Experiment and simulation configuration schema

SimulationConfiguration is the versioned JSON schema for reproducible model compilation. Version 1 contains:

• true_experiment: concrete Optics, an Illumination variant, and optional optical background;
• reconstruction_experiment: the separately assumed experiment description;
• image_shape and reconstruction_shape, always (height, width);
• compiled_models: both derived ImagePlaneModel values;
• optional concrete CameraModel and IlluminationAcquisitionErrors;
• deterministic random_seed.

All physical distances are metres, wave vectors are radians/metre, array data is row-major, and source-weight values are intensity weights. Illumination uses Serde's externally tagged representation. For example:

{
  "illumination": {
    "LEDArray": {
      "grid_shape": [3, 3],
      "pitch": 0.004,
      "distance": 0.09,
      "center": [1.0, 1.0],
      "wavelength_override": null,
      "illumination_order": null,
      "intensity_weights": null,
      "rotation_radians": 0.0
    }
  }
}

The additional externally tagged variants are LEDSphere, SphericalLEDArm, and RotatingLEDArc. Their coordinate conventions, fields, equations, and identifiability constraints are documented in Spherical illumination geometries.

Serialized PupilAberration values are direct coefficients used by Pupil::circular: astigmatism, coma, and spherical terms are radian weights for the sampled radial polynomials documented on Optics, and edge_apodization sets the radial amplitude decay. They are not normalized Zernike coefficients.

Use ExperimentDescription::compile for one model or SimulationConfiguration::new for a validated true/reconstruction pair. Use save and load instead of calling serde_json directly: these methods enforce the format version and validate that serialized compiled geometry, pupil, sampling, gains, background, and multiplexing still agree with their concrete descriptions.

Both constructors take ReconstructionShape. Exact((height, width)) validates a prescribed grid, while Minimum, Smooth, and PowerOfTwo resolve a concrete aspect-preserving grid from the compiled illumination geometry. For paired simulation configurations, automatic sizing covers the union of the true and assumed geometries. Only the resolved tuple is serialized; the selection variant does not alter the versioned schema.

Known uniform camera response is deliberately not baked into compiled_models.reconstruction_model. Call reconstruction_model_for_counts() when constructing a problem directly from detector counts loaded from a serialized configuration. In contrast, Simulator::simulate returns a SimulationResult.reconstruction_model that has already been adjusted for the known linear camera response, so simulated detector counts can be used directly.