feat(examples): add simple-sampling server and client

[trentisiete]

Adds an end-to-end sampling example to examples/, addressing #1205.

Two new workspace members:

• examples/servers/simple-sampling/ — minimal server with one tool (write_story) that calls session.create_message(...) populating every advisory field of CreateMessageRequestParams: model preferences with hints + cost/speed/intelligence priorities, system prompt, temperature, stop sequences, include_context, metadata.
• examples/clients/simple-sampling-client/ — wires sampling_callback to a real LLM via any OpenAI-compatible endpoint. Default is Groq (free tier). Switching to OpenAI / OpenRouter / Ollama (/v1) / vLLM is 1-3 env vars; no extra dependencies.

Context

There's already #1436 from @yarnabrina addressing this issue. Their PR has been waiting on review for months and they've explicitly given the green light for an independent PR (see the recent comment thread on #1205). This one tries to land the same goal while incorporating the feedback their PR received and never got a chance to resolve:

• No dependency on the openai SDK; uses httpx against the OpenAI-compatible /chat/completions schema.
• LLMClient abstraction matching the simple-chatbot pattern that was requested in their review — the sampling callback knows nothing about httpx, and the LLM wrapper knows nothing about MCP.

Design notes

• Model selection: first usable ModelHint.name is treated as a soft override, falling back to LLM_MODEL. The numeric priorities (cost/speed/intelligence) get logged but not used for routing — picking a model from those would require a provider-specific catalog, which is out of scope for an example.
• LLM failures return types.ErrorData(code=INTERNAL_ERROR, ...) rather than raising, so the server gets a readable error instead of a transport-level one.
• Non-text content blocks (image/audio) become [<type> content omitted] placeholders rather than being silently dropped. A production client would forward or explicitly reject them.

Verification

End-to-end check against a local ThreadingHTTPServer stub of /chat/completions that captures the request body and returns a canned story. Asserts that:

• The selected model is the first hint (llama-3.1-8b).
• max_tokens=200, temperature=0.8, stop=["THE END"], the system prompt, the user message and the metadata round-trip correctly.
• The canned response surfaces in the client stdout.

The verification harness itself is not part of the PR. Repo checks pass:

• uv run ruff format --check: clean
• uv run ruff check: 0 issues
• uv run pyright: 0 errors, 0 warnings, 0 informations

Try it

export LLM_API_KEY=<your_key>
uv sync --all-packages
uv run mcp-simple-sampling-client --topic "a lighthouse keeper"

Full env var table in examples/clients/simple-sampling-client/README.md.