Add nvidia-alchemi agent (NVIDIA ALCHEMI Toolkit) 1.1.2

agents/nvidia-alchemi/.dockerignore

@@ -0,0 +1,4 @@
+__pycache__/
+**/__pycache__/
+**/*.pyc
+.git/

agents/nvidia-alchemi/CHANGELOG.md

@@ -0,0 +1,130 @@
+# Changelog
+
+All notable changes to the NVIDIA ALCHEMI Toolkit agent are recorded here.
+The format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/), and the
+agent uses [Semantic Versioning](https://semver.org/). Versions match `version` in
+`metadata.yaml` and `tools/nvidia-alchemi/tool.yaml` and the container image tag.
+
+## [1.1.2] - 2026-10-01
+
+Catalog-policy and security-hardening release. Scientific algorithms and bundled
+checkpoint bytes are unchanged from 1.1.1.
+
+### Security
+
+- Updated GitPython, hydra-core, multidict and urllib3 to patched versions after
+  dependency review identified published advisories.
+- Updated certifi to 2026.10.19 to remove the compromised emSign root CAs
+  (GHSA-6q48-6c7w-9gw2).
+- Updated aiohttp 3.14.4, filelock 4.2.1, GitPython 3.2.1, hydra-core 1.3.8 and
+  Pillow 13.0.0 for high and critical advisories reviewed after the previous
+  update, and setuptools 84.1.0 for a low-severity advisory. yarl moves to 1.25.1,
+  the minimum aiohttp 3.14.4 requires.
+- Checkpoints are downloaded, size/hash verified, and only then deserialized.
+- Runtime learned-model aliases resolve only to hash-verified checkpoints baked
+  into the image; explicit existing local paths remain available for training
+  checkpoints.
+- Cached learned-model files are read-only, and `logs.out` is reserved from
+  artifact replacement.
+
+### Changed
+
+- The pinned dependency list is renamed from `requirements-build.lock` to
+  `requirements-build.txt`, a file type the catalog accepts.
+- CPU execution must be explicitly requested for validation; deployed workflows
+  fail when CUDA is unavailable.
+- Experimental workflows outside completed validation coverage are identified in
+  catalog descriptions and require explicit user confirmation.
+- Catalog tags use the catalog vocabulary: `machine-learning`, `gpu` and
+  `simulation` replace `machine-learned-potential`, `gpu-accelerated` and
+  `high-throughput`.
+
+### Fixed
+
+- Image-build tests can import the checkpoint downloader from `/app/build`
+  without adding that directory to the runtime import path, so
+  `python /app/release_smoke.py` runs unchanged in the deployed image. Image-build
+  tests now run with the runtime `PYTHONPATH` to catch import-path regressions.
+- Screening validation tests explicitly use the selected test device, and the
+  release-manifest test checks the shared AIMNet cache under `/opt/cache/aimnet`.
+
+### Validation
+
+- Rebuilt from a clean image: 43 regression tests passed on CPU and on one H100 NVL
+  with no skips, and the GPU release smoke passed.
+- MACE `small-0b`, MACE `medium-0b2` and AIMNet2 inference, MACE relaxation and
+  100-step nonperiodic NVT were revalidated on H100 with independent numerical
+  checks. See the README verification status.
+
+## [1.1.1] - 2026-10-01
+
+### Added
+
+- `single_point(batch, model)`: the supported way to evaluate energies and forces for
+  fixed geometries. It builds the neighbor list itself and rejects missing or
+  non-finite outputs.
+- `step_recorder()`: a ready-made dynamics hook that records energy, kinetic energy,
+  temperature and maximum force for every step and structure.
+- The release manifest now names each cached checkpoint (`loader` and `name`) next
+  to its path, size and SHA-256, so models can be looked up by name.
+- Distribution license and notice files are copied into the image under
+  `/app/licenses/`, and their per-package counts are recorded in the manifest.
+- Regression tests for the new helpers and the manifest, bringing the suite to 32.
+
+### Changed
+
+- The image is built from the public `nvidia/cuda:13.0.1-devel-ubuntu24.04` image,
+  pinned by digest, instead of a private parent image. Anyone can now rebuild it.
+  It installs the same 198 package versions as 1.1.0 from
+  `requirements-build.lock`, without further dependency resolution.
+- MACE `small-0b`, MACE `medium-0b2` and AIMNet2 checkpoints are downloaded at build
+  time, and the build fails unless every file matches its recorded SHA-256.
+  Jobs still run offline.
+- The release manifest records the actual base image, build inputs and release
+  version. The smoke test reports the release and the helper-library version
+  separately.
+- Agent instructions document the new helpers, the custom-hook requirements, batch
+  field names (`batch_idx`, `atomic_masses`) and checkpoint lookup by name.
+- The README is self-contained for the catalog: support links point to the
+  Microsoft Discovery community, and build and deployment steps are documented.
+- The release smoke test is renamed from `hardening_smoke.py` to `release_smoke.py`.
+- The build fails if any source file has Windows line endings, so the recorded
+  source hashes always match the repository.
+- Sample structures in `/app/example-input-files` are documented for users and the agent.
+
+### Fixed
+
+- `write_structures()` crashed when given a single `ase.Atoms`; it now accepts a
+  `Batch`, a single `Atoms`, or a list.
+- The instructions no longer suggest that model wrappers have a `compute()` method.
+
+### Validation
+
+Validated on one H100 NVL through the agent: 32 regression tests with no skips; all
+three cached models selected by name, hash-checked and evaluated on CUDA; and a
+100-step MACE NVT run whose final kinetic energy, recomputed independently, matched
+the recorded value to $6\times10^{-9}$ eV.
+
+## [1.1.0] - 2026-09-22
+
+### Added
+
+- `alchemi_validators`: NumPy reference calculations for kinetic energy and energy
+  drift, finite and shape-aware numeric comparison, and hook-event auditing. Results
+  distinguish `pass`, `fail`, `unverified` and `unexercised`.
+- `screen()` results carry `input_index`, the original input position, separately
+  from the toolkit's admission-order `system_id`.
+
+### Changed
+
+- Declared outputs of any file type, including nested directories, are copied to
+  the output folder and verified by SHA-256 and size before results are published.
+  An artifact manifest lists every delivered file.
+- Results JSON is strict: non-finite values are encoded explicitly and unsupported
+  objects are rejected.
+- The tool reports the script's real exit status.
+
+### Fixed
+
+- Outputs such as NPZ, Python and text files were silently left out of results.
+- Screening results could be matched to the wrong inputs for mixed-size batches.