Behaviour for executing an eval prompt against a target (spec §4).

A runner receives the tool as already rendered through the §0/§2/§3 resolution pipeline for the permutation under test (its description and field descriptions resolved for the requested verbosity/runner/model), the eval's prompt, and a context map, and returns a transcript for the judge to grade.

The context map carries the permutation:

%{
  server: module(),
  tool: String.t(),
  eval: String.t(),
  verbosity: 0..9 | nil,
  runner: atom() | nil,
  model: String.t() | atom() | nil,
  render_ctx: Noizu.MCP.RenderCtx.t()
}

The transcript is opaque to the harness — whatever shape the paired judge understands (a string, a list of messages, a structured map, ...).

Select the active runner via the :noizu_mcp :eval_runner application env; it defaults to Noizu.MCP.Eval.Runner.Stub. Real LLM runners (driving an actual model to call the tool) are app-layer follow-ups.