Qualify and enable MTP 2 with paired workload evidence and fallback

compose.yaml

@@ -30,10 +30,11 @@ services:
         --host 0.0.0.0 --port 8000
         --tensor-parallel-size 1 --dtype bfloat16
         --engram-config '{"cpu_offload":true}'
-        --kv-cache-dtype fp8 --gpu-memory-utilization 0.96
+        --kv-cache-dtype fp8 --gpu-memory-utilization 0.985
         --max-model-len 131072 --max-num-seqs 1 --max-num-batched-tokens 2048
         --enable-chunked-prefill --enable-prefix-caching
-        --compilation-config '{"mode":0,"cudagraph_mode":"FULL","cudagraph_capture_sizes":[1]}' --no-enable-flashinfer-autotune
+        --speculative-config '{"method":"mtp","num_speculative_tokens":2}'
+        --compilation-config '{"mode":0,"cudagraph_mode":"FULL","cudagraph_capture_sizes":[1,3]}' --no-enable-flashinfer-autotune
         --load-format safetensors
         --reasoning-parser qwen3 --tool-call-parser qwen3_xml --enable-auto-tool-choice
         --api-key "$$(cat /run/secrets/api-key)"