v0.18.0

Added

• mlx-serve is a supported backend (ADR-0031), adopted on the task bench and
  a server measurement with the same Qwen 3.6 weights under mlx-serve and LM
  Studio: the speed comes from the MTP head of mlx-serve's own model builds, not
  the runtime; a 122k-token prompt completes. The ADR records the residual risk
  of running it — chiefly that 26.10.1 listens on every interface with no key
  by default — and that its default 2 GB prefix cache re-reads long sessions
  every turn: with an 8 GB cap a 122k resend took 1.1 s instead of 207 s.
• bench serve measures a model server: cold prompt reading per size,
  reuse on a resend and the next turn, decode speed, alternating conversations
  and two streams at once, logged raw; bench serve-report recounts a run.
• [llm].provider = "mlxserve" for
  mlx-serve. The context window is
  read from its /v1/models at startup, so leave [model].context_window at
  0 for it — a set window skips the lookup. A [llm].model that list does not
  carry is reported at startup: mlx-serve answers a chat request for any model
  name with the model it has loaded, so a mistyped id otherwise runs silently
  against whatever is resident.

Fixed

• A completion of only blank lines is asked again, as an empty one is
  (ADR-0007, amended by ADR-0031). Under LM Studio, Qwen 3.6 starts its answer
  with the blank lines after its thinking; one bench run's entire final answer
  was \n\n, which was returned as the answer.