feat(eon-pet-neb): run eOn via pip pyeonclient + rgpot engines (#293)

* feat(eon-pet-neb): run eOn via pip pyeonclient + rgpot engines

Drop conda-forge eOn/eonclient. Use pyeonclient 0.3 with RGPOT metatomic
dlopen engines from rgpot>=2.4.1 (parity with native Metatomic on this
recipe: same force-call counts, energies within ~1e-5 eV).

* style(eon-pet-neb): ruff format and drop unused chdir

CI lint uses ruff format --check; chdir unused after pyeonclient helper.

* fix(eon-pet-neb): require pyeonclient>=0.3.1 manylinux_2_28

0.3.0 was manylinux_2_43-only and unusable on GHA glibc; 0.3.1 is
manylinux_2_28 for CI installability.

* style(eon-pet-neb): avoid Path default arg (flake8 B008)

* feat(eon-pet-neb): drive client via pyeonclient steppers

Compose ClientEON stages in Python (load_parameters → run_job →
potcalls → timing; minimize_workdir for Matter.relax) instead of
run_job_in_directory. Require rgpot>=2.4.2 for nested-safe engines.

* feat(eon-pet-neb): use Matter path API, not workdirs

Drive pyeonclient as list[Matter] + NudgedElasticBand.compute /
Matter.relax with live Parameters attributes. Drop config.ini,
run_job workdir composition, and minimize_workdir. Disk is only
an optional export for rgpycrumbs plots.

* style(eon-pet-neb): ASE-shaped Matter API for eOn NEB and min

Present pyeonclient like ASE: make_potential ≈ calculator, from_ase /
to_ase for Atoms↔Matter, NudgedElasticBand.compute ≈ LBFGS(neb).run,
Matter.relax for endpoints. Drop workdir/path-file scaffolding from
the narrative; keep .con export only for plot tools.

* docs(eon-pet-neb): use eOn-native IDPP and pyec alias

Build the eOn band with neb_idpp_path / NEBInit.IDPP instead of ASE
interpolate into Matter. Prefer import pyeonclient as pyec.

* refactor(eon-pet-neb): use pyeonclient make_backend and collapse NEB knobs

Build the shared PET-MAD potential with backends.make_backend("rgpot_metatomic"),
load ASE MetatomicCalculator via the PET-MAD-safe helper, and group energy-weighted
CI+MMF settings into configure_neb_ew_mmf so the eOn half reads as Matter → path →
compute rather than pot plumbing.

* refactor(eon-pet-neb): call make_backend directly, drop make_potential

Teach the real registry API (rgpot_metatomic / metatomic / ase_metatomic)
instead of a one-off wrapper that hid how the pot is built.

* chore(eon-pet-neb): pin pyeonclient 0.3.2 + eon-schema 0.2.1

Use released NebSpec for EW/CI/MMF knobs and require the published
pyeonclient[models] floor that ships make_backend and NEB helpers.

* docs(eon-pet-neb): appendix force-cost ASE vs pyeonclient

Time single-point PET-MAD forces for ASE MetatomicCalculator,
ase_metatomic, and rgpot_metatomic so the backend overhead is separate
from the NEB algorithm comparison. Also prefer write_neb_results and
share Parameters with the min backend.

* fix(eon-pet-neb): rattle-style force timing like metatomic profiling

Use atoms.rattle(1e-6) for the ASE path and analogous Gaussian position
noise for Matter, plus force_calls deltas, so the appendix cannot report
cache hits as force-call cost.

* style(eon-pet-neb): tighten force-cost appendix

Shorter prose and timing helper; keep rattle + force_calls checks.

* fix(eon-pet-neb): do not pass params into ase_metatomic factory

CI failed after the ASE force timing line: MetatomicCalculator rejected
params=. Share Parameters only with from_ase / Matter.

* fix(eon-pet-neb): restore gallery; ASE vs rgpot force appendix only

Drop ase_metatomic from the microbench (params leak into MetatomicCalculator
on pyeonclient 0.3.2). Keep rattle + force_calls checks; ruff format.

* fix(eon-pet-neb): appendix is ASE vs rgpot only

Remove ase_metatomic microbench path that crashed gallery generation
on PyPI pyeonclient 0.3.2 (params leaked into MetatomicCalculator).

* style(eon-pet-neb): wrap long comment lines for flake8 E501

* style(eon-pet-neb): shorten comparison comment block for E501

* fix(eon-pet-neb): force appendix does not claim pot speedup

CI: ASE ~17 ms/call vs rgpot_metatomic ~15 ms/call. Drop "speedup"
wording; NEB advantage is EW/MMF, not force wrapper magic.

* chore(eon-pet-neb): upper-bound pip deps for gallery stability

Cap pyeonclient, eon-schema, rgpot, ase, metatrain, and readcon so the
example does not silently break on the next major/minor release.

* docs(eon-pet-neb): plain wording for make_backend and NebSpec

Drop defensive registry/packaging asides; state the pot load and knobs.

* docs(eon-pet-neb): shorter force-appendix note

* docs(eon-pet-neb): drop force-cost appendix

NEB comparison is enough; the microbench added bulk without payoff.

* docs(eon-pet-neb): remove force-cost appendix body

Previous commit dropped the import and README note; finish removing
the timed force microbench.

* docs(eon-pet-neb): neutral wording for deps and dual NEB paths

State what the example uses; drop "not conda / not ASE" asides.

* fix(eon-pet-neb): drop force appendix and finish prose cleanup

Remove the timed force microbench entirely (NameError: time). Neutral
wording, SIDPP citation (Schmerwitz et al.), upper-bound deps already in
tree.

* docs(eon-pet-neb): full NEB trace, parallel prose, in-process plots

Report the write_movies band evolution (neb_NNN.dat), final neb.dat,
and results.dat alongside the status/E_ref summary. State parallel
per-image force evaluation honestly. Drive 1D/2D NEB and min figures
through rgpycrumbs/chemparseplot imports instead of a CLI/uv subprocess.

* fix(eon-pet-neb): pin chemparseplot and ruff-format report helper

CI pulled chemparseplot 1.9.12, which raises NameError on SurfaceFitConfig
under Python 3.13. Cap at <1.9.10 (last known-good). Apply ruff format so
the lint session stays green.

* fix(eon-pet-neb): avoid Path default for flake8 B008

flake8-bugbear rejects function-call defaults; use cwd=None and Path.cwd().

* chore(eon-pet-neb): pin chemparseplot>=1.9.13

Drop the temporary <1.9.10 cap; 1.9.13 fixes the SurfaceFitConfig
NameError on Python 3.13.

* style(eon-pet-neb): drop changelog notes from environment pins

* style(eon-pet-neb): strip remaining pin commentary from environment.yml

* docs(eon-pet-neb): drop redundant load-compat call; cite OCINEB Frontiers

make_metatomic_ase_calculator already installs the PET-MAD load patch.
Reference (6) is the Frontiers in Chemistry OCINEB paper (Goswami, Gunde,
Jónsson).

* docs(eon-pet-neb): drop verbose NEB report helper

Keep a short status/ΔE/written printout; the 1D profile figure is the
band evolution, not a hand-rolled snapshot table or client log dump.

* docs(eon-pet-neb): use pyeonclient.plot_neb / plot_min

Drop cookbook argv wrappers around rgpycrumbs CLI. Figures go through
pyeonclient[plot] (0.3.3): plot_neb, plot_min, thin_min_movie.

* docs(eon-pet-neb): use rgpycrumbs.eon.plot_neb library API

Drop argv shims and pyeonclient.plot_*. NEB figures call plot_neb from
rgpycrumbs 1.10.2; min figures use chemparseplot load/plot helpers.
Pin rgpycrumbs>=1.10.2 and chemparseplot[neb,plot]>=1.9.13.

* docs(eon-pet-neb): use rgpycrumbs.eon.plot_min for min figures

Drop local thin_min_movie and chemparseplot low-level min plot calls.
plot_min (1.10.3) with auto_thin covers landscapes; pin rgpycrumbs>=1.10.3.

* fix(eon-pet-neb): declare adjustText for rgpycrumbs plot_neb

CI failed at import of rgpycrumbs.eon.plot_neb because adjustText is
required at module load but was not listed in environment.yml. Pin
pyeonclient to >=0.3.3 for path_frames / retain_frames.

* fix(eon-pet-neb): floor rgpycrumbs 1.10.4 for library plot_neb

1.10.3 applied Click to plot_*_from_settings, so keyword library
calls fed settings keys into argv and failed after a good NEB.
Match chemparseplot plot extra at 1.9.14.

* fix(eon-pet-neb): install xyzrender for structure strip plots

plot_structures with strip_renderer=xyzrender needs the xyzrender
binary on PATH; chemparseplot raises at render time without it.

* fix(eon-pet-neb): declare jax for grad_imq landscape surfaces

plot_neb landscape with surface_type=grad_imq loads rgpycrumbs
surface models that require jax at import time.

* fix(eon-pet-neb): lazy-load jax via RGPYCRUMBS_AUTO_DEPS + uv

Library plot_neb/plot_min use ensure_import; CLI dispatch turns on
AUTO_DEPS by default. Gallery imports need the same flag so jax and
adjustText stage through uv into the dep cache instead of env pins.

* fix(eon-pet-neb): floor rgpycrumbs 1.10.7 for library AUTO_DEPS

plot_neb enables RGPYCRUMBS_AUTO_DEPS like CLI; drop need for host jax
pin. keep setdefault in the notebook as belt-and-suspenders.

* style(eon-pet-neb): place AUTO_DEPS before plot import only

Keeps RGPYCRUMBS_AUTO_DEPS=1 for ensure_import without E402 on the
whole import block; match other gallery env-then-import patterns.

* docs(eon-pet-neb): one gallery cell per minimization figure

Four plot_min outputs in a single sphinx-gallery block were tiled into
an unreadable multi-image grid. Split reactant/product landscapes and
profile/convergence into separate cells, and size show_png from the
PNG aspect ratio so structure-strip figures are not squashed.

* fix(eon-pet-neb): drop chemparseplot/xyzrender host pins

rgpycrumbs>=1.10.8 + AUTO_DEPS stages chemparseplot, jax, adjustText,
and xyzrender via ensure_import; only recipe-native deps stay in env.yml.

* docs(eon-pet-neb): show min landscapes side by side

Render reactant and product plot_min landscapes as two PNGs, then one
gallery figure with equal-height panels so endpoints can be compared.

* docs(eon-pet-neb): drop show_png_row; inline side-by-side min panels

Keep a single show_png helper; reactant/product landscapes use a plain
1x2 subplot for the gallery comparison.

* fix(eon-pet-neb): replace leftover show_png_row call with 1x2 subplot

Previous commit removed the helper but left a call site (F821).

* chore(eon-pet-neb): recipe-only env; drop host plot pins and AUTO_DEPS set

rgpycrumbs>=1.10.8 library plot enables AUTO_DEPS and ensure_import for
jax, adjustText, chemparseplot, and xyzrender (Python API, not PATH).
Keep uv for the dep cache; drop manual setdefault and E402-split imports.

* docs(eon-pet-neb): polish example, trim env and redundant prose

Drop redundant eon-schema pin, unused Matter returns, empty min
stub section, and duplicated landscape text. Share N_INTERMEDIATE_IMGS
for ASE NEB, hoist show_png/MPLBACKEND, tighten IRA and README.

* ci(eon-pet-neb): re-verify after eOn 2.17.2 / pyeonclient 0.3.4

* chore(eon-pet-neb): floor release pins (pyeonclient 0.3.5, rgpot 2.5.2)

Match the published stack: torch 2.10 engine ABI, vesin 0.6 VesinOptions,
and an absolute model_path captured before min_* chdirs. Keeps the PR's
library plot_neb/plot_min path; only the floors from the gallery run.

* chore(eon-pet-neb): floor rgpycrumbs to 1.10.10

Stack already at PyPI latest for pyeonclient 0.3.5, rgpot 2.5.2,
vesin 0.6, readcon, and metatrain. Raise the plot library floor to the
current release and cap <2. Torch stays on 2.10 for the rgpot engine ABI.

* chore(eon-pet-neb): floor rgpot to 2.5.3

PyPI now has the 2.5.3 abi3 wheel (was GitHub/crates-only after the tag).

* chore(eon-pet-neb): floor rgpot to 2.5.4 multi-ABI

2.5.4 ships metatomic engines for torch 2.7–2.13. Recipe torch pin
stays 2.10 for gallery reproducibility.

* chore(eon-pet-neb): allow torch 2.10–2.13 with rgpot multi-ABI

rgpot 2.5.4 ships engines through 2.13. Drift check on PET-MAD XS
(climb.con) was bit-identical for 2.10/2.11/2.13; widen the pin.

* style(eon-pet-neb): drop commentary from environment pins

* chore(eon-pet-neb): upper-bound remaining pip pins

Close the last unpinned gallery dependencies so minor upstream bumps
cannot break the recipe.

* build(eon-pet-neb): require rgpot>=3.0.2 for the metatensor-torch 0.10.3 load fix

Author: HaoZeke

examples/eon-pet-neb/README.rst

@@ -1,5 +1,5 @@
-Finding Reaction Paths with EON and a Metatomic Potential
+Finding Reaction Paths with eOn and a Metatomic Potential
 =========================================================
 
-Describes the usage of the Nudged Elastic Band with energy weighted springs and
-single ended dimer searches using EON and PET-MAD. Contrasts with ASE.
+Nudged elastic band with energy-weighted springs and OCI-MMF using
+``pyeonclient`` and PET-MAD (ASE climbing-image NEB for comparison).

examples/eon-pet-neb/environment.yml

@@ -3,27 +3,20 @@ channels:
   - conda-forge
   - https://repo.prefix.dev/rg-forge
 dependencies:
-  # eOn 2.16 writes CON via readcon-core (con_spec_version=2 metadata).
-  - eon>=2.16.0
   - ira
   - python=3.13
   - pip
   - uv
   - pip:
-    - torch>=2.9,<2.10
-    - ase>=3.23
-    - matplotlib
-    # eOn 2.16 is built against libmetatomic-torch >=0.1.15; metatrain
-    # 2026.3.1 is the first release that requires the same floor.
-    - metatrain>=2026.3.1
-    - requests
-    - metatomic-ase
-    # Notebook CON I/O only. Plot deps (chemparseplot, jax, adjustText, …)
-    # Prefer library API when plot pipelines are public; today full NEB
-    # plots still go through CLI uv PEP 723 / ensure_import — not host pins.
-    - "readcon>=0.13.1"
-    # rgpycrumbs 1.9.21+ plot-script headers require chemparseplot>=1.9.10,
-    # whose 1.9.10-1.9.12 releases fail to import on Python <=3.13. Drop the
-    # cap (and UV_EXCLUDE_NEWER in the recipe) once a fixed release is out.
-    - "rgpycrumbs>=1.9.16,<1.9.21"
+    - "pyeonclient[models]>=0.3.5,<0.4"
+    - "rgpot>=3.0.2,<4"
+    - "vesin>=0.6.0,<0.7"
+    - torch>=2.10,<2.14
+    - ase>=3.23,<4
+    - matplotlib>=3.8,<4
+    - metatrain>=2026.3.1,<2027
+    - requests>=2.28,<3
+    - metatomic-ase>=0.1,<0.2
+    - "readcon>=0.13.1,<0.14"
+    - "rgpycrumbs>=1.10.10,<2"
     - atomistic-cookbook-utils >=0.1,<0.2