Dev

.github/workflows/compatibility.yml

@@ -21,6 +21,7 @@ on:
 
   # Also run on PRs that modify dependencies
   pull_request:
+    branches-ignore: [dev]
     paths:
       - 'pyproject.toml'
       - 'tox.ini'

.github/workflows/dev-pr.yml

@@ -0,0 +1,47 @@
+name: Dev PR Tests
+
+on:
+  pull_request:
+    branches: [dev]
+
+permissions:
+  contents: read
+
+concurrency:
+  group: dev-pr-tests-${{ github.event.pull_request.number }}
+  cancel-in-progress: true
+
+jobs:
+  cpu:
+    name: CPU (Python 3.12)
+    runs-on: ubuntu-latest
+    timeout-minutes: 90
+    env:
+      TORCHREF_DEVICE: cpu
+      NUMBA_CACHE_DIR: /tmp/numba_cache
+
+    steps:
+      - uses: actions/checkout@v7
+
+      - name: Set up Python
+        uses: actions/setup-python@v7
+        with:
+          python-version: '3.12'
+          cache: pip
+          cache-dependency-path: pyproject.toml
+
+      - name: Install dependencies
+        run: |
+          python -m pip install --upgrade pip
+          python -m pip install torch --index-url https://download.pytorch.org/whl/cpu
+          python -m pip install -e ".[dev]"
+
+      - name: Run tests on CPU
+        run: |
+          python -m pytest tests/ \
+            -m "not gpu and not slow" \
+            -v --tb=short -rf --durations=20
+
+  mps:
+    name: MPS
+    uses: ./.github/workflows/accelerator.yml

.gitignore

@@ -18,6 +18,8 @@ _temp.mtz
 __pycache__/
 *.pyc
 *.pyo
+# macOS finder metadata
+.DS_Store
 # Large binary files
 *.ccp4
 *.png
@@ -84,4 +86,19 @@ graphify-out/
 # Large binary scratch dirs & squashfs images
 *.sqsh
 anisotropic/
-torchref_refine_optimization/runs/
+torchref_refine_optimization/runs/
+
+# Alignment lab: keep lab/, diagnostics/, analysis/ and tests/ tracked;
+# ignore run outputs (CSVs, Phaser working dirs) and scheduler logs.
+alignment_lab/runs/
+alignment_lab/slurm/
+alignment_lab/**/__pycache__/
+
+# Third-party Phaser source copy (from PHENIX 1.20-4459) for FRF map instrumentation
+phaser_src/
+
+# scaler_investigation lab outputs: the scripts are tracked, the outputs are not.
+scaler_investigation/cache/
+scaler_investigation/metrics/
+scaler_investigation/figures/
+scaler_investigation/slurm/

AGENTS.md

@@ -186,13 +186,13 @@ Black, 88 columns, `isort` with the black profile. Ruff lint with
 |---|---|
 | `base/` | Low-level math and crystallography. `coordinates/` (Cartesian↔fractional), `reciprocal/` (basis, HKL, d-spacing, interpolation, symmetry), `direct_summation/` (F_calc by summation; eager + Triton), `electron_density/` (real-space splatting with CPU/CUDA/MPS kernels, solvent mask, radius policy), `fourier/` (FFT and grids), `scattering/` (form-factor and anomalous tables), `metrics/` (R-factors, binwise scale, loss), `targets/` (the *kernels* behind refinement targets, eager + `triton/`), `french_wilson.py`, `math_torch.py`, `alignment/` |
 | `io/` | `ReflectionData`, `DatasetCollection`, `FcalcDataset`; MTZ / PDB / CIF / IHM readers and writers; `read_mtz` / `read_pdb` / `read_cif` |
-| `model/` | `Model` (refinable atomic parameters), `ModelFT` (adds F_calc via `SfFFT` or `SfDS`), `MixedModel`, `ModelCollection`, and the parametrizations in `parameter_wrappers.py` / `rigid_xyz.py` that decide what is refinable |
+| `model/` | `Model` (refinable atomic parameters), `ModelContext` (the cell, space group, atom table, links and provenance a model is loaded with — `model.cell` / `.spacegroup` / `.pdb` forward to it, the rest is `model.ctx.*`), `ModelFT` (adds F_calc via `SfFFT` or `SfDS`), `MixedModel`, `ModelCollection`, and the parametrizations in `parameter_wrappers.py` / `rigid_xyz.py` that decide what is refinable |
 | `refinement/` | Drivers (`Refinement`, `LBFGSRefinement`, `RigidBodyRefinementStep`), `targets/` (`xray/`, `geometry/`, `adp/`, `collection/`, `combined.py`), `weighting/`, `optimizers/` (annealing, Langevin, preconditioned/seeded L-BFGS), `model_error_estimation/` (σ_A, σ_M), `loss_state.py`, `logger.py` |
 | `restraints/` | Bonds, angles, torsions, planes, chirals, VDW. Built from the CCP4 Monomer Library, resolved lazily via `get_library_manager()` — importing this package must not trigger a library download |
 | `scaling/` | `ScalerBase` (model-independent), `Scaler`, `CollectionScaler`, `SolventModel` (k_sol, B_sol) |
-| `symmetry/` | `SpaceGroup` (buffers on `nn.Module`), `Cell`, `MapSymmetry`, `ReciprocalSymmetry`, grid utilities |
+| `symmetry/` | `Symmetry` (operations plus everything derived from them), `SpaceGroup` (adds the crystallographic identity and the CCP4 ASU verbs), `Cell`. All dataclasses over `DeviceMixin`, not `nn.Module` — they hold no refinable parameters. Map and reciprocal-grid operators are private, reached through `Symmetry` |
 | `maps/` | `Map` (2Fo−Fc, Fcalc), `DifferenceMap` |
-| `cli/` | Entry points: `torchref.refine`, `torchref.difference-refine`, `torchref.mtz2map`, `torchref.validate-ded`, `torchref.phased-difference-map`, `torchref.add-metadata`, `torchref.strip-altlocs` |
+| `cli/` | Entry points: `torchref.refine`, `torchref.difference-refine`, `torchref.mtz2map`, `torchref.validate-ded`, `torchref.difference-map`, `torchref.add-metadata`, `torchref.strip-altlocs` |
 | `experimental/` | APIs that may change without notice: `alignment/` (Patterson MR), `kinetic/` (time-resolved), `ensemble/`, `monolithic_refinement/`, `targets/` (AMBER/GAFF2, real-space, sampled-ML phase) |
 | `utils/` | See §5 |
 | `config.py` | See §4 |