feat(mesh): DomainMeshZarrSink emitting the physicsnemo mesh-zarr schema

[negin513]

Summary

Adds DomainMeshZarrSink: a mesh-domain sink that writes physicsnemo.mesh.DomainMesh (and plain Mesh) objects as Zarr group trees in the physicsnemo mesh-zarr schema — the chunked, compressed, random-access counterpart of MeshSink's .pdmsh/.pmsh memmap output. Together with the reader side (NVIDIA/physicsnemo#1798), this makes raw CFD data → training-ready zarr a one-hop curation: DrivAerMLSource (multi mode) already yields DomainMesh objects, so the sink is a drop-in alternative to MeshSink in the DrivAerML example pipeline.

Design

• DomainMesh schema mirrors the .pdmsh tree: case-level global_data at the group root (single source of truth, merged at read time by the reader — never copied into member meshes), interior/ and boundaries/<name>/ as mesh-schema subgroups.
• Verified layout attr: with soup_boundaries=True, boundary meshes are denormalized to a per-cell vertex soup so training-time contiguous block subsamples map to sequential, chunk-aligned reads (recommended for cell-centric surface training on meshes without point-index locality, e.g. DrivAerML). The attr is verified from the written cells, never caller-asserted; connectivity-dependent consumers use the default indexed layout.
• Cell-axis chunking (chunk_cells) aligned with the training reader's subsample size; blosc-zstd compression (level 0 disables).
• Curator conventions throughout: params() descriptors (wizard/serialization-visible), registered in the mesh-domain registry, atomic temp-then-rename writes for safe resume, {run_id} naming via set_source.

Schema contract

physicsnemo-domainmesh-zarr / physicsnemo-mesh-zarr, version 1, specified in physicsnemo/datapipes/MESH_ZARR_SCHEMA.md (see NVIDIA/physicsnemo#1798). The sink implements the schema self-containedly (no dependency on unreleased physicsnemo APIs); parity is enforced by a cross-repo round-trip test — sink output read back through physicsnemo's ZarrMeshReader with subpath navigation and read-time global_data merge — which skips gracefully on physicsnemo releases without the reader.

Validation

• 8 tests in test/domains/mesh/test_mesh_domain_zarr.py (schema shape, soup relayout geometry preservation, compression, atomicity, naming, plain-Mesh items, registry, cross-repo round-trip).
• End-to-end on real DrivAerML: converted cases (all 5 boundaries + 166M-point interiors + in-group full-resolution STL, 6.1 GB/case zstd vs 6.5 GB .pdmsh) train GeoTransolver surface and volume through physicsnemo's unified external aero recipe with losses matching the .pdmsh path; dataloading is ~2.5x faster (details and benchmarks in the companion PR).

[codecov-commenter]

Codecov Report

❌ Patch coverage is 94.40559% with 8 lines in your changes missing coverage. Please review.
✅ Project coverage is 63.11%. Comparing base (4111fce) to head (59b8b31).
⚠️ Report is 2 commits behind head on main.

Files with missing lines	Patch %	Lines
...emo_curator/domains/mesh/sinks/mesh_domain_zarr.py	94.24%	8 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #64      +/-   ##
==========================================
- Coverage   63.39%   63.11%   -0.29%     
==========================================
  Files          88       89       +1     
  Lines        9488     9630     +142     
==========================================
+ Hits         6015     6078      +63     
- Misses       3473     3552      +79     

Flag	Coverage Δ
integration	63.11% <94.40%> (-0.29%)	⬇️
unit	63.11% <94.40%> (-0.29%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.