APEX workflows, reliable cloud startup and L20 tongqi qualification

.github/workflows/publish-docker.yml

@@ -6,7 +6,7 @@ on:
       tag:
         description: Docker image tag
         required: true
-        default: "1.3.0"
+        default: "1.3.2.post1"
 
 permissions:
   contents: read

.gitignore

@@ -49,6 +49,8 @@ confs
 # Generated MatMaster skill archive
 apex-flow.zip
 
-# Large DeePMD checkpoints (fetch with scripts/fetch_models.py; keep frozen *.pb in git)
+# Exclude arbitrary DeePMD training/source checkpoints, but keep the one
+# authoritative single-task DPA4 model bundled by apex-flow.
 apex/skills/apex-flow/models/**/*.pt
+!apex/skills/apex-flow/models/DPA4-alloytongqi/model.pt
 apex/skills/apex-flow/models/**/*.partial

.pre-commit-config.yaml

@@ -0,0 +1,36 @@
+# Local pre-commit checks (no remote hook repos; works behind restricted proxies).
+# Install once:
+#   pip install pre-commit && pre-commit install
+repos:
+  - repo: local
+    hooks:
+      - id: check-merge-conflict
+        name: check merge conflict markers
+        entry: python3 -c
+        args:
+          - |
+            import pathlib, re, sys
+            pat = re.compile(r"^(<<<<<<< |>>>>>>> )")
+            bad = False
+            for raw in sys.argv[1:]:
+                path = pathlib.Path(raw)
+                if not path.is_file():
+                    continue
+                try:
+                    text = path.read_text(encoding="utf-8", errors="ignore")
+                except Exception:
+                    continue
+                for i, line in enumerate(text.splitlines(), 1):
+                    if pat.match(line):
+                        print(f"{path}:{i}: merge conflict marker")
+                        bad = True
+            raise SystemExit(1 if bad else 0)
+        language: system
+        types: [text]
+
+      - id: apex-unit-smoke
+        name: APEX unit smoke (gamma + ticket + failed-result)
+        entry: bash -c 'cd tests && python3 -m unittest -v test_gamma.TestGamma.test_compute_lower test_skill_scripts.TestGenerateConfigHelpers.test_get_bohrium_ticket_success test_ops.TestTaskStatusHelpers.test_is_failed_task_result_accepts_dict_and_mapping_like'
+        language: system
+        pass_filenames: false
+        types: [python]

CHANGELOG-1.3.0.md

@@ -15,6 +15,8 @@ Major additions include:
 - Finite-temperature lattice workflow improvements
 - Finite-temperature elasticity workflow support and example documentation
 - Annealing workflow support for LAMMPS
+- Melting-point workflow support for solid-liquid coexistence calculations
+- Image-resident DPA4/PT2 runtime support with a fail-closed T4 qualification profile
 - Shape-controlled automatic supercell generation
 - `req_calc`-based relaxation/property selection for joint workflows
 - GUI-side batched submission for large configuration sets
@@ -74,6 +76,21 @@ Major additions include:
 - Added LAMMPS annealing workflow support.
 - Improved generated annealing input scripts for heating, cooling, RDF output, and holding stages.
 
+### Melting Point Workflow
+
+- Added a LAMMPS solid-liquid coexistence workflow for estimating melting points across multiple temperatures and replicas.
+- Added restart-file staging, interface-axis controls, per-temperature restart validation, and generated skill templates.
+- Added input validation and regression coverage for melting-specific temperature, replica, restart, and backend constraints.
+
+### DPA4/PT2 Runtime and Qualification
+
+- Added image-resident DeePMD PT2 model handling so immutable runtime artifacts are not copied into task upload packages.
+- Added DPA4-aware LAMMPS input generation, including the required atom maps, runtime plugin auto-loading, and rejection of incompatible legacy plugin commands.
+- Added an audited `DPA4-alloytongqi` source checkpoint, container wrapper templates, runtime manifest template, and a reproducible CPU/GPU and phonoLAMMPS benchmark harness.
+- Added a single-source DPA4 runtime profile used by generation, recommendation, standalone validation, and submission. The bundled profile remains fail-closed until an immutable image `ref@digest` passes the recorded post-snapshot T4 qualification.
+- Restricted the qualified production contract to one MPI rank on one `c4_m15_1 * NVIDIA T4`, with exact image, wrapper, model hash, and dispatcher validation.
+- Added complete automatic `type_map` expansion across every resolved structure and every effective LAMMPS `overwrite_interaction`.
+
 ### Workflow Selection and Submission
 
 - Added `req_calc`-based workflow selection for relaxation and property calculations.
@@ -94,6 +111,7 @@ Major additions include:
   - `.debug.log`
   - `.debug.stdout`
   - `.debug.stderr`
+- Added bounded retries for transient LAMMPS remote-startup failures, including header-only logs, retry evidence preservation, and explicit retry classification in task status files.
 
 ### Reporting
 
@@ -126,4 +144,5 @@ Major additions include:
 - Added runnable examples for RSS, GammaSurface, and finite-temperature elasticity.
 - Added GUI developer documentation.
 - Added `monty` to the package dependencies and constrained supported Python versions to `<3.13`.
-- Added support for Phonopy v4 in terms of phonon calculation.
+- Added Phonopy v4-compatible setup, force-constant, and band-generation fallbacks.
+- Expanded the bundled `apex-flow` skill with backend-aware generation and validation for Bohrium, local debug, local cluster, VASP, LAMMPS, DPA4/PT2, melting-point, gamma, and finite-temperature workflows.

CHANGELOG-1.3.1.md

@@ -0,0 +1,60 @@
+# APEX 1.3.1 Changelog
+
+Release date: 2026-10-13
+
+APEX 1.3.1 consolidates workflow, runtime, and validation improvements made
+after `1.3.0.post5`. It adds two-phase melting-point calculations, restores
+finite-temperature VASP execution, expands cloud execution support, and
+hardens LAMMPS, phonon, Gamma, and agent-assisted workflows.
+
+## Highlights
+
+- Added a LAMMPS two-phase coexistence melting-point workflow with restart
+  transport, replica diagnostics, post-processing, reporting, and bracket
+  determination.
+- Added OpenAPI Sandbox execution support alongside Bohrium and local paths.
+- Restored finite-temperature VASP workflows with stage-aware execution,
+  output validation, failure evidence, and automatic `vasp_gam`/`vasp_std`
+  selection from generated k-points.
+- Integrated the validated DPA4 phonoLAMMPS runtime for GPU workflows while
+  retaining CPU-safe image routing for CPU potentials.
+- Added the bundled single-task DPA4 alloytongqi checkpoint and a fail-closed
+  immutable PT2 runtime contract with image wrappers and qualification tools.
+- Strengthened Gamma and GammaSurface generation, validation, and preview for
+  difficult and disordered parent structures.
+- Expanded the bundled `apex-flow` agent skill with safer preflight checks,
+  execution profiles, and configuration generation.
+
+## Workflow and Runtime Updates
+
+- Added configurable preparation, conditioning, production, restart, and
+  replica controls for `melting_point`.
+- Added solid-fraction analysis, tidy CSV/JSON outputs, trajectory plots,
+  task-status evidence, and periodic restart files.
+- Updated phonon processing for Phonopy 4 while retaining compatibility with
+  older force-constant generation paths.
+- Removed obsolete explicit `plugin load libdeepmd_lmp.so` commands for the
+  integrated USER-DEEPMD runtime.
+- Added primitive-axis propagation and phonoLAMMPS command improvements.
+- Added DeepMD atom maps where required by the DPA4 PT2 runtime.
+- Added retries and per-attempt evidence for transient LAMMPS termination,
+  timeout, and signal-like failures.
+
+## Validation and Configuration
+
+- Added shared Gamma/GammaSurface geometry and parent-lattice mapping helpers.
+- Added overlap, atom-count, and geometry guards before Gamma/GammaSurface
+  submission and preview generation.
+- Added multi-structure `type_map` inference and validation, including property
+  overwrite interactions.
+- Added image/machine compatibility checks, CPU/GPU image routing, confirmed
+  property profiles, image overrides, and model staging validation.
+- Added adaptive DFT k-spacing and stronger VASP/ABACUS input preflight checks.
+- Corrected Bohrium ticket expiration handling to use hours.
+
+## Compatibility
+
+- Supported Python versions remain 3.10 through 3.12.
+- Existing 1.3.0 parameter files remain supported; new workflow and runtime
+  controls are opt-in unless selected for a validated backend.
+- Package and container release version: `1.3.1`.

CHANGELOG-1.3.2.md

@@ -0,0 +1,30 @@
+# APEX 1.3.2
+
+Release date: 2026-10-14
+
+- Generated cloud jobs reuse a working image-bundled APEX 1.3.2 or newer stable
+  version. Only missing/older versions trigger installation of the pinned
+  `apex-flow==1.3.2` package. CLI startup failures stop submission.
+- Runtime checks and submission use the same Python interpreter; the packaged
+  `ensure_apex.py` helper travels with `run.sh`.
+- Correct the ordinary Bohrium L20 default to `1 * NVIDIA L20_48g` while retaining
+  the distinct OpenAPI Sandbox SKU `c16_m120_1 * NVIDIA L20`.
+- Update the outer client and make/post image to APEX 1.3.2. Preserve the
+  separate CPU/GPU calculator images and the OpenAPI dispatcher image.
+- Align Sandbox ABACUS/VASP calculator and dispatcher image addresses, refresh
+  Sandbox access-key configuration without using the legacy ticket API, and
+  respect explicitly selected contexts when merging account defaults.
+- Align cloud skill instructions, examples, templates, and regression coverage
+  with image reuse and the separate machine catalogs.
+
+Existing generated jobs keep their saved scripts and machine settings. Generate
+a fresh job directory using the updated skill before resubmitting; preserve
+existing task outputs and check the original workflow state first.
+
+## 1.3.2.post1
+
+Use the independently snapshotted `apex-flow:1.3.2-post1` client image in
+Bohrium project 397953. The earlier `1.3.2-r1` image reference has an active
+platform build task and must not be used as the release runtime. Missing/old
+clients bootstrap the pinned 1.3.2.post1 package; working 1.3.2 runtimes remain
+compatible and are reused. Runtime functionality is unchanged.

CHANGELOG-1.3.4.md

@@ -0,0 +1,11 @@
+# APEX 1.3.4
+
+- Use the updated APEX/CPU image for workflow preparation, post-processing and classical LAMMPS, and a separate general DeePMD/L20 image for compatible TensorFlow and PyTorch models.
+- Preserve the configured calculator image for phonon and Grüneisen tasks. The dedicated alloytongqi PT2 contract still selects its pinned image and wrappers.
+- Validate the preinstalled APEX runtime instead of downloading or upgrading packages at job startup. Regenerate configuration for existing jobs to select the new image.
+- Synchronize skill templates, examples, interpreter paths and machine recommendations. MACE, NEP, GAP, RANN and SNAP require an explicit compatible custom image.
+- Fix the skill YAML frontmatter so MatMaster can load the updated instructions.
+
+Runtime validation and scientific scope are recorded in `images/runtime-20261016/qualification.json`. Runtime success does not establish convergence or compatibility with every model and GPU.
+
+- Use absolute phonoLAMMPS executable paths for the new images so sandbox login-shell PATH resets cannot break phonon/Gruneisen submission.

MANIFEST.in

@@ -0,0 +1,6 @@
+include LICENSE
+include README.md
+include CHANGELOG-1.3.0.md
+include CHANGELOG-1.3.1.md
+include CHANGELOG-1.3.2.md
+recursive-include apex/default_config *