test(e2e): prove MCP tool-catalog sync end-to-end

Four cross-target scenarios (cloud + selfhost) driving only public surfaces
(typed API + sandbox executions) against real MCP servers whose catalogs
mutate mid-scenario:

- list_changed: a tool call renames the server's catalog mid-call; the
  notification rides the open connection and the very next tools read
  re-lists — no refresh click, no failing retry.
- unknown-tool self-heal: the catalog mutates outside any call window (no
  notification available); invoking the retired tool fails with the typed
  mcp_tool_unknown error telling the agent to re-list, and that failure
  alone heals the catalog for the next read.
- pagination: a server paging tools/list via nextCursor registers its whole
  catalog, not just the first page.
- outage resilience: a refresh against a dead server keeps the working
  catalog instead of wiping it, and converges to the server's new catalog
  once it recovers.

The mutable-catalog test fixture now also sends list_changed through the
tool handler's request-scoped notifier: the SDK's RegisteredTool.update
emits it with no relatedRequestId, which the streamable-http transport
routes to the standalone GET SSE stream — a stream a request-scoped client
may never hold open. Stamping the request id routes it onto the in-flight
call's response stream, matching how a real server notifies mid-call.

Author: RhysSullivan