- Measure Qwen3-Coder-30B-A3B, Qwen2.5-Coder-32B, and Tiel-Coder-35B-A3B at 64k and 32k context
- Verify long-context degradation profile and MoE attention scaling
- Add benchmark suite and results to BENCHMARK_MOE_CANDIDATES.md and benchmark_moe_results.json
