Recipe command snippets point to nonexistent runtime modules

[nficano]

Many recipe pages tell users to start examples with commands such as uv run python -m examples.agent_versions.runtime & in docs/recipes/agent-versions.md:11, but the example packages use server.py and client.py files rather than runtime.py. The same examples.*.runtime pattern appears across recipe pages for list jobs, lease violation, heartbeat, vendor extensions, progress, delegation, cancellation, subscribe, resume, cost budget, ack backpressure, result chunks, idempotent retry, submit and stream, custom auth, lease expiry, and host tracing. These commands fail before users reach the SDK behavior being demonstrated.

Fix prompt: Replace every uv run python -m examples.<name>.runtime command in docs/recipes/*.md with the corresponding importable server module, such as uv run python -m examples.agent_versions.server, and verify the matching client command still exists. Add a lightweight documentation command check that imports or runs python -m with --help or a dry start mode for each recipe entry so stale module names are caught in CI.

[Keesan12]

This kind of docs/runtime drift is worth catching with an executable recipe smoke test. If the docs say python -m examples.foo.runtime, CI should prove that module exists and starts.

A simple matrix over recipe command snippets would prevent a lot of trust loss: parse fenced commands, run import/module existence checks, and fail when examples point at nonexistent entrypoints.

For an agent runtime/control SDK, examples are part of the API surface. Broken recipe commands make the control story feel less reliable than the protocol itself.