made from this offline fixture.

Boundaries

• no broker interaction
• no live DeepSeek calls by default
• no test-period outcome leakage
• no generated artifacts committed