agent-platform release v4.96.0

nemotron-3-super-nvfp4 serves with CUDA graphs: --enforce-eager is gone from its arguments (giantswarm/giantswarm#37941). Measured on one DGX Spark (128 GB unified memory) with the preset’s other arguments unchanged, 134 requests, no error: a single stream decodes at 15.7 instead of 15.1 tok/s, aggregate throughput at 8 to 32 concurrent requests is unchanged (57, 81, 114 tok/s), and the time to first token drops by 10–15 % for short prompts and by 29 % for 5.8k-token prompts (3.3 to 2.3 s). Split across two Sparks the gain is larger: 16.3 to 26.3 tok/s single stream once the tensor-parallel all-reduce also runs over RoCE. The costs: the captured graph pool (2.2 GiB) counts inside --gpu-memory-utilization=0.72, so one node keeps 287k instead of 372k KV-cache tokens (35 full 8k contexts), and a start takes about 40 seconds longer on one node (torch.compile 18 s, cached on the model cache claim for the next pod, and 24 s of graph capture), next to the 6 minutes of weight loading. The tool-calling turn with the qwen3_coder parser passes unchanged.