v0.2.5

• Add decide.policies with four alternatives to thresholding the top
  probability: CalibratedGate rescales by a fitted temperature,
  EntropyGate reads the whole distribution, MarginGate reads the gap
  between first and second place, and ConformalGate accepts only when one
  candidate survives a fitted cut. Each is a drop-in for Gate, which is
  unchanged, and all of them leave Score answers ungated.
• Add decide.calibration with fit_temperature and fit_conformal_quantile,
  which set those parameters from answers whose correct label is known, plus a
  Calibration record that saves and loads as JSON. Pure Python, so the core
  install gains no dependency.
• The llm backend now sends max_tokens, defaulting to 512. Without a cap, a
  model that loses the JSON format generates until its context runs out, and a
  hosted API bills for all of it. Pass None for the old behaviour.
• Document running the typesafe backend against other servers that speak
  POST /v1/systemone, with a table of the ones actually tested.
• Split the reference sections of the README into docs/.