Anchor prefers NVIDIA Build over local inference when it's live

Local Ollama/llama.cpp is the right default for offline work and tiny swarm edits, but it's often the slowest option on a laptop. Plenty of operators already have an NVIDIA Build key sitting in their environment for something else — Anchor now checks for it and prefers it automatically when it's actually live, without touching an operator's existing choice otherwise.

Configured and reachable, not just configured

A key existing in the environment doesn't mean the endpoint is reachable right now — a typo, an expired key, or a network hiccup all look identical from the outside. scripts/nvidia_build_probe.py does a real, short-timeout check against the endpoint before ever preferring it:

python scripts/nvidia_build_probe.py --probe

Missing key, invalid key, or no network: it reports local as the backend and moves on — never a crash, never a loud failure blocking an Aider session from starting.

Operator override always wins

If you've explicitly configured local inference — a specific model in .aider.conf.yml, a CLI flag, prefer_local in your conventions — a live Build key never silently overrides that. The preference only kicks in when nothing else was already decided.

Hosted, not self-hosted

This is NVIDIA's hosted developer API (build.nvidia.com) — no container, no GPU of your own. It's a different product from the self-hosted NIM containers Anchor already documents (platforms/nvidia-nim/NEMOTRON.md), and the two docs pages now cross-link explicitly so nobody confuses which one they're reading. See NVIDIA Build for the full configuration and Aider wiring.