Polarized channels

.github/workflows/pypi-test.yml

@@ -24,6 +24,8 @@ env:
 jobs:
   # Job 1: Test installation from PyPI on multiple OS
   pypi-package-tests:
+    # Only run the package tests on the master branch
+    if: github.ref == 'refs/heads/master'
     strategy:
       matrix:
         os: [ubuntu-latest, windows-latest, macos-latest]
@@ -45,18 +47,10 @@ jobs:
       - name: Init pixi project
         run: pixi init easyreflectometry
 
-      - name: Set the minimum system requirements
-        working-directory: easyreflectometry
-        run: pixi project system-requirements add macos 14.0
-
       - name: Add Python 3.13 from Conda
         working-directory: easyreflectometry
         run: pixi add "python=3.13"
 
-      - name: Add other Conda dependencies
-        working-directory: easyreflectometry
-        run: pixi add gsl
-
       - name: Add easyreflectometry (with dev dependencies) from PyPI
         working-directory: easyreflectometry
         run: pixi add --pypi "easyreflectometry[dev]"

.github/workflows/python-ci.yml

@@ -13,17 +13,6 @@ name: CI using pip
 on: [push, pull_request]
 
 jobs:
-  Code_Consistency:
-    runs-on: ubuntu-latest
-    steps:
-      - uses: actions/checkout@v4
-      - uses: chartboost/ruff-action@v1
-      - name: Suggestion to fix issues
-        if: ${{ failure() }}
-        run: |
-          echo "::notice::In project root run 'python.exe -m ruff . --fix' and commit changes to fix issues."
-          exit 1
-
   Code_Testing:
     strategy:
       max-parallel: 4

.github/workflows/test.yml

@@ -174,6 +174,8 @@ jobs:
   # Job 3: Test the package
   package-test:
     needs: source-test # depend on previous job
+    # Only run the package tests on the master branch
+    if: github.ref == 'refs/heads/master'
 
     strategy:
       fail-fast: false
@@ -217,9 +219,6 @@ jobs:
             echo "Adding Python $py_ver"
             pixi add "python=$py_ver"
 
-            echo "Setting macOS 14.0 as minimum required"
-            pixi project system-requirements add macos 14.0
-
             echo "Looking for wheel in ../dist/py$py_ver/"
             ls -l "../dist/py$py_ver/"
 

CHANGELOG.md

@@ -1,3 +1,176 @@
+# Unreleased
+
+All four polarization channels (pp, pm, mp, mm) are now available from
+the refl1d calculator. Previously only the non-spin-flip pp channel was
+returned.
+
+- New `LayerMagnetism` sample element. `Layer` takes an optional
+  `magnetism` with fittable, serialized `Parameter`s `rho_m` (magnetic
+  SLD) and `theta_m` (in-plane moment angle). Adding a magnetic layer
+  turns on `include_magnetism` on the calculator, or raises
+  `NotImplementedError` if the backend cannot do magnetism. Removing the
+  last magnetic layer turns it off again. `Model.has_magnetism`,
+  `CalculatorBase.supports_magnetism` and
+  `Project.calculator_supports_magnetism` report the current state.
+- New `PolarizedDataSet` groups per-channel `DataSet1D` objects (one
+  file per channel; NSF experiments use 'pp'/'mm' only, spin-flip
+  channels are optional) into one experiment that shares a single model.
+  `Project.load_polarized_experiment(paths)` loads from an explicit
+  channel-to-file mapping.
+  `Project.suggest_polarized_channel_assignment(paths)` fills that
+  mapping from the ORSO header polarization (`pp`/`mm`/`pm`/`mp` only).
+  Partially analysed observables such as `po`/`mo` (channel sums) and
+  `op`/`om`/`unpolarized` are left for the user. For plain text files
+  the mapping comes from filename tokens (`_uu`/`_up`/`_pp` → pp,
+  `_dd`/`_down`/`_mm` → mm, `_ud`/`_pm` → pm, `_du`/`_mp` → mp).
+- Experiment and model accessors are channel-aware.
+  `Project.experimental_data_for_model_at_index(index, channel=...)`
+  returns the `DataSet1D` of one spin channel. `channel=None` (the
+  default) still returns the stored experiment.
+  `Project.model_data_for_model_at_index(index, q_range, channel=...)`
+  calculates one spin cross-section.
+  `Project.experiment_is_polarized_at_index(index)` and
+  `Project.experiment_channels_at_index(index)` report the polarization
+  state. A channel that was not measured raises `KeyError`. An unknown
+  channel, or any channel on an unpolarized experiment, raises
+  `ValueError`.
+- Summary/report figures now plot one measured series per spin channel
+  of a polarized experiment, each in its channel colour, plus the
+  matching calculated cross-section. Channels that cannot be calculated
+  (for example spin-flip on a non-magnetic model) are shown without a
+  calculated overlay. Previously a polarized experiment made the report
+  figures fail on `PolarizedDataSet.x`.
+- The summary experiments table lists one row per spin channel of a
+  polarized experiment, named `<experiment> (<channel>)`. It previously
+  raised
+  `AttributeError: 'PolarizedDataSet' object has no attribute 'x'` and
+  crashed anything that read the summary while a polarized experiment
+  was loaded.
+- New `Project.calculators_supporting_magnetism` lists the available
+  calculators that can model magnetic samples, without switching the
+  active one. `Project.models_have_magnetism` reports whether any model
+  has a magnetic layer. Use these to pick a suitable engine, or to
+  refuse one that cannot carry the sample's magnetism, instead of
+  hitting an error inside the binding.
+- New `Project.magnetic_sld_data_for_model_at_index(index)` returns the
+  depth profiles of a magnetic model as `DataSet1D`s keyed `'sld'`,
+  `'rho_m'`, `'theta_m'`, `'spin_up'` and `'spin_down'`. The last two
+  are the potentials each spin state sees, rho +/- rho_m\*cos(theta_m -
+  A). The guide-field angle A is the new module constant
+  `GUIDE_FIELD_ANGLE` (270 degrees, refl1d's default and the only value
+  the library can currently model). A non-magnetic model raises
+  `ValueError`. `Project.model_has_magnetism_at_index(index)` reports
+  whether the model is magnetic.
+- The magnetic depth profile is now built by smoothing the two in-plane
+  components of the moment and converting back, rather than smoothing
+  magnitude and angle separately as refl1d does channel by channel. At
+  an interface where moments differ by a few degrees across 0/360, the
+  smoothed _angle_ used to take the long way around the circle, pass
+  through the guide-field direction, and report the full moment as
+  longitudinal. That produced a spurious spin-up/spin-down splitting
+  exactly at the interface (a 2-degree difference gave the full 2*rho_m
+  splitting; it is now the correct ~0.02*rho_m). Collinear samples are
+  unaffected. The reported `theta_m` profile is restricted to depths
+  that carry a moment (the angle of a zero-length vector is arbitrary)
+  and is made continuous within each magnetic region. A profile turning
+  from 359 to 1 degree is a 2 degree turn; the wrapped values would plot
+  as a full sweep. If the installed refl1d does not expose the microslab
+  data the component-safe profile needs, the calculator now raises
+  `NotImplementedError` instead of falling back to the angle-smoothed
+  profile.
+- New `Project.spin_asymmetry_for_experiment_at_index(index)` returns
+  the measured spin asymmetry (R++ - R--)/(R++ + R--) of a polarized
+  experiment, the matching model curve when the model is magnetic, and
+  the number of points dropped. `ye` holds the SA **variance**, as
+  everywhere else in the library. Channels measured on different q grids
+  are interpolated onto the pp grid (values with the linear weights,
+  variances with their squares) only inside the q range both channels
+  cover. Outside that range `np.interp` would clamp to the edge value.
+  Dropped points are reported as `out_of_overlap_points`. Points where
+  R++ + R-- is not above `SPIN_ASYMMETRY_SIGNIFICANCE` (3) times its own
+  uncertainty are also dropped. A second, uncertainty-independent guard
+  drops points whose denominator is non-positive or smaller than
+  `SPIN_ASYMMETRY_CANCELLATION_FRACTION` (1e-3) of |R++| + |R--|.
+  Without it, a file with no uncertainties (two columns, or a malformed
+  uncertainty array) had no guard, and background-subtracted data could
+  put values of +/-1e3 on the axis. Points with a non-finite
+  reflectivity or a negative/non-finite variance are dropped rather than
+  treated as having no uncertainty. Dropped points are reported by
+  reason (`low_significance_points`, `small_denominator_points`,
+  `invalid_points`).
+- Both channels of a spin asymmetry are validated before use. Empty,
+  length-mismatched, non-finite or duplicated q grids are rejected, and
+  `experiment_supports_spin_asymmetry_at_index` reports False for them.
+  A descending grid is sorted before pairing; `np.interp` silently
+  returns nonsense for one.
+  `Project.experiment_supports_spin_asymmetry_at_index(index)` reports
+  whether both non-spin-flip channels were measured.
+- New `calculate_channel(q, model, channel)` on the wrapper (and
+  `reflectivity_profile_channel` on the calculator,
+  `fit_func_for_channel` on `CalculatorFactory`) evaluates one explicit
+  spin channel without touching the global `polarization_channel` state.
+- New `MultiFitter.for_experiments(experiments)` builds a fitter with
+  one fit function per dataset (one per measured spin channel for a
+  polarized experiment, one for an ordinary one) across any number of
+  experiments and models, and returns without running the fit.
+  `fit_datasets` and `fit_channels` give the flat dataset list in
+  fit-function order, so an application can prepare the data arrays and
+  drive `easy_science_multi_fitter.fit(...)` from a worker thread.
+- New `MultiFitter.record_fit_results(results)` adopts results from such
+  a caller-driven fit, so `chi2` and `reduced_chi` describe it instead
+  of reporting that no fit was performed. The classical metrics need the
+  original data arrays and stay None.
+- `rho_m` now takes part in the project's default-limit policy. It is
+  created with `default_limits_pending`, and
+  `Project._sync_parameter_states` gives it the shared SLD window (-1
+  to 10) unless an explicit `Parameter` with its own bounds was passed.
+  `theta_m` keeps its explicit 0-360 bounds. Previously both stayed
+  unbounded.
+- New `MultiFitter.fit_polarized(data)` fits all measured channels of a
+  `PolarizedDataSet` simultaneously against the shared model: one fit
+  function per channel, common structural parameters, magnetic
+  parameters constrained by all channels at once. Returns per-channel
+  `FitResults`.
+- The refl1d wrapper now caches the four polarized cross-sections per
+  model state and (q, dq) grid. They come from a single kernel
+  evaluation, so a simultaneous N-channel fit costs about one evaluation
+  per iteration instead of N.
+
+- New `polarized_reflectivity_profiles(x_array, model_id)` on the
+  calculator (and on `CalculatorFactory`) returns the reflectivity of
+  all four spin channels in one calculation as a dictionary keyed
+  `'pp'`, `'pm'`, `'mp'`, `'mm'` (in that order). Requires
+  `include_magnetism = True`.
+- New `polarization_channel` property (accepts
+  `'pp'`/`'pm'`/`'mp'`/`'mm'` or the new `PolarizationChannel` enum)
+  selects which channel `reflectity_profile` (and therefore fitting)
+  returns, so fits can target spin-flip or mm data. Default `'pp'`;
+  disabling magnetism resets it to `'pp'`. The channel belongs to the
+  currently active calculator instance, not to a model or dataset. It
+  affects every subsequent calculation with that calculator.
+  `interface.switch(...)` constructs a fresh calculator and resets both
+  this and `include_magnetism`.
+- New `magnetic_sld_profile(model_id)` on the calculator (and on
+  `CalculatorFactory`) returns the nuclear and magnetic scattering
+  length density profiles as a tuple `z`, `sld(z)`, `rhoM(z)` (magnetic
+  SLD) and `thetaM(z)` (magnetic angle). Requires
+  `include_magnetism = True`; refl1d only.
+- Magnetic calculations now always build all four refl1d cross-sections,
+  so they may take somewhat longer than before. pp results are
+  unchanged.
+- Bug fix: `include_magnetism = True` on a refnx-backed calculator now
+  raises `NotImplementedError`. Previously it was silently accepted (the
+  guard sat on a property the calculator never called) even though refnx
+  magnetism is not supported.
+- Bug fix (pre-existing): disabling magnetism after layers were created
+  with it enabled used to leave refl1d `Magnetism` objects on the slabs,
+  and a later unpolarized calculation raised `AttributeError` inside
+  refl1d. Disabling magnetism now strips the magnetic state from
+  existing layers. Magnetic parameters (`rhoM`/`thetaM`) are kept in a
+  per-layer store inside the wrapper, so they survive a
+  disable/re-enable cycle and are re-attached when magnetism is enabled
+  again. `update_layer` also accepts the magnetism keys one at a time.
+
 # Version 1.7.0 (1 Aug 2026)
 
 Restored the measured per-point resolution on data load (issue #368).

docs/docs/api-reference/calculators.md

@@ -0,0 +1,29 @@
+# Calculators
+
+The calculator translates an EasyReflectometry model into a backend
+engine (refl1d or refnx) and computes reflectivity.
+
+## Polarized reflectivity
+
+With the refl1d calculator and `include_magnetism` enabled, all four
+spin channels are available:
+
+- `polarized_reflectivity_profiles(x_array, model_id)` returns the
+  reflectivity of all four channels as a dictionary keyed `'pp'`,
+  `'pm'`, `'mp'`, `'mm'` (in that order).
+- `polarization_channel` selects which channel `reflectity_profile` —
+  and hence fitting — uses (default `'pp'`).
+- `magnetic_sld_profile(model_id)` returns the nuclear and magnetic
+  scattering length density profiles as a tuple `z`, `sld(z)`, `rhoM(z)`
+  (magnetic SLD) and `thetaM(z)` (magnetic angle).
+
+Note that `polarization_channel` is state of the currently active
+calculator instance, not of a model or dataset: it affects every
+subsequent calculation using that calculator, and switching calculators
+via the factory constructs a fresh instance, which resets the channel
+(along with `include_magnetism`). Calculators without magnetism support
+(refnx) raise `NotImplementedError` when magnetism is enabled.
+
+::: easyreflectometry.calculators.polarization
+
+::: easyreflectometry.calculators.calculator_base

docs/docs/tutorials/advancedfitting/bayesian_bumps.ipynb

@@ -217,12 +217,18 @@
     "# We deliberately start with a *very* short run.  It finishes in seconds but is\n",
     "# far too short to trust — which is exactly the situation the \"extend the chain\"\n",
     "# section below exists to fix.  In production you would ask for 20 k+ samples.\n",
+    "#\n",
+    "# ``thin=1`` here (rather than the ``10`` used later in ``extend()``) works\n",
+    "# around a BUMPS DREAM bug: its outlier-chain replacement indexes the thinned\n",
+    "# sample buffer using the un-thinned generation counter, which overruns the\n",
+    "# buffer whenever outlier removal fires before enough thinned generations\n",
+    "# have accumulated — reliably the case for a run this short with thin > 1.\n",
     "\n",
     "posterior_dict = fitter.mcmc_sample(\n",
     "    data,\n",
     "    samples=500,  # Deliberately too short — extended later in this notebook\n",
     "    burn=100,\n",
-    "    thin=10,\n",
+    "    thin=1,\n",
     ")\n",
     "\n",
     "print('DREAM sampling complete.')\n",