Change Log
==============

0.2.0 (unreleased)
----------------------------
Packaging and compatibility
- Set the supported runtime to Python >=3.10,<3.15 (Python 3.10–3.14), and
  raised dependency floors to NumPy 1.24, pandas 1.5, SciPy 1.10, PyYAML 6,
  and requests 2.28. Other runtime floors remain typing-extensions 4 and
  colour-science 0.4.6.
- Replaced the unversioned Git dependency with
  `radphen-iadpython>=0.7.0,<0.8`, retaining the `iadpython` import
  package and `iadp` command. Upstream and fork installations are mutually
  exclusive; empylib rejects missing fork metadata and conflicting installs. This fork must be published before
  empylib 0.2.0.
- Tabulated phase-function integration now reports an actionable error when
  the required fork backend is unavailable; it no longer silently falls back
  to the less accurate g-only path.
- Updated SciPy Simpson integrations to pass sample coordinates with `x=`,
  restoring compatibility with SciPy versions where `x` is keyword-only.
- Removed the ten placeholder .pyi stub files so VS Code/Pylance can use the
  Python implementations for function signatures, parameter suggestions and
  docstrings, without stale stubs hiding public functions.
- Removed .pyi package-data patterns from setup.py and MANIFEST.in.
- Retained py.typed, explicit __all__ exports and private dependency aliases.
  Runtime behavior is unchanged; editor suggestions now follow the source API.
- Verified the source tree, built wheel, and rebuilt sdist, including
  installation and import checks. Regression tests are tracked in the
  repository and run in CI across supported Python versions.

0.1.0 (02/08/2026)
----------------------------
First organized release of the substantially revised empylib package.

Accuracy
- cross_section_ensemble/adm_sphere: csca and gcos under dependent_scatt=True are
  now integrated on a Gauss-Legendre grid whose node count scales linearly with
  the size parameter, instead of Simpson's rule on 101 uniform theta points.
  The previous scheme lost accuracy as particles grew -- measured csca error
  against a converged reference was 4.4e-3 at x~47, 2.44 at x~149 and 12.8 at
  x~361 (i.e. >1000% for micron-scale particles in the visible); the asymmetry
  parameter g was off by 0.80 absolute at x~361. It is now ~6e-5 across that
  whole range. Results for large particles with dependent scattering will change
  substantially, toward the correct values.
- _phase_function_single: the per-size-bin qsca renormalization now accepts
  quadrature weights (mu_weights). Simpson's rule was being applied to
  Gauss-Legendre abscissas, a ~10% error that, for a polydisperse ensemble,
  reweighted the size bins against one another.

Performance
- Riccati-Bessel rho = xi/psi is obtained from a recurrence over the Dn/Gn
  arrays instead of scipy's complex-argument Bessel backend, with an exact
  fallback only where the seed is ill-conditioned (real x near k*pi). ~30x
  faster on that step; agreement with the Bessel reference is 1e-13..1e-9.
- psi_n(y) is no longer computed for lossless hosts, where its only consumer
  (the Johnson absorbing-host Qext/Qsca formulas) is discarded anyway.
- adm_sphere now makes a single pass over the ensemble phase function instead
  of two on two different angular grids, halving the per-call evaluations of
  the phase function and the Percus-Yevick structure factor. Net effect on a
  representative UV-vis fit (6 thicknesses x 100 wavelengths x 16 size bins):
  4.03 s -> 2.49 s per residual.

Internal
- New miescattering._ensemble_optics owns the Mie coefficient loop and the one
  angular grid; cross_section_ensemble, phase_function_moments and
  phase_scatt_ensemble are thin wrappers over it.
- scatter_efficiency and _scatter_efficiency_state share one implementation;
  _phase_function_single has a single code path (via the coefficient state),
  removing an overflow-prone fallback that raised FloatingPointError for large
  complex host size parameters.
- utils._estimate_theta_npts is deprecated (superseded by
  miescattering._quadrature_order) and no longer used internally.

API
- Hardened module API visibility by adding explicit __all__ exports across empylib modules.
- Standardized internal dependency imports to private-style aliases (leading underscore) to reduce autocomplete leakage of third-party/internal symbols.
- Kept only intended public functions/constants visible in module-level star imports.
- Added .pyi stub files so VS Code/Pylance autocomplete surfaces only the supported public API and hides private module helpers.
- Added NumPy 2.5 compatibility for trapezoidal integration and scalar wave-optics results.

0.0.2 (19/05/2022)
----------------------------
- Previous PyPI release.

0.0.1 (17/05/2022)
----------------------------
- First release.
