v0.9.0

Added

• Public in-process transport surface on MCPServer — two methods that let
  an in-process consumer (a bridge, a Pydantic-AI toolset, a management command)
  run tools exactly as a remote MCP client would, without the HTTP / JSON-RPC hop
  and without reaching into handler internals:

  • MCPServer.list_tools(cursor=None, *, user, request=None) returns one
    tools/list page with the same merged inputSchema (serializer fields plus a
    selector tool's filter / ordering / pagination arguments and the
    additionalProperties policy), the same FILTER_LISTINGS_BY_PERMISSIONS
    per-caller filter, and the same opaque-cursor pagination as the wire.
  • MCPServer.acall_tool(name, arguments=None, *, user, request=None) (async)
    invokes a tool with the full transport applied — the transport-level MCP
    permissions and rate limits, the selector post-fetch pipeline (filter / order
    / paginate), a selector binding's MCP-only input_serializer, chain tools, and
    the output format — everything the spec-core call_tool deliberately omits.
    Returns the wire's dict payload (content / structuredContent / isError)
    or a JsonRpcError for a protocol fault.

  Both build the call context internally from user + request (synthesising a
  minimal request when request is None).

• JsonRpcError and JsonRpcErrorCode re-exported from the package root, so a
  consumer of acall_tool / list_tools can branch on protocol faults without a
  leaf-module import.