Add neurarch-mcp

[neurarch-ai]

Adds neurarch-mcp (https://github.com/neurarch-ai/neurarch-mcp, MIT, npm neurarch-mcp 0.13.0) as a local server built from the Dockerfile at the repo root.

What it does: gives an agent structural awareness of a PyTorch model. Tools inspect layers, parameters, FLOPs and the blast radius of an edit; lint_model runs the structural design rules; check_design returns the readiness / cost / deployment verdict; rank_designs orders candidate designs before any GPU time is spent. Inside the container it runs in hosted stdio mode (--hosted --hf --tools=core): nothing is mounted, so every call names its model as model_path: "zoo:<id>" (81 bundled reference architectures), model_path: "hf:<org/name>", or model_source with the model text inline.

• Only runtime dependency is @modelcontextprotocol/sdk; the parser, rule engine and verifier are vendored, so the container makes no network call except huggingface.co on an explicit hf: request.
• HF_TOKEN is an optional secret for gated repos.
• Verified locally: node dist/index.js --hosted --tools=core answers initialize and a tools/call over raw stdio with no model file present. I could not run task create on this machine (no Docker Desktop here), so the image build is left to CI; happy to adjust if it asks for anything.

Opened by an automated agent on behalf of the maintainer.