flowTorch 1.10.1

Fixes

• Tecplot now honors the requested tensor dtype for snapshots, time series, vertices, cell locations, and weights.
• Image normalization now runs in the requested dtype, avoiding intermediate float32 rounding when requesting float64.

Precision warnings

• EnSight loading with dtype=torch.float64 warns that binary files store float32 values and the current VTK ASCII reader also reads fields and coordinates as float32. Native binary chunk loading shares the file format's precision limit.
• EnSightWriter warns once per writer when float64 fields (including complex128 components) are converted to binary float32. Binary remains the default. ASCII preserves double-precision field values in the file, although the current VTK EnSight reader reads them as float32.

Compatibility

Tecplot snapshots and coordinates now follow the default torch.float32 dtype even when the reader returns float64. Pass dtype=torch.float64 explicitly to retain available reader precision. SCube is unchanged.

Validation

832 tests passed, 5 skipped locally; external-data integration tests excluded. Formatting, lint, distribution validation, and packaged version checks passed.

Cross-platform CI passed on Linux, macOS, and Windows with Python 3.10, 3.12, and 3.14; distributed Gloo tests passed. The optional type-check job retains the same 32 baseline errors as the previous release.