v1.1.0

Added

• Production CLI workflows: Support named sources for joint and mixed datasets, mirror source and entry-selection overrides onto validation datasets, apply per-module checkpoint overrides, and export a fully composed model checkpoint without initializing data I/O.
• Materialized GrapPA training: Run GrapPA directly from cached node features, edge features and graph indexes, and optionally cache resolved node and edge supervision targets with static validity masks while preserving iteration-dependent selection.
• Graph augmentation: Add reciprocal-aware edge dropout, node and truth-group dropout with optional class selection, and independent feature masking and noise for materialized GrapPA inputs.
• Detector-response augmentation: Add configurable signal response, noise and recombination distortions alongside stochastic calibration-parameter throws for lifetime, gain, signal shape, noise and recombination.
• Transactional staged caches: Store internal stage products in HDF5 V2, publish independently produced stages through atomic sidecars, preserve source provenance, and replace incomplete or explicitly selected stages without rebuilding completed siblings.
• Cached full-chain entry points: Declare externally supplied full-chain products, propagate cached PPN points as canonical state, and resume particle or interaction aggregation from cached upstream results.
• Metric reporting workflow: Add the standalone report CLI and streaming reports for segmentation, PPN, clustering and node-level tasks, including configurable semantic mappings, bounded-memory reductions, JSON summaries and rendered figures.
• Track-breaking transform: Add a configurable full-chain transform for splitting track-like fragments before downstream aggregation.
• Barycenter flash matching: Add configurable optical-response selection, shared OpT0Finder model integration, hypothesis reporting and a maintained SBND configuration.
• Early-stopping progress: Report the monitored validation value, change from the previous best, minimum delta, status and patience progress inside each checkpoint section.

Changed

• Staged-cache format: Make the internal staged HDF5 reader and writer exclusively V2. Staged caches produced with the earlier disposable layout must be rebuilt; the general HDF5 reader retains its V1 and V2 compatibility contract.
• HDF5 minibatch throughput: Coalesce contiguous V2 product reads across loader minibatches and keep the established V1 handle-reuse behavior.
• Analysis serialization: Support directional truth-to-reco or reco-to-truth matched exports, serialize object collections in bulk through the canonical data-class schema, and use efficient system buffering by default for analysis CSV files.
• Clustering evaluation: Share contingency tables across requested metrics, represent undefined comparisons with NaN, retain legitimate finite values such as ARI -1, and report truth, reconstruction and comparable support overall and by semantic class.
• Metric histogram semantics: Apply custom clustering ranges consistently to histograms and moments while recording finite values excluded from the selected range; increment the report schema to 1.5.0.
• Validation presentation: Give on-the-fly validation its own inference-style log stream and group validation metrics, checkpoint paths and early-stopping state into clear bounded stdout sections.
• Composite loss outputs: Remove duplicated parent-stage namespaces from cached GrapPA supervision products.
• Production documentation: Expand CLI, configuration, workflow, troubleshooting and package API documentation with warning-strict documentation audits.

Fixed

• On-the-fly validation logs: Write checkpoint validation batches to dedicated validation_log-* segments, keep intermittent validation metrics out of fixed-schema training CSV files, let TrainDrawer discover standalone inference and on-the-fly validation logs together, and present each completed training step before its bounded validation and checkpoint section.
• Cached GrapPA supervision: Reapply dynamic shower-purity selection at training time, normalize cached shape and target representations, accept cached orientation targets without endpoints, and convert Torch-backed grouping inputs before Numba evaluation.
• Cached aggregation coordinates: Prefer canonical cached PPN points before falling back to truth coordinates when resuming GrapPA stages.
• Distributed cache execution: Restore RNG state across devices, route mixed-dataset stage writes through sidecars, validate explicit stage maps and preserve cumulative post-processing provenance across repeated cache passes.
• CUDA cluster association: Keep batch-index tensors on the same device while expanding particle associations.
• Model-only export: Apply configuration and weight overrides before requiring an io section, making model-only checkpoint composition reachable from the CLI.
• Documentation and lightweight CLI checks: Restore dependency inspection at its canonical CLI utility location and cover configuration failures and optional-runtime imports in CI.

Full Changelog: v1.0.4...v1.1.0