Inference routing follow-ups: default NVIDIA fallback and sandbox agent instructions #148

Context

Follow-up items from #133 (simplify inference routing) that were deferred from the initial implementation in #146.

Items

1. Default route: managed NVIDIA inference when no local model is present

Currently, inference.local requires explicit configuration via nemoclaw cluster inference set --provider <name> --model <id>. When no cluster inference is configured, requests to inference.local return a 503 with a hint to run the set command.

The original issue envisioned that inference.local would default to managed NVIDIA inference when no local model is deployed (e.g., on Brev/CPU). This would improve the out-of-box experience — a freshly deployed cluster would route inference to NVIDIA endpoints automatically if an nvidia provider exists.

2. Drop a skill/instructions in the sandbox telling agents about inference.local

Agents running inside sandboxes currently have no way to discover inference.local unless told externally. We should inject instructions (e.g., a file or environment hint) into sandboxes so agents know they can use https://inference.local/v1/chat/completions for inference without credentials.

Note: inference.local only works over HTTPS (the proxy intercepts CONNECT requests).