Changelog • fmristore

fmristore 0.1.0.9000 (Development)

Added metadata-only frame-store inspection, conservative rollback of proven uncommitted write artifacts, and an atomic streaming migration from the real 0.1-provisional HDF5 frame layout to FDS v1.
Added atomic sharded-frame directories. New observation shards are committed as immutable FDS v1 HDF5 files before a small root-manifest swap, so appends never rewrite prior assay data and reopen as canonical row-sharded sources.
HDF5 frame writers now support balanced, imagewise, and featurewise chunk layouts under explicit byte targets. h5_read_plan() and validate_h5_read_amplification() expose enforceable, metadata-only I/O amplification gates, including exact partial-edge chunk accounting.
Added reconstructible h5_array_source() descriptors implementing the fmridataset lazy source protocol without serialized file handles.
Replaced the provisional frame format with an atomic FDS v1 codec. write_frame_h5() now streams separately bound assays and aligned blocks under a hard memory budget; open_frame_h5() validates semantic and storage manifests while leaving every numerical array lazy.

HDF5-backed neuroimaging data structures:
- H5NeuroVol: 3D brain volumes
- H5NeuroVec: 4D time series data
- H5NeuroVecSeq: Sequences of 4D scans
Parcellated data storage:
- H5ParcellatedScan: Single scan with cluster-based organization
- H5ParcellatedScanSummary: Summary statistics per cluster
- H5ParcellatedMultiScan: Multi-run experiments
Latent representation I/O:
- Read/write LatentNeuroVec objects to HDF5
- Spec-compliant BasisEmbedding format
Labeled volume sets:
- LabeledVolumeSet: Named brain region collections
- Efficient HDF5 storage with compression
Utility functions:
- read_dataset(): Auto-detect and load HDF5 neuroimaging data
- write_dataset(): Generic writing interface
- as_h5(): Convert in-memory objects to HDF5 format