Bound pre-lifecycle bootstrap attempts in server initialization #1315

Follow-up from #1248.

After #1248, serve_server stays in its bootstrap loop until a request selects a lifecycle. Malformed or unsupported requests get an error response and the loop continues. Discover probes also keep the loop open. This lets clients recover, for example after a bare tools/list without _meta. But a client can also keep the connection in the pre-lifecycle state forever by sending requests that never select a lifecycle.

Open question: should rmcp cap the number of pre-lifecycle attempts (or apply a timeout)? Or should the transport that owns the connection handle this?

Options:

• Count rejected/bootstrap-neutral requests and return ServerInitializeError after N attempts.
• Add a configurable bootstrap timeout.
• Leave this to transports and document that expectation.