Benchmark DeepEP-LL donor Monte Carlo cost

[jasonqinzhou]

Context

AISimulate PR #80 introduces Stage 1 DeepEP low-latency decode modeling. Local expert compute remains measurement-backed through MoeExpertCompute, including the SGLang DeepEPMoE.run_moe_core path. The deterministic Monte Carlo model applies only to DeepEP-LL dispatch/combine endpoint imbalance.

This non-blocking follow-up owns the operational cost of that communication model after Stage 1 lands.

Follow-up

• Benchmark uncached donor Monte Carlo queries for representative EP8/32/64/128 configurations in an optimized production build.
• Measure end-to-end predict and recommend cost across realistic candidate counts, including cache-hit rate.
• Set an acceptable cold-query and sweep-time budget.
• If the budget is exceeded, evaluate adaptive trial counts, precomputation, or broader cache reuse while preserving deterministic results.

Acceptance criteria

• A reproducible optimized benchmark reports cold and warm timings for representative exact and donor queries.
• The measured predict/recommend impact and agreed performance budget are documented.
• Remediation work is filed if the budget is missed.

Related: #80