v3.0.0

3.0.0 - 2026-07-26

Added

• Continuous ASV dashboard: results history on the asv-results orphan branch, asv-tachyon UI under gh-pages/bench/, optional Netlify deploy for bench.eondocs.org.
  Windows CI: robust MSVC activation (VS 18 runners) and strip GNU link.exe from PATH for meson test. (#378)

• PHVA mobile/active sets for dense Hessian and matrix-free min-mode via
  phva_atoms (default All = all free atoms). Free/fixed stays the
  optimizer mask; Krylov dimension is 3 * N_active. Shared
  resolveMobileAtoms drives HessianJob, Lanczos, and Davidson.
  pyeonclient: Lanczos/Davidson.compute(..., atoms=),
  resolve_mobile_atoms, free_atom_indices, and
  Parameters.hessian_phva_atoms / lanczos_phva_atoms /
  davidson_phva_atoms. INI keys: [Hessian|Lanczos|Davidson] phva_atoms. (#379)

• Cap'n Proto schema/eon_job_result.capnp defines typed JobRequest /
  JobResult / flat Geometry envelopes for the in-process control plane
  (kill-file-IPC). eon_schema.jobs provides results.dat adapters for
  legacy paths. (#381)

• Vendored vesin updates to 0.6.0 with three local extensions carried as
  upstream candidates: a CPU implementation of the declared-but-missing
  VesinBruteForce algorithm (nearest-image MIC pair search), lazy
  thread-pool worker spawn (serial consumers stop paying
  hardware_concurrency() thread creations per process), and a fused pair
  visitation API (vesin_neighbors_visit plus the header-only
  vesin_visit.hpp).

  Every Fortran-backed potential draws its neighbours from vesin rather than
  from pot-local scaffolding: the EDIP and Lenosky cell/ghost machinery is
  gone, Tersoff trades its O(N^3) sweeps for per-atom lists, and SW's silent
  MAXNEI overflow and FeHe's 800-neighbour gather ceiling are hard errors
  instead of quiet truncation. Those kernels moved to rgpot in the same
  release (see below), so the vesin Fortran interface is vendored there
  rather than here; eOn's vendored copy is the C++ translation unit only. (#389)

• Documentation systems tutorials (Morse Pt NEB, LJ minimization, Pt saddle) with
  built-in potentials and current rgpycrumbs / plt-neb / plt-min conventions
  (1:1 reaction-valley landscapes, full structure strips, one min landscape per
  endpoint).

• Optimized hyperplanar TST (OH-TST) job (Johannesson and Jonsson, J. Chem. Phys.
  115, 9644 (2001)): a job = oh_tst mode that progresses a hyperplanar dividing
  surface by reversible work with thermostatted, plane-constrained sampling, and
  reports the free-energy barrier and crossing rate. Supports Andersen or GLE
  colored-noise (Ceriotti-Bussi-Parrinello) thermostats and symmetry-restricted
  sampling of equivalent product minima. Configured under [OH_TST].

Developer

• Golden masters for CPython server atoms helpers vs pre-#368 scalar code; eon.fileio load/save/round-trip via live readcon fixtures; document chemfiles as optional and unused on server .con path. (#370)
• Windows metatomic CI follows the metatomic/metatensor torch workflow:
  windows-2022, setup-python, pip CPU torch (PIP_EXTRA_INDEX_URL), and MSVC,
  with pixi for C++ deps and flang. In-tree Fortran pots and CuH2 stay enabled
  via feedstock-style flang_rt LIBPATH and MSVC AR=lib. Catch2 runs on the basic
  multi-OS matrix including windows-2022; ConFileIO tests close temp streams
  before remove (Windows file locks); inih example tests strip CR for MSVC
  text-mode stdout. (#377)
• The Windows CI step that used to load eon_sw.dll and assert it exports a
  bare sw_ now audits librgpot instead: the kernels ship inside it with
  hidden linkage, so the check is that no legacy Fortran name reaches the
  export table, which is the PE counterpart of the Linux FortranSymbolAudit.
  It accepts a --default-library=static build, where the kernels land in an
  archive that has no export table at all. (#390)

Changed

• Client public headers live under include/eon/ (numpy/fmt layout). Sources
  stay in client/. Use #include "eon/Potential.h" (and
  eon/fpe_handler.h, eon/potentials/...) with -I$prefix/include.
  Relative ../ includes are gone; headers install via install_subdir. (#379)

• Classical C++ pair pots (LJ, Morse, LJCluster, QSC) use the shared
  eonc::VesinNeighbors wrapper for neighbor lists instead of pot-local
  Verlet / O(N^2) loops. Vesin is always linked into eoncbase so Metatomic
  and classical pots share one NL backend. Client wall time continues to be
  tracked by the existing ASV suite (TimePointMorsePt,
  TimeMinimizationLJCluster, Morse saddle/NEB); see benchmarks/README.md. (#386)

• The Fortran-backed potentials (Stillinger-Weber, EDIP, Lenosky, Tersoff,
  EAM aluminium, FeHe, CuH2, and TIP4P-H) are Fortran 2018 kernels inside
  librgpot and evaluate through RgpotAdapter like the classical
  pots. Each kernel was rewritten rather than wrapped: modules with
  implicit none, kinds from iso_fortran_env checked against the C
  types at compile time, derived-type parameters in place of COMMON
  blocks, intent on every argument, pure kernels, structured
  control flow, and status returns instead of stop. Neighbours come
  from vesin, and the pair sums are restated as gathers so the atom loops
  run under do concurrent.

  eOn no longer builds, installs, or dlopens Fortran: the eon_*.so
  plugin modules, their Windows .def export files, and the flang
  runtime handling are gone. FortranPotLoader becomes PluginLoader,
  which still finds engine plugins (the rgpot metatomic and xtb backends)
  across EON_POTENTIALS_PATH and [Potential] potentials_path.

  Configuration is unchanged: the same potential names select the same
  physics, pinned by the existing reference energies in SiPotTest,
  EAMAlTest, FeHeTest, and cuh2Test. (#390)

• [Hessian] phva_atoms replaces atom_list for the PHVA mobile set
  (default still All). The C++ client still accepts the legacy
  atom_list key when phva_atoms is absent. Lanczos and Davidson
  gain the same phva_atoms key in INI, schema, and config.yaml.

Fixed

• eOn no longer requires a Fortran compiler. It compiles no Fortran since the
  kernels moved into librgpot, but client/meson.build still asked for
  the language with required: true whenever with_fortran or
  with_cuh2 was set -- both default on -- so every consumer had to supply
  an unused toolchain. That bit hardest on builds against an installed
  rgpot, which need none at all. The language is detected rather than
  required now; a wrap build still finds it for rgpot's kernels, and the
  Windows flang runtime discovery is gated on actually having one.

  with_fortran and with_cuh2 keep their meaning as potential
  selectors: they gate the RgpotAdapter arms, not any compilation.

• Local communicator writes client stderr to stderr.dat (no undrained
  PIPE deadlock). The FPE continue handler masks the fault class in the
  restored MXCSR so a single divide-by-zero cannot re-storm. The LAMMPS pot
  worker demotes floating-point traps after fork and forceLocal uses
  eat_fpe (same external-pot contract as ASE/Metatomic); SafeMath guards
  remain on CG/dimer/min-mode bare divisions. (#379)

• Server-side process search and superbasin amsel gating receive a ConfigClass
  (no bare config / missing self.config). Superbasin recycling passes
  config into Recycling. Restored atoms.identical for
  indistinguishable-atom matching. LocalInProcess unpacks
  Matter.relax as (Matter, converged). LAMMPS forceLocal restores
  FE traps after eat_fpe. Tip4p implements the full Potential::force
  signature (variance parameter). (#380)

• ClientEON and pyeonclient append_results_timing write the
  results.dat timing footer as <value> <key> (same contract as
  job writers and parse_results), so time_seconds /
  user_time / system_time parse as floats under those keys. (#382)

• AMS_IO writes the XC functional name into the run script: the
  fprintf("xc %s\\n") call was missing its xc argument (undefined
  behavior). (#384)

• ASE potential import and force failures throw std::runtime_error
  instead of exit(1), so pyeonclient / in-process callers can recover
  instead of killing the whole process. (#385)

• LAMMPS worker path survives a bad geometry: non-finite or failed evaluations
  reject the structure without ending the client, with bounded respawns,
  SIGPIPE ignored on the pipe, serialised concurrent exchanges, and process
  search reporting of how far a rejected endpoint landed from the reactant. (#387)

• Classical pair pots (LJ, Morse, LJCluster, QSC) use eonc::PairListCache:
  a process-global pool of Verlet-skin cached pair lists. The candidate list
  builds at cutoff + skin once; force evaluations on geometries whose atoms
  have moved less than skin/2 since the build re-use the cached pairs and
  derive exact vectors from current positions, with the true-cutoff filter
  keeping results identical to a fresh build. On a cache miss in the MIC regime
  the pair kernel inlines into the build's single brute-force scan, so one-shot
  evaluations (point jobs) pay one pair sweep like the pre-list code did.
  Proximity-matched pool slots keep NEB's per-image force path both race-free
  and cache-warm, whether images are evaluated serially or in parallel — the
  pool survives NEB's per-iteration worker threads. Fixes the ASV regressions
  from the #386 vesin port (point / min / saddle / NEB wall times); minimization,
  saddle-search, and NEB fixtures run faster than the pre-#386 baseline. (#389)

• A build that resolves rgpot through the subproject wrap now takes vesin
  from rgpot instead of compiling eOn's vendored copy beside it. Both trees
  carry the same upstream release, but rgpot's carries local patches its
  Fortran interface binds to, so linking both left those objects calling a
  symbol eOn's copy did not define. The vendored translation unit still
  serves builds against an installed rgpot.

  pyeonclient wheels need no shared object beside librgpot: the Cap'n
  Proto schema moved inside it, so importing pyeonclient._core no longer
  depends on finding a separate libptlrpc.so at run time. (#390)

• Process registration no longer freezes the process-id counter when the
  processtable contains duplicate ids. New process ids are
  content-addressed with xxHash (allocate_process_id / xxh64 over the
  saddle payload and barrier), not len(procs) or max(id)+1, so
  procdata files are not overwritten. Duplicate appends raise.