Add the NVIDIA B300 page and the sandboxed-agent demo

[seyeong-han]

Two NVIDIA contributions that reference each other: the B300 platform page, and the sandboxed-agent demo it enables. python platform/validate.py passes with 0 errors and 0 warnings; every relative link resolves.

platform/nvidia.md — B300

Contributed by Faradawn Yang, with a bf16 concurrency sweep from 1 to 256 concurrent requests.

The sweep is single-GPU, now confirmed by NVIDIA. The draft said both "fits on a single 8xB300 node" and "no tensor parallelism flags are needed for single-GPU serving"; Anusha confirmed the cluster node has eight cards but only one was used. That's stated above the table, since it changes how a reader sizes hardware.

Reconciled against what this cookbook already established — no new claims:

Draft	Published	Why
Pre-release model name throughout	Muse Glimmer	The draft predated the public name
--tool-call-parser onyx / --reasoning-parser onyx	muse_glimmer	Matches what upstream registers
--chat-template …onyx.jinja	dropped	The checkpoint ships the template at its root
pip install vllm==0.1.dev19055	vllm/vllm-openai:muse-glimmer container	Support is not in a released wheel (open PR), so the pip line would not reproduce
hf download <meta-models/…>	no angle brackets	A shell reads <…> as redirection
Memory observed: TBD	empty cell	We don't publish placeholders

NVIDIA's build string stays in the snapshot and methodology line, so the measurements keep their provenance even though the reader path is the container.

recipes/sandboxed-agent/ — the demo

An agent is asked to find a PRODUCTION_API_KEY, display it, and upload it. OpenShell blocks it in two independent places: credential brokering keeps the raw key outside the sandbox, and deny-by-default network policy refuses the egress. Terminal logs confirm both.

Written the way NVIDIA suggested — point at their setup docs, then show the demo as a practical example, with the recording. It is explicitly not a copy-paste recipe, and the status note says so: the setup authority is NemoClaw's own documentation, linked at each step. All four external links verified 200.

Includes the architecture diagram and the 5.2 MB recording. Both filenames were rewritten; the diagram's contents were already clean.

Not included

The RTX 5090 figure (190 tok/s, DFlash on). Holding it for two reasons:

1. We already publish a 5090 DFlash number — the model card reports 74.9 → 233.4 tok/s, 3.1× on K-Quant-17GB, batch 1, greedy, measured with llama.cpp. Publishing 190 beside it without explaining the gap would leave a reader unable to tell which applies to them.
2. The figure arrived without a quantization or a server. bf16 is ~60 GB and a 5090 has 32, so it has to be a quantized build — and DFlash appears on our llama.cpp and ExecuTorch pages, not vLLM.

Both differences are probably legitimate — different quantization, server, or prompt set will do it. We just can't say which yet. Worth one more round with NVIDIA rather than shipping two numbers for the same card.

[seyeong-han]

@sid-rp Could you please review the vLLM + NVIDIA inference server part?

[anushapant]

Looks good to me @seyeong-han

[sid-rp]

LGTM