docs: serve a model in "Your own mesh", cover the three service types

[aojea]

What

The "Your own mesh" walkthrough declared a static file server as type: mcp and fetched it through the raw proxy path, which taught the mesh as a generic HTTP passthrough. The type is a contract: the node verifies that the backend speaks the declared protocol (model list, MCP initialize, agent card) before advertising it, and the docs should say so and show a real service.

• your-own-mesh: node A publishes Ollama as an inference service; node B calls it with curl over its socket (/v1/models, then /v1/chat/completions). The section explains what the node checks for each type and why a plain web server is never advertised. A new "MCP servers and A2A agents" section shows the other two declarations and links to the quick start's MCP call, Exposing services, A2A chat (feat(node): add A2A support #347) and Gemini Buddy. mcp-client is no longer a prerequisite for this page.
• quickstart: after the MCP call, note that MCP is one of three types and list the testnet's models with curl --unix-socket … /v1/models.
• node-api reference: curl examples for inference and for fetching an A2A card through the proxy path, next to the existing MCP one.
• sandboxed-agents: the boundary table said <service>.mcp.sam.alt; ParseMeshHost/dialMeshService are type-generic, so it now reads <service>.<type>.sam.alt for mcp, inference and a2a.

Testing

tests/e2e/standalone.bats mirrored the old passthrough example (python3 -m http.server declared as mcp, fetched via /sam/<peer>/mcp/smoke/hello.txt). It now follows the doc: a stdlib fake OpenAI backend (tests/e2e/fixtures/fake_openai.py) declared as inference on node A, asserted through node B's /v1/models (owner is A's peer ID) and /v1/chat/completions. The same server is also declared as type: mcp; the test asserts the node logs not advertising it and that B's discovery never lists it, so the type contract cannot regress silently.

Passes locally in ~6.5s. Ollama itself is not installed here, so the gemma3:1b steps were checked against the code path (/v1/models on the backend root) and the e2e rather than a live Ollama.

[gemini-code-assist]

Code Review

This pull request updates the documentation and end-to-end tests to transition from a generic HTTP smoke service to an OpenAI-compatible inference service type, utilizing a new fake_openai.py mock backend. It also documents the a2a and mcp service types. Feedback is provided on the mock server's request parsing logic to prevent potential crashes or hangs from malformed requests.

[gemini-code-assist]

Parsing Content-Length directly using int(self.headers.get("Content-Length") or 0) can raise a ValueError if the header is present but invalid (e.g., contains non-numeric characters). Additionally, if a negative value is supplied, self.rfile.read() will block indefinitely waiting for EOF. Finally, if the request body is not a JSON object (e.g., a list or a primitive), req.get will raise an AttributeError and crash the handler.\n\nUsing a robust try-except block with max(0, ...) and validating that req is a dictionary ensures the server does not crash or hang due to malformed requests.

        try:\n            length = max(0, int(self.headers.get("Content-Length", 0)))\n        except ValueError:\n            length = 0\n        try:\n            req = json.loads(self.rfile.read(length) or b"{}")\n        except ValueError:\n            req = {}\n        if not isinstance(req, dict):\n            req = {}