Changelog

3.6.0 - 2026-10-05

Added

• A per-image potential such as ext_pot keeps its own exchange directory on each dimer image and on each process-search product minimization. An empty checkpoint lets a finite-difference Hessian share RgpotPot calculator groups. In-process results carry job_result, and cancel_state stops only a running batch. An even half-ring instanton with an odd unstable mode continues on the full ring. Structure comparison pairs atoms one to one, and Langevin and Nose-Hoover leave a frozen Cartesian component fixed.

• An RgpotPot run on calculator groups prints, when the driver stops the workers, each group’s evaluated systems, busy seconds and idle share, with the POP load balance, communication efficiency and parallel efficiency of the run.

• Dynamics movies load reactant frames from a read-only readcon-db corpus built with ingest_directory. The simulation .con files stay the store.

• In-process jobs return a JobResult (scalars plus ConFrame saddle and product). results.dat is only the job_result_to_results_dat adapter for cluster and HPC workers. parse_results still reads legacy files.

• JobResult carries optimizer backend, xtsci ABI, eindir ABI, and rgpot identity. Legacy results.dat lines stay value-key records; the classic parser is unchanged.

• Optional rgsaddle sessions step an NEB band and rotate a min-mode when built with -Dwith_rgsaddle=true and opt_method or min_mode_method set to xtsci.

• Structure comparison removes a rigid minimum-image shift when remove_translation is on and check_rotation is off. No atom may be fully fixed. A Hessian drops eigenvalues at or below zero_freq_value. It warns when the dropped count is not 6, 5, 3, or 0.

• The user guide has a glossary of method names and short forms, with links to the pages that use them.

• The user guide now has pages for tad, monte_carlo, replica_exchange, global_optimization (minima hopping), finite_difference, and oh_tst. The hot temperature for temperature accelerated dynamics is the main temperature, and exchange_trials is replaced by the replica count after the file is read.

• Tier-2 libeon_relax_engine.so exports eon_relax_create / eon_relax_run / eon_relax_step / eon_relax_reset / eon_relax_destroy. NULL create is NEB with Parameters NSDMI defaults. A Cap’n Proto RelaxEngineParams flat-array selects neb or saddle. Matter keys energy and variance caches on a surface epoch.

• Potential::surfaceEpoch() and Matter::setSurfaceEpoch() key the energy and variance caches on a surface generation, so identical positions after a surrogate refit recompute. Matter::getEnergyVariance() evaluates the potential when the cache is stale, and the variance of a direct force() call is kept.

• [RgpotPot] backend = uma loads libuma_engine.so through the generic rgpot engine interface (engine_c_abi.h), with model_path, task_name, charge and multiplicity as its configuration.

Developer

• Basin hopping, parallel replica, prefactor, bond boost, and dynamics tests cover job-runner branches that the coverage run was missing.

• The GP dimer links a checkout at subprojects/gpr_optim. That checkout and the gpr_optim, rgmin and anneal wraps stay untracked, so a clone of eOn carries no private wrap.

• The rgpot suite runs the calculator-group cases only when that build links MPI. The launcher preloads librgpot_pot.so from the build tree only when the loader would otherwise pick another copy, such as one in the prefix libdir; an unneeded preload corrupted the heap before main under EESSI 2026.06. The single and uneven cases run real CPMD calls on N2 in a 6 Å cell and time out on a build where calculator groups make different numbers of engine calls.

Changed

• A Lanczos saddle search on a batching potential evaluates the moved centre and the first Krylov product’s displaced image in one batch, as the improved dimer does with its forward image.

• Against rgsaddle ABI minor 5 the rgsaddle band resends the path every step (unchanged rows are free), restarts only the optimizer after a reparameterization or MMF, and takes the session’s evaluation of the accepted band for any image it would otherwise evaluate again.

• An improved-dimer saddle search or climbing-image MMF step on a batching potential (RgpotPot calculator groups, a batched model) evaluates the moved centre and its forward image in one batch instead of two serial calls. The trajectory is unchanged; a converged search can spend one extra forward image.

• Periodic cutoff lists use linkcell pairs_within from v0.3.8 (atom index, image index, and the caller’s shift). An open boundary stays on vesin. knearest is only the k-nearest list.

• RGPOT calculator groups schedule each batch so that no group takes more than ceil(M / G) of its M systems while every image or bead stays on the group that holds its orbitals, and send a single request to the group that last evaluated the nearest geometry instead of always to group 0. The RGPOT page explains how to choose the number of groups.

• RGPOT with backend = cpmdc names each system’s key before its SCF (the NEB image, the ring-polymer bead, the batch position, or one key for single requests), so a libcpmdc that keeps orbitals per key starts each image or bead from its own orbitals of the previous step. One calculator group now batches too, so the keys hold with ranks_per_image = 0. Ring-polymer force calls pass the bead index as the owner of each system.

• RgpotPot options are projected from the schema into the INI reader, the JSON codec, and the default assignment. Alias order and the cpmd and XTBPot overlays stay on that projection.

• The CPMD ExtPot wrapper (examples/akmc-cpmd-slurm/potfiles/cpmd_extpot.py) takes the orbital convergence from CPMD_CONVERGENCE, refuses a zero cell instead of handing CPMD an invented one, and documents its unit chain; tests/test_cpmd_extpot.py checks energy, force sign, units, atom order and cell against a stand-in cpmd.x.

• The NEB image pool keeps its threads for the life of the process and sizes itself to the cores the process may run on (the Slurm or taskset affinity mask) instead of every core of the node.

• The finite-difference cell stress of a solid-state band sends the 12 strained copies of every image to the potential as one batch, so calculator groups or a batched model evaluate them together instead of 12 serial calls per image.

• The rgpot wrap moves to rgpot main 0047072. librgpot no longer links MPI, and eOn’s calculator-group MPI code is a separate librgpot_pot_mpi loaded only under an MPI launcher, so a plain eonclient maps no MPI library and skips its load-time cost (UCX memory hooks, PSM2). The same rgpot keeps a CPMD force call’s collectives inside its calculator, so calculator groups with unequal work per group no longer deadlock.

• The rgsaddle band (opt_method = xtsci) resends the band to its session only after a reparameterization, and the surface callback and the step read images that already hold the requested geometry from their own evaluation. A FIRE step on the LJ test band costs at most two evaluations of the interior images and none of the endpoints, against about 3.2 evaluations of the whole band.

• The rgsaddle min-mode (min_mode_method = xtsci) passes the finite-difference length ([Main] finite_difference, Angstrom) as its displacement, not the dimer rotation angle in radians. With rgsaddle ABI minor 5 one session serves the whole search: the refreshed mode seeds the next estimate, the host’s gradient spares the centre call, and the converged angle and the [Hessian] fd_scheme difference reach the session.

• The rgsaddle min-mode rotation reads the centre from its cached evaluation and runs its displaced probes on a copy, so the host centre is neither re-evaluated inside the rotation nor after it.

• With [Main] parallel on, the NEB projects its images (tangent, spring and projected force) on the image pool as it evaluates their forces.

• -Dwith_xtb=true builds potential = xtb on rgpot XTBPot ([XTBPot] maps onto XTBConfig). The in-tree XTB kernel is gone. The metatomic engine installs as librgpot_metatomic_engine; libmetatomic_engine remains a compat symlink for this release, and the xtb/metatomic loaders try librgpot_* then the previous soname.

• makePotential returns exclusive ownership. Point and minimization keep that unique_ptr until one Matter takes it. Shared ownership is explicit via sharePotential at multi-owner boundaries.

• pyeonclient 0.4.2 publishes the client from the 3.5.0 tree.

Fixed

• A /usr pkg-config nlohmann or Highway is not used; the wraps supply them.

• A half-ring instanton keeps the reaction coordinate monotone between the turning points, so the search stays on one bounce. The index-1 step shortens inside the trust radius when the quadratic model is off.

• A one-dimensional rate instanton uses each bead’s own curvature and solves the requested temperature directly. An index-1 step is kept when its predicted and actual energy changes both sit under the energy resolution, and near-zero saddle modes are held at a spring-sized curvature so a rigid displacement does not singularize the chain.

• A rate-instanton search stops as soon as B_N falls below 1e-4 of the starting ring’s, since a ring of coincident beads is no instanton: RateInstanton::collapsed is set, the job logs where the beads fell and writes instanton_collapsed to results.dat. A 16-bead Al slab ring that slid onto a minimum ran 1749 Newton iterations and 3.5 h before this. Each Newton iteration is traced at debug level (U_N, residual, B_N, climb curvature, negative count, trust radius).

• A readcon-core subproject that does not define readcon_order adds no order-stamp sources instead of failing configure.

• A ring-polymer step evaluates the beads once: the forces at the end of one step serve the start of the next, which halves the force batches of pile and piglet dynamics and of the PI-QTST planes.

• A safe-hyperdynamics run that continues after a transition keeps the bond boost on the trajectory, and dephasing keeps it too. The continued steps recompute the bias after each move.

• A solid-state band on a potential that reports stress (LAMMPS, SocketNWChem, xTB through rgpot, RgpotAdapter) reads the stress that came with each image’s force call instead of evaluating every image a second time.

• An ImprovedDimer compute() on a Lanczos, Davidson or LOR rotation backend, and a mode-loss restore, leave the dimer centre’s evaluation in place, so a read of the centre before the next step (an energy-accepting optimizer, the nonnegative-displacement check) costs no force call.

• An MPI build of eonclient reads its command-line flags (--version, --help, one-shot jobs) unless an eOn server launched the rank. It used to ignore every flag and run as a client.

• An MPI build of eonclient started outside an eOn server (no EON_SERVER_PATH) runs config.ini standalone instead of exiting with an error; EON_CLIENT_STANDALONE is no longer needed for that.

• An uphill basin hop uses exp(-de/(kB*temperature)) only at a positive temperature in kelvin. Client replacements are finite_difference, dynamics, and safe_hyperdynamics for finite_differences, molecular_dynamics, and hyperdynamics, with displacement_sampling outside the client list and akmc on the server.

• Client startup defers the Quill backend until a job path runs. Help and version return before the log thread starts.

• Climbing image is applied only when it is active and the highest interior image is strictly above the higher endpoint. A monotonic band keeps the spring.

• Economised spring frequencies stop at the stationary fit. The Nyquist sample is that minimum, so the same bead count gives the same frequencies on every libm.

• Harmonic TST and quantum harmonic TST choose the saddle’s rigid modes among its bound modes, so a barrier softer than a rigid mode’s finite-difference residue no longer leaves that rigid mode in the product.

• Highway is looked up as hwy, the name highway.wrap provides, so --force-fallback-for=hwy selects the wrap.

• In-process relax, NEB, and saddle search poll a CancelToken and stop when cancel_state is called during the batch.

• OH-TST moves the guideline through the sampled average only after the reversible work has passed two kT. A force sign change in the reactant basin keeps the original reactant-to-product line.

• Path-integral dynamics (pile, piglet) draws thermal ring momenta at the start, averages the centroid-virial kinetic energy over the second half of the run and logs it; Dynamics::pathKineticEnergy returns it. The run used to start the ring at rest and report nothing.

• Quantum harmonic TST evaluates ln 2 sinh(x / 2) through expm1, so a soft mode no longer loses its zero-point factor to cancellation (3e-8 in ln k at beta hbar omega = 1e-9).

• Rate searches with initial_hessians = finite_difference no longer write the bead blocks into an empty vector; on a multi-dimensional ring the option was undefined behaviour.

• Release pages open on a title, and notes in the install and user guides stay on the page. C++ API sections use one label per potential.

• Rotational zero modes of the reactant are classified by the rotation’s curvature r^T H r / r^T r against the softest vibration of the Hessian once the rigid motions are projected out (zero at or below 0.1), instead of ||H r|| against the Frobenius norm, which grows with the atom count: a 16 x 16 periodic slab’s in-plane rotation passed as a zero mode at 0.0074 of the Frobenius norm under the 1e-2 threshold. tunneling::rotationZeroModes is the classifier.

• Runs with [RgpotPot] ranks_per_image above 0 no longer hang on their first single-structure force call. rgpot 3.4.0 agrees on engine errors across every MPI rank after each CPMD call; eOn ran a single structure on calculator group 0 alone, and a batch split unevenly gave groups different call counts. Every group now evaluates a single structure, and a group that owns fewer systems in a batch repeats one.

• Samples name the client binary eonclient. The replica loop includes do, the install command passes -C, and Python runs the displacement scripts.

• TIP4P uses the Carney-Curtiss-Langhoff equilibrium OH length and angle. The out-of-line members take those table values instead of initializing from their own names.

• The Local communicator starts eonclient with UCX_MEM_EVENTS=no unless the environment sets it. An eonclient linked with MPI loads UCX in every run, and its memory hooks cost about 0.4 s per process with conda’s OpenMPI against 0.02 s for a one-call job.

• The MPI potential example launches eight VASP ranks and one client with potential = mpi. Parallel-replica and basin-hopping MPI examples call eonclient through eon.server. A LAMMPS input is read from the client working directory, and a displacement script path takes no flags.

• The Python API reference keeps the version string and leaves out the generated version module and the test scripts. C++ parameter notes name the function arguments.

• The Windows relax-engine DLL links the readcon import library.

• The Windows xtb import library directory is on the linker search path.

• The climbing image stays on its band index while it is a strict interior local maximum, instead of moving to the highest image at every force update.

• The cutoff list sizes its row buffer from the pairs found. A shell denser than the ideal-gas estimate no longer writes past that buffer.

• The docs build loads libenchant-2.so for spelling, and it stops when sphinx-build fails.

• The guide records that -Dwith_xtb, -Dwith_water, -Dwith_metatomic, and -Dwith_vasp default false, and that -Dwith_lammps and -Dwith_dftd3 are not eOn options. MPI runs launch eonclient and read EON_CLIENT_STANDALONE plus EON_NUMBER_OF_CLIENTS; the version record is 3.5.0, and the metatomic extra depends on rgpot>=3.2.0.

• The installed eon Python package includes eon/job_result.py. eon.explorer imports it, so the aKMC server failed to start from a meson install.

• The instanton rate leaves the ring Hessian’s near-zero eigenvalue out of det’ exactly. It lifted the central difference along the ring and divided by the lift alone, which is det’ only for an exact zero mode with that eigenvector; on a discrete ring the eigenvalue is not zero (-0.19 on a 12-bead Eckart ring at T_c / 2, 0.07 in ln k) and the central difference is not quite its eigenvector. Inverse iteration now finds the eigenvector and the determinant is divided by the lift plus the eigenvalue.

• The instanton rate of a free cluster leaves out the ring’s own rotations, built from its beads about the ring’s centre of mass, instead of the reactant’s rotations copied to every bead, and the ring’s bead Hessians keep their rotational curvature. The reactant’s rotations are null vectors of the ring Hessian only when no bead moves, so the rate of a cluster whose atoms move along the ring was off by the overlap of the two. instantonRate takes the atoms through RingRigidBodies.

• The instanton splitting decides that the path minimises the action from the inertia of the action Hessian instead of its determinant’s sign, so a path with two negative modes is refused.

• The nudged elastic band objective reports one mass per free atom on each interior image.

• The rate job gives a converged ring a rate only when it crosses the seeded saddle’s dividing plane within its own span of the saddle and its turning points lie within 60 degrees of the unstable mode; straddling the plane alone let a ring around a neighbouring saddle through. results.dat reports the turning points along the mode, the chord overlap and the crossing offset.

• The rate job seeds its ring at the requested period on flat-topped and kinked paths. The orbit energy was found by one bisection over the whole range, which assumed the period falls monotonically with energy and resolved the top only to the first of 400 geometric levels; on the Al slab of examples/neb-al it returned the same orbit (0.103 eV above the reactant, period 103.8) for beta hbar 129.1 and 91.2, a ring for neither temperature. The scan now resolves both ends of the range, refines every crossing of beta hbar, discards jumps of the turning points, and keeps the crossing that minimises theta(E) + beta E. The job logs the orbit energy, its period and the height of the path’s ends; the path itself is traced at debug level.

• The rate-instanton Newton search rebuilds its bead Hessian blocks from finite differences (at most four times) when the Bofill-updated blocks show more negative curvatures than the last exact blocks did. Blocks copied from the saddle and carried by Bofill updates grew from 2 to 11 spurious negative curvatures on a 1203-coordinate Al slab, and every flip then turned the ring downhill.

• The relax engine is a shared library, so Windows can load it. The prefactor frequency test closes its file before deleting the directory.

• The relax engine passes each image’s stored cell to the host surface in row-major order, including when the image is not periodic.

• The relax engine schema compiles as .cpp, so MSVC links libeon_relax_engine.

• The saddle frame stored in readcon.db carries the process mode in its displacements section. A suggestion reads that section. mode_<id>.dat remains the client sidecar, and amsel.KdbProcess.mode keeps the same components.

• The saddle-search catalog accepts min_mode_method = xtsci. That rotation runs in a rgsaddle session when the build enables rgsaddle.

• The server reads [AKMC], with confidence_scheme defaulting to old, and the default opt_method is cg (QuickMin is qm). A file the server loads exits on [Refine] and on the unknown keys parallel, climbing_image_band_slack, match_endpoints, and match_method.

• Up to denseLimit ring coordinates (4096 by default) instantonRate takes the determinant, the negative-mode count and the lowest eigenvalue from the dense spectrum of the lifted ring Hessian and checks the block-chain determinant against it; the limit was documented but not read, so every ring used the block chain and a Lanczos estimate of the lowest mode.

• Windows copies the relax engine’s imported DLLs beside the test, and the Metatomic job searches the pixi runtime before torch.

• -Dwith_rgsaddle=true builds against rgsaddle 0.2: the band and min-mode surface callbacks return rgsaddle_status_t, and the band answers the interior-only and one-image requests rgsaddle sends after its first evaluation, evaluating the carried images as one batch. The rgsaddle floor is 0.2.0.

• [RgpotPot] task_name is listed in config.yaml and on the schema model. The default is omol.

• eon_relax_run returns MAX_UNCERTAINTY when any image variance is above the uncertainty limit.

• tie_lifetime calls keep_alive_py on nanobind 3 and nb::detail::keep_alive on nanobind 2, so a returned path keeps the Parameters owner alive.

• job_result_legacy_dict maps a JobResult onto the historical results.dat keys. job_result_to_results_dat writes that dict for cluster and HPC adapters.

3.5.0 - 2026-10-01

Removed

• The tsase kdb path for aKMC (eon/eon_kdb.py and its import of the PyPI kdb package) is removed. use_kdb = true stores and suggests processes through amsel.KdbStore and the run’s readcon-db corpus.

Deprecated

• -Dwith_parallel_neb is deprecated and has no effect. The NEB image pool uses std::thread, so [Main] parallel = true needs no TBB.

Added

• Above the crossover, mode = rate writes parabolic_factor and rate_parabolic, the factor (pi*T_c/T)/sin(pi*T_c/T) times the harmonic TST rate from the reactant and saddle Hessians. The ring search still runs only below the crossover. At the crossover the factor diverges and that temperature records no rate.

• An even bead count evaluates the rate-instanton potential from one turning point to the other and copies that half onto the closed ring. The rate is the same as a full evaluation, and each bead still contributes its Hessian.

• Path-integral quantum transition-state theory after the rate instanton: [Instanton] pi_planes samples a ring polymer with its centroid held on parallel planes from behind the reactant to the saddle, integrates the centroid mean force to a free-energy profile with block-average errors, and writes the quantum free-energy barrier and the PI-QTST rate beside the classical and instanton barriers, with rate_piqtst.dat and one centroid-plus-spread frame per plane.

• Path-integral trajectories on the dynamics job, with normal-mode PILE, a normal-mode GLE read from a matrix file and a separate centroid Langevin thermostat, and economised ring-polymer springs. Bead forces are one batch. A centroid hyperplane reports its mean force. Economised springs are refused with the normal-mode GLE and with the instanton.

• Ring-polymer MD transmission factor for PI-QTST. [Instanton] pi_recrossing_parents samples parent rings on the top plane and launches momentum-reversed pairs of thermostat-free RPMD children from each. The job reports the Bennett-Chandler plateau kappa with a jackknife error and k_RPMD = kappa k_PI-QTST in results.dat and rate_piqtst.dat. kappa_piqtst.dat holds the kappa(t) curve.

• Rings of at most 4096 active coordinates take an index-1 Newton step with a Bofill update. Below three quarters of the crossover that search starts at 0.85 of the crossover. An even ring runs from one turning point to the other. Larger rings follow the minimum mode and copy one half of an even ring. The rate uses the bead Hessians.

• Setting [amsel] discover_decide = true runs on the current state’s process table while use_mcamc stays off. A barrier below e_min_init stays in the basin (0.20 eV on Si6N8 isomer 1, against a 0.30 eV flip out). The exit time and channel come from the mean-rate method or first-passage-time analysis. The [amsel] keys are AmselConfig.

• Solid-state NEB relaxes the lower-triangular cell of each interior image together with the atoms. The tangent and spring use the Jacobian of doi:10.1063/1.3684549. A potential without a stress tensor is differentiated on the cell. OCINEB, zoom, and the action-based springs stay off when solid_state is set.

• The Hessian job writes modes.con: one frame per normal mode, with the unit Cartesian mode as a readcon displacements section and mode_eigenvalue, hbar_omega and wavenumber in the frame metadata. [Hessian] write_modes = false turns it off.

• The finite-difference Hessian evaluates its displaced structures through forceBatch when the potential batches, so under [RgpotPot] ranks_per_image its columns spread over the CPMD calculator groups. Prefactor and instanton Hessians gain the same. A column checkpoint keeps the one-column path.

• The instanton rate on a long ring comes from the cyclic block fluctuation product. A small ring still uses the dense product, and the two have to agree.

• The rate instanton’s Newton step, ring determinant and inertia count run through a block LU of the open bead chain plus a low-rank Woodbury correction (closure, cycle, rigid modes, eigenvector-following flips), O(N f^3) with no dense ring matrix; bead curvature blocks start from the saddle Hessian and take Bofill updates ([Instanton] initial_hessians), and the Eckart flux is checked against the exact transmission.

• The search accepts a list of temperatures and starts at the highest. Each temperature writes one row of rate_instanton.dat. A supplied band seeds the ring and reports a one-dimensional semiclassical rate.

• [RgpotPot] params_path loads a CPMDParams file, and ranks_per_image splits the launch into calculator groups. The user guide renders the model and a 7-image band on 28 ranks.

• [RgpotPot] ranks_per_image splits the MPI world into cpmdc calculator groups, one CPMD session per group, and a NEB spreads its images over the groups through forceBatch. Each rank receives every image’s energy and forces, so all ranks keep the same band; single force calls run on group 0 and are shared the same way. Needs rgpot built with -Drgpot:with_mpi=enabled and libcpmdc with cpmdc_bind_calculator.

• job = instanton computes the ring-polymer instanton between two minima and its tunnelling splitting, beyond one-dimensional WKB along a band. It writes the beads to instanton.con and the splitting, action and diagnostics to results.dat. The beads are evaluated in one batch per iteration, so CPMD calculator groups run them in parallel.

• job = instanton with mode = rate computes the thermal rate through a saddle below the crossover temperature, from the closed ring-polymer instanton of Richardson and Althorpe (doi:10.1063/1.3267318). It reads the reactant and saddle.con, takes temperature in kelvin, and writes rate_instanton to results.dat. Above the crossover it writes the parabolic barrier factor times the harmonic TST rate. At the crossover the factor diverges and the job stops. The reactant Hessian shows which rotations to omit for a cluster inside a periodic cell. The cubic-well check is the decay of Caldeira and Leggett (doi:10.1016/0003-4916(83)90202-6).

• job = instanton writes instanton_centroid.con: the beads’ centroid with the readcon spreads section (per-atom root-mean-square spread in Å), the delocalised configuration as centroid plus spread in both the splitting and the rate mode.

• neb.con frames carry reaction_coordinate_mw, the mass-weighted arc length along the band, and the first frame carries the band’s one-dimensional tunnelling estimate: hbar_omega_reactant, hbar_omega_product, tunnel_action, tunnel_splitting (WKB, Landau and Lifshitz prefactor), tls_energy and tunnel_deep_wells, the inputs a two-level-system screen of glass minima needs. On a quartic double well sampled by 21 images the splitting is within 5 percent of the exact two-level gap once the barrier stands three quanta high.

Developer

• The metatomic setupeon pixi task and scripts/run_coverage_cpp.sh pass -Dwith_gprd=enabled and -Dwith_gprd=disabled. with_gprd is a feature option, and meson rejects True and false for it at meson setup.

Changed

• A library force call writes no RESTART.1, LATEST, GEOMETRY, or GEOMETRY.xyz. The orbitals for the next call stay in memory.

• Configuration models render as a field list. The PDF title page carries the logo, and chapters open on the next page.

• On a potential that batches (CPMD calculator groups), the NEB evaluates its two endpoints in one call, the improved dimer evaluates its two ends together, a band update routes image i to the same calculator whether or not the images before it moved, and an OCI restore puts the saved forces back instead of asking the potential again. L-BFGS seeds its first step from the min-mode curvature when the objective knows one, so the MMF dimer inside OCI-NEB spends no force call on the finite-difference probe. NEBObjectiveFunction::getEnergy goes through the batched band update.

• Ring-polymer dynamics transforms the beads to normal modes as one matrix product and holds a centroid hyperplane without transforming the ring, about eight times faster per step at 32 beads.

• The build requires readcon-core 0.16.0 or newer, from pkg-config or the subproject, and the wrap pins v0.16.0. ConFileIO writes spreads with set_spreads_from_flat, which 0.15.x does not have, so a 0.15 host package used to pass configure and fail to compile. Packagers raise the host readcon-core floor to >=0.16.

• The rgpot subproject pins the 3.4.0 release (calculator groups, MPI finalised at exit, CPMD session reuse and the abort after a failed engine call).

• -Dwith_mpi=enabled requires an embedded Python (python3-embed): the server rank of the MPI client runs EON_SERVER_PATH through Py_Main. Configure stops when the embed dependency is not found.

• [cpmd] holds the scalar CPMD message when [RgpotPot] params_path is empty. cutOffRy there overrides cutoff_ry. A loaded CPMDParams file keeps its sections, and input_block is appended to that file’s blocks.

• python -m eon.server --reset also removes the [Paths] kdb catalog directory, wherever it points. A shared or curated catalog needs a copy outside the run. The readcon-db corpus readcon.db is not removed. The kinetic database guide describes both.

• with_mpi is a feature option and stays disabled unless set to enabled. auto does not link MPI. -Dwith_mpi=enabled builds the client/server program; calculator groups stay -Drgpot:with_mpi=enabled.

• aKMC with use_kdb = true writes each good process into amsel.KdbStore when it is registered: the barrier in eV, the prefactor, the mode, and the readcon-db frame keys. The reactant, saddle, and product frames go into the run’s readcon.db. The next search of a matching state refines from the stored saddle, then uses a random displacement when no suggestion remains. The kdb module is not imported. kdb_nf (fraction, default 0.2), kdb_dc (angstroms, default 0.3), and kdb_mac (minimum mode cosine, default 0.7) are read from the ini. Paths.kdb is the catalog directory.

Fixed

• A calculator-group run whose rgpot engine call failed (rgpot requested MPI_Abort at exit) now aborts the MPI world on exit. The driver skips the stop broadcast and eOn’s exit handler calls MPI_Abort instead of MPI_Finalize, so ranks left inside a CPMD collective no longer hold the job until the walltime kill.

• A calculator-group run with the rgpot wrap (-Drgpot:with_mpi=enabled) constructs its groups again. rgpot 3.4.0 reports MPI only after MPI_Init, and the construction agreement that calls MPI_Init was gated on that report, so every ranks_per_image greater than 0 stopped with “ranks_per_image needs rgpot built with MPI”. The agreement runs whenever the process was started as one of several ranks.

• A cpmdc job under mpirun calls cpmdc_finalize on every rank, then MPI_Finalize, then _Exit, so CPMD and MPI library destructors do not run on a finalized world. Rank 0 prints the engine error from another calculator and exits non-zero. A rank that cannot read params_path stops every rank before MPI_Comm_split. eOn’s -Dwith_mpi build is the client/server program; calculator groups are mpirun -np N eonclient with -Drgpot:with_mpi=enabled.

• A failed evaluation in one [RgpotPot] ranks_per_image calculator raises on every rank after all results are shared. Other ranks no longer wait in a broadcast the failed calculator never joins, and workers keep serving the driver.

• A half-ring instanton that has converged, or that is stationary at the wrong index, is probed for an unstable mode odd under the ring mirror. That mode marks two copies of the instanton on one ring, and the search finishes on the whole ring. A cooling schedule probes only its last temperature.

• A process whose product column is -1 is an amsel exit when its barrier is at or above e_min_init. The absorbing label is a 32-bit id. The hop creates the product state from the process id.

• In the MPI communicator a client whose job fails (a potential error, a bad config.ini, an unknown job, a missing job directory) logs the error, stages its logs into the job directory and hands the directory back without results.dat. The server skips that result and the rank takes the next job, so one failed CPMD call no longer ends the allocation.

• JSON configuration (the Python binding and serve mode) reads every [Hessian] key that to_json writes, including write_modes, with atom_list accepted for phva_atoms as in the ini. JSON loading also derives the internal path_pile_tau and Andersen collision period from the femtosecond inputs when the keys are absent, so a pile or piglet run loaded from JSON uses the same damping time as one loaded from config.ini instead of stopping on a zero time.

• NEB image forces with [Main] parallel = true run on at most std::thread::hardware_concurrency() threads, each taking the next image, instead of one thread per image. An exception from one image’s potential is rethrown after every thread joins; it used to reach std::terminate. GCC and Clang builds no longer need TBB; nvc++ keeps -Dstdpar=cpu|gpu.

• On the Morse Pt cell, gprdimer and dimer both stop at -1462.008706 eV. gprdimer uses 17 force calls and dimer uses 38. A Linux build links that method under -Dwith_gprd=auto when a gpr_optim checkout at b55c89e2 sits in subprojects/gpr_optim.

• The Hessian accepts the trivial-mode count the structure’s symmetries give: 6 for a free cluster, 5 for a linear one, 3 for a periodic cell and none once any atom is fixed. It used to report every periodic bulk cell as an error for having 3 instead of 6.

• The MPI communicator reads the job path a client returns with ndarray.tobytes. tostring is gone in NumPy 2, so the server stopped with an AttributeError at the first returned job.

• The Python server accepts a potential name in any case. An exact listed spelling wins.

• The Python server accepts the CPMD cutoff and functional spellings the client reads: cutOffRy, cutoff_ry and cpmd_cut_off_ry, and cpmd_functional, on [cpmd] and on [RgpotPot]. A config.ini with [cpmd] cutoff_ry or [RgpotPot] cutOffRy used to stop python -m eon with “unknown option”. The eon-schema Cpmd and RgpotPot models accept the aliases and leave them out of model_dump, so a written file carries one cutoff key.

• The amsel superbasin gate returns its configured fallback when an amsel call fails, instead of raising NameError on an unbound config; the policy is [amsel] on_error (fallback_single, unavailable_mcamc or raise). Processes whose product is not yet linked to a state (-1) are left out of the graph handed to amsel, and the discover cutoff stays at [amsel] e_min_init instead of being lifted above every known barrier, which left the basin without an absorbing border.

• The calculator-group safety code in the RGPOT potential (the agreement before MPI_Comm_split when a rank cannot read params_path, and the cpmdc_finalize, MPI_Finalize, _Exit sequence at exit) compiles whenever rgpot links MPI: -Drgpot:with_mpi=enabled with eOn’s with_mpi off, the documented calculator-group build, or an installed rgpot whose rgpot.pc defines RGPOT_HAS_MPI. It used to compile only with eOn’s -Dwith_mpi=enabled.

• The economised ring-polymer spring fit iterates to the least-squares minimum, with the 10000-iteration cap of Zeng and Manolopoulos, and raises an error when it does not converge. The fit stopped after 500 Newton steps, partway along a flat valley of the objective, and returned frequencies that differed between compilers and platforms by up to 5% on the high-frequency modes (48 beads at a maximum reduced frequency of 20).

• The installed eon server package ships cancel.py and process_id.py. An installed AKMC server no longer stops with ModuleNotFoundError: No module named 'eon.process_id' on the first registered saddle.

• The installed eonclient finds libeonclib with a multiarch or other non-default libdir. Its run path is $ORIGIN plus the relative path from bindir to libdir, and the installed libraries carry $ORIGIN, where the potential modules are installed. Meson’s own libdir entry is not added once install_rpath is set, so such an install used to exit 127.

• The minimum-mode rate-instanton search on half of an even ring checks a converged ring for an unstable mode odd under the mirror and, when it finds one, finishes on the whole ring. The half-ring search no longer returns two copies of the instanton on one ring with two negative modes.

• The rgpot potential links the Message Passing Interface (MPI) when with_mpi is enabled. A system mpic++ on the default path no longer adds its include directory to a build that left MPI off.

• The server catalog lists the path-integral keys on [Dynamics] and springs on [Instanton], with the same names and defaults as the client. JSON load reads those keys, including an instanton section.

• The solid-state band keeps the Cauchy stress a calculator already computed. xTB, the i-PI virial, and LAMMPS pressure are read on the force call. A cell difference is used only when that tensor is absent.

• With [amsel] discover_decide on, a kinetic Monte Carlo step no longer waits for the old repeat-count confidence. A lone barrier at or above e_min_init leaves through the mean-rate method or first-passage-time analysis, and a faster edge stays in the basin. A missing amsel package logs status=unavailable and does not step early.

• Without initial_path, the rate instanton seeds its ring from a steepest-descent path out of the saddle. A ring that converged onto another saddle is refused. The climb follows the last climb by overlap. A converged ring is classified with finite-difference bead Hessians. The rigid motions of the whole ring come from the current beads.

• [RgpotPot] input_block reaches the cpmdc backend as CPMD &SECTION text ahead of the generated sections, so an in-process CPMD run can be periodic instead of the isolated cold deck. A new permanent_dir key sets the CPMD FILEPATH for RESTART files.

• eon.atoms loads without minimage again. The Cell.wrap_many binding for readcon-ops ran at import time, so the aKMC server stopped with ModuleNotFoundError: No module named 'minimage' wherever minimage, which is not on PyPI or conda-forge, was not installed by hand. The binding now runs just before the readcon-ops calls that need it, and is skipped when minimage is absent.

• pip install eon declares PyYAML, which eon.config imports at start-up. Without it python -m eon failed with ModuleNotFoundError: No module named 'yaml' in any environment other than the pixi and conda ones, which already listed it.

• with_gprd=auto leaves the GP dimer off when the gpr_optim fetch fails. with_gprd=enabled still stops configuration.

  A build directory that stored with_gprd as true or false rejects meson setup --reconfigure (Option "with_gprd" value auto is not boolean). python scripts/migrate_with_gprd_option.py <builddir> maps true to enabled and false to disabled, and the reconfigure then runs.

• pyeonclient 0.4.1 carries the eOn 3.4.0 client. PyPI already held a 0.4.0 built from July sources, and the 0.4.0 publish of the 3.4.0 client left those files in place; the wheel publish now refuses a version the index already has.

• pyeonclient builds with the default GP dimer: the min-mode bindings include AtomicGPDimer.h when WITH_GPRD is set.

3.4.0 - 2026-09-29

Added

• Cutoff-colored finite differences on a four-atom line take 19 central force calls, against 25 for one column per coordinate, and match the serial matrix to 1e-8. The fourth-order stencil stays within 1e-10 of the complex-step derivative of z^4 + 0.3 z^2 at 0.8, where the central stencil is more than 1e-3 away.

• Hessian fd_scheme = fourth builds the assembled Hessian, and Lanczos and Davidson products, with a real fourth-order central stencil.

• In-process jobs poll a cancel token before each job and before each dispatch. cancel_state returns 1 and stops the next job only while a batch is running; an idle call returns 0 and does not stick. A compiled relax still finishes the current call.

• New example examples/akmc-cpmd-slurm: adaptive kinetic Monte Carlo on a silicon vacancy with CPMD on several MPI ranks behind ext_pot, one Slurm job per search. The ExtPot guide explains how the wrapper reads CPMD forces from GEOMETRY.

• The in-process record stores termination, energy, force calls, and job type on job_result. results.dat is written from that dict.

• The production neighbor list stays on vesin. neighbor_list_linkcell compares that cutoff list with linkcell k-nearest pairs.

• Wire Geometry carries forces, atom ids, and a per-axis fixed bitmask at ordinals 8, 9, and 10. Empty lists mean those fields are absent.

• Zoom-NEB packs NEB images onto a window around the climbing image and keeps OCINEB available on that band.

• eon_schema.jobs writes an eon.trajectory.v1 manifest of exact frame geometry digests and the rgpot potential identity. Landfold embedding and FES inputs are read from that manifest, not from results.dat.

• opt_method = xtsci can borrow an eindir objective when eindir-core is installed. The adapter checks the ABI stamp and the eV/angstrom descriptor, and minimization results record the optimizer name.

Developer

• Docs CI minimizes the LJ13 cluster and regenerates the plt-min tutorial figures.

• The readcon-core wrap pins v0.14.11. The v0.14.10 tag recorded its version as 0.14.9 in Cargo.toml and meson.build, so pkg-config reported 0.14.9.

Changed

• Finite-difference and saddle-search jobs log through the client logger instead of printf, and the finite-difference step list is a fixed array rather than a sentinel. Monte Carlo writes its final structure once.

• Indistinguishable structure matches, rotational matches, and crystal space groups go through readcon-ops. The space-group result includes the Hall number. identical is the readcon-ops one-to-one map, and it binds Cell.wrap_many when the installed minimage only has displacement.

• Lanczos, Hessian, prefactor, bundling, IDPP, and the BGSD and basin-hopping saddle searches use standard library parsing and drop unused helpers. Basin-hopping refuses a band that has no interior tangent, and both searches record the status they return.

• LocalInProcess takes a Structure or ConFrame on the job and returns saddle and product as ConFrames. In-process job dicts no longer carry .con text, and those frames are not written to disk.

• Per-atom force convergence (Matter::maxForce and the NEB max_atom metric) reduces row norms with Highway when the library is available. Fully fixed atoms stay out of the maximum.

• Periodic wrapping uses Highway for the minimum-image floor (x - floor(x + 0.5)) and the legacy unit-interval wrap. Builds without Highway keep the same scalar formulas.

• The direct in-process RGPOT arm is linked on every non-Windows build, and Cap’n Proto is required there. -Dwith_rgpot is deprecated and ignored, so existing invocations still configure. Windows builds still omit the NWChem/CPMD frontends.

• Parameters stores load state and option groups in a private Impl. sizeof(Parameters) is that pointer. Const accessors and ParametersLoadAccess are the read and write surface. Option-group types stay in ParametersOptions.h. Matter still exposes Eigen.

• eOn builds against rgpot 3.3.0, and the pyeonclient wheel check installs rgpot>=3.3.0, the first rgpot wheel that imports without a system OpenBLAS or OpenMP runtime.

Fixed

• A LAMMPS worker waitpid interrupted by a signal is retried. A failed fork closes the pipes it just opened.

• A process search or dimer on ext_pot with parallel = true evaluated its two images on one instance at the same time, so two external programs shared one exchange directory; CPMD then read a corrupt RESTART.1. The images now keep separate instances, cloned from the job’s potential, and a force call that fails in the second thread stops the job with its error instead of std::terminate.

• A signal during waitpid no longer makes a live VASP job look dead. The child closes the extra vaspout descriptor after redirecting it.

• GLE apply leaves velocities unchanged when any mass is non-positive, so a zero-mass DOF cannot write inf.

• LAMMPSPot locks the fixed-atom mask and the in-process force call on one mutex, including the Windows and MPI paths.

• LBFGS, conjugate gradients, and steepest descent allocate zero history vectors when the optimizer is constructed.

• Least-coordinated AKMC displacements accept per-coordinate free masks. Atoms with at least one free coordinate remain eligible, and fully frozen atoms are excluded.

• NEB skips the initial L-BFGS finite-difference curvature probe while preserving automatic scaling from successive steps. Climbing-image forces apply only when an interior image exceeds both endpoint energies; paths whose highest energy lies at an endpoint retain their spring forces. The convergence tests exercise the default climbing-image setting.

• The AKMC server honours remove_translation the way the client does: with no fixed atoms, a saddle or state that differs only by a rigid drift of the whole periodic cell matches its earlier copy, so repeated processes count as repeats and the state confidence rises. Every repeat was counted as a new process before.

• The Hessian job opens pos.con through getRelevantFile, so pos_cp.con and pos_in.con are honoured like the other jobs.

• The MPI potential sends the working directory as int character codes, matching the MPI_INT it declares, and sleeps mpi_poll_period seconds between polls instead of spinning. A run with no potential ranks, or a rank count that does not divide by the number of clients, stops with a message.

• The MSVC xtb import library aliases lowercase MinGW exports to the camel-case names in the xtb header.

• The NEB maximum-image tangent, the dynamics saddle search modes, and the Lanczos and Davidson eigenvectors no longer turn a collapsed vector into NaN. They stay zero.

• The Slurm scripts in tools/clusters/slurm read the account, partition and job shape from EON_SBATCH_ARGS and the client command from EON_CLIENT. queued_jobs.sh lists the current user’s jobs and fails when squeue fails.

• The cluster communicator cancels only the jobs still queued when a state reaches its confidence. Finished results stay for harvest, and a failed cancel is logged instead of stopping the server. The MPI communicator no longer stops idle clients at that point.

• The local communicator starts each client in its own session and kills the whole process group on exit, so ExtPot wrappers and the MPI launchers they start do not outlive the server.

• The rgpot wrap builds ExprPot and the client defines RGPOT_HAS_EXPR, so PotType.EXPR works in wrap builds instead of stopping with the generic error.

• The saddle-search job opens pos.con through getRelevantFile.

• Windows MSVC builds of the native xtb potential link a generated xtb.lib from the conda-forge libxtb DLL instead of the MinGW import library.

• identical keeps a one-to-one map. A crossed pair within the tolerance still matches. Two atoms cannot both match the same partner.

• identical no longer maps two atoms onto one site. internal_motion translates atom 0 onto the reference atom. point_energy_match forwards use_identical into the match.

• identical rematches a close index pair whose elements differ. internal_motion turns a reversed bond around and removes the rotation about that bond.

• internal_motion places atom 0 on the reference atom before it removes the rotation.

• lammps_logging keeps client_lammps-N.log. The worker child writes that file. The parent copies new lines into the process log after each force call. A new LAMMPS open truncates that file, and the copy starts over. The child does not call the process logger.

• eonclient -c compares a copy so Matter::compare cannot translate the first structure.

• resolveMobileAtoms / freeAtomIndices throw on a null Matter.

• pyeonclient Linux wheels are manylinux_2_28 again and import on a system without gfortran or OpenMP: the private librgpot SONAME step now rewrites RECORD and runs before auditwheel, which vendors those runtimes. A failed repair fails the build instead of shipping a linux_x86_64 wheel.

• pyeonclient wheels bundle librgpot as libeon_rgpot.so.3 so a side-by-side rgpot install does not share that SONAME.

3.3.1 - 2026-09-27

Removed

• Drop leftover AMS debug cout dumps.

• Drop leftover Eclipse Created-on banners from Obs, CuH2, AtomsConfiguration, and GPRDimer unit tests.

• Drop leftover Eclipse auto-generated ctor/dtor TODOs in Obs, CuH2, AtomsConfiguration, and GPRDimer unit tests.

• Drop leftover MPI Recv and Table.write debug prints.

• Drop leftover Meantime debug stdout from each AKMC step.

• Drop leftover Python 2 from builtins import input in AKMC. Live reset/restart input() prompts stay.

• Drop leftover canary os.system tee next to the live redirect. Keep the redirect and pass/fail stdout.

• Drop leftover commented AMS debug readFile and validate_order. Live force and real TODO notes stay.

• Drop leftover commented BasinHopping force-call and recentRatio debug dumps. Live basin-hopping accept and displacement adjust stay.

• Drop leftover commented Broken AMS test_one_pt_ams_dimer. Live Morse test_one_pt stays.

• Drop leftover commented BytesIO alias, pylab import, and Python 2 canary energy prints. Live StringIO I/O and pass/fail prints stay.

• Drop leftover commented DEBUGGER recieveFromSystem and old AMS force bodies. Live AMS force path stays.

• Drop leftover commented EONC_LOG_TRACE in GPSurrogateJob. Live traces stay.

• Drop leftover commented Matter operator==/!= and unused force/max variance stubs. Live compare() and getEnergyVariance stay.

• Drop leftover commented Matter::setPotentialEnergy stub. Live getPotentialEnergy stays.

• Drop leftover commented MinimizationJob include, unused status/returnFiles/epot_hop/earr, and fSPDLOG/fprintf debug dumps in GlobalOptimizationJob.cpp. Live hop and escape path stays.

• Drop leftover commented PR search-result and superbasin basin calls. Live allocate_process_id, make_basin_from_sets, and _get_filtered_states stay.

• Drop leftover commented Python 2 debug prints and dead max-rate blocks in AKMC state connect and Water displace.

• Drop leftover commented Python 2 prints in eon-state-stats. Live table output stays.

• Drop leftover commented dead branches in the eon server. Live saddle, basin-hopping, superbasin, match, and water-displace paths stay.

• Drop leftover commented dead statements and the rescaleVelocity stub in GlobalOptimizationJob.cpp. Live hop and escape stay.

• Drop leftover commented debug prints in ASE_NWCHEM client.py and blah.py. Live socket force and energy path stays.

• Drop leftover commented debug prints in disconnectivity_graph, gpaw_sp, and mk_cuh2_vid. XXX notes stay.

• Drop leftover commented energy and force printf dumps in MPIPot.cpp. Live MPI pot send/recv path stays.

• Drop leftover commented map_out_pes and plot helpers from the extpot 2D PES test fixture. Live _calculate_landscape stays.

• Drop leftover commented print_help/sys.exit in eon-minimize and the unused compareStru body in disconnectivity_graph.

• Drop leftover commented try/except around AKMC search-result append and Main random_seed. Live append_search_result write and getint stay.

• Drop leftover commented try/except around live dg.draw_minima in disconnectivity_graph. Live draw_minima stays.

• Drop leftover commented unit_system include and ERGS_PER_ANGSTROM2 notes from Water CCL and SPC/E. Unit-system fallbacks stay.

• Drop leftover commented unused members in GlobalOptimizationJob, ReplicaExchangeJob, Hessian, and BondBoost. Live members stay.

• Drop leftover debug stdout of cwd in akmc-gui and bh-gui. GTK unused imports, pathfix, and os.fork daemonize stay.

• Drop leftover pot rank / my_client_rank debug stdout in tools/emt-sp.py. Commented GPAW constructor and live MPI rank split stay.

• Drop leftover register_process debug prints and the commented rate = cur_rate line. Live rate assignment, eq-rate clamp, and process-table writes stay.

• Drop leftover unused PoissonSolver from the GPAW import in tools/emt-sp.py. The commented GPAW constructor stays.

• Drop leftover unused algorithm includes in Obs, CuH2, AtomsConfiguration, and GPRDimer unit tests.

• Drop leftover unused fileio cell reexport and unused imports in server, state, and superbasinscheme.

• Drop leftover unused imports and debug prints in explorer and communicator. MPI Recv and harvest stay.

• Drop leftover unused imports in ASE_NWCHEM blah.py, client.py, and server.py. Live socket energy and force path stays.

• Drop leftover unused imports in analyze, basinhopping, config, eon.init, and eon_kdb.

• Drop leftover unused imports in tools disconnectivity_graph, emt-sp, evt, eon-ini, eon-config-docs, and mk_cuh2_vid. pathfix side-effect imports and the commented GPAW constructor imports stay.

• Drop leftover unused imports in unit tests. Keep pytest where mark, raises, approx, fixture, or importorskip is used.

• Drop leftover unused numpy in tools/modedot.py and unused PoissonSolver plus ase.utils.devnull in tools/gpaw_sp.py.

• Drop leftover unused sys, atoms, and numpy imports from tools.

• Drop the leftover commented eonclient path in eon-minimize. The live call is already eonclient on PATH.

Added

• [Xtsci] method = lbfgs honours [LBFGS] lbfgs_step and lbfgs_precon. The pair matrix stays in eOn; xtsci applies it as (H_0) or as a Newton/RFO step.

• [Xtsci] qn_step and precon are first-class (Cap’n Proto, INI, JSON, YAML, Pydantic). They drive xts_solver_set_qn_step and the host pair Hessian. Native [LBFGS] is unchanged; lbfgs_step / lbfgs_precon remain a fallback when the Xtsci fields are left at their defaults.

• ci_mmf_restore_unhelpful (default false) restores the climbing image after an alignment reject or force increase. The Frontiers OCI-NEB article (doi:10.3389/fchem.2026.1807063, Algorithm 1) restores only on positive curvature; that is the default.

• opt_method = xtsci selects the xtsci-optimize engine. [Xtsci] method picks the solver inside it (L-BFGS, BFGS, SR1/SR2, Newton, RFO, NLCG conjugacies, Adam, PSO). YAML, JSON, pydantic, and SSOT defaults carry the same catalog.

Developer

• The NEB regression tests run between two real LJ13 minima joined by a saddle. The climbing-image case converges at -O2 with NDEBUG as well as at -O3.

• The manylinux wheel build retries the Cap’n Proto download and uses the GitHub tag archive when capnproto.org drops the TLS handshake.

• Unit tests alias eonc types, including EigenmodeStrategy, from the shared test header.

Changed

• AKMC displacement sampling, eon-minimize, and GPAW single-point leftover path I/O uses pathlib.

• AKMC leftover path I/O uses pathlib. Reset and restart share remove_tree_and_empty_parents and info_txt_path.

• AKMC movie leftover path I/O uses pathlib for POSCAR, graph.dot, and dynamics.txt.

• AKMC, BH, PR, and escape-rate pass the config path into ConfigClass.init_from_cli instead of rewriting sys.argv. Dead commented RandomStructure in basin hopping is gone.

• AKMC, PR, and escape-rate write info.txt through fileio.write_info_txt. The unused CatLearnPot variance member is gone.

• AKMC/BH helper leftover path I/O uses pathlib for search-stats, state-neighbors, fillkdb, and GTK glade loaders.

• AKMCState leftover path I/O uses pathlib for badprocdata, jobs.tbl, search results, and saddle/mode files.

• ARTnSaddleSearch accepts an injected IARTnResource. Production still uses the process-default ARTnResource singleton.

• AS-KMC leftover path I/O uses pathlib for askmc_data.txt, askmc_processtable, and recycle_path.

• Amsel superbasin split cache leftover path I/O uses pathlib.

• Basin hopping leftover path I/O uses pathlib for states, wuid.dat, bh.log, and the lockfile.

• CI concurrency groups use the PR number or git ref so develop pushes cancel superseded runs.

• CNA helpers drop the leftover unused brute argument. Importing atoms no longer mutates sys.getrecursionlimit.

• Communicator leftover path I/O uses pathlib: harvest, unbundle, scratch, script submit, and client lookup.

• Config leftover path I/O uses pathlib for config.yaml, config.ini, and the PRNG state file.

• Develop package version is 3.3.1.dev0 after the 3.3.0 release, not 3.2.1.

• Explorer leftover path I/O uses pathlib for debug results, KDB scratch, jobs.tbl, incomplete saddles, and explorer.pickle.

• GPRHelpers remaps Matter Z to dense 0..n-1 because gpr_optim pairtype is indexed that way. Stale Fortran meson and statelist TODOs are gone.

• Helper tools leftover path I/O uses pathlib for pathfix, INI dump, GTK glade paths, CNA walks, and state-stats.

• IRACompare accepts an injected IIRAResource. Production still uses the process-default IRAResource singleton.

• In-process communicator leftover path I/O uses pathlib for config.ini basename.

• LockFile leftover path I/O uses pathlib. Stale and missing locks unlink with missing_ok.

• MetatomicDynPot accepts an injected IMetatomicLoader. Production still uses the process-default MetatomicLoader singleton.

• Parallel replica and escape-rate leftover path I/O uses pathlib. info.txt goes through info_txt_path.

• Path-based con reads and writes keep a copy in a readcon-db corpus next to the run. readcon still parses the file, and the con file remains the copy the client reads.

• Potential accepts an injected IPotRegistry. Production still uses the process-default PotRegistry::get() singleton.

• Queue estimates use the communicator’s construction-time bundle size. KDB insert on a finished state writes a marker so a second explore does not insert again. Table float precision is a constructor argument.

• Replace leftover Python 2 file() in config docs and xrange in mastereqn. Keep the live open() write; do not convert tools/toykmc.

• Replace leftover Python 2 print and SafeConfigParser in tools and dynamics analyze with configparser.ConfigParser and print().

• Repo tests, get_version, and Sphinx cache leftover path I/O use pathlib.

• RgpotAdapter accepts an injected IPluginLoader. Production still uses the process-default PluginLoader singleton.

• Server leftover path I/O uses pathlib for potfiles, cwd listing, and output/output_old.

• State leftover path I/O uses pathlib for procdata, reactant, processtable, and staging.

• Superbasin and KDB path I/O uses pathlib. Dead commented get_superbasins is gone.

• Superbasin merge always archives basins to storage. The MPI client stop path is STOPCAR. ASE-NWChem warns when a non-identity cell is ignored.

• Superbasin recycling leftover path I/O uses pathlib for recycling_data.txt, current_sb_states, and saddle_suggestions.

• The server package imports numpy as np. Optional annotations in modules that already postpone evaluation are written as X | None, and the config loader’s section local is section_name.

• load_potfiles, make_bundles, and the leftover reset/KDB path helpers use pathlib instead of os.path.

• make_bundles loads potfiles. load_potfiles skips subdirectories under the pot directory, not same-named entries in the CWD.

• clang-format wrap leftover after the GlobalOptimization debug drop.

• eon package tests leftover path I/O uses pathlib for canaries and ConfigClass injection tests.

• eon-schema MainConfig leftover path I/O uses pathlib for jobs, states, potfiles, kdb, and superbasin defaults.

• eonc::Runtime is a move-only composition root. ClientEON constructs one and moves it into Job; Job builds the Potential from runtime.pots() rather than PotRegistry::get(). Catch2 uses a local Runtime. NEB image-force std::execution::par is unchanged.

• fileio leftover path I/O uses pathlib for atomic_write, savecon, Dynamics, and Table.

• gitversion and wheel-repair leftover path I/O uses pathlib. os.path.relpath stays for the meson dist printout.

• gitversion leftover meson-dist printout uses Path.relative_to.

• pyeonclient Session uniquely owns Runtime. make_job(params, session) borrows it with nanobind keep_alive so Jobs cannot outlive the Session. write_potcall_summary uses session.pots(); CLI Job borrows a stack Runtime. Stock nanobind, not jaxlib nb_class_ptr.

Fixed

• NEB now rejects reactant and product structures with different atom counts before the Eigen path interpolation, instead of aborting inside the subtract. (#540)

• A blank or non-integer nperp_limitation token returns an ARTn error and calls artn_destroy instead of throwing out of the search.

• AKMC movie fastest_path(full=True) sorts graph nodes with a Python 3 key instead of a Python 2 cmp.

• ARTn sends each lattice row to pARTn as a box column. A non-orthogonal cell is no longer transposed, so the periodic minimum image uses the cell eOn stored.

• ARTnSaddleSearch::run returns an error when has_sad is set but tau_sad was not retrieved, so the pre-convergence geometry is not recorded as the saddle.

• AtomicGPDimer restores floating-point traps if execute throws, so a saddle search that catches the failure does not leave later force calls running with traps masked.

• AtomicGPDimer::compute copies the Matter it is given before it builds the GP midpoint. A saddle search that moves the geometry after the solver is constructed is the geometry that is searched and written back.

• Basin hopping total_normal_displacement_steps subtracts only quench displacements that ran. A stop_energy break no longer counts configured quenching_steps that never executed, which could make that total negative.

• Basin hopping includes the last interior NEB image when it chooses the climb and when it writes neb_initial_band.con. An image count below two no longer reads the bead before the reactant.

• Basin hopping keeps the minimized energy as the Metropolis reference when a jump does not relax. With significant_structure off, the raw jump energy is no longer stored and that geometry is not saved as the global minimum.

• Basin hopping linear and quadratic displacement keeps the configured step when every atom sits at the center, instead of dividing by a zero radius.

• Basin hopping skips displacement adjustment when adjust_period is zero, instead of taking the Monte Carlo step modulo that period.

• Basin hopping takes the dimer direction from the minimum-image of (r_3 - r_1) / 2. A bead that crosses the cell no longer starts the climb along a full box jump.

• BasinHoppingJob::getElements records atomic number 118 instead of writing past the element table.

• BasinHoppingJob::run applies exp(-dE/(kB T)) only when the hop is uphill and temperature is positive. Temperature at or below zero rejects that hop instead of dividing by temperature.

• BasinHoppingSaddleSearch::describeStatus labels a Metropolis rejection as Basin hop rejected instead of Initialized.

• BasinHoppingSaddleSearch::run applies exp(-dE/(kB T)) only when the quenched hop is uphill and temperature is positive. Temperature at or below zero rejects that hop instead of dividing by temperature before the test.

• BasinHoppingSaddleSearch::run returns the min-mode climb status, so an unconverged or aborted climb is no longer reported as a good saddle.

• Bond-boost forces follow the minimum-image bond from distance(), so the stored bias force is the gradient of the bias energy in a non-orthogonal cell.

• Classic dimer no longer applies rotation_angle after the torque check accepts a direction, so Dimer::compute returns that orientation with the curvature measured there.

• Classic dimer rotation rejects non-finite forceBatch energies and forces in Dimer::calcRotationalForceReturnCurvature before setComputedPotential. A NaN torque no longer skips every exit check in that loop.

• ClientEON starts the wall clock at each job, so time_seconds is that job’s duration instead of the time since the process began.

• Default L-BFGS takes the two-loop step (ASE). Energy and Grippo accept are opt-in.

• Doubly nudged projection now leaves the perpendicular spring remainder unscaled unless doubly_nudged_switching is set.

• Drop leftover dead rotation comments and the TShacked marker in atoms.py. point_energy_match returns False on mismatch.

• Dynamics saddle search keeps the detecting configuration when refineTransition has no MD snapshots, instead of indexing an empty list.

• Dynamics saddle search no longer takes a modulus of a state-check interval that rounds to zero steps. The interval is at least one dynamics step, so the state check still runs.

• Dynamics saddle search with bias_potential set to bond_boost now calls setBiasPotential and BondBoost::advance once per MD step, so the trajectory includes the bond-boost force.

• DynamicsSaddleSearch::refineTransition returns the first snapshot that does not minimize to the reactant. The reported transition time stays at or after the preceding reactant frame.

• Energy-weighted NEB now sets E_ref to the higher endpoint energy, so springs stay at k_min until an image rises above that minimum.

• FiniteDifferenceJob steps only free axes from Matter::getFixed(atom, axis). A column-4 mask that freezes one axis no longer moves that axis or includes it in the normalized curvature direction.

• Fix leftover TabError in tools/akmc-search-stats.py mixed tabs and spaces.

• GPR dimer check_derivatives is passed to gpr_optim as the strings “true” and “false”. Assigning the bool stored one character, so derivative checks never turned on.

• GPSurrogateJob takes sub_job and linear_path_always from gp_surrogate_options and builds the surrogate with helpers::create::makeSurrogatePotential. Later bands turn climbing images off with neb_options().climbing_image.enabled and restart from copied Matter images.

• LBFGS compiles again: leftover OCINEB helpers::maxAtomMotionAppliedV calls now use eonc::geometry:: like FIRE, CG, SD, and Quickmin.

• Langevin dynamics no longer moves a coordinate frozen on only one axis. Dynamics::langevinVerlet applies noise and the position step per free axis, not only when the whole atom is fixed.

• MSVC Cap’n Proto builds force-include a guard that undefines windows.h interface after Meson’s cl.exe sanity check.

• Matter passes a zero cell into force() when periodic boundaries are off. GFN2 treats a stored .con box as PBC and dies on multipoles.

• MonteCarloJob lists out.con once. Bundling can rename that file instead of failing on a second copy that is already gone.

• NEB L-BFGS no longer runs the auto_scale finite-difference H0 probe. That probe treats the projected band force as a potential gradient and never built memory. The LJ13 job fixtures also pin climbing_image off, matching the unit tests.

• NEB minimize_endpoints defaults to false, matching SVN and the pyeonclient NEB API. The old true default added endpoint relaxations to total_force_calls on fixtures that never asked for them.

• NEB file initialization keeps the endpoint structures passed to NudgedElasticBand. Frames listed in initial_path_in still fill the interior images, and minimize_endpoints_for_ipath is no longer overwritten by those files.

• NEB peak modes on the first and last spline intervals now use the geometric endpoint tangent. setup_mmf_peaks no longer writes the neighboring interior tangent or a zero mode for those peaks.

• Nose-Hoover (noseHooverVerlet) counts each unfixed Cartesian axis in nFreeCoords. A partly fixed atom is no longer thermostatted as three degrees of freedom.

• OCINEB restores a climbing image that finishes a min-mode walk at positive curvature before a band force under force_tolerance can mark the NEB converged.

• OCINEB scores MMF help on the walked image so a CI-index hop cannot rewind a finished saddle step.

• OH-TST planes_used now counts every sampled plane, including the plane where the progression converges.

• OH-TST keeps the unwrapped free coordinates in samplePlane when the next hyperplane is loaded. A reload after setPositionsFreeV shifted a coordinate by a lattice vector, and the projection then left the constrained restart.

• OH-TST symmetry products now minimum-image and remove rigid drift with the same fixed point as the primary guideline. symmetryReflect compares those rays in one frame.

• OHTSTJob::symmetryReflect measures distance to each product half-line, so a configuration behind the reactant uses the distance back to that endpoint instead of the perpendicular distance to the infinite line.

• On Apple x86_64, the SIGFPE handler now sets MXCSR exception masks in the saved context, and disableFPE writes those masks back instead of restoring the unmasked environment from enableFPE.

• PMF scan places OH-TST planes on a uniform grid from the reactant to the product. The scan no longer starts at s_init and then steps backward.

• Parallel replica dephase keeps the bond-boost pointer when it copy-assigns the trajectory back to the saved state. Later dynamics steps still apply that bias while the boosted clock advances.

• Parallel replica keeps its bond boost across the copies it makes of the trajectory: the dephase reset and, when stop_after_transition is false, the transition structure. Matter::assignKeepingBias does the copy.

• Parallel replica reads stop_after_transition. When it is false, the dynamics keep the full step count after the first new state instead of stopping.

• PotentialBase::computePt includes a platinum Lennard-Jones pair when one atom is fixed, so the fixed atom still contributes energy and force on its movable neighbor.

• Same-size Matter::resize keeps .con column-5 atom ids. A leftover duplicate resize body was resetting them to 0..N-1.

• Session supports Python weak references, so a Job that borrows it keeps the Session alive after the Python name is dropped. The installed ARTn saddle header no longer names the feature build flag.

• Shared plugin TUs include Parameters.h so accessors compile when Potential.h only forward-declares Parameters. Potential constructors that take Parameters live in eoncbase so plugin .so files do not need eonclib. Duplicate leftover field-style option blocks are gone. readcon is resolved before plugin subdirs so SocketNWChem links rkr_*.

• The MPI build links the MPI potential against MPI.

• The MPI client copies client_quill.log and client_traceback.log into the job directory before return_files.dat names them. Harvest keeps those logs instead of dropping names that exist only in the launch directory.

• The MPI client stops when enterJobDirectory cannot enter the job directory. It no longer runs that job in the launch directory or sends the path back as a finished job.

• The MPI client waits for the job-path send to finish before that buffer is released.

• The NEB no longer puts a projected force on fixed atoms. When the endpoints placed a fixed atom differently, the spring force along the tangent leaked onto it, and the norm convergence metric counted it.

• The Windows floating-point exception filter stores the exception-mask bits in ContextRecord->MxCsr before EXCEPTION_CONTINUE_EXECUTION. _controlfp_s alone does not survive that restore, so the faulting instruction trapped again.

• The Zhang-Xu LBFGS quadratic test writes optimizer options through ParametersLoadAccess, matching the private Parameters layout.

• The release dry-run accepts a development version when pyproject.toml and pixi.toml match. A .devN tip does not need a changelog section yet.

• Wheel and sdist builds pin nanobind to >=2.2,<3, matching pixi. Unconstrained pip install -U nanobind pulled 3.0, which dropped nb::detail::keep_alive used by pyeonclient tie_lifetime.

• Wheel repair searches for librgpot.so* next to libreadcon_core, so pyeonclient can import without a host librgpot.so.3.

• When NEB max_iterations is not positive, DynamicsSaddleSearch::run passes the initial-band tangent to MinModeSaddleSearch instead of an empty mode.

• With no equilibration samples (hyperdynamics rmd_time of zero), BondBoost::advance and BondBoost::boost measure the current tagged-bond lengths before BondSelect. The bias is no longer stuck at dvmax, so a stretched bond can turn the hypertime factor off.

• AtomicGPDimer passes a zero cell into the GP force box when Matter::getPeriodic is false. XTB no longer treats a stored .con box as periodic boundaries on those force calls.

• ConjugateGradients::line_search clamps a negative max move by its length, so a forward secant step stays forward when bowl breakout passes that value.

• Lanczos::compute diagonalizes the current Krylov block when the residual vanishes, instead of returning the Ritz pair from the previous subspace.

• NudgedElasticBand::updateForces passes a zero cell into forceBatch when Matter::getPeriodic is false. Interior images no longer look periodic to a batch evaluator that infers boundaries from the stored box.

• Parameters::load accepts an AMS or AMS_IO file whose engine is DFTB with resources, or FORCEFIELD with only the engine.

• Parameters::load, Parameters::load(FILE*), and Parameters::load_ini_text return failure when the INI parser reports an error, including a missing ‘=’ or a line past the maximum line length. A partial config is not applied.

• ARTnSaddleSearch::run holds library_mutex from artn_create through artn_destroy, including force calls between artn_step invocations, so another search cannot touch the process-global Fortran state mid-run.

• structure_to_matter treats Structure.periodic as a 3-vector. bool(array) is a ValueError; use np.any.

• eonclient -p no longer starts serve mode unless a serve flag is set. A structure file on that command is read as a normal client run.

• geometry::identical searches for a one-to-one pairing inside distanceDifference instead of locking an index-aligned pair that blocks the only valid permutation.

• matter2xyz append inserts a newline when the existing file does not end with one, so the next frame atom count starts on its own line.

• opt_method = xtsci holds an xts_solver_t session. Each step() is one outer iteration, so L-BFGS pairs and NLCG conjugacy survive the host loop.

• pushApart steps along the minimum-image separation. A coincident pair takes a finite step instead of dividing by zero.

• unbundle restores compressed names that bundle writes, so results_3.con.gz is copied back to results.con.gz instead of being skipped.

• findSplineExtrema stores one stationary point when the cubic coefficient is zero or the cubic root is repeated, instead of writing that point twice.

• fpe_signal_handler no longer returns into an x86 integer divide. FPE_INTDIV and FPE_INTOVF step past the faulting instruction instead of only masking floating-point traps, so sigreturn does not re-execute that divide.

• get_param and get_runparam now take a void** out pointer. pARTn allocates the value and writes that pointer through the argument, so a void* out parameter stored it at the wrong address.

• pyeonclient wheels carry the vendored rgpot as libeon_rgpot.so.3. Under the shared librgpot.so.3 SONAME, the pip rgpot wheel bound to that copy. Importing rgpot after pyeonclient then failed on a missing rgpot::D3Pot symbol, and the metatomic engine was not found.

• sortedR compares neighbor shells by the number of distances in the shell, not by the atom count. An identical geometry, including a homonuclear diatomic compared with itself, now matches.

• xtsci rematch now uses the native per-atom max_move clip, rigid-mode projection, cautious pair filter, and extra-updates instead of a Euclidean cap on the whole 3N step.

• xtsci sessions evaluate energy and forces in one potential call. [Xtsci] accept (none / energy / nonmonotone) replaces a mandatory energy backtrack that spent extra force calls on every Newton / pair step.

3.3.0 - 2026-09-15

Removed

• Drop the unused with_pybind11 meson option and the dead EONMPIBGP include in ClientEON. Feature-gated public headers remain.

• OCI-NEB MMF triggers on the adaptive threshold only. The floor is 2 * force_tolerance; ci_mmf_after is no longer an extra OR.

Added

• AKMC CI builds with -Dwith_pyeonclient=true. XTB CI includes windows-2022. Lanczos docs name stiff EAM/Tersoff. load_ini_text and AMSEl on_error / fallback_single are in.

• Document Potential/Matter thread-safe interface. matter_to_ase copies atom ids. Testing guide names Baker 05/17/19/20 as the reduced parameter-check set.

• External C++ includes eon/api.h for Matter, Parameters, NEB, IDPP/SIDPP, and dimer. Job headers stay internal.

• Hash-pin GitHub Actions (zizmor), add ci_zizmor.yml, and run zizmor from prek.

• Hessian, Prefactor, and Dynamics jobs write results.dat from JobResultEnvelope. LocalInProcess builds CON text only when a caller reads min.con.

• IRACompare.cpp is always compiled so NEB can call match_endpoints when IRA is off (the methods stub). matchArrays plus pyeonclient.ira_match let eon.atoms.rot_match use IRA when the wheel has it, else Kabsch.

• In-process process_search puts saddle.con on the result record so explorer can register the saddle without an eonclient subprocess.

• JobResult Cap’n Proto names historical termination_reason integers as TerminationCode (0–21). statusCode stays the results.dat integer.

• JobResult has a body union (minimization / neb / processSearch). Shared scalars stay on the outer struct.

• Jobs register themselves into a table; makeJob looks up the factory. AMSEl persist_split writes a sidecar so a resume does not re-split the basin.

• LAMMPSPot accepts an injected ILammpsLoader. Production still uses the process-default LammpsLoader singleton.

• LocalInProcess also dispatches dynamics, Monte Carlo, basin hopping, Hessian, prefactor, and finite-difference. Those C++ jobs write results.dat from JobResultEnvelope.

• LocalInProcess dispatches minimization, point, process_search, and saddle_search. PointJob writes results.dat from JobResultEnvelope. JobType.OH_TST is bound.

• Meson -Dstdpar=cpu|gpu turns on nvc++ -stdpar=multicore|gpu for the NEB image-force std::execution::par path without TBB. -Dstdpar=gpu also passes -gpu=cc80,cc90 (override with -Dstdpar_gpu_cc=). GCC/Clang still use -Dwith_parallel_neb=true. Morse atom loops stay on the host. QMC_GPU is a QMCPACK flag, not an eOn option. eOn skips the cmake Highway wrap on nvc++ so Meson does not import hwy_list_targets (PGICompiler has no get_pie_args).

• Metatomic forceBatch tries a single model.forward over N systems and falls back to sequential force() if that path throws.

• MetatomicPotential can clone the already-loaded Torch module for per-image NEB instances without a second disk load.

• MinimizationJob fills an in-memory JobResultEnvelope and writes results.dat from it. The Cap’n Proto schema is unchanged.

• NEB writes peakNN_pos.con and peakNN_mode.dat for every interior spline maximum (setup_mmf_peaks, default on). Those files are the dimer seeds for a follow-up saddle search.

• NEB, dimer, and process search query pot thread-safety through potAllowsSharedInstance in eon/PotCapabilities.h. Virtual isThreadSafe stays the override point.

• Optional ASE batch_calculate sets supportsBatchEvaluation. Without the hook the pot stays on sequential force(). From #409.

• Parameters can load INI from a memory string (load_ini_text). LocalInProcess uses that instead of a temp file. Docs floors drop the invalid rgpycrumbs analysis extra.

• RgpotAdapter rejects a kernel whose caps().reentrancy is not SharedInstance, PerInstance, or ProcessSerial. OCI-NEB hybrid-dimer defaults stay ci_stability_count=5 and max_steps=1000.

• SaddleSearchJob and ProcessSearchJob accept an in-memory Matter via runFromMatter, so pyeonclient does not have to write pos.con first.

• With write_movies = true, saddle search writes mode_000.dat, mode_001.dat, … so dimer mode-evolution plots have per-iteration eigenvectors.

• Potential::layoutFlags() reports in-process vs cwd vs subprocess execution. ExtPot sets Subprocess and NeedsWorkingDirectory.

• RGPOT_NWCHEM_ENGINE is accepted as an alias of RGPOT_NWCHEMC_ENGINE. Devdocs list which pots live in rgpot vs in-tree.

• [Nudged Elastic Band] match_endpoints = true rigid-aligns and permutes the reactant onto the product with IRA before interpolation.

• [Potential] thread_safe = false keeps a shared Potential serial. EAM and EMT refuse shared-instance threading (cell lists / ASAP state).

• eon.geometry.pbc uses minimage wrap_many for packed displacements. neighbor_list_pairs returns vesin ijS rows without unique-index MIC reduction.

• eon_schema.jobs encodes and decodes JobResult dicts (job_result_dumps / job_result_loads). pycapnp is used when installed; otherwise JSON of the wire field names.

• job = test now constructs TestJob. The implementation is linked into eonclib and writes results.dat (OK/FAIL/SKIP per pot). Missing pos_test.con is SKIP, not a link error.

• pyeonclient exports ProcessSearchJob with min1/min2/saddle/prefactors. LocalInProcess uses that job when present. NEB writes results.dat from JobResultEnvelope plus image keys.

Changed

• AMS and AMS_IO look up symbols with readcon::z_to_symbol. GPR dropped an unused private element table.

• ASE and CatLearn pot headers no longer inject using namespace pybind11::literals. The _a literal is function-local in the TUs.

• ASE-NWChem basis and memory come from [ASE_NWCHEM] instead of hard-coded 3-21G / 2 gb. Defaults are unchanged.

• ConjugateGradients, FIRE, LBFGS, Quickmin, and SteepestDescent headers no longer inject using eonc::…. Implementations live in namespace eonc.

• Dimer opt_method is OptType (CG/LBFGS/SD) like the other optimizers. INI still accepts cg/lbfgs/sd.

• Dimer, ImprovedDimer, Lanczos, Davidson, LOR, LowestEigenmode, MinModeSaddleSearch, and EigenmodeStrategy headers no longer inject using eonc::….

• IRA CShDA marks unassigned pairs with huge() instead of 999.9. A 38-atom core-plus-outlier pair that used to return a non-bijective permutation now assigns every index.

• Matter, Parameters, NEB, and the RPC server TUs live in namespace eonc. NudgedElasticBand.h, Optimizer.h, and ServeRpcServer.h no longer inject using eonc::….

• Minimization, Point, Dynamics, Hessian, FiniteDifference, and Test job headers no longer inject using eonc::… into the global namespace. The job factory qualifies eonc:: at the register site.

• More public headers (dynamics, Hessian, IDPP, IRA, saddle methods) no longer inject using eonc::…. Implementations live in namespace eonc.

• NEB image forces use std::execution::par when built with -Dwith_parallel_neb=true (TBB), instead of one raw thread per bead.

• Option structs live in ParametersOptions.h. Parameters.h is the loader aggregate and keeps Parameters::neb_options_t aliases.

• Parameters option groups now expose only const accessors. INI/JSON loaders and bindings write through ParametersLoadAccess; MPI comm/rank use dedicated setters so ParametersMpi.h no longer takes the address of a temporary.

• Pot headers inherit eonc::Potential directly. Potential.h no longer injects using eonc::Potential.

• Production client sources no longer use using namespace eonc::helpers. Call sites are qualified.

• Public headers no longer include mpi.h or gate EigenmodeStrategy on WITH_GPRD. Potential and Matter forward-declare Parameters. Optional MPI helpers live in ParametersMpi.h.

• Public headers no longer inject using eonc::Matter, Parameters, or the BaseStructures enums. Helper rng/geometry names are called as eonc::rng and eonc::geometry. Potential::get_ef is defined on eonc::Potential.

• Python CNA now builds unique-index adjacency from vesin ijS pair arrays. The local element table is a Z-keyed radius/color overlay; symbol and Z lookups stay on readcon.

• Python CNA uses the same 0/1/2 labels as the C++ client (fcc/hcp/other). Common-neighbor lookup is a set membership test that keeps neighbor-list order.

• Remaining job headers no longer inject using eonc::…. Implementations live in namespace eonc. Job.h still has the using eonc::Job alias.

• ServeMode, BondBoost, Rgpot, and AMS use std::ranges::transform for case folding instead of std::transform.

• SocketNWChem looks up symbols with readcon::z_to_symbol instead of a private element table. The last owned file-scope using directives are gone.

• The ASE potential maps incoming C arrays with Eigen::Map, matching ASE-ORCA, instead of building pybind11 shape vectors by hand.

• The client Fortran compile no longer passes -w. Named -Wno-* flags stay. Water displacements are no longer a fake Displace subclass.

• The command-line parser qualifies Argum types instead of a file-scope using-directive. Helper file helpers use std::string / std::ifstream.

• The eon-schema rgpycrumbs extra matches the tree pin (>=1.10.4). Dimer docs mark LOR as added in 2.17.2.

• Water and Water_Pt potential TUs wrap implementations in namespace forcefields instead of file-scope using namespace.

• Job.h and ServeMode.h no longer inject using eonc::… into the global namespace.

• Parameters::load, load_ini_text, and load_json take std::string_view instead of owning std::string.

• Parameters now has a private load-state Impl (last_load_source / last_load_error). Option-group layout stays in the installed header and is not ABI-stable. Matter still exposes Eigen.

• Parameters option groups are private. Callers use main_options() and the other accessors. load / to_json still round-trip.

• Potential::force has a std::span overload that checks sizes before the raw C-array virtual. Matter, get_ef, and surrogate get_ef_var use it. Fortran/FFI loaders keep the pointer API.

• eon.atoms.atomic_number / symbol_for_z call readcon’s Python symbol_to_atomic_number / atomic_number_to_symbol (0.14.9+). The ctypes probe of rkr_symbol_to_z is gone.

• eon.atoms.atomic_number / symbol_for_z prefer readcon for Z/symbol. The local table keeps radius and colour for viz.

• eon.geometry.neighbor_list takes vesin’s per-axis periodic flag. Structure.periodic defaults to all-periodic and is used when the argument is omitted.

• eon.geometry.pbc accepts a bool or length-3 mask so a free axis is not minimum-image wrapped. Partial masks use the numpy wrap.

• import eon no longer imports the AKMC server. eon.server loads on first access.

Fixed

• NEB minimize_endpoints defaults to false, matching SVN and the pyeonclient NEB API. The old true default added endpoint relaxations to total_force_calls on fixtures that never asked for them.

• A failed hessian.dat write now fails the Hessian calculation instead of returning a successful eigen solve.

• AKMCState stores the search count in MetaData and increments it on each append, instead of rereading the search log.

• AtomicGPDimer throws on a null Matter and returns a zero mode if the GP orientation size does not match the free-atom count.

• BGSD uses safe_normalize on the trial mode so an all-fixed Matter does not write NaN.

• Bond-boost returns zero bias when QRR or the selected bond count is not positive, and skips a bond whose length or equilibrium is zero.

• Collective IDPP lastMaxForce uses free-atom residuals only, so frozen pair forces no longer pin SIDPP at max_iterations.

• Collective IDPP leaves a collapsed-image tangent as zero instead of normalize() to NaN.

• Cookbook tests skip unless EON_PET_NEB_ROOT / EON_PET_MAD_MODEL are set. ASE NWChem examples take NWCHEM_COMMAND.

• Dephase and INI dynamics step counts no longer divide by a zero time_step. A non-positive dt throws in the dephase paths and yields 0 steps in INI load.

• Document that XTB NEB is serial: isThreadSafe and needsPerImageInstance are both false because of Fortran restart units.

• Documentation PRs no longer get contents: write. The build job is read-only; deploy is a separate job that runs only on push.

• Dynamics throws on a null Matter. Kinetic temperature, thermal velocities, and Andersen redraws skip when there are no free atoms or a mass is not positive.

• EpiCenters throw on a null Matter instead of dereferencing it.

• Finite-difference displacements use structure_comparison.neighbor_cutoff and reject a zero displacement instead of normalizing it to NaN.

• Fix Matter size/cache bugs, dimer rotate/PBC/fixed-atom mode, NEB/Helper/min-mode signed max-component, FIRE zero-force/max-move, Hessian failed-cache, CG extra step, Dynamics extra MD step, BondBoost 0/dt and listed-atom range, Prefactor frozen atoms in the Hessian filter, GP surrogate target layout, OCINEB/GP-dimer eigenvector size, min-mode status/forcecalls/confine loop, ARTn direction.dat load, process-search pot/mode/displacement, fCallsMin double-count, replica-exchange config swap and per-thread ran2, basin-hopping accept/compare/swap, Nose-Hoover G2/Q2 plus post-Verlet KE, EpiCenters empty-set throw, MonteCarlo Metropolis sign, Quickmin zero-force, HessianJob failed-freq status, TAD/SafeHyper clock clamp and refine bounds, SafeHyper bias binding, BGSD NaN/iteration cap, OCINEB converged_only plus SIDPP collapse, TAD empty product, ParallelReplica first-hop stop, NEB path size, ASE position alias, SD two-point PBC, Prefactor empty moved set, ASE mixed-PBC throw, ASE unknown-constraint throw, readonly forces view, non-finite force reject, LAMMPS extract null and missing in.lammps, setForces mask persist, Structure-Matter atom_id copy, job catch of runtime_error, Lanczos/Davidson max_iterations floor, LBFGS negative s·y skip, LAMMPS neighbor rebuild and restricted cell, and Structure id uniqueness. Morse Pt saddle energy is unchanged; the dimer-rotate fix shifts the FD curvature from -1.014995 to -1.010564 and process-search force calls from 67 to 69.

• GPRD CI no longer reports Success when SUBMODULE_PRIVATE is missing. A gprd-gate job records availability; gprd-integration is Skipped on forks instead of a fake pass.

• IRA compare returns an error on a zero-atom Matter instead of indexing cand1[0].

• LAMMPS setforce follows Matter’s per-axis freeze mask. An atom frozen only in z is no longer free to move in z inside LAMMPS.

• LAMMPS applies fix setforce 0 0 0 to atoms with Atom.fixed set on all three axes before run 1, so buffer atoms stay put during post-saddle minimization.

• MPI stop_clients sends STOPCAR to ready client ranks. AKMC uses that instead of Abort.

• Matter atom accessors throw out_of_range on a bad atom index or axis instead of indexing Eigen unchecked.

• Meson accepts readcon-core 0.14.9. The v0.14.10 wrap tag still reports crate version 0.14.9, so a cargo-c install failed the old >=0.14.10 check.

• Metatomic seeds the Torch RNG from main.randomSeed when rotation averaging is on. NC force blocks must have nAtoms*3 values before reshape.

• Min-mode bowl_breakout does not index atom 0 on an empty Matter and caps the bowl set at nAtoms.

• Minima-hopping Boltzmann accept/reject compares the hop to the current minimum and uses exp(-dE/(kB T)) with params.constants.kB.

• NEB isUncertain now includes the product endpoint (path[numImages+1]).

• NEB spline extrema leave a collapsed endpoint tangent as zero instead of normalize() to NaN.

• Process search opens pos.con through getRelevantFile, throws on a failed load, and compares a copy so Matter::compare cannot translate initial.

• Removed the empty CUH2_POT ifdef from Potential.cpp.

• Replica exchange rejects replicas < 1 and applies the same temperature fallback as the Python bind (low from main T or 300 K, high at least 1.5 * low).

• SIDPP throws if adjacent images collapse. climbing_image_converged_only no longer reports GOOD while the rest of the band exceeds climbing_image_band_slack times the force tolerance.

• Same-size Matter::resize and displacement.con loads keep the reactant’s .con atom ids instead of restamping 1..N.

• SocketNWChem drops a dead i-PI connection and accepts a new NWChem client once, so process_search / AKMC can continue after task scf optimize exits.

• Superbasin make() also calls connect_states on the new basin so intra-basin reverse processes register.

• TAD and SafeHyper rewind to mdBuffer[0] when refine returns 0. SafeHyper skips the boost exponential when T or kB is not positive.

• The SIGFPE continue path now demotes ARM FPCR trap-enable bits (Linux fpsimd_context and Apple neon state). enableFPE on Apple aarch64 sets __fpcr, not __fpsr. x86 MXCSR masking is unchanged.

• The client calls getBundleSize() again and enables bundling when a numbered bundle is present. A thrown int from a job is a failed exit, not a silent continue.

• The point job opens pos.con through getRelevantFile, so pos_cp.con / pos_in.con are used when present.

• The prefactor job writes prefactor_reactant_to_product and prefactor_product_to_reactant, honors getPrefactors failure, and does not assert on an empty moved set.

• Unbundled results name each slot {state}_{first_wuid+slot} so explorer does not add result['number'] to recover the wuid. Reset paths share fileio.remove_tree_and_empty_parents.

• When a superbasin is active, previous_state stays the last KMC hop. Only explore_state is the lowest-confidence basin member.

• XTB sets periodicity from the box diagonal.

• Matter::pbc is the raw difference when periodic boundaries are off. OCINEB does not resample a non-periodic climbing image, so a 3-atom vacuum cell cannot fold to a 60 A reaction coordinate.

• Parameters::load(FILE*) returns 1 on a null FILE, a failed seek, or a negative ftell size.

• Prefactor::getPrefactors returns -1 if any of min1, saddle, or min2 is null.

• WeightedSpring::compute throws if the image index is 0 or past the spring table.

• [ASE_ORCA] charge is read from the INI and passed to ase.Atoms.

• SaddleSearchConfig.displace_atom_list accepts the scalar int -1 used by examples/akmc-al/run_akmc_al.py, as well as [-1] and "-1".

• SaddleSearchJob and ProcessSearchJob apply client_displace_type = listed_atoms (and random / last_atom / least_coordinated) through listedAtomEpiCenter plus a displace_radius / displace_magnitude kick instead of copying the reactant unchanged.

• listedAtomEpiCenter now picks with randomDouble(size) so the last listed (or last free, for lone -1) atom can be chosen. The previous size-1 interval always dropped that last index.

• accuratePES compares max |pred-true| and restores each image’s incoming potential. The unused sqrt(pred^2-true^2) leftover is gone.

• debug_keep_all_results writes results.dat and con payloads from the result dict into debug_results_path.

• eon_matter_to_atmconf throws on a null Matter or when no atoms move.

• eonclient -m does not write a .con when no output path is given.

• geometry::identical will not assign two left atoms to the same right partner.

• getMidSlice keeps the CatLearn order (endpoints, then a two-thirds interior image) and throws if the path has fewer than three images. It is not n/2.

• getRelevantFile no longer throws on a name without a .. _cp / _in suffixes still win when those files exist.

• gh release edit failures are no longer ignored. A stale notes update now fails the release job.

• job = dynamics writes a results.dat envelope (termination_reason, job_type, potential_energy, total_force_calls) and still returns final.con.

• job = structure_comparison loads matter1.con and matter2.con, compares a copy, and writes results.dat (match, distance, per_atom_norm, energies).

• min_mode_method=gprdimer constructs AtomicGPDimer in place in the eigenmode variant. A WITH_GPRD Catch2 case covers that path.

• removeNetForce no longer runs on a one-atom Matter. Subtracting the mean force was identically zero and made a 1-atom NEB report immediate GOOD.

• rgpot_pot now compiles with the same _args as the other pot plugins (WITH_RGPOT). Dead Metatomic n_avg and the empty .cargo/config.toml are gone.

• rotm, get_rotation_matrix, and internal_motion call numpy.cos / sin / arccos. They used to call bare cos/sin/acos and raise NameError. get_mappings uses vector PBC distances.

• unbundle drops only those bundle slots that lack a results* file.

• unbundle only copies *_N.ext when N is all digits and matches the bundle number. Non-numeric suffixes such as pos_final.con are left alone.

• eonclib compiles after the Parameters accessor split: pot TUs include Parameters.h, write-holes use ParametersLoadAccess, and Dimer/ConFileIO qualify helpers and Matter.

• pyeonclient ConFrame export uses mkstemps (exclusive) and releases the GIL while writing. The guessable world-writable temp name is gone.

3.2.1 - 2026-09-13

Added

• Exposed opt_method, neb_opt_method, refine_opt_method, and refine_threshold on pyeonclient Parameters so dimers and NEBs can pick FIRE/LBFGS/CG without a minimization-only path. (#406)

• Added PotType.EXPR so rgpot 3.1 ExprPot can compose named kernels (0.5*lj + d3) from [ExprPot] expression and terms.

• Agreement tests for minimage and linkcell against eOn’s numpy/C++ wrap oracles, LAMMPS minimum_image (ortho and restricted triclinic), and GROMACS pbc_dx.

• Exposed rgpot 3.1 Grimme DFT-D3 and DFT-D4 as PotType.DFTD3 / DFTD4 (potential = dftd3 / dftd4, [D3Pot] / [D4Pot]).

• Exposed rgpot 3.2 MOPACPot as PotType.MOPAC (potential = mopac, [MOPACPot], Expr term mopac). Default model is AM1.

• Standalone doxyYoda C++ API HTML is published at /api-cpp/ next to the Sphinx book.

• The C++ API landing page links to the Sphinx book. pixi run -e docs makedocs writes /api-cpp/ next to the book.

• eon.geometry.pbc uses minimage when installed. neighbor_list_linkcell compares the vesin neighbor list to linkcell.knearest on the same Structure.

Developer

• Pinned potentials-schema wrap to e825c207 so it matches rgpot 3.2.0’s vendored Potentials.capnp.

• Updated wrap pins to current releases: rgpot v3.1.2, readcon-core v0.14.10, Highway 1.4.0, potentials-schema v1.15.1, xtb v6.7.1, plus current IRA and ARTn heads.

Changed

• PotentialConfig now uses parseable pot tokens (socketnwchem, ase_pot, catlearn) and exposes emt_rasmussen and potentials_path. (#414)

• Docs footer credits antics and loads only antics.js.

• Raised Python floors to readcon 0.14.9, vesin 0.6.1, and rgpot 3.1.2.

Fixed

• The C++ API mainpage no longer prints undefined Makefile Doxygen aliases, and the theme header version matches 3.2.0. (#401)

• Fixed IDPP, collective IDPP, and SIDPP path initialization moving frozen atoms when a mover sat next to a constraint. (#410)

• Fixed ASE from_ase wrapping coordinates before PBC was applied, and stopped treating every get_indices constraint as FixAtoms. Positions views are not writeable in place. to_ase attaches SinglePoint energy/forces. NEB keeps the GIL when the pot is not thread-safe. (#414)

• Write hessian.dat even when [Main] quiet = true. Quiet still suppresses the log line; the Hessian job’s artifact is no longer gated on it.

• LammpsLoader::require_loaded now says whether liblammps.so is missing, failed dlopen (glibc), or opened without lammps_open_no_mpi (the eOn plugin is not LAMMPS).

• displace_atom_list is CON file-order. ListedAtoms remaps those rows through the atom_id sort when the raw list is all frozen, so a movable-first active-volume .con no longer raises “Listed atoms are all frozen”.

3.2.0 - 2026-08-16

Added

• The basic-eon workflow runs the tests/ suite on Unix after the client install, including the sh package those process tests import. Job modules import version from the eon package so a source-tree checkout without generated version.py still dispatches.

• User guide for the AiiDA plugin (pip install aiida-eon) at docs/source/user_guide/aiida.md. The communicator page now links there.

• RgpotAdapter reports caps().batched and forwards a band of images to rgpot::forceBatchImpl, so a kernel that evaluates several systems in one call sees the whole band. The wrap tracks the rgpot branch that adds ForceBatch until a tag carries it.

Changed

• Appending a frame to a .con no longer reads and rewrites the whole movie. matter2con(filename, append=true) serializes the new frame on its own and concatenates it, so writing an N-frame trajectory costs N frame writes instead of N(N+1)/2, and per-step movie writers such as dynamics.con, movie.con, and the basin hopping trial movies stop growing quadratically with step count. Output bytes are unchanged: a movie built by repeated appends is identical to the same frames written in one call. A target eOn did not write, or whose size or modification time moved since eOn wrote it, is still parsed once before anything is added, so an unparseable file yields an append error with its bytes intact. Gzip and zstd targets keep the read-and-rewrite path because a compressed member cannot be extended in place.

• C++ wrap and Python pins move to readcon 0.14.5 together, so the client can read the spec-3 files the 0.14 writer emits. x-only constraints round-trip on the client and the Python Structure writer (readcon-core #25). 0.14.5 ships a win_amd64 wheel, so Windows CI does not build the sdist.

• Rewrite the Sphinx user guide, tutorials, install notes, developer docs, and older release pages so they no longer trip the house prose rules (robust, comprehensive, below is, out of the box, highlighting, first-class, ship-as-verb, and the rest of that list). Drop remaining X-not-Y contrast frames: state what each path is.

Fixed

• ASE-NWChem and ASE-ORCA calculators each get a private work directory instead of directory='.', so concurrent LocalInProcess jobs do not clobber each other’s scratch files. Calculator errors throw rather than abort the process.

• Bond-boost hyperdynamics advances the equilibration counter once per MD step. boost() only evaluates the current bias, so ParallelReplica (which also installs the bias potential on the trajectory) and SafeHyper share the same rmd_time schedule.

• Constructing a VASP potential no longer deletes the working directory’s contents. The constructor removed fifteen files, among them WAVECAR, CHGCAR, TMPCAR and OUTCAR, so a client restarted in a directory lost the wavefunction and charge density an ISTART/ICHARG restart reads, along with the previous run’s record; and a potential built after the first force call, as AtomicGPDimer does, deleted FU and NEWCAR out from under a VASP process that was still running. A run now clears only the three handshake files it owns, FU, NEWCAR and STOPCAR, once, immediately before it starts VASP. STOPCAR was never removed although eOn writes it at shutdown, so a client restarted after a clean shutdown handed the new VASP an abort instruction on its first ionic step. Discarding the results and restart files is available as VASP::removeStaleFiles() for a caller that means to start from an empty directory; the examples/akmc-vasp-slurm scripts already do the same removals in shell.

• Copying a Matter leaves biasPotential null. Copy-assign used to skip that pointer, so getBiasForces read an indeterminate BondBoost and the Python suite died on main.

• Library code no longer calls std::exit on a bad convergence_metric, FIRE time-step collapse, or basin-hopping / global-optimization enum typo. Those errors throw, so a Python caller inside gil_scoped_release gets an exception instead of a silent interpreter kill. Unknown enumerated strings are rejected when Parameters is loaded.

• On Windows, ExtPot launches a suffix-less ext_pot wrapper with python. cmd.exe does not honor a shebang, so a quoted path to the script was not a runnable command.

• Potential loaders probe plugin libraries on disk before dlopen, and LammpsLoader no longer loads liblammps just to ask whether it is present. Availability checks therefore skip the banner-printing static initializers those libraries run on load.

• Setting params.write_con_forces on a pyeonclient Parameters object now writes force sections for that Matter. A ConFrameMetadata.write_con_forces value overrides the process-wide flag so two writers can disagree.

• The external-program potential backends (AMS, AMS_IO, ExtPot, VASP) now check every file open, read and shell command instead of trusting them. A missing, truncated or stale result file previously left the force and energy arrays holding whatever they held before, and those values went straight into the optimizer; they now raise an error naming the file and how many atoms were read. AMS_IO also truncates ams_output rather than appending to it, so a failed run cannot hand back the previous run’s forces, and VASP rejects a structure whose species are not in contiguous runs rather than writing a POSCAR whose counts disagree with its coordinates.

• The readcon wrap overlay still cargo-builds both crate types. A generated .c that depends on that custom_target is the ninja order edge, so eonclib does not link before the outputs exist and does not record NEEDED libreadcon_core.so or pass the DLL to MSVC.

• ExtPot and VASP no longer claim to be safe to call from several threads on one instance. Both exchange structures and forces through files at fixed names, so the parallel image evaluation in NudgedElasticBand, ImprovedDimer and ProcessSearchJob had every thread writing one input file and reading one result file, and images took each other’s forces. ExtPot also asks for a potential instance per image, which gives each thread an exchange directory of its own and keeps the evaluation parallel; VASP cannot, since every instance drives the same VASP process through the same files, so its image evaluation runs sequentially. The other external-program backends (ASE, ASE_NWCHEM, ASE_ORCA, LAMMPS) already declared both.

• ExtPot runs the external program in a private exchange directory named extpot_<pid>_<n> instead of the client’s working directory. The exchange files keep their names, from_eon_to_extpot and from_extpot_to_eon, so a wrapper that opens them by relative name needs no change; two clients started in one directory can no longer read each other’s structures and forces. A relative ext_pot_path that names an existing file, such as the default ./ext_pot, is resolved before the command runs; a command line with arguments needs an absolute path to the script, and EON_EXTPOT_RUN_DIR names the directory eOn runs in for a wrapper that has to reach it. On Windows a resolved Python wrapper (shebang or .py) is invoked through python.exe, because cmd.exe will not run an extensionless file. The previous result file is removed before each call, so an external program that exits successfully without writing one raises an error rather than handing back the previous call’s forces.

• PotRegistry is a process-lifetime heap singleton, so a Potential destroyed during interpreter finalization no longer calls into a destroyed registry or locks a destroyed mutex.

• loadposcar accepts both VASP 4 (integer counts after the cell, as the kdb tool writes) and VASP 5 (species names then counts, as saveposcar and the server movie files write). eOn can read back movie.poscar, dynamics.poscar, and the other movie.py outputs it wrote.

• matter2xyz writes extended XYZ: the comment carries Lattice="..." so a reader can reconstruct the cell, and coordinates use 17 significant digits to match the CON path. Appending a frame whose atom count differs from the last frame in the file is rejected and leaves the file unchanged.

• min_mode_method = gprdimer without -Dwith_gprd=true is an error. The search no longer falls through to ImprovedDimer and reports a nonnegative-mode abort.

• savecon(..., w="a") appends one serialized .con frame instead of reading the whole movie back and rewriting it. Gzip and zstd targets still rewrite, because a compressed member cannot be extended in place.

• pyeonclient Structure/Matter conversion looks up every element (H–Og) without importing the server package, and raises on an unknown atomic number instead of writing a fabricated Z79 symbol.

3.1.0 - 2026-08-08

Changed

• pyeonclient 0.4.0: the potentials it exposes come from librgpot now rather than eOn’s in-tree kernels, so the wheels bundle that library and no longer carry per-pot plugin objects. The Python API is unchanged – the same potential names select the same physics.

  Its wheel workflow gained a target input so an upload can be rehearsed against TestPyPI before the irreversible one; a pyeonclient-v* tag still goes straight to PyPI.

Fixed

• Fat MetatomicPotential loads exported PET-MAD (and similar) models on metatensor-torch 0.10.3: scripted modules without _mts_buffer_names no longer trip the mixed-dict .to() walk.

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.

2.17.10 - 2026-07-20

Fixed

• Optimizer file logs (_lbfgs.log etc.) use a durable temp directory so fixture workdir cleanup no longer races Quill file sinks.

• Packaging: Catch2 --allow-running-no-tests for the optional rgpot embed suite so all-SKIP without nwchemc/cpmdc is a clean success.

2.17.9 - 2026-07-20

Fixed

• EDIP OpenMP: zero shared energy/force accumulators under !$omp single before partial reduction (fixes wrong PointJob energies with multi-thread OMP).

• Basin-hopping force-call reference updated for default LBFGS auto_scale path (1692); energy and acceptance assertions unchanged.

2.17.8 - 2026-07-20

Fixed

• Evaluate EDIP serially under OpenMP (Fortran energy reduction race under OMP_NUM_THREADS>1 gave wrong PointJob energies with plausible forces).

• Basin-hopping integration test pins LBFGS auto_scale off so force-call budget matches the SVN reference path.

2.17.7 - 2026-07-20

Fixed

• Un-nest SW CG minimization TEST_CASE in SiPotTest (was inside LBFGS case after packaging SKIP refactor, breaking with_tests packaging builds).

2.17.6 - 2026-07-20

Fixed

• Keep JobIntegrationFixture::runJob a class method after SKIP macro refactor (stray brace broke TEST_CASE_METHOD inheritance under packaging builds).

2.17.5 - 2026-07-20

Fixed

• Catch2 SKIP only via macros in job integration tests (helpers cannot call SKIP).

2.17.4 - 2026-07-20

Fixed

• Fix Catch2 SKIP usage in unit-test helpers (macros / fixture require path) so packaging builds compile with -Dwith_tests=true.

2.17.3 - 2026-07-20

Fixed

• Pin metatomic builds to vesin>=0.6 and rgpot>=2.5.2 so RGPOT libmetatomic_engine and fat Metatomic share a matching VesinOptions ABI (skin/n_threads). Bump the rgpot wrap to the 2.5.2 fix commit.

• Unit tests SKIP when packaging fixtures or optional engines are missing (EON_TEST_SYSTEMS_DIR / EON_POTENTIALS_PATH / nwchemc·cpmdc), so meson test succeeds for installable client builds without local test data.

2.17.2 - 2026-07-17

Added

• pyeonclient 0.3.3: in-memory path_frames / to_conframes for NEB (same stamps as writePathCon), Matter.relax(retain_frames=…) movie frames, and MinModeSaddleSearch.run_retain_frames climb frames. (#pyeonclient-path-frames)

• Dimer supports rotation_backend = classical|lanczos|davidson|lor (Leng et al. JCP 2013 LOR for softest-mode with force translation; Lanczos/Davidson FD min-mode; classical constrained rotation). Climb remains on the dimer path.

• Optional Superbasin gate via amsel.discover_decide_status ([amsel] / amsel_discover_decide) before MCAMC; rejected_no_metastable_basin falls back to AKMC.

Fixed

• HessianJob FD eigen solve: ColMajor copy for SelfAdjointEigenSolver (avoids RowMajor segfaults on partial VTST blocks), finite-force/index guards, optional [Hessian] atom_list intersected with free atoms, and a stable mobile VectorXi copy so moving-atom Hessians match PHVA-class active sets. (#357)

• Release CI installs quill/capnp so eon-akmc sdist and pyeonclient wheels configure cleanly on GitHub runners. (#418)

• Republish pyeonclient as 0.3.4 with the full wheel matrix (abi3 + freethreading) after incomplete 0.3.3 tag CI.

2.17.1 - 2026-07-17

Fixed

• Restore with_pyeonclient and install_eon_server meson options dropped from meson_options.txt in the v2.17.0 merge (configure failed with Option with_pyeonclient does not exist).

• RGPotEngine accepts rgpot >= 2.5 operator() return type std::tuple<double, AtomMatrix, double> (energy, forces, variance).

2.17.0 - 2026-07-17

Added

• Distribution packaging: prefer installed rgpot via pkg-config (wrap pin v2.5.0) and installed nlohmann_json (module with wrap fallback, same pattern as readcon). RGPOT backend=metatomic / backend=xtb dlopen libmetatomic_engine.so / libxtb_engine.so so packaging keeps -Dwith_xtb=false (native XTBPot deprecation warning) and avoids fat metatomic links into eOn.

• SafeMath accepts Eigen 5 EIGEN_CORE_MODULE_H as well as Eigen < 5 EIGEN_CORE_H.

• Add eon_schema.config INI helpers (write_ini, hydrate_ini, unknown_ini_keys, write_models_ini) so tooling can author validated config.ini from L0/L1 without importing eon-akmc. (#eon-schema-ini)

• Parameter field graph for Main, Potential, Optimizer, Structure Comparison, and Process Search is authored in schema/eon_params.capnp (Cap’n Proto L0 SSoT) with INI/JSON adapters; parity tests gate config.yaml / eon.schema against the catalog. (#params-ssot)

• Public step composition for Matter-first NEB/min: write_neb_results, pot_registry_total_force_calls at package root; NEB.find_extrema; GIL release on forces_free / max_force. NebSpec covers EW/CI/OCI-MMF. (#pyeonclient-step-compose)

• Add packages/eon-schema (PyPI eon-schema): vendored Cap’n Proto SSoT copy, optional pydantic API models; full-tree/conda-forge eon path unchanged. (#eon-schema-0.1)

• Add pyeonclient nanobind module (Matter/Parameters/Potential); stable ABI abi3 on CPython 3.12+, free-threaded when Py_GIL_DISABLED; no pybind11. (#372)

• pyeonclient Matter end-to-end: LocalInProcess communicator (comm_type local_lib/inprocess); NbGuard without pybind11; ASE embed polarity documented. (#373)

• Expand pyeonclient: full enums, Job/make_job/run_job_in_directory, Potential.get_ef, ASE to_ase/from_ase and Structure helpers. (#374)

• Standalone PyPI package pyeonclient (pyproject-pyeonclient.toml): abi3 + cp313t + cp314t wheels via pyeonclient-wheels.yml. (#375)

• RGPOT backend metatomic loads libmetatomic_engine.so (thin host path). Native potential = Metatomic is unchanged for conda-forge packaging. (#377)

• Document fat vs RGPOT-dlopen vs ASE metatomic backends; benchmark figure is generated from committed JSON at sphinx-build time (not checked in).

• Add pixi environment mta-bench and mta-backend-bench task to reproducibly build fat/ASE pyeonclient, refresh the metatomic backend compare JSON, and regenerate the docs figure.

Developer

• Multi-flag Codecov coverage for Python, C++, and Fortran (OIDC uploads). (#367)

• CPython server entry points fixed (eon-server → eon.server:main); vectorized atoms neighbor/free-atom hot paths. (#368)

Changed

• Move full job-config pydantic models into eon-schema (eon_schema.config). eon.schema and pyeonclient.models re-export the shared package so both eon-akmc and pyeonclient share one schema surface. (#eon-schema-l1)

• Structure is readcon-backed (ConFrame bridge); geometry PBC/NL via vesin; process-atom selection uses vesin shells. Retires aselite-era atoms container as storage source of truth. (#371)

Fixed

• Release the GIL around Matter energy/force accessors and NEB construction, update_forces, and per-image energy reads so metatomic/torch autograd can run from C++ without deadlocking. Optional torch/cuda.h and torch/mps.h includes (__has_include) so CPU-only pip torch builds compile without CUDA/MPS headers. (#gil-torch-autograd)

• get_process_atoms returns plain Python ints so recycling metadata repr/eval round-trips (no np.int64). (#369)

• Standalone pyeonclient wheels omit eon server package and eonclient binary (install_eon_server=false). (#376)

2.16.0 - 2026-07-03

Added

• Matter exposes PbcConvention (Legacy vs MinimumImage) for position wrapping, selectable without merging the historical v3c_pbcs branch (issue #176). (#176)

• ASE NWChem calculator takes mpi_launcher (mpirun default or srun) for Slurm-friendly execution (issue #193). (#193)

• Metatomic accepts explicit energy_output / energy_uncertainty_output keys when models use non-default output names (issue #215). (#215)

• Metatomic can apply a random SO(3) rotation per evaluation with random_rotation (issue #287), rotating forces back to the lab frame. (#287)

• Metatomic can average energy and forces over n_symmetry_rotations random rotations for approximate O(3) symmetrization (issue #292). (#292)

• Metatomic supports non-conservative forces via non_conservative and variant_force / force_output (issue #296), matching metatomic-ase. (#296)

• In-process RgpotPot for NWChem/CPMD via rgpot (optional build); user guide and examples for BLYP/DFT XC blocks.

• Metatomic exposes deterministic and deterministic_strict knobs for reproducible PyTorch evaluation (JIT profiling off, deterministic algorithms).

• Minimum-mode finding supports min_mode_method = davidson: Ritz subspace with FD Hessian-vector products and residual correction (alternative to dimer rotation minimization and Lanczos).

• The rgpot potential is selectable from the Python configuration layer: eon/config.yaml gains an [RgpotPot] section and eon.schema a matching RgpotPot model, mirroring the C++ INI keys (backend, basis, theory, scf_type, functional, cutoff_ry, charge, multiplicity, engine paths, title, memory_mb, scratch_dir, input_block).

Changed

• Matter::getForcesFree / getForcesFreeV are const, and setPositionsFree applies PBC wrapping like setPositions so free-atom optimizers stay consistent for TIP4P/SPCE and other PBC pots (issue #171). (#171)

• Metatomic potential requests per-atom outputs via sample_kind == "atom" instead of the removed get_per_atom / set_per_atom API. (#362)

• Build against metatomic-torch >=0.1.15 and metatensor-torch >=0.10, including the renamed pip layout (metatensor_torch/) and sample_kind on ModelOutput.

• Morse pair force hot path inlines the potential, uses inv-r scaling, and accumulates energy in a register (cachegrind: ~60% Ir in Morse::force on NEB Morse workloads).

• ConFileIO targets readcon-core 0.13 with a nanobind-ready IoStatus enum (replacing bool returns), bulk/zero-copy builder maps (positions_data / set_*_from_flat / set_atom_velocity), and writeNebPath using ConFrameBuilder::clone() so NEB bands write in one pass without re-reading the movie file per image. Matter exposes atom-index and CON header accessors for bindings.

Fixed

• Min-mode saddle search no longer reports success (STATUS_GOOD) when the climb objective is still unconverged (issue #20). (#20)

• Saddle and process search synthesize a missing displacement.con from pos.con plus saddle_search.displace_magnitude times a unit direction.dat mode (issue #79). (#79)

• Molecular QM potentials (NWChem socket, ASE-ORCA, ASE-NWChem) auto-disable PBC on Matter attach and hard-fail if PBC is re-enabled, preventing wraps from tearing non-centered molecules (issue #188). (#188)

• MCAMC energy_level superbasin scheme tracks a true run-wide minimum energy for increment scaling and no longer passes a generator into min() (issue #212). (#212)

• NEB with initializer = file and initial_path_in loads endpoints from the first and last path frames and no longer requires (or crashes on) separate reactant.con / product.con (issue #278). (#278)

• .con outputs are readable by classic con parsers again: “Forces of Component” sections are opt-in via [Main] write_con_forces (default off) instead of always written. ASE’s eon reader rejects frames carrying force sections, which broke every ASE-based consumer of eOn-written con files (including the atomistic-cookbook eon-pet-neb example). Enable the knob to keep force+energy co-loading for warm restarts; reading force-bearing frames works regardless.

• RgpotPot engine-path environment overrides are backend-scoped: NWCHEMC_LIBRARY / RGPOT_NWCHEMC_ENGINE apply only to the NWChem backend and CPMDC_LIBRARY / RGPOT_CPMDC_ENGINE only to the CPMD backend, so setting both no longer sends the NWChem engine library into a CPMD configure.

• The examples/rgpot_cpmd_blyp example and the rgpot-vs-SocketNWChem comparison scripts now use the shipped in-process [RgpotPot] configuration (backend / functional / cutoff_ry / engine_library) instead of the abandoned potserv host/port keys, which eonclient never read.

2.15.0 - 2026-06-24

Added

• MPI build revived on the C API (with_mpi) after retiring the obsolete C++ MPI bindings path. (#339)

• Runtime-loadable Fortran potentials via dlopen and [Potential] potentials_path (no rebuild to swap pot modules). (#342)

Developer

• Documented and automated the full maintainer release path (cog/towncrier lockstep assert, GitHub tarball release, PyPI, conda-forge/eon-feedstock checklist, incomplete-release recovery). (#343)

Fixed

• Windows/conda-forge builds enable in-tree Fortran including CuH2 (m2w64 gfortran + MSVC C++, 16 MiB stack, xtb MinGW import lib) per feedstock#15. (#15)

• NEB isolates per-image LAMMPS instances on private MPI communicators so parallel image evaluation does not corrupt neighbor state. (#340)

• Process search restores fixed-atom rows after loading displacement.con, avoiding drift of constrained atoms. (#341)

2.14.0 - 2026-04-24

Added

• Structured per-iteration metadata is now embedded directly in minimization, saddle-search, and NEB movie .con files via readcon-core. The legacy minimization and saddle-search sidecar tables remain available temporarily behind write_deprecated_outs = true. (#trajectory-output)

• Added the filin option under [ARTn] to expose pARTn’s input-file name. Empty (the default) leaves pARTn with its post-artn_create sentinel so no file is read; setting it to a path wires that file into pARTn setup and aborts early with a clear error if it is missing. (#334)

Fixed

• Fixed a regression in threaded force-evaluation paths by keeping legacy Fortran-backed potentials out of shared-instance parallel execution. (#332)

2.13.0 - 2026-04-18

Removed

• Removed 9 dead legacy pointer+size math functions from HelperFunctions (dot, length, add, subtract, multiplyScalar, divideScalar, copyRightIntoLeft, normalize, makeProjection), all superseded by Eigen. (#dead-code)

• Removed dead code: INIFile.cpp/h (superseded by inih), legacy CMakeLists.txt (36 files), Makefile/buildRules.mk, old unittest framework (unittests/). Total: ~2600 lines of dead code removed. (#dead-code-removal)

• Removed dead TADJob::saddleSearch() function and associated dimerSearch member (dead since 2013). (#dead-tad)

Added

• Release builds now use -O3 -flto=auto. Optional native_arch and fast_math meson options for -march=native and -ffast-math. (#compiler-flags)

• Added ARTn (Activation-Relaxation Technique nouveau) saddle search method via the pARTn Fortran library, as a complementary explorer for pushing away from minima. Supports two modes: method = artn (standalone, pARTn drives push + internal eigenmode search + relaxation) and min_mode_method = artn (drop-in, eOn’s displacement seeds the mode). Configurable via [ARTn] section with push_step_size, force_threshold, max_iterations, ninit, nperp_limitation, lanczos_min_size, nsmooth, and nnewchance parameters. Requires -Dwith_artn=true at build time. For automated saddle point refinement in production workflows, prefer OCINEB (ci_mmf = true). (#feat-artn-integration)

• Added IRA (Iterative Rotations and Assignments) structure comparison via the libira Fortran library. Provides IRACompare::match() (CShDA + SVD alignment), matchPBC() (periodic boundary conditions), and findSymmetry() (SOFI point group detection). Requires -Dwith_ira=true at build time. (#feat-ira-integration)

• Added Highway SIMD as optional subproject (cmake wrap). When available, -DWITH_HIGHWAY is set so future potential kernels can opt in. Hand-written SIMD kernels for Morse/LJ/EAM are staged for a follow-up release (see the separate highway-simd-potentials developer note); today the main benefit of enabling Highway is compile-time availability of the subproject for those downstream kernels. (#highway-simd)

• Highway SIMD subproject builds and is detected at configure time. SIMD kernels for Morse/LJ/EAM force loops are planned but not yet implemented (core algorithm files already benefit from Eigen’s auto-vectorization). (#highway-simd-potentials)

• IDPP (Image Dependent Pair Potential) path initialization for NEB. (#neb-idpp)

• NEB decomposed into modular strategy-pattern components: tangent, projection, spring force, OCINEB controller, spline extrema, initial paths, and objective function. (#neb-modularize)

• OCINEB (Off-Path Climbing Image NEB): recommended hybrid CI-NEB + Min-Mode Following with Hessian eigenmode alignment for automated saddle point refinement (ci_mmf = true). See Goswami, Gunde, Jónsson, Enhanced Climbing Image Nudged Elastic Band Method with Hessian Eigenmode Alignment, 2026, arXiv:2601.12630. (#neb-ocineb)

• Onsager-Machlup action-based NEB for minimum action paths (onsager_machlup = true). (#neb-om)

• Parallel image force evaluation for NEB (requires TBB, -Dwith_parallel_neb=true). (#neb-parallel)

• Parallel improved dimer gradient evaluation via std::thread. The two dimer replicas evaluate forces concurrently when the potential is thread-safe. std::thread rather than std::jthread for Apple Clang libc++ compatibility. (#parallel-dimer-stdthread)

• Parallel NEB force evaluation via std::thread. Each image evaluates its potential concurrently when the potential is thread-safe. Achieves 2.5x speedup over SVN baseline on 5-image NEB with Morse potential. No external dependencies (replaces TBB-based std::execution::par). Apple Clang’s libc++ does not yet ship std::jthread, so the implementation uses std::thread with explicit exception-safe joins. (#parallel-neb-stdthread)

• ReplicaExchangeJob now runs replica MD steps in parallel via std::thread when parallel=true and the potential supports it. Each replica gets its own potential instance when needsPerImageInstance() is true (e.g. ML potentials). (#parallel-replica-exchange)

• JSON serialization for Parameters via nlohmann/json <https://github.com/nlohmann/json>_. New load_json() and to_json() methods enable programmatic configuration for library usage and RPC transport. (#params-json)

• Added readcon-core v0.7.1 as a Meson subproject for .con/.convel file I/O. The Rust FFI library replaces the hand-written C FILE*-based parser with an mmap-based reader and a type-safe ConFrameBuilder/ConFrameWriter for output. Requires Rust >= 1.88 and cbindgen >= 0.29 at build time.

Developer

• Planned: approval tests (snapshot-based regression) and fuzz testing for parser robustness (INI, .con format, command line arguments). (#approval-fuzz-testing)

• Competitive NEB benchmarking framework (private repo HaoZeke/eon_benchmarks). Compare eOn against ASE, ORCA, OPTIM, ARTn, NWChem, Sella on Baker test set and Pt surfaces. Snakemake workflow with reproducible environments. (#competitive-benchmarks)

• Added developer design docs for Parameters decomposition and NEB modularization. Updated testing inventory, user guides (minimization, dynamics, NEB, saddle search), and added algorithm selection guide with rgpycrumbs examples and atomistic-cookbook links. (#docs-design)

• Replace remaining new[]/delete[] in EMT/Asap (NeighborList.cpp, EMT.cpp) with std::vector. 5 allocations in Asap library internals. (#emt-asap-raii)

• Added Fortran column-major layout conversion helpers in Eigen.h: AtomMatrixF type alias, zero-copy to/from_fortran_layout(), and map_from_flat_colmajor/rowmajor() for interfacing with Fortran libraries. (#feat-eigen-fortran-helpers)

• Modernize Fortran source files (SW, Tersoff, EDIP, Lenosky, Aluminum, CuH2). Convert to free-form F90, add intent declarations, replace common blocks with modules. Requires SVN regression verification. (#fortran-modernization)

• Highway SIMD kernels for Morse, LJ, and EAM pair-potential force loops. Subproject builds and -DWITH_HIGHWAY is set; needs actual vectorized inner loop implementations using HWY_DYNAMIC_DISPATCH. (#highway-potential-kernels)

• SafeHyperJob integration test with SVN reference data. Needs Morse Pt system with element-specific BondBoost parameters (SIGFPE on generic LJ). (#safe-hyper-test)

• Add SVN-verified integration tests for remaining unused reference data: neb_morse_pt, global_optimization_lj, minimization_eam_fire, minimization_sw_cg, min_lj_sd, replica_exchange_lj, bh50. (#svn-reference-coverage)

• Migrated test suite from GoogleTest to Catch2. Added 5 new tests (DimerTest, SaddleSearchTest, OptimizerTest, ConFileIOTest, HessianTest) and revived 4 (MatterTest, PotTest, ImpDimerTest, StringHelpersTest). Total: 16 tests. (#tests-new)

• Removed abandoned v3c TOML migration artifacts from 11 documentation files. (#docs-v3c-cleanup)

Changed

• Eigenmode methods (Dimer, ImprovedDimer, Lanczos, GPRDimer) now use std::variant instead of an abstract base class, eliminating virtual dispatch overhead. (#eigenmode-variant)

• LAMMPS potential now uses runtime dynamic loading (dlopen/LoadLibrary) instead of compile-time linking. A single eOn binary can use LAMMPS potentials if liblammps is installed, without requiring LAMMPS at build time. Install via conda install -c conda-forge lammps. (#lammps-runtime)

• Extracted RandomNumbers, GeometryAnalysis, and ConFileIO modules from Matter and HelperFunctions. Matter.cpp reduced from 1253 to ~500 lines, HelperFunctions.cpp from 838 to ~300 lines. (#module-extract)

• Removed all using namespace std; from the entire client codebase (60+ files). All standard library symbols now explicitly qualified with std::, preventing ADL-related bugs and improving code clarity. (#namespace-cleanup)

• INI parsing extracted from Parameters.cpp into ParametersINI.cpp with a validate_and_link() function for cross-group dependency resolution. (#params-ini-extract)

• Replaced vendored 724-line CIniFile INI parser with inih <https://github.com/benhoyt/inih>_ (r62) via meson wrap. (#params-inih)

• Narrowed parameter passing: Matter, Potential base, Optimizer hierarchy (OptimizerConfig), and Dynamics (DynamicsConfig) no longer require the full Parameters object. NEB no longer mutates its Parameters copy. (#params-narrow)

• All 33 Parameters option-group structs now use C++20 NSDMI (Non-Static Data Member Initialization) for defaults. The constructor shrunk from 481 to 3 lines. (#params-nsdmi)

• Fixed pass-by-value of VectorXd in ObjectiveFunction, LBFGS, and optimizer interfaces. Dynamic Eigen types now passed by const reference per Eigen documentation, eliminating unnecessary 40-160KB copies per optimizer step. (#perf-eigen-passbyref)

• Matter::getForces() now returns const AtomMatrix& to a cached masked-force result instead of copying and zeroing fixed atoms on every call. (#perf-masked-forces)

• NEB tangent and projection strategies cached as class members (built once in constructor). Spring strategy still rebuilt per iteration as it depends on per-step energy data. SIMD-optimized Eigen::Map<VectorXd>.dot() replaces (a.array() * b.array()).sum() in all NEB force projections. (#perf-neb-strategy-cache)

• Vectorized PBC wrapping: replaced scalar fmod loop with floor-based Eigen array operation. Single x86 vroundsd instruction instead of expensive fmod library call per element. (#perf-pbc-vectorize)

• Eliminated std::pow() with integer exponents from all hot-path force loops: LJ (pow(x,6) -> xxx), EAM (pow(r,5/6), simplified Morse pair from 3 exp to 1), IDPP (pow(r,4/5)), NEB spline (Horner’s method), Water_Pt (11 pow calls replaced with explicit multiplies). (#perf-pow-elimination)

• Extracted ReplicaDynamicsJob base class from TADJob and SafeHyperJob, deduplicating ~190 lines of shared code (checkState, refine, dephase, saveData). (#replica-base)

• All job classes now use std::shared_ptr<Matter> and RAII (ForceCallTimer, std::ofstream). No more raw new/delete for Matter objects or unchecked fopen calls. (#smart-ptrs)

• Converted commented-out SPDLOG_LOGGER_DEBUG calls to active QUILL_LOG_TRACE_L1 (compiled out in release builds). (#spdlog-quill)

• Replaced C-style headers (math.h, string.h, time.h, etc.) with C++ equivalents (cmath, cstring, ctime) in 10 client files. Added using enum JobType in Job.cpp. (#cpp20-headers)

• Python eon/fileio.py now uses the readcon package (PyPI) for loading and saving .con files, replacing ~60 lines of hand-written parsing. The loadcon, loadcons, and savecon functions delegate to readcon.read_con() / readcon.write_con().

• Replaced all FILE*-based con/convel I/O in the C++ client with readcon-core. Reading uses readcon::read_first_frame() (mmap). Writing uses ConFrameBuilder + ConFrameWriter with 17-digit precision for positions. All FILE* overloads removed; callers now pass filenames directly.

Fixed

• Fixed bare abs() calls on double values in BondBoost, LBFGS, and Hessian that resolved to the C integer abs(int) overload on non-MSVC compilers, silently truncating floating-point values. (#bugfix-bare-abs)

• cellInverse now recomputed in Matter::setCell() (was stale after cell changes). (#bugfix-cellinverse)

• Fixed confine_positive Eigen indexing (flat 3*i+0 replaced with proper (i,k) row-column access) and replaced raw new[]/delete[] with std::vector. (#bugfix-confine-positive)

• Fixed convergenceForce() loop variable mutation that could skip images. (#bugfix-convergenceforce)

• Fixed LBFGS aborting on degenerate curvature updates. The abs() -> std::abs() fix exposed that the s0.y0 < LBFGS_EPS check was previously disabled by integer truncation. Now resets L-BFGS memory (standard restart strategy) instead of aborting the optimization. (#bugfix-lbfgs-curvature)

• Fixed uninitialized cuttOffU member in LJ potential (caused garbage energy with MALLOC_PERTURB_). (#bugfix-lj-cutoffu)

• Fixed maxAtomMotionV out-of-bounds read when vector size < 3 elements (e.g. 2D optimizer objectives). Valgrind caught segment<3> reading past buffer, causing silent data corruption with MALLOC_PERTURB_ enabled. (#bugfix-maxatommotionv)

• maxEnergyImage now default-initialized to 0 (was uninitialized). (#bugfix-maxenergyimage)

• Added null FILE* guard in NEB write_movies to prevent crash when movie writing is disabled. (#bugfix-neb-write-movies)

• Added bounds check for numExtrema to prevent out-of-range access in NEB spline extrema. (#bugfix-numextrema)

• Fixed Prefactor Hessian size validation that checked min1 vs saddle on both sides of the OR condition, never validating min2 frequency array size. (#bugfix-prefactor-hessian)

• Fixed DynamicsSaddleSearch MD snapshot recording using shared_ptr aliasing instead of deep copy, causing all snapshots to point to the same live object and making transition time refinement unreliable. (#bugfix-snapshot-aliasing)

2.12.0 - 2026-03-08

Removed

• Remove _potcalls.log text logger and QUILL_LOG_TRACE_L3 from Potential::get_ef(). Remove static counters Potential::fcalls, fcallsTotal, wu_fcallsTotal, totalUserTime. Output is now _potcalls.json with structured per-instance records.

Added

• Add PotRegistry singleton for thread-safe, enum-indexed force call tracking. Replaces per-instance FileScoped text loggers and dead static counters (Potential::fcalls et al.). Tracks per-instance lifecycle (created_at, destroyed_at, force_calls, unique ID) with JSON output (_potcalls.json). Restore force call delta tracking in SafeHyperJob, TADJob, ParallelReplicaJob, ReplicaExchangeJob, NudgedElasticBandJob, BasinHoppingJob, HessianJob, and PrefactorJob.

• Add SafeMath.h utility header with guarded arithmetic functions (safe_div, safe_recip, safe_acos, safe_sqrt, safe_atan_ratio) and an Eigen-aware safe_normalized template. These prevent floating-point exceptions (SIGFPE) from division-by-zero and domain errors in numerical code without changing results for valid inputs.

• Drop spdlog and fmt for quill and cpp20

Changed

• Replace cxxopts with argum for command line parsing. This change updates the CLI argument handling library and requires C++20 support. (#320)

• Replace per-site scoped_interpreter guards with lazy singleton eonc::ensure_interpreter() in PyGuard.h. Python interpreter is only started when a Python-based potential is actually used. (#324)

• Migrated from spdlog/fmt to quill logging library with std::format for modern C++20 logging infrastructure. Quill provides lock-free asynchronous logging with lower latency and better performance characteristics. All LOG_* macros now use quill backend with configurable formatters and sinks. (#327)

• Migrate to modern C++20 logging API (EonLogger.h).

  Replaced verbose quill logger initialization throughout codebase with new eonc::log::Scoped RAII helper and eonc::log::get_file() utility. Eliminates 14-line boilerplate for file loggers and manual initialization for default loggers. Net reduction of 74 lines while improving code clarity and safety.

• Optimize quill backend for improved logging performance.

  Configure BackendOptions with reduced sleep duration (100us to 10us), larger initial transit buffer (256 to 2048), zero timestamp ordering grace period (single-threaded, SPSC guarantees ordering), faster flush interval (200ms to 100ms), and disabled printable char checking (numeric data only).

• Switched all logging call sites from bare LOG_* to QUILL_LOG_* prefixed macros and enabled QUILL_DISABLE_NON_PREFIXED_MACROS to prevent macro collisions when eOn is compiled alongside other libraries.

• Updated metatomic ecosystem dependencies: torch 2.10, metatomic-torch 0.1.9+, metatensor-torch 0.8.4+, vesin/vesin-torch 0.5.2+, metatrain 2026.2.1+. Ensures compatibility with latest machine learning potential infrastructure.

• Updated vesin to v0.5.2: adapted to new API with per-dimension periodicity (bool[3] instead of single bool) and VesinDevice struct syntax. Ensures compatibility with latest metatomic/metatensor ecosystem.

• Wrap all client classes, enums, and helper namespaces under namespace eonc. Rename helper_functions namespace to helpers. BaseStructures.h enums (PotType, JobType, AtomState) are now scoped under eonc. Backward-compatible using aliases are provided for all classes and enums at global scope. Remove using namespace std from all client headers to prevent symbol leakage into downstream translation units.

Fixed

• Fix ASE_POT compilation errors (wrong constructor, PotType, FPE calls, undeclared variable) and rename -DASE_POT to -DWITH_ASE_POT to avoid macro-enum collision. (#321)

• Fix [IDimerRot] column misalignment: widen force placeholder from 10 to 18 dashes, change angle precision from {:6.2f} to {:6.3f}, and add missing “Align” column specifier to the [Dimer] header. (#322)

• Suppress FPE trapping during libtorch operations in MetatomicPotential constructor and force(), preventing SIGFPE from benign NaN/Inf produced by SiLU (sleef) and autograd internals. Follows the existing FPEHandler pattern from ASE_ORCA, ASE_NWCHEM, and AtomicGPDimer. (#323)

• Change uncertainty_threshold default from 0.1 to -1 (disabled) in both C++ and Python. Most models lack uncertainty outputs, so the previous default triggered a noisy exception+catch in the metatomic constructor for no benefit. (#325)

• Fixed quill migration test failures: added logger initialization to all test fixtures (XTBTest, ASEPotTest, ServeSpecParseTest, EpiCentersTest, MetatomicTest) to prevent segfaults from uninitialized quill backend. Restored correct ConfigParser defaults in config.yaml (49 path interpolations accidentally replaced during migration). Added -DNOMINMAX for Windows builds to fix MSVC compilation errors in quill headers. (#327)

• Added safe_normalize_inplace to SafeMath.h and guarded remaining unprotected .normalize() / .normalized() calls in Dimer, ImprovedDimer, ConjugateGradients, and LBFGS that could trigger FPE on zero vectors.

• Guard unprotected floating-point divisions and domain-error-prone math across 12 source files using eonc::safemath utilities. Eliminates spurious SIGFPE signals during saddle search (Dimer, ImprovedDimer, Lanczos), optimization (LBFGS, CG, SteepestDescent), and infrastructure (Matter, Hessian, HelperFunctions, NEB, ReplicaExchange). Fallback values preserve existing branch/skip/reset behavior so valid inputs produce identical results.

• Make the POSIX FPE signal handler async-signal-safe by replacing std::cerr (undefined behavior in signal context) with write(STDERR_FILENO, ...). Windows SEH handler switched from std::cerr to fprintf(stderr, ...) for consistency.

• Use -isystem instead of -I for pip-installed metatomic and vesin include paths to suppress third-party compiler warnings when building with -Wall -Wextra.

2.12.0 - 2026-03-04

Changed

• Replace cxxopts with argum for command line parsing. This change updates the CLI argument handling library and requires C++20 support. (#320)

• Replace per-site scoped_interpreter guards with lazy singleton eonc::ensure_interpreter() in PyGuard.h. Python interpreter is only started when a Python-based potential is actually used. (#324)

Fixed

• Fix ASE_POT compilation errors (wrong constructor, PotType, FPE calls, undeclared variable) and rename -DASE_POT to -DWITH_ASE_POT to avoid macro-enum collision. (#321)

• Fix [IDimerRot] column misalignment: widen force placeholder from 10 to 18 dashes, change angle precision from {:6.2f} to {:6.3f}, and add missing “Align” column specifier to the [Dimer] header. (#322)

• Suppress FPE trapping during libtorch operations in MetatomicPotential constructor and force(), preventing SIGFPE from benign NaN/Inf produced by SiLU (sleef) and autograd internals. Follows the existing FPEHandler pattern from ASE_ORCA, ASE_NWCHEM, and AtomicGPDimer. (#323)

• Change uncertainty_threshold default from 0.1 to -1 (disabled) in both C++ and Python. Most models lack uncertainty outputs, so the previous default triggered a noisy exception+catch in the metatomic constructor for no benefit. (#325)

2.11.1 - 2026-03-01

Added

• External potential (ext_pot) documentation with protocol spec, DeePMD and ASE wrapper examples, and conda-forge availability badges on all potential pages. (#318)

Developer

• Add ExtPotTest unit test verifying the file-based ext_pot protocol with a harmonic spring calculator. (#318)

Fixed

• Rename PotType::EXT to EXT_POT so magic_enum matches the ext_pot config string. Previously potential = ext_pot was silently mapped to UNKNOWN. (#318)

2.11.0 - 2026-02-24

Added

• Add eonclient --serve mode that wraps any eOn potential as an rgpot-compatible RPC server over Cap’n Proto. Supports four serving modes: single-potential (--serve-port), multi-model (--serve "lj:12345,eam_al:12346"), replicated (--replicas N on sequential ports), and gateway (single port with round-robin pool via --gateway). All options are also available through a [Serve] INI config section. Requires -Dwith_serve=true at build time. (#316)

• Add dictionary-style configuration examples using rgpycrumbs to the user guide, demonstrating programmatic config generation alongside INI files. (#317)

Developer

• Switch benchmark PR comment workflow from hand-rolled scripts to the asv-perch GitHub Action, and parallelize benchmark execution with a matrix strategy for main and PR HEAD. (#315)

• Add rgpot subproject wrap, with_serve meson option, serve pixi environment, CI workflow for serve mode builds, and Catch2 unit tests for serve spec parsing. (#316)

Fixed

• Skip torch_global_deps on Windows where the conda-forge libtorch package does not ship it. (#314)

• Fixed serve mode segfault caused by AtomMatrix type collision between eOn’s Eigen-based type and rgpot’s custom type. Replaced the rgpot::PotentialBase virtual interface with a flat-array ForceCallback, eliminating the name collision entirely. The serve code now only links the capnp schema dependency (ptlrpc_dep) from rgpot, not the full library. (#316)

2.10.2 - 2026-02-22

Fixed

• Fixed a significant performance regression in NEB calculations caused by incorrect Eigen matrix storage order mapping. Added a regression test and updated CI to automatically mark PRs as draft if benchmark regressions exceed 10x. (#310)

• Absorbed conda-forge Windows patches upstream: replace C99 VLA in XTBPot with std::vector, guard empty-string indexing in INIFile, decouple xtb from Fortran requirement, add Windows library search paths for libtorch/metatensor/vesin, guard POSIX headers, and replace shell commands in IMD with std::filesystem. (#312)

2.10.1 - 2026-02-18

Developer

• Added a CI-NEB XTB regression test (CINEBXTBTest.cpp) that runs a 10-image climbing-image NEB with GFN2-xTB on a 9-atom molecule. The test completes in under 2 seconds and guards against storage-order regressions that corrupt force projections.

Fixed

• Replaced the EIGEN_DEFAULT_TO_ROW_MAJOR preprocessor macro with explicit row-major type aliases in client/Eigen.h. The macro made eOn’s Eigen types binary-incompatible with other Eigen-based libraries; removing it without updating bare MatrixXd types caused NEB force projections to silently corrupt and the optimizer to diverge from the first step.

• Use datetime.timezone.utc instead of datetime.UTC in get_version.py for Python 3.10 compatibility (the datetime.UTC alias was added in 3.11).

v2.10.0 - 2026-02-15

Added

• Added ASV benchmark CI workflow with asv-spyglass for PR performance comparison

• Added adsorbate_region.py example script for identifying adsorbate atoms and nearby surface atoms by element or z-coordinate

• Added displacement scripts tutorial with worked examples for vacancy diffusion (PTM) and adsorbate-on-surface scenarios

• Added displacement strategies prose section to saddle search docs explaining epicenters, weight-based selection, and dynamic atom lists

• Expose gprd_linalg_backend option for selecting GPR-dimer linear algebra backend (eigen, cusolver, kokkos, stdpar)

Developer

• Added macOS arm64 to metatomic CI matrix using Homebrew gfortran (conda-forge gfortran_osx-arm64 wrapper is broken)

• Cleanup to build on windows

• Expanded ASV benchmark suite with point evaluation, LJ minimization, and NEB workloads

• Use internal pick output helper

• bld(meson): reduce build times by linking to xtb by default

Changed

• Eliminated unnecessary Eigen matrix copies in Matter, Potential, and HelperFunctions hot paths

• Replace per-typedef Eigen::RowMajor with a single eOnStorageOrder constant in client/Eigen.h

• Enriched schema descriptions for displace_atom_kmc_state_script, displace_all_listed, displace_atom_list, and client_displace_type

• Refactored MetatomicPotential variant resolution to use upstream metatomic_torch::pick_output

• Updated pinned gpr_optim commit with new linear algebra backends and performance improvements

Fixed

• Fix Windows STATUS_STACK_OVERFLOW crash caused by large Fortran local arrays in the EAM Al potential (gagafeDblexp.f) exceeding the 1 MB default stack; request 16 MB via linker flags

• Fix Windows silent client failure by using non-color spdlog sink when stdout is redirected

• Use Goswami & Jonsson 2025 for removing rotations through projections

v2.9.0 - 2026-01-27

Added

• Add support for ‘charge’ and ‘uhf’ (multiplicity) parameters in the xTB potential

• Introduce custom Catch2 Eigen matchers and add comprehensive regression tests for GFN2-xTB forces

• Setup Collective-IDPP path generation for NEB runs

• Setup IDPP path generation for NEB runs

• Setup sequential Collective-IDPP path generation for NEB runs

• feat(mtapot): handle variants for energy and energy uncertainty within Metatomic models

• feat(neb): add a zbl+sidpp penalty for initial path generation

• feat(neb): implement the OCI-NEB/RONEB/enhanced CI via MMF

• feat(neb): implement the onsager machlup action logic

• feat(neb): write out peaks and modes for subsequent dimer runs

Changed

• Optimize xTB potential performance by persisting internal state and using coordinate updates between force calls

• Update installation guide to recommend Pixi and clarify dependency management

Fixed

• bug(ewneb): do not turn on if cineb threshold is not met!

• fix(mtapot): stop double counting mta calls

v2.8.2 - 2025-12-01

Added

• Metatomic is now uncertainty aware

• Metatomic variance reports per-atom uncertainty mean

Changed

• Reworked metatomic to use torch 2.9

Developer

• Update to use metatensor_torch::Module

• Use metatomic::pick_device correctly

v2.8.1 - 2025-11-03

Added

• Enable minimization for given initial paths

Changed

• Reworked metatomic to use torch 2.8

Fixed

• Generate neb.dat correctly without clobbering neb_000.dat

2.8.0 - 2025-09-04

Added

• Potentials & Interfaces

  • Expanded potential interfaces to a variety of new quantum chemistry and ML potentials via an embedded Python interpreter:

    • NWChem: A high-performance, socket-based interface. (#244)

    • ORCA: Interface to the ORCA quantum chemistry program via ASE.

    • AMS: Interface for the Amsterdam Modeling Suite.

    • XTB: Interface for semi-empirical GFN-xTB methods.

    • ASE: A general-purpose interface to any calculator supported by the Atomic Simulation Environment.

  • Added the Ziegler-Biersack-Littmark (ZBL) universal screening potential, useful for collision cascade simulations. (#241)

  • Integrated support for metatomic machine-learned potentials via the vesin library, enabling high-performance simulations with models from the metatensor ecosystem. (#201)

• Nudged Elastic Band (NEB)

  • NEB calculations can now pre-optimize the initial and final states, improving path quality and convergence. This feature is fully compatible with restarts. (#221)

  • NEB calculations can now be initialized from a user-provided sequence of structures, offering greater control over the initial reaction pathway.

  • Introduced energy-weighted springs to improve the stability and quality of paths with high energy barriers.

  • Enabled the use of dual optimizers (e.g., a starting with QuickMin and switching to LBFGS after a convergence threshold).

  • Implement the novel RO-NEB-CI (Rohit’s Optimal NEB with MMF CI steps) method (#239)

Developer

• Consistent formatting and counting

• Support for M1 MacOS machines

Build & Tooling

• Build System

  • Overhauled the build system, migrating from legacy Makefiles/CMake to Meson for a faster, more reliable, and truly cross-platform build experience. This change also lays the groundwork for a future pure Python eon-server package. (#124)

• Dependency Management

  • Adopted pixi and conda-lock for robust, reproducible dependency management across all platforms.

• Cross-Platform Support & CI

  • Established a full Continuous Integration (CI) pipeline, testing on Linux, Windows, and macOS (Intel & Apple Silicon).

  • The Command Line Interface (CLI) is now fully compatible with Windows environments.

Code Quality & Refactoring

• C++ Modernization

  • Modernized the C++ backend to the C++17 standard, improving code clarity and performance.

  • Enhanced memory safety by replacing raw pointers with smart pointers (std::unique_ptr, std::shared_ptr).

  • Adopted the <filesystem> library for platform-independent file I/O.

• Logging

  • Replaced the internal logging system with spdlog for high-performance, asynchronous, and more informative configurable output.

• Code Style

  • Enforced a consistent code style and formatting across the entire C++ and Python codebase.

Documentation

• Configuration & Schema

  • Implemented a comprehensive Pydantic schema for all configuration files, providing automatic input validation and clear error messages. This forms the foundation for automated API documentation.

• User Guides

  • Added detailed user documentation for the Nudged Elastic Band (NEB) module, covering theory, keywords, and practical examples.