Fix calibration bugs and clean up solvers, UDFs, UI and packaging

.github/workflows/ci.yml

@@ -37,14 +37,13 @@ jobs:
           CODECOV_TOKEN: ${{ secrets.CODECOV_TOKEN }}
 
   pinned-requirements:
-    if: false  # skip job
     runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@v7
-      - name: Choose Python version ${{ matrix.python_version }}
+      - name: 'Choose Python version 3.14'
         uses: actions/setup-python@v7
         with:
-          python-version: '3.11'
+          python-version: '3.14'
       - name: install the package
         run: python3 -m pip install uv && uv pip install --system .[test,common,diffraction,gui] -r packaging/requirements.txt
       - name: run pytest

CLAUDE.md

@@ -60,7 +60,7 @@ third-party dependency you introduce.
 The whole library is organized around one central object, **`Model4DSTEM`**
 (`common/model.py`), a `jax_dataclasses.pytree_dataclass` holding the full geometric
 calibration of a 4D-STEM experiment (overfocus, scan pixel pitch, scan center/rotation,
-camera length, detector pixel pitch/center/rotation, semi-convergence angle, flip factor, and
+camera length, detector pixel pitch/center/rotation, semi-convergence slope, flip factor, and
 a 12-parameter `DescanError`). Because it is a JAX pytree, it is both differentiable and
 traceable.
 
@@ -88,8 +88,8 @@ Key layers, bottom-up:
 4. **`udf/stem_overfocus.py`** — LiberTEM UDFs that apply (3) over a real dataset:
    `OverfocusUDF` (reconstruct an overfocus/real-space image) and `CorrectedPickUDF`. Models
    are passed wrapped in a `dict` so they can be mutated in place (LiberTEM issue #1780); the
-   transformation matrices are cached via `lru_cache` and only recomputed when the model
-   changes.
+   transformation matrices are calculated in `get_task_data()` and cached per model via
+   `lru_cache`.
 
 5. **`util/stem_overfocus_sim.py`** — forward simulator (e.g. `smiley()` test object,
    `project()`) that generates synthetic 4D-STEM data from a model, used heavily in tests and
@@ -104,9 +104,10 @@ Key layers, bottom-up:
    structure (uses the `diffraction` extra); feeds known-geometry references into the
    optimizers.
 
-8. **`ui.py`** — interactive Panel/Bokeh calibration GUI (`gui` extra). `CalibratedDataset`
-   pairs a dataset with a model; `CoordinateCorrectionLayout` wires sliders/tables to live
-   re-projection so a user can tune calibration against a displayed dataset.
+8. **`ui.py`** — interactive Panel/Bokeh calibration GUI (`gui` extra).
+   `CoordinateCorrectionLayout` wires sliders/tables to live re-projection so a user can tune
+   calibration against a displayed dataset. `CalibratedDataset` (`dataset.py`, also importable
+   from `ui`) pairs a dataset with a model.
 
 ### Mental model for making changes
 
@@ -115,6 +116,11 @@ Key layers, bottom-up:
   you see `try/except TracerBoolConversionError` and `_one`/`flip_factor` style tricks
   (constants threaded through so operations stay differentiable instead of using Python
   conditionals). Preserve that pattern.
+- Model functions are always SymPy-compatible and use SymPy math (`sym.cos`, `sym.pi`, ...);
+  called directly on floats they may return SymPy numbers. For NumPy or JAX, wrap the call
+  site with `@lambdify(modules=np|jnp, recurse_for=...)` (or use `lambdify_trace_for`). Create
+  the lambdified function once, e.g. at module level: each new function object triggers a new
+  symbolic derivation, so don't decorate inside loops or frequently called functions.
 - `trace()` is the single source of truth. Derive relations from it symbolically instead of
   hard-coding them: register the derivation in a `GENERATED_FUNCTIONS` registry (collected in
   `common/generate.py`) and run the generator. Never edit `common/_generated.py` by hand.

examples/stem_overfocus.ipynb

@@ -1997,7 +1997,7 @@
    "outputs": [],
    "source": [
     "# We assume it is constant over the series\n",
-    "acceleration_voltage = jeol_dm4.acceleration_from_dm4(descan.dm4)"
+    "acceleration_voltage = jeol_dm4.acceleration_from_dm4(descan.tags)"
    ]
   },
   {

packaging/overfocus.def

@@ -1,5 +1,5 @@
 bootstrap: docker
-from: python:3.11-slim
+from: python:3.14-slim
 
 %files
         ./Microscope-Calibration        /Microscope-Calibration