v1.2.0

Added

• Graceful training completion: Treat SIGUSR1 or a submission-specific marker file as a request to finish the current optimizer update, synchronize all distributed ranks, run configured validation, write a complete resumable checkpoint, flush outputs and exit successfully.
• File-aware entry filtering: Add the spine-filter scan/build workflow, fingerprinted per-source measurement caches, transactional dataset manifests, LArCV product-size inspection, and composable training/validation filters for raw, HDF5 and SPINE cache inputs.
• Sharded cache repositories: Add CacheReader, CacheWriter and CacheDataset around atomic manifests and immutable per-source/per-stage HDF5 V2 shards, with lineage validation, transparent mixed-dataset source projection, parallel array publication, retry fencing, atomic stage replacement and conservative recovery garbage collection through spine-cache.
• Sparse CNN lattice phases: Randomize each training image's alignment with the sparse convolution lattice while retaining detector-frame inputs and outputs, with automatic encoder-depth periods or explicit scalar/per-axis periods.

Changed

• Standalone configuration composition: Load report and entry-filter configurations through the standard SPINE resolver, including search paths, nested includes, overrides/removals, environment expansion, and cycle detection.
• PPN particle supervision: Restrict PPN targets with raw particle associations aligned to the final sparse coordinates, keep reconstructed-domain label adaptation for downstream objectives, and allow segmentation-only full chains to disable adapted-label production.
• Learning-rate scheduler cadence: Add explicit optimizer-step, epoch, and scheduled-validation timebases, deprecate the operational checkpoint spelling in favor of validation, warn when the historical step default is implicit, and prevent graceful-completion snapshots from advancing the schedule.
• Cache production contract: Replace monolithic staged HDF5 output and end-of-job sidecar merging with *.spine-cache repositories. Parallel replacement contributions share a launcher-provided publication identity and are fenced by an explicit begin operation so delayed tasks cannot mix submission attempts.
• Rotation augmentation sampling: Accept either one quarter-turn value or a list of choices through k, and optionally gate the complete rotation with p.
• Shared artifact permissions: Create SPINE output files as collaboration-readable and shared output directories as group-writable and group-inheriting, independent of restrictive temporary-file defaults.

Fixed

• LArCV entry-filter counts: Inspect products through direct ROOT branch access and invalidate stale version-1 scan records that could report false zero counts.
• Crop augmentation coordinates: Preserve continuous subvoxel coordinates, validate quantized crop bounds deterministically, intersect explicit sampling bounds with the source volume and support activity-biased selection from one coordinate product.
• Cached input alignment: Preserve original source-entry provenance in compact caches and publish upstream sparse-index mappings so raw segmentation labels align with cached model inputs.
• Distributed checkpoint failures: Synchronize ordinary rank-zero serialization failures before any worker rank can leave checkpoint collectives.
• Geometry-aware weight export: Initialize a configured detector geometry before constructing a model-only full-chain export.

Removed

• Staged HDF5 sidecars: Remove the experimental stage_hdf5 reader/writer surface and its expensive sidecar merge path in favor of the cache repository abstraction.

Full Changelog: v1.1.0...v1.2.0