v3.1

Second release of the 3.x series (mucoll-stack@3.1). It consolidates the layered build to two
images, removes the Marlin/ILCSoft/LCIO reconstruction chain in favour of the native key4hep path,
brings back the GNN tracking pipeline, and pins the stack to tagged key4hep/MuColl releases instead
of tracking main.

Added

• Event generators in the sim image. New +gen variant pulling in whizard +openloops,
  madgraph5amc, and pythia8. The published sim layer is now built as +sim+gen — both the
  mucoll-layered environment and the CI (amd64/arm64).
• k4gaudipandora in the +sim layer, providing the Gaudi ↔ PandoraPFA interface that replaces
  the dropped DDMarlinPandora path. It is required as @0.3.0 ~ddkaltest: ~ddkaltest takes
  the track states already present on the input tracks instead of recomputing them at the
  calorimeter, which avoids the k4reco +conformal_tracking dependency (and with it LCIO).
• ACTS overlay recipe (packages/acts) so the material-mapping workflow of
  k4ActsTracking/doc/material_mapping.md actually works against acts@main: it passes the renamed
  ACTS_BUILD_PYTHON_BINDINGS / ACTS_BUILD_EXAMPLES_ROOT cmake options (the builtin package's
  names are silent no-ops on main), installs the Examples/Scripts/Python helper scripts, and
  exports PYTHONPATH / ACTS_EXAMPLES_SCRIPTS. ACTS is now required as
  +dd4hep+json+edm4hep+examples+geant4+hepmc3+python cxxstd=20.
• GNN tracking pipeline back in the default build. The stack requires
  k4actstracking@00-05 +gnn, which pulls acts +gnn+onnx+torch together with py-torch and
  py-onnxruntime. The pipeline is now selected by a variant on a tagged release rather than by
  a branch pin, and needs no ACORN (GNN4ITk) dependency.
• Concretization CI now exercises the real image bootstrap. concretize-template.yaml builds
  Docker/Dockerfile.base --target concretize instead of running inside the key4hep externals
  image, so it uses exactly the same pinned spack / spack-packages commits, key4hep-spack and local
  .cherry-pick, LLVM-20/rust externals and buildcache mirror that the published images use.
  Nothing is installed, so the check stays cheap while failing on exactly the concretization
  problems a real build would hit; the concretized spec is uploaded as spec-<target>.log.
• Geant4 examples variant via cherry-pick. .cherry-pick now cherry-picks the key4hep-spack
  commit that adds the examples variant to geant4, so the build can turn the examples off
  without vendoring a full local geant4 recipe.

Changed

• Simplified the layered build from three images to two. Collapsed the
  analysis ⊂ sim ⊂ ml chain into analysis ⊂ sim: the machine-learning tools (+ml) are now
  folded into the base analysis layer instead of shipping as a separate mucoll-ml image, and both
  the analysis and sim roots carry +ml. Removed the dedicated +analysis variant — the
  edm4hep/podio analysis stack is now the always-installed base of every layer. Moved hepmc3
  into the +sim layer. The roots are now mucoll-stack~devtools+pytools+ml~sim and
  mucoll-stack~devtools+pytools+ml+sim+gen.
• Pinned the stack to tagged releases rather than tracking upstream main: edm4hep@1.1,
  k4fwcore@1.7, k4geo@0.26, k4gen@0.1pre16, k4gaudipandora@0.3.0, k4actstracking@00-05,
  k4reco@0.3 (new version added to the recipe and made preferred). k4simgeant4 still tracks
  @main.
• Physics-validation plotting moved to the mucoll-benchmarks python studies. The plot step now
  runs analysis/python/edm4hep/study_{tracks,seeds,hits}.py (and study_photons.py for photons)
  instead of the retired TrackingPlots submodule (RunAnalysis.C -> PlotAll.C), which no longer
  exists upstream. The studies fuse ntuple writing and plotting into a single RDataFrame event loop,
  so there is no intermediate ntuple stage. validation/RunAnalysis.conf is replaced by
  validation/plot_settings.sh, whose settings can all be overridden from the environment.
• Pinned numpy (@:2.2, the numba-compatible ceiling), py-sympy@1.13.3, eigen@3.4,
  py-fsspec +http and valgrind ~boost in packages.yaml so the analysis and sim roots stay
  shareable under unify: when_possible — each of these otherwise forks the expensive
  py-torch/onnxruntime subtree across the two roots.
• Disabled Geant4 examples (geant4 ~examples) to trim the build, and dropped +lcio from the
  dd4hep and Whizard specs (now whizard +openloops).
• k4geo is still required as ~beampipe_stl — it skips the network download of the FCC-ee MDI
  beampipe CAD, which MuColl does not use and which intermittently breaks the build. The variant
  itself now comes from upstream, so it no longer needs a local overlay to exist.

Removed

• Removed Marlin, ILCSoft, and LCIO from the stack. Dropped the entire Marlin/ILCSoft
  reconstruction chain (marlin, marlinreco, marlintrk, marlinutil, marlindd4hep,
  marlinfastjet, marlinkinfit*, pandorapfa/pandoraanalysis, gear, kaltest/ddkaltest,
  kitrack*, clicperformance, fcalclusterer, generalbrokenlines, aidatt, raida, sio,
  ced/cedviewer, garlic, lcfiplus/lcfivertex, …) together with k4marlinwrapper and the
  lcio pin from the stack's direct dependencies, along with the now-unused muoncvxddigitiser and
  mybibutils recipes. Reconstruction is now driven entirely by the native key4hep/k4reco path
  (k4geo still pulls upstream lcio. The changes needed to avoid this will land later).
• Dropped the version requirements on k4simdelphes and k4edm4hep2lcioconv.
• Dropped the local k4geo, k4actstracking and k4gaudipandora recipes in favour of the
  upstream Spack packages, which now carry everything the overlays were adding (the beampipe_stl,
  gnn and ddkaltest variants). packages/ is down to acorn, acts, k4reco, mucoll-stack,
  pelican and the python/ML helpers.
• Removed the mucoll-release-debug environment (and its entry from the concretization matrix).

Full Changelog: v3.0...v3.1