V1.2.0 Major Release

PortPy 1.2.0

Highlights

Beamlet dose prediction (portpy.ai) — a full pipeline for predicting individual
beamlet dose columns with a 3D deep-learning model, and assembling them into a predicted
influence matrix A_pred. This lets you build a plannable influence matrix without a
dose engine. Includes preprocessing, dataset, training/testing, and matrix assembly, with
an end-to-end example (examples/beamlet_dose_prediction.ipynb).

New portpy.ai.inference subpackage — the inference half of the above, usable on its
own: beam-frame geometry (gantry/collimator/couch), ray-aligned BEV patch extraction,
GPU resampling via torch.grid_sample, and two predictors (RayUNetPredictor,
PatientGridPredictor) behind build_ray_unet_predictor / build_patient_grid_predictor.
Importing the package pulls no torch; torch loads lazily inside the predictors.

Two sampling grids — beamlet dose can be predicted on a world-aligned patient grid or
a ray-aligned BEV patch, selectable with grid="patient" | "bev" through preprocessing,
training and inference.

Plan quality scorecards — Evaluation.compute_total_quality_score() scores a plan
against a scorecard and returns a per-criterion table. Two scorecards ship with the
package (Lung 60Gy and Prostate SBRT), plus adapt_scorecard_prescription() to
rescale a scorecard written for a different prescription.

New models — UNet3D alongside an expanded networks3d
registry (beamlet_unet, ray_unet3d, ray_attention_resunet3d, mednext).

New protocol — Prostate_5Gy_5Fx, with clinical criteria and IMRT/VMAT optimization
parameters.

Bug fixes

• Influence matrix: crash on some of the shipped patients. preprocess_beams raised
  IndexError while slicing MLC_leaf_idx for patients whose stored leaf table is
  inconsistent with the beamlet grid. The field is no longer needed: the VMAT RT-plan
  writer now derives leaf pairs per row from leaf_pair_row_match, which is also correct
  for rows spanning different numbers of leaves.
• Leaf sequencing worked for only some patients. leaf_sequencing_siochi matched
  beamlet rows against the stored leaf table with an exact int() comparison, which only
  succeeded when that table happened to be whole-millimetre. It now resolves leaf rows
  from the MLC leaf edges and works for every patient.
• Influence matrix assembly was quadratic in beam count. get_influence_matrix grew
  the matrix one beam at a time; it now concatenates once. Output is bit-identical.
  Largest measured effect: a 191-beam breast case went from 306 s to 5.4 s.
• Windows: import portpy.photon broke a later import torch. The VMAT SCP utilities
  loaded the Intel MKL runtime eagerly; that import is now deferred until first use.
• Clinical criteria: negative or unresolvable dose limits and goals are dropped instead
  of raising, and "coldest" (CV/DC) dose-volume constraints are supported.

Installation fixes

pip install -r requirements.txt previously failed because the file was UTF-16 encoded.
It is now UTF-8 and mirrors the core dependencies in pyproject.toml.

• Removed ecos and qdldl — C extensions that required a compiler when no wheel matched
  the user's Python. They are unused by PortPy.
• Added clarabel to core dependencies, so a fast free solver is available by default.
• Relaxed the exact jinja2==3.1.5 pin, which conflicted with other packages.
• requirements.txt no longer forces the commercial Mosek solver on source installs;
  it stays an optional extra (pip install portpy[mosek]).

Testing

New tests/test_ai_module.py (26 tests) covering beam geometry conventions, the 3D UNet
forward pass, and GPU resampling accuracy. CI installs CPU-only torch so these run on
every push.

Full changelog: v1.1.4...v1.2.0